WorksPhysics and Astronomy
Energetic Particle Confinement In Black Hole Accretion Disks
Abstract
Using tracer particles propagating in the environment provided by general relativistic magnetohydrodynamic simulations, this paper addresses fundamental questions related to the confinement and acceleration properties in accretion disks around compact objects. By varying the spin parameter of the central black hole and the accretion regime of the disk, a comprehensive picture is obtained for particles whose Larmor radius is comparable to the gravitational radius of the system. It is found that such particles can escape the system at the outermost boundary of the simulations, though, depending on the accretion regime, the time it takes them to escape, and the energetic dynamics along the trajectory can vary significantly. Dynamics near the black hole are even more sensitive to environmental details as black holes with higher spin values generate stronger magnetic fields, which influence particle motion in non-trivial ways. The primary acceleration mechanism is identified as turbulent scattering, with the jet-sheath emerging as the dominant site for particle energization
Introduction
The environment surrounding massive black holes (BHs) is a leading candidate site for the acceleration of energetic charged particles, as revealed by their multi-messenger signatures [e.g. 2, 3, 4] (Boulanger et al. 2018). A recent detection of particular interest was reported by the IceCube Collaboration [5], which identified an excess in the neutrino flux in the energy range TeV originating from a region of the sky coincident with the Seyfert II active galactic nucleus (AGN) NGC 1068 [6, 7]. Analysis of this source [7], in particular with regard to its bright X-ray emission but weak gamma-ray counterpart, suggests that a compact () region around the BH, commonly referred to as the “corona”, is the most likely origin of the observed neutrino flux.
The overall properties of such a region can be constrained through phenomenological arguments, but the details of its composition remain uncertain, with a growing body of literature exploring these unresolved aspects [e.g. 7, 8, 9, 10] (Das et al. 2024; Lemoine and Rieger 2025). Given the experimental challenges with directly probing such environments, numerical simulations provide a fundamental instrument for investigating the phenomena occurring within these systems.
The present paper is one such study, in which accretion disks with Advection Dominated Accretion Flows (ADAFs) [13, 14, 15], the Magnetically Arrested Disk (MAD) [16] and the Standard And Normal Evolution (SANE) [13] disk configuration, are simulated using the equations of General Relativistic Magnetohydrodynamics (GRMHD) [17, 18]. SANE and MAD states represent limiting magnetization regimes of magnetized accretion flows in which the horizon-threading magnetic flux remains respectively subcritical or saturates at dynamically important values, respectively. The distinction is most clearly realized in geometrically thick, advective flows, where magnetic-flux advection is efficient, although analogous magnetically arrested states may arise in radiatively efficient disks provided sufficient net poloidal flux is supplied.
It is worth noting that the nomenclature ”Magnetically Arrested Disk” can be somewhat misleading, as accretion is only ever temporarily suppressed and resumes after a transient time in a periodic fashion. While we use the SANE and MAD labels for convenience in the following, we also highlight which fundamental properties of the flows or of the magnetic fields, including the dependence on BH spin, are at play in each instance, so that the results are not strictly tied to the specific type of simulation, but provide deeper insight into the underlying physics. Once the accretion rate and angular momentum transport in our simulations reach a quasi-steady regime, energetic tracer particles are injected and evolved self-consistently with the disk in order to investigate their transport and propagation properties. These particles respond to the electromagnetic fields generated by the plasma flow, and obey the laws of General Relativity (GR) to account for strong gravitational effects. At the moment, these particles are completely passive and do not exert any feedback on their environment. In GRMHD simulations, the normalization length is set by the gravitational radius (, with the gravitational constant, the BH mass and the speed of light). The ratio of the Larmor radius (, defined below) of the energetic particles to the gravitational radius, is what determines the scaling of the results of our simulations. Namely, the dynamics observed in these simulations are representative of any system in which particles move with GR dynamics and satisfy the ratio set by our parameters.
Overall, we aim to provide some further insight into the physics of particle acceleration and neutrino production in AGN environments. Indeed, reproducing the measured spectra of such particles requires modeling the emission by the distribution function at the source with certain assumptions (e.g., proton escape times in Fokker-Planck type diffusive modeling (e.g., Lemoine and Rieger 2025, Fraschetti 2021)), that can be rendered more robust by supplying information from numerical simulations. At the same time, the simulations presented in this paper seek to provide a baseline for exploring the particle confinement properties of ADAFs, potentially identifying shortcomings and critical points that will need to be addressed by future studies. For instance, classical ADAFs do not easily reproduce corona-like features, as the flows tend to be geometrically thick, with little temperature variations, except once the jet region [20, 21] is reached. For the purposes of this paper, the “corona” is defined as a region of gravitationally bound plasma near the black hole (within ) that satisfies the Bernoulli parameter criterion defined later and that is strongly magnetized. This definition allows flexibility in identifying the corona’s structure but explicitly excludes magnetized outflows (winds and jets) that are more common in ADAF simulations and that do not match observational constraints from NGC 1068 [7]. Thus, the analysis in this paper will help clarifying the potential of ADAFs as environments for energetic particle confinement, and their associated multi-messenger signatures. The paper is structured as follows: Sec. II and Sec. III describe the numerical setup and simulation parameters. Sec. IV presents the analysis of particle escape and infall times, including comparisons with standard estimates, and examines the mechanisms of particle acceleration, with a focus on turbulent scattering. Sec. V discusses the implications for high-energy astrophysics, including the potential role of these flows in accelerating cosmic rays and producing high-energy neutrinos. Finally, Sec. VI summarizes the key findings and outlines directions for future work.
Numerical methods
The code employed for the simulations shown in this paper is a re-implementation of the Athena++ code [22] based on the use of the KOKKOS library [23] for portability, named AthenaK [24]. The advantages of having access to a multitude of architectures with good scalability, in particular GPUs, are manifest, and crucial when dealing with simulations that involve particles, since the computational cost can increase significantly (see App. B). The porting of codes to GPUs has also significantly contributed to making simulations of radiative flows more realistic [25, 26, 27]. AthenaK uses Cartesian Kerr-Schild coordinates () to define the lab frame in GR, which suits well the Cartesian grid employed by the code to discretize the physics.
The integration of the equations in GR is frequently done by employing a “3+1” decomposition (e.g., [28, 29]), performing a so-called “foliation” of spacetime labeled by coordinate time. This approach allows one to clearly distinguish temporal and spatial components of quantities measured by a “fiducial” (also called “normal”) observer (FIDO) moving with 4-velocity (measured in the coordinate frame) and , with spacetime metric being expressed in the form
where is the lapse function, is the shift three-vector, and is the spatial part of the spacetime metric , with the algebraic inverse of . As usual, Greek indices run from 0 to 3, thus including “temporal” components, while Latin indices run from 1 to 3.
As we show below, the equations of motion for a charged particle can be incorporated as a Hamiltonian pair in terms of the shift vector , the inverse of the spatial three-metric and the covariant particle four-velocity in coordinate basis . Note that from (1), the relation between and is . The FIDO measures the particle Lorentz factor and three-velocity to be and , respectively, with “coordinate” (i.e., measured in a frame of reference at infinity) velocity . The FIDO and coordinate frame Lorentz factors are related as , with the one measured in the FIDO frame.
GRMHD in AthenaK
AthenaK discretizes the fluid on a cartesian grid and applies Riemann solvers to solve the integral flux conservation equations, i.e., it uses a Finite Volume Method [18, 22, 24, 30, 31] and a Constrained Transport algorithm for the evolution of the magnetic field [32]. For an in-depth analysis of the ideal GRMHD equations and their solutions, the reader is referred to the literature, as only some fundamental points will be recalled here.
The manifestly covariant equations can be summarized as follows:
Here, is the determinant of the spacetime four metric, with , where is the determinant of the spatial three metric, is the fluid four-velocity, is the proper rest mass density, is the dual of the antisymmetric electromagnetic tensor, is the energy-momentum tensor, and the Eulerian magnetic field in coordinate basis. One can notice how the left-hand side of these equations obeys a flux conservation law, thus making the problem suitable for solution using Godunov-type schemes with Riemann solvers [e.g. 33]. The magnetic field is evolved using a constrained transport scheme [32] to ensure that the divergence-free constraint is enforced.
The GRMHD simulations are run by initializing a Fishbone-Moncrief [34] torus to which a dynamically negligible magnetic field is added in the form of poloidal loops encircling the torus’s cross-section, cf. eq. (9). The torus is initially in hydrostatic equilibrium, presenting an axisymmetric laminar flow with constant angular momentum, but the presence of the weak magnetic field allows the development of the Magneto-Rotational Instability (MRI) [35] that enables accretion by providing a means for the outward transport of angular momentum. At the same time, MHD turbulence will develop as a consequence of the MRI, and such turbulence is bound to interact with energetic particles in the fluid, possibly leading to their stochastic acceleration (e.g., Lemoine and Rieger 2025), and/or to changes in the magnetic field configuration [35], that will affect the confinement of these particles. The self-consistent development of MRI-driven turbulence from an initial magnetic field provides the physical mechanism that has been otherwise modeled phenomenologically by the parameter in Shakura and Sunyaev [36, 37]. The basic idea is always the same: stresses in the flow lead to accretion by effectively converting azimuthal motion into radial inflow, with the MRI offering a fundamental mechanism for this process that involves the Maxwell (magnetic) stress-tensor, whereas the classical model in Shakura and Sunyaev [36, 37] summarizes the effect of the Reynolds (fluid) stress-tensor with the parameter . Depending on the amount of magnetic flux accreted at horizon scales and the mechanisms for its dissipation, the system will evolve into either the MAD or the SANE state, both of which are considered in this work. Ultimately, the initial conditions for the two systems differ only in the size of the initial torus. The magnetic field is kept dynamically negligible (hence the use of a hydrodynamic solution for the torus, with a weak, high- field added) and serves only to trigger the MRI and break the otherwise stable, constant angular momentum laminar flow, enabling accretion. The fully developed turbulent flow that follows exhibits notable differences from the initial conditions in both SANE and MAD regimes: magnetically dominated regions form, and in the MAD regime, accretion is modulated by the competition between gas and magnetic pressure. Accretion dynamics also self-consistently lead to the formation of outward flows, i.e., winds, that do not require strong magnetic fields to exist but can be enhanced and shaped by them, evolving into jet-like structures: highly collimated, strongly magnetized, relativistic outflows ejected along the poles of spinning black holes. These can feedback on scales far beyond those probed by the simulations. The fully developed system thus exhibits a non-trivial structure, including vertical stratification, sheared and turbulent flows that connect spatially distant regions, and strongly varying magnetizations, among other key features. While this general picture of the fully-developed flows is valid for both SANE and MAD regimes, the most notable, though certainly not the only, differences between the two are (i) the presence of bursty behaviour in MAD disks, with periodic ejections of matter that limit the build-up of magnetic pressure at horizon scales, (ii) the formation of jet-like structures in MAD disks [13, 21, 38, 39, 40, 41], defined as regions of highly-magnetized outflows located at the poles of a BH, when the central BH has a sufficiently large spin parameter, and (iii) the development of non-axisymmetric modes in MAD regimes associated with the bursty behavior. Note that SANE disks with high BH spin can also exhibit similar structures, but the magnetization and stability (persistence over time) in MAD disks are higher, thus one refers broadly to such regions as “funnel” regions Hawley and Villiers [42]. How these different accretion regimes affect energetic particle confinement is one of the core questions addressed in this paper. Although systems such as NGC 1068 do not seem to exhibit the dynamics and properties that match those of radiatively inefficient ADAFs Narayan et al. [13], as these flows obey different radiative transport dynamics and operate at accretion rate well below 1% of the Eddington limit, ADAF models have proven effective for studying many features of supermassive black holes (SMBHs) Porth et al. [43]. This suggests that the results of this study are relevant to real astrophysical scenarios, despite the GRMHD dynamics not fully matching those of any particular AGN one-to-one.
AthenaK uses normalized code quantities, including lengths, times, velocities, densities, and magnetic fields, to characteristic properties of the system. Physical values are recovered by specifying a mass, length, and density scale, so that, for example, velocities scale with and lengths with .
Particles in AthenaK
Particles in AthenaK are a recent development, and in order to investigate the physics relevant to this work, both GR and electromagnetic effects have to be accounted for in the particles’ equations of motion. Since the spacetime metric in our simulations is an analytical, static function of space only, it is possible to use the Hamiltonian formulation of the geodesic equation of motion in “3+1” decomposition to derive a conservative pusher for GR effects. Electromagnetic effects, on the other hand, are computed using discrete fields stored in the computational grid, and thus are not suitable for a Hamiltonian treatment, requiring other types of approaches. Following the work done in Bacchini et al. [44], Ripperda et al. [45], Bacchini et al. [46], Zou et al. (2024), we implemented an implicit mid-point scheme with the Newton-Raphson method to compute the updated values of position and velocity. The equations to solve are:
where the geodesic part is derived from the free-particle Hamiltonian:
where is the covariant particle 4-velocity in the coordinate frame, and for massive particles. The electromagnetic terms are obtained by expressing the terms related to the Faraday tensor as a function of the quantities available in the code (fields in the coordinate frame and 3 + 1 metric).
The relation for the electric displacement field , as measured by a FIDO, is
with Levi-Civita (pseudo-)tensor .
At each time-step, for each particle, non-linear iterations are performed to advance the values for the position and velocity at the next time-step until a tolerance of on the (square of the) residuals is met. The six-dimensional system of equations (5)-(6) is inverted by building the full 6-dimensional Jacobian matrix. While this is more expensive than a GMRES-like approach Saad and Schultz [48] the straightforwardness of the implementation, the small size of the system and the small number of iterations typically needed to reach convergence (the average count is around 15, with recorded maxima below 60) do not warrant, at present, the employment of more sophisticated algorithms.
The particles’ diagnostics consist of the position at each time-step of interest and the Lorentz factor measured by the normal observer (). Particles are removed from the domain over time as they cross either the outer boundary of the simulation or the inner boundary, set at for reasons explained in Sec. III. At present, particles are injected at a single specific time and no mechanism is implemented to have particles re-injected in the system, meaning that the total number of particles in the system decreases monotonically over time.
The details of the numerical scheme are available in the references, and salient points of the validation work done during development are reported in Appendix A.
Simulation setup
Accretion disk
The accretion tori are initialized as is usual for the SANE and MAD regimes, i.e., with parameters listed in Tab. I, where the main difference between the two initializations lies in the size and distance of the torus from the BH, and in the different dependence of the magnetic field on the hydrodynamic quantities. A SANE case with spin parameter was not included due to the excessively weak magnetic field for the parameters of interest. The Fishbone-Moncrief torus [34] is a hydrodynamic self-similar solution on top of which a weak (high plasma ) magnetic field is imposed. The field is defined through a magnetic potential that has only a azimuthal component, producing poloidal loops in the magnetic field, with the expression (in Boyer-Lindquist, BL, coordinates):
where is the radius in BL coordinates, is the fluid density in the fluid rest frame, and , and are parameters used to shape the magnetic potential. The exponential and the -dependence only apply to the MAD case. The hydrodynamic Fishbone-Moncrief solution is characterized by two main parameters, and , that determine the distance of the innermost edge of the torus from the central BH and the radial position of the maximum of the enthalpy (defined in Sec. IV A 2), respectively.
The resolution of the grid is set to in the region (in units of ) , with Static Mesh Refinement (SMR) that halves the grid resolution each time the domain extent is doubled, with the total box size set to .
The GRMHD parts of the MAD (SANE) simulations are allowed to evolve for (), until a quasi-steady regime in the accretion rate, angular momentum and magnetic flux transport is reached (see Figure 1a). For the resolution chosen (i.e., in the innermost region), it has been shown ([43]) that the macro-scale dynamics of the disk have saturated, in the sense that increasing the resolution further does not lead to significant changes in the accretion rates or the magnetic fluxes. While increasing the resolution would allow the study of particles with lower energies, the limited gain in the dynamic range of particle energies does not justify the substantial rise in computational cost at this exploratory stage. Moreover, maintaining a lower resolution enables running multiple simulations to explore the parameter space more effectively. The choice of resolution is also informed by the properties of the MRI, as it can be shown ([35]) that for a given rotation profile (assumed Keplerian for simplicity), the maximum growth rate of the MRI in the disk is given by (the reference position being chosen to lie within the disk of the SANE cases, that are much smaller than the MAD cases, and thus have higher growth rates), with wavenumber (for the innermost boundary of the jet, where , ) which is well resolved in the present simulations of high-, thick tori ().

Figure 1. Summary of the most relevant global properties of the GRMHD simulations analyzed in this paper.
Injection of tracer particles
Particles are injected at in the SANE case and at in the MAD case, within a spherical shell at , with velocities aligned with the local fluid flow. Initial Lorentz factors (see Tab. I) are drawn from a uniform distribution, and particles are assigned a realistic charge-to-mass ratio of (in SI units: C/kg) (for protons). Because of the realistic ratio chosen, the gyration frequency of particles can quickly grow to values that impose excessively small time steps to resolve the gyration. Focusing on particles with higher factors increases the Larmor frequency, rendering the computational task manageable and provides an interesting scenario to investigate the physics of cosmic-ray acceleration.
The ranges of factors used to initialize particles are chosen such that, at their (spherical) injection location , their Larmor radii satisfy , assuming initial velocities purely perpendicular to the field. Here, is measured in the coordinate frame, in which is defined. This choice ensures that the Larmor orbits are resolved well for the spatial resolution of the fluid grid, even as particles enter the innermost regions where the magnetic field increases significantly. The broad range of Larmor radii further reflects the intention of maintaining a consistent energy distribution across all injection locations. Incidentally, since both the magnitude of the metric coefficients and of the magnetic field magnitude scale approximately as , in the absence of dynamics other than gravity, one only needs to ensure resolution of the orbits in the innermost regions with largest and will automatically be resolving the Larmor motion at all position. For realistic applications, where there are additional dynamics, a margin of safety should be considered, and additional points in regions where the magnetic field is strong but the resolution weaker than closest to the BH should be monitored for under-resolution, as done here.
Indeed, with these initialization parameters, assuming constant, or slightly decreasing, particle energy during motion, the spatial resolution of orbits is ensured, for the high-spin MAD case, up until . The high-spin MAD case is used as reference since it is the one with highest magnetic field closer to the event horizon. This requires that the artificial particle sink be set at this radius, and for ease of comparison we set the sink at the same radius across all simulations. At this radius, for the parameters chosen, particles have Larmor radii satisfying . Future studies will aim at using finer resolutions for the GRMHD grid, that will allow to move the artificial sink closer to the actual event horizon, and thus probe dynamics even closer to the BH.
For the simulations presented in this paper, a resolution of or better (the inner most region has a resolution of ) is maintained until the domain reaches . Crucially, as one moves further away from the BH, the magnetic field drops in magnitude, leading to larger orbits and compensating for the weaker resolution. Notably, in the SANE simulation, the fall-off of the magnetic field is quicker than in the MAD cases, thus interesting dynamics in this simulation end up being mostly focused in the innermost region, with particles quickly switching to an almost free-streaming regime.
The chosen Larmor radius range for initialization also sets the scaling for the simulation results, as the particle dynamics scales with the ratio . While the Larmor radius must be resolved by the grid and the Larmor frequency by the time step, the Larmor radius must also be large enough to allow interaction with turbulent structures in the flow, as these are the ones that can impact the propagation properties. This is why the ratio was chosen for this study, as it provides a balance of resolution and interesting dynamics. Simulations where the initialization was chosen to satisfy the criterion directly (rather than having energies within a certain range) were also run, with no remarkable differences from those presented here. However, setting the Larmor radius rather than the energy leads to a bias where the most energetic particles are located preferentially at regions of higher magnetic field, which is not desirable for obtaining a comprehensive picture of the particle dynamics. The Larmor frequency of the injected particles has to be resolved, thus we implement an adaptive time-step solution that ensures the condition is satisfied, from the time-point when particles are injected onwards, i.e., the Larmor orbits are resolved with at least two points for all particles (the worst-case scenario of two points likely occurring for only a few particles in the most extreme regions). Since AthenaK sets a global time-step, this then becomes the time-step for the evolution of the system ([24]). Since the pusher we are using is implicit, its stability is guaranteed, regardless of how many points are used to resolve the orbits. However, in order to sample the interaction of particles with the fluid, the more points one can use to resolve the orbit, the better. As particles leave their injection locations towards regions of higher magnetic field, the time step can decrease by up to two orders of magnitude. The chaotic nature of the particle dynamics also introduces unpredictability in the evolution of the adaptive time step; jumps of more than one order of magnitude between cycles can, for example, be observed when a particle that was extremely close to the horizon ends up being swallowed.
For reference, the time step used to evolve the GRMHD simulations alone is , the injection of particles leads to an average time step , but temporary drops to are observed. Particles and GRMHD fluid are evolved together for an additional (a little over one orbital time at the original position of the pressure peak in the MAD tori) in the MAD simulations and for the SANE cases, after which time most particles have left the innermost region of interest to our study.
Injecting particles with velocities exceeding those of the background flow means that an injection mechanism capable of generating particles of such energies is assumed, but not directly simulated, a limitation of the current approach that will need to be addressed in future work. On the other hand, it is highly interesting to examine the dynamics of energetic particles, as it is not trivial to determine their fate solely on heuristic scale arguments (e.g., the Hillas criterion ([49]).). In general, injected particles should not have excessively large Larmor radii, as the magnetic field generally satisfies (outside the jet region and some very localized areas) and thus provides little contribution in confining particles that have energies (scaled to the Larmor radius) significantly exceeding that of the bulk fluid flow. Intuitively, the initial torus is in hydrostatic equilibrium, in a stable orbit with Keplerian velocity. If its velocity were any larger, it would escape from its orbit, and likewise particles with energies significantly larger will be almost free to escape the system without feeling any confinement effect from the disc. Furthermore, analytical approaches, tested through simulations, to the interaction of particles and MHD-like turbulence ([50]) suggests that scales much below the Larmor radius will not have significant effects on the transport of particles.
A total of particles is injected in each simulation. A summary of the simulation runs and their parameters is shown in Table I.
Analysis of results
We first perform a statistical analysis of the particle dynamics to identify general trends and guide the subsequent analysis, that consists of an in-depth look at the histories of individual particles to understand how the interaction of the particles with the MHD accretion flow determines their dynamics.
| Parameter | MAD (M0) | MAD (M9) | SANE (S9) |
| 20.0 | 20.0 | 6.0 | |
| 41.0 | 41.0 | 12.0 | |
| 100.0 | 100.0 | 100.0 | |
| 3.0 | 3.0 | 1.0 | |
| 400.0 | 400.0 | – | |
| 3.0 | 3.0 | – | |
| 0 | 0.9375 | 0.9375 | |
Table 1. Summary of the torus (eq. 9) and particle injection parameters used in the GRMHD simulations. Energetic particles are injected in a spherical shell with . A zero-spin BH () is explored only for the MAD case.

Figure 1a. Comparison of the spherically integrated angular momentum transport rate (), the magnetic flux (), the mass accretion rate () and the energy transport rate () in code units at . Black arrows indicate the time of injection of the particles.

Figure 1b. Slice at , averaged over in the quasi-steady phase, of the magnetization (top half) and the radial (in the spherical sense) 3-velocity (bottom half), comparing M9 (right half) and S9 (left half). The dashed black line in the bottom half indicates the zero value where the direction of inverts, with an additional dotted line on the side of the negative values.
Different particle initializations were tested (e.g., initialization based on a density threshold in the fluid or at different radii). These did not lead to significant changes in the properties reported here, but made it more difficult to assess quantities like the escape times, as it becomes less straightforward to compare escape times for particles initialized at the random location where the fluid meets certain criteria. Furthermore, not having control over where particles are initialized makes it more difficult to identify a set of parameters that ensure proper resolution of the orbits for all particles involved in the study.
Statistical Analysis
In order to carry out the statistical analysis of the results, a new, spatial (cylindrical) grid, independent of the one used to carry out the simulations, is introduced during post-processing. Such a grid is used to ensure that enough particles are being sampled when estimating statistical quantities, and to facilitate the mapping of the particle quantities to the fluid dynamics, as the fluid quantities can also be mapped to this new grid by coarsening or interpolation, depending on the relative resolution of the simulation and post-processing grid. The grid is also discretized in energy space for the particles. The spatial domain of the simulation within a box is projected to cylindrical coordinates with a resolution . Initially, 25 linearly spaced bins in the energy domain (MAD) and (SANE) are used to study the overall particle behavior. To capture acceleration to very high energies and avoid artificial accumulation near the injection upper limit, a broader analysis with 25 logarithmically spaced bins over is also performed (see Fig. 5b), enabling the observation of a power-law regime in the distribution.
Additional analysis is carried out on a grid with resolution , covering the domain , in order to examine the dynamics further away from the BH.
Figures use colormaps to represent the quantity that is the subject of the figure, which is, for the smaller domain, an average involving hundreds to thousands of particles for “single-contribution” quantities (e.g., the escape time, since each particle can contribute only once to only one cell) and thousands to tens of thousands for “multiple-contribution” quantities (e.g., the trapping time or the , as each particle can contribute at multiple locations with different energies at different times). Having a sufficiently large number of particles contributing to the averaging is particularly important for energy-related quantities, since particle gyration induces rapid, small-amplitude oscillations in the Lorentz factor (see Fig. 5a). These oscillations can be neglected only when averaging over enough particles for their net contribution to cancel out. To facilitate interpretation of the results, contour levels of the magnetization (where is the magnetic field in the fluid rest frame, obtained from the Hodge dual of the electromagnetic tensor ) and the rest-frame density , determined from the fluid diagnostics, are also shown (based on data averaged over ). These values provide a reference to identify where the magnetic energy becomes dominant over the fluid rest mass, as well as separating the disk from the low-density regions.
The analysis of the particles’ trajectories is done with a temporal resolution of (diagnostics are saved every ).
Escape Times
Estimating escape times from numerical simulations
Having access to the full history of the particles in the system allows for the study of many characteristics of their dynamics. In this section, the focus will be on the escape and infall times ( and respectively) of the particles injected as described in Sec. III. Here, is defined as the difference between the last time-point before a particle leaves the region of interest at the outer boundary (i.e., at and ) and the time the particle was injected. Similarly, is defined as the difference between the last-time point before a particle leaves the domain at the artificial sink at and the time it was injected. This allows to reconstruct maps that indicate the distribution of and within the disk, that can be averaged over the energy of the particles or further broken down by energy range. does not take into account particles that might re-enter the region of interest at a later time (see Sec. IV B), and is thus more applicable to the usual approach (e.g., Lemoine and Rieger 2025, Fraschetti 2021) of assuming particles are injected at one location and eventually diffuse out of the region of interest after a typical time .
The choice of the outermost point at and for the more detailed analysis carried out here is somewhat arbitrary, but it can be justified. The dynamics that are specific to a GR environment, such as the generation of powerful outflows (Blandford and Znajek [38]; Blandford and Payne [51]), and that are likely to have the strongest effect on particles, occur in the innermost parts of the accretion flow (see Figure 3), as this is where the metric deviates the most from the flat-space metric and the dynamics of the flow are most affected by non-Newtonian effects. Secondly, with the long-term goal of applying these results to interpret the neutrino detection by IceCube (Collaboration et al. [5]), the region of the accretion flow that is of greatest interest for this purpose is the so-called “corona” (Padovani et al. [7]; Liang [52]; Murase [53]; Middei et al. [54]; Dash et al. [55]; Dhaka et al. [56]), that is inferred to have an extension of a few tens of . Lastly, as is usual for GRMHD simulations, the initial condition of the weakly magnetized torus should be sufficiently evolved to have dynamics that are not strongly dependent on the initial setup (e.g., Narayan et al. [13]). In the outer parts of the torus, in particular along the mid-plane where the outflows from the BH do not have direct effects, this requires much longer evolution times than those considered in these simulations, and thus some artifacts are likely to appear (see Sec. IV B).

Figure 3. Map in the plane of the base 10 logarithm of absolute value of the gradient of the fluid velocity magnitude (right half) and of the current (left half). Overlaid with a solid line is the contour for the Bernoulli number , and with a dotted line for the magnetization . These contours are nearly coincident in most cases, motivating our use of alone to identify the funnel region in the remainder of the paper. Localized maxima in the funnel () region are due to the density floor used to treat the vacuum regions. The left colorbar refers to the current, the right one to the velocity gradient.
There is also a technical reason, as mentioned in Sec. III, since the resolution used in these simulations only allows to resolve the dynamics of particles with up to and , thus dynamics further out are less trustworthy, even though the drop in magnetic field means that Larmor radii can grow significantly and it will be shown (see Sec. IV B) that interesting dynamics can occur even further from the BH. Note that at sufficient large distances the lapse function , meaning that the Lorentz factor measured at infinity is captured by the quantity , which will be convenient when analyzing the particle distributions in Sec. IV A 2. We often use the Lorentz factor as a proxy for the energy of a particle at infinity, , with the latter given by . Far enough from the BH and , whereas closer in, the presence of a gravitational well requires care in distinguishing changes in the total energy from changes in the kinetic energy (the latter captured by ).
tbl:2 summarizes the main results regarding the escape times for all simulations broken down by the spawning energy of the particles. In SANE disks, weaker magnetic fields that fall-off more rapidly with distance from the BH lead to larger gyro-radii, limiting the interesting dynamics to low-energy particles. Overall, of the particles initialized within the spherical shell, those whose poloidal coordinate positions them in the funnel-sheath regions (i.e., in regions where ) have the largest , which may seem surprising: in these simulations, particles are initialized with velocities aligned with that of the fluid, that close to the BH, and away from the mid-plane, is directed as a strong outflowing wind. One could thus expect particles in this region to leave the system quickly. However, the funnel-sheath region is where the strongest turbulence is located, and, as will be shown, turbulent scattering plays a major role in the dynamics of the particles, including several deviations of the particles’ trajectory, that increase the time they remains in the region.
tbl:2 shows, somewhat unsurprisingly, that the highest energy particles mostly escape at infinity in these simulations. This is readily explained by the fact that the particles in question have large Larmor radii, and are thus less confined by the magnetic field. Other than the magnetic field, the only confining factor is the gravitational potential, that has weak effects far from the BH horizon on particles at such high energies.
Nevertheless, a significant fraction of particles are captured by the artificial sink at (see Table II), even though their energies may lie in the higher end of the spectrum. This may reflect the acceleration of lower-energy particles through interactions with the disk as they approach the innermost regions of the accretion flow.
Comparing simulation results to analytical estimates
To explore the overall transport behavior of particles in the system, their statistically averaged properties are compared to simple estimates from three canonical transport models: advection, free-streaming, and uniform diffusion. If the particle dynamics match the advection estimate, then particles effectively behave as ”frozen-in” to the fluid, with their transport governed by fluid motion. In contrast, free-streaming represents the opposite extreme, with particles moving independently of the fluid, responding only to gravity. Since the GRMHD simulations show that the poloidal velocity is, on average, one order of magnitude smaller than the azimuthal velocity (and thus , see Tab. III), our free-streaming estimate assumes that a particle has only azimuthal and radial velocity at injection, such that the ratio between the two is sufficient to characterize the velocity considering that . The diffusive regime, often modeled via Fokker-Planck approaches, can incorporate varying levels of complexity (e.g., Lemoine and Rieger 2025, Fraschetti 2021, Lemoine 2023, Effenberger and Petrosian 2018), with the simplest case assuming constant (isotropic) diffusion in space and momentum. In the present work, we adopt an idealized diffusion coefficient as given by
with values for provided in Table IV, and the geometrical factor for diffusion in 2D.
| Simulation | |||||||
| M9 | |||||||
| M0 | |||||||
| S9 | |||||||
Table IV. Summary of different properties for particles in different energy regimes related to the analysis within the region and . refers to the time a particle stays in one of the cells defined for the post-processing grid (cells have size , , ), without considering azimuthal motion, i.e., considering only motion that might lead to particle loss. The acceleration rate is defined as . The diffusion coefficient is defined as .
Since the particles in our simulations have large Larmor radii, a well-defined diffusive behavior may, however, not be reached. A uniform diffusion approach with the nominal coefficient above, in fact, breaks down quickly when comparing the average measured particle properties, summarized in Tab. II, to the estimates in the tenth column of Tab. III. While a more sophisticated choice for the diffusion coefficient might have given more satisfactory comparisons the significant gaps between the values in Tab. II suggest that uniform diffusion alone may not sufficiently capture the transport dynamics. The exploration of alternative models, such as space- and energy-dependent or anisotropic diffusion, is however left for future work.
The high-spin MAD case M9 has infall times that can be significantly larger than those determined from advection, depending on the poloidal region considered, and exceed the free-streaming estimate by more than a factor of ten. Because of the smallness of (and thus ), the area considered for diffusive transport of particles is a circular sector covering . The escape times for particles in the higher energy range (see Tab. II) can approach free-streaming estimates, indicating that particles with large Larmor radii interact weakly with the magnetic field in the accretion disk.
This behavior is also observed in M0. M0 also shows estimates for infall times from advection that are compatible with values measured directly from simulations, especially further away from the midplane, though advection severely overestimates the escape times. This suggests that if particles do not spend sufficient time interacting with the disk, advection may provide a reasonable estimate of their motion. Similarly, the (uniform) diffusive estimate is reasonable for in M0, primarily because particles, initialized with velocity aligned with the fluid, have the largest (relative)
| Simulation | |||||
| M9 | (81644/94138) | (267956/618493) | |||
| (4243/94138) | (154996/618493) | ||||
| M0 | (157011/258111) | (242827/612119) | |||
| (40211/258111) | (158271/612119) | ||||
| S9 | (5018/34773) | (80363/1091104) | |||
| (7398/34773) | (258854/1091104) |
caption: TABLE II: Summary of different properties for particles in different energy regimes related to the analysis within the region and . refers to the time it takes particles to leave this domain at the outer boundaries after being spawned. refers to the time it takes particles to fall inside the BH region after being spawned. and refer to the fraction of particles that either escape at the outer boundary or fall into the black hole region at a given energy range with respect to the total number of particles that have (respectively) escaped and fallen in.
:::
radial velocity in this case, making poloidal diffusion a more accurate proxy than in other configurations.
In S9, escape times are compatibles with free-streaming estimates, as expected from the distinct magnetic field scaling in MAD and SANE regimes (Sec. III). Particles in S9 interact significantly with the magnetic field only at radii smaller than their injection radius, but are almost freely streaming further out, as their Larmor radii become too large for interaction. This is supported by infall times, that are significantly longer than the free-streaming estimate, while also not matching the advection estimate, indicating more complex interactions between particles and fluid close to the BH. Although escape times are compatible with free-streaming, they are systematically smaller, suggesting that interaction between particles and fluid close to the BH tends to deflect particles radially outwards, giving them additional radial velocity. This effect might be aided by gravitational assists from the BH, though particles in the simulations rarely complete undisturbed orbits due to frequent scattering events.
Overall, M9 provides the clearest example of particle dynamics deviating from simple estimates based on fluid properties or uniform diffusive models.
There seems to be a “trend” (not shown) across the simulations where particles initialized at the lowest energies tend to escape along the jet (or, more broadly, the central region of higher magnetization, i.e., the funnel), whereas particles with higher energies escape more isotropically, with a slightly over-density around the mid-plane. The relevance of this result is, however, uncertain, as not enough realizations of the simulations have been carried out to assess whether this is an effect related to the particular initial conditions chosen.
Overall, Table II points to M9 (high spin) as being the simulation where particles stay longest in the system at the lower end of the energy spectrum. This is in line with the magnetic field being the dominant driver for the dynamics of the particles as explored in the following. In M9, the magnetic field magnitude decreases more slowly with radius ([58]) due to the bursty behavior that effectively ejects magnetic field into the disk with magnitude obtained by expanding the flux that was “frozen-in” at horizon scales. This means that, for the energy range chosen, interaction between the magnetic field and the particles extends for larger regions of the domain compared to the other simulations.
The parameter range chosen for M9 seems to identify a threshold in the particle energy at where particles initialized in the disk region (with fluid density ) have a significantly decreased likelihood of falling into the BH. This is probably a combined effect of the higher energy and the more effective scattering against the turbulence in the region for particles with a particular , but more detailed studies than those presented here are needed to draw firm conclusions. It will be particularly interesting to extend the energy range of injection to potentially identify other similar thresholds, also for the other cases, as well as regarding escape.
Expanding the region of interest for this analysis to and , the average exceeds in the MAD cases, which is still larger than the naive expectation , and shows less variance with respect to particle energy. Infall times can also increase, but this effect is only significant in M9, where a dramatically increased is recorded. This is related to the much stronger effect of turbulence on particles in M9, and to the phenomenon analyzed in Sec. IV B of particles being reflected multiple times while traveling along the funnel, which makes it more likely for them to end up inside the BH.
These results highlight the necessity of stochastic models that capture both coherent transport and turbulent scattering to accurately describe particle dynamics in magnetized accretion flows.
| Simulation | in/esc | ||||||||
| M9 | in | 0.0 | 0.05 | 110 | 2.74 | 0.94 | 6 | ||
| M9 | in | 0.349 | 0.103 | 53 | 3.47 | 0.96 | 6 | 0.255 | 28 |
| M9 | in | 0.378 | 14 | 18.21 | 0.99 | 5 | |||
| M9 | esc | 0.0 | 0.05 | 1.2 | 0.77 | 72 | |||
| M9 | esc | 0.349 | 0.103 | 544 | 3.47 | 0.96 | 58 | 0.318 | |
| M9 | esc | 0.378 | 148 | 18.21 | 0.99 | 56 | |||
| M0 | in | 0.0 | 0.05 | 110 | 1.56 | 0.84 | 6 | ||
| M0 | in | 0.349 | 0.14 | 39 | 5.89 | 0.98 | 5 | 0.37 | 20 |
| M0 | in | 0.272 | 20 | 92.74 | 0.99 | 5 | |||
| M0 | esc | 0.0 | 0.02 | 1.86 | 0.88 | 63 | |||
| M0 | esc | 0.349 | 0.14 | 400 | 5.89 | 0.98 | 57 | 0.343 | |
| M0 | esc | 0.272 | 205 | 92.74 | 0.99 | 56 | |||
| S9 | in | 0.0 | 0.03 | 180 | 0.27 | 0.26 | 21 | ||
| S9 | in | 0.349 | 0.018 | 305 | 0.38 | 0.36 | 15 | 0.335 | 24 |
| S9 | in | 0.051 | 108 | 0.88 | 0.66 | 8 | |||
| S9 | esc | 0.0 | 0.1 | 560 | 0.54 | 0.47 | 119 | ||
| S9 | esc | 0.349 | 0.018 | 0.38 | 0.36 | 155 | 0.348 | ||
| S9 | esc | 0.049 | 0.63 | 0.53 | 105 |
Table 3. Summary of mean transport properties estimates for energetic particles obtained with different methods. For each simulation the infall (“in”) and escape (“esc”) properties are determined. The infall transport covers a distance (from the injection position at to the artificial sink at ). The escape properties are determined over a distance (from the injection position at to the outer boundary of the region of interest ). Inflow and outflow could not be distinctly separated at all locations. For regions exhibiting a single behaviour, the sole available value was adopted. is the radial component of the fluid 3-velocity, from which the “advection” estimate is derived. For the outward (“esc”) directed properties, this velocity reflects the average wind velocity in the funnel-sheath region, being used to e.g. evaluate . The estimate for the free-streaming time-scales are obtained from the initial ratio of radial to azimuthal velocity of the fluid , which determines the radial component of the particle . The diffusive parameters , are obtained as the averages of the values in Tab. IV, using values for for and values for for , and used to evaluate . The angle corresponds to .
Particle Acceleration
Acceleration behavior in different accretion regimes
In discussing particle acceleration, we emphasize that the particles considered here are purely passive tracers and thus exert no feedback on the disc. While a more complete treatment would be desirable, the results already provide valuable insight into the underlying acceleration processes.
In analyzing the acceleration, we rely on , the Lorentz factor measured by the local FIDO/normal observer. This quantity is easily accessible from the code and in principle well suited to describe the local radiation physics. However, it has the drawback of depending on the metric explicitly, thus the values reported closest to the compact object require care when being interpreted. The analysis follows each particle as it travels through the disk and records the variations of its Lorentz factor as it enters and leaves the cells of the post-processing grid (though motion in the azimuthal direction is neglected, i.e., only crossing cells in the and directions is considered), thus obtaining a picture of which regions of the disk result in larger variations of the particle energy, and in which “direction” (acceleration or deceleration).
As might have been expected, the presence of more strongly magnetized regions, leads to a noticeable localization and intensification of the interaction of the particles with the fluid (through the electromagnetic fields generated by its dynamics), clearly shown in Figure 2, in the region around the contour, encompassing the region of the funnel, across all simulations. We note here that these regions are where the transition from the disk to the low-density zone at the center of the funnel occurs. In this area, the numerical density and pressure floors (, ) required for the GRMHD computations—which become particularly relevant in the innermost part of the funnel [18, 30, 43, 59]—may influence the structure of the flow. In particular, they could affect the magnitude and spatial extent of the gradients in the fluid quantities. Consequently, while the qualitative trends seem robust, their precise quantitative characterization may depend, at least in part, on the adopted floor prescription. The region where the floors impact the fluid quantities is visible in Figure 3 as the noisy central part of the velocity gradient.

Figure 2. (a) M9, analysis done in , . (b) M0, analysis done in , . (c) S9, analysis done in , . Azimuthal average of the average energy change for particles entering and leaving a specific cell. Since the magnitude is shown as the of its value, the red (blue) colorbars indicate positive (negative) values of the change in . The dotted and solid lines indicates the contour levels and respectively, highlighting the role played by regions with strong gradients and turbulence in the acceleration of particles.
Going into detail for each of the simulations run, M9 shows the largest relative acceleration (or deceleration), both closer and further from the BH, with increasing intensity along the funnel region. The disk can still affect particles’ energies, though its importance is secondary to the funnel region. Particles are unlikely to stay in the innermost regions of the funnel as they propagate outwards, and instead move along the region of transition between the funnel and the disk, where most of the acceleration takes place. It is also worth noting that, although the magnetization attains its maximum in the funnel core, the magnetic field strength peaks close to the denser funnel sheath, which may help to explain the localization of the acceleration.
M0 shows similar behavior, though the innermost region of the funnel seems to be less important than in M9, possibly due to the overall lower values of magnetic field reached when the BH has lower spin. However, the relative acceleration achieved by particles is lower than in the case with larger spin, which suggests a relevant role played by the magnitude of the magnetic field (and its fluctuations) in the acceleration mechanism. Crucially, however, the behavior close to the BH, is very different between the two MAD cases: M9 shows interaction between particles and disk inside the funnel region close to the BH, while M0 exclusively shows interaction outside the region. Associating this behavior with the dynamics of the accretion disk across the two simulations may suggest that the stronger magnetic field and outflows in the M9 case close to the BH play a role in the localization of the interaction. Alternatively, it is possible that there is a common acceleration mechanism at play in the region where the funnel (the jet in M9) transitions to the disk (the sheath), and an additional one inside the jet of M9.
Indeed, Figure 2c shows that even in S9 the funnel region is the seat of intense interactions between particles and fluid, though S9 is the least efficient accelerator of the cases analyzed.
These acceleration regions are most effective for particles at the lower end of the energy spectrum, whereas particles with higher energies experience almost exclusively decreases in energy.
We remind that changes in in the region closest to the BH can be attributed to the role of the metric, such that the only effect visible in a simulation run (not shown here) for a SANE case with 0 spin is the change in due to the particles leaving the gravitational potential (with an almost exact dependence of the energy change).
We also note the “arch-like” features in the localization of the acceleration in Figure 2c, that match rather well the regions of strong gradient in Figure 1b, and are a strong indication that the acceleration mechanism, also in the disk, is strongly correlated with gradients in the properties of the fluid. Note that the MAD cases show negligible acceleration features in the denser regions of the disk, a difference that might be attributable to the different flow structures in MAD and SANE disks, as the SANE cases are much closer to an axisymmetric flow ([13, 16]), that can thus act “consistently” (systematically), rather than stochastically, on the particles.
Indeed, in the absence of fluid-particle interactions, would simply decrease as particles move outward due to receding gravitational time dilation and kinetic deceleration, explaining the large regions of average negative , especially at higher energies, where interactions with the fluid become less frequent. Only when interactions are frequent enough does turn positive, as is e.g. the case in the funnel sheath regions of all the simulations. At the same time, this means that one needs to ensure that a sufficient number of particles are involved in the averages, as the negative changes in from free-streaming are the default behavior for most particles, especially at high energies. Alternatively, the differing degrees of magnetization (see also Figure 1b) may be the key factor determining the distinct acceleration behaviors, a point that will be further discussed at the end of this section. It is possible that the low efficiency of S9 in accelerating particles stems from the quick transition of particles to the free-streaming regime, which could be mitigated by choosing different injection energies, with all the challenges, mentioned in Sec. III that come with resolving those orbits reliably across the whole domain.
Tentative extrapolation would suggest that particles with even smaller Larmor radii would then be more efficiently accelerated by the fluid dynamics in all regimes, though this remains to be verified. Still, the results presented here show that even when particles have reached significantly larger energies than the background fluid, both MAD and SANE disks can continue to accelerate them further. This, in turn, suggests that the distribution of energetic particles is not only shaped by local injection and acceleration mechanisms, but also by their longer term dynamics and interaction with the disk, as will be explored further in Sec. IV B.
The average distribution over the whole domain over all times for M9 is shown in Fig. 5b to form an extremely hard tail, through dynamics that are explored more in detail in Sec. IV B. The tail can be fit with a power-law index close to (see Fig. 5b, which is compatible with previous studies ([60, 61]), although in the present study we use a fully GR system and our results have the strong limiting factor that the injection is not continuous, and thus does not establish a proper steady-state condition.
This type of analysis could be used as a method to validate predictions of spectral shapes from theoretical arguments (e.g., (Lemoine and Rieger 2025)), although to carry out this task more accurately a mechanism for continuous injection of particles and feedback of the particles on the fluid are crucial components that limit the applicability of the results shown here.
Understanding the acceleration dynamics
The two key candidates to explain the observed acceleration behavior are intense local electric fields from magnetic reconnection [30, 62, 63, 64] and scattering on structures of the magnetic field, whether it be turbulence, curvature or shear [50, 65, 66, 67]. In the present case, reconnection is due to grid-scale numerical dissipation, since no resistive terms are implemented. Distinguishing the two mechanisms can prove difficult, since the currents required to drive reconnection can be due to turbulence, or be the result of an enhancement of a local shearing of the magnetic field by the flow dynamics, while the consequence of magnetic reconnection can be an enhancement of turbulence. Exploring which of the two is at play here will be the goal of the remainder of the section.
Figure 3 show the current density and the velocity gradient distribution in the system. Since the areas we are focusing on are far from the compact object, we use the non-relativistic definition and compute the current as (i.e., we neglect the displacement current), and the gradient as . In principle, computing the current this way mainly provides a visualization of irregularities in the magnetic field. As the maps are obtained from data averaged over , they do not capture the effects of transient reconnection events. However, one can expect that most reconnection sites will coincide with regions of strong average currents, reflecting the fact that magnetic reconnection is due to current sheaths (usually already present) becoming excessively elongated due to the dynamics of the flow. Such current sheath are then “converted” to electric fields capable of accelerating particles by the resistivity. It is noticeable that the region with the most intense current density is the inner () mid-plane, and its localization aligns well with the disk regions highlighted by the analysis in Figure 2b and even more Figure 2c, where significant acceleration takes place in the areas of the disk closest to the BH.
However, looking more closely at Figure 3a, one notices that the current in the jet region (enclosed by the isocontours ) tends to weaken away from the BH, while the region of acceleration within the jet in Figure 2a seemed to be rather uniform where present, though we stress that the innermost part of the jet is prone to numerical artefacts, and drawing conclusions from what is observed in this region should be done cautiously. Crucially, if the main acceleration mechanism were currents related to magnetic reconnection, one should not expect a threshold in the energy of the particles where these stop being accelerated and are instead decelerated. If anything, one could expect a dependence of the acceleration and deceleration on the region of the domain, but constant across particle energy, which is not observed here. Finally, if reconnection were the main driver of the acceleration, one should expect, on average a change in the particle energy of the order , which can be seen not to be the case by comparing the maps in Figure 2 and Figure 1b.
The energy dependence of acceleration and deceleration is much more consistent with the alternative explanation of scattering: particles with lower energies than that available in the turbulence, or whose Larmor radius is one the same scale as the structures in the magnetic field will, on average, gain energy by interaction with the field, whereas particles with higher energies will lose energy (even though in these simulations we do not actually subtract or deliver this energy from and to the fluid). This latter point is of interest also for future work, where the feedback of particles on the disk could be included in the simulations, and high-energy particles losing energy to the fluid could modify the dynamics on larger scales. Finally, from the spectra of the magnetic field fluctuations (see App. § 1.13) one can see how further from the BH the fluctuation level quickly exceeds , indicating much stronger fluctuations at the same locations where the acceleration intensifies. Considering these points, and the ones that will be further provided in Sec. IV B, the possibility that magnetic reconnection is the main source of acceleration seems rather unlikely. The currents shown in Figure 3 are then a useful way of visualizing and quantifying these structures in the magnetic field, rather than a means to look into reconnection.
Interaction of the particles with the dynamical structures of the magnetic field is the main alternative explanation for the acceleration behavior. In particular, the region of the funnel-sheath is, as already mentioned, a region of strong and chaotic velocity gradients (see Figure 3), and can thus drive instabilities like the KH mode, that in turn result in turbulent structures that can interact with the particles, leading to rich dynamics, and, through scattering, to stochastic acceleration [61, 65, 68, 69, 70]. Another strong indication that turbulence is the main acceleration mechanism is the similarity in the interaction patterns in Figure 2c (they look like spherical shells localized around the mid-plane, that become more or less visible depending on the binning on the particles) and the structure of the inflow (radial ) velocity shown in Figure 1b. The acceleration localizes strongly in regions where the gradient of the inflow velocity is the strongest. Thus the localization of the acceleration regions in Figure 2, together with the analysis presented in Figure 3, provides the first indication that turbulence is the primary acceleration mechanism.
In order to explore this further, we perform coarse-graining [68, 71] on the MHD fluid, even though the dynamic range of the resulting spectrum is limited in range. The need to use coarse-graining arises from the strongly anisotropic and inhomogeneous nature of the system we are analyzing, that also lacks clear periodic directions, where Fourier modes work best. The details of the coarse-graining can be found in App. § 1.13, here we mention the definition of the fluctuation amplitude for the magnetization and the “wave-number” at scale , .
Here, is a typical scale of the system, and the “high-pass-filtered” magnetization is obtained from the coarse-grained through . The spectra shown here are obtained with isotropic convolution, without distinguishing particular directions. Coarse-graining allows to extract the spectra shown in Figure 4. Comparing these spectra to Figure 3, Figure 2a and 2b, one can notice how they can easily explain the localization of the acceleration: the regions where the turbulence is stronger in Figure 4a (i.e., at locations (15, 0, 20) and (20, 0, 40), where in this notation the numbers refer to the cartesian coordinates of the point and no longer to the cylindrical ones) match rather well the most intense acceleration regions along the jet-sheath. Furthermore, regions with similar turbulent spectra across the two MAD cases (i.e., spectra at (30, 0, 40) and (30, 0, 20)) are also the regions that showed similar acceleration behavior in Figure 2, whereas the current was varying significantly (Figure 3) in regions where the acceleration behavior was similar. Note that the funnel region is known to be the seat of intense turbulent heating across SANE and MAD cases [72], providing further strength to our interpretation.

Figure 4. High-pass filter spectra of the magnetization at the locations indicated in the heatmaps at the bottom right of each sub-figure, for simulation M9 and M0. The heatmaps show the high-pass filtered field. The dashed vertical lines show the Larmor radii for particles with at each location, while the solid vertical lines show the computational grid size at that position. The squares in the bottom right portion of the figures indicate the maximum extent of the coarse-graining around a given point.
The different behavior in the funnel region closest to the BH can be understood by noting that in the funnel region the magnetic field in M0 is much weaker than in the disk, which is not the case in M9. As a results, the Larmor radii in the central region of M0 are excessively large, preventing particles from interacting with the turbulent structures, a further indication that turbulence is the driver of the acceleration. At the same time, this suggests that the choice of as the quantity for coarse-graining might not have been optimal, or that the normalization of the spectra could warrant more careful considerations. Nevertheless, we believe that the analysis presented offers a satisfactory explanation of the dynamics.
The statistical analysis presented here already provides a strong and general understanding of the underlying processes, suggesting that particle dynamics are driven more by large-scale turbulent field structures than by magnetic reconnection. Further evidence for this picture will be discussed in Sec. IV B.
If turbulent scattering is considered as the dominant mechanism for particle energization, it is possible to correlate the behavior of the energy change to the underlying process that causes such changes. Given the location of the most intense interaction in Figure 2, two candidates emerge: (classical) second-order Fermi acceleration and flow-shear acceleration. The distinction can be probed via the expected relative energy change . For second-order Fermi acceleration , where is the Alfvén velocity, is the enthalpy for a magnetized gas (computed using the adiabatic index for a perfect relativistic fluid) and is the internal energy. For flow-shear acceleration [66, 67, 73] , where is the gamma factor of the flow. One can immediately notice how the quick increase in visible in Figure 1b, which can even span several orders of magnitude going from the sheath to the center of the funnel, does not correspond to a significantly enhancement in Figure 2 in the central part of the funnel region compared to the sheath. If anything, intense acceleration, indicated by the darker red colors in Figure 2a and Figure 2b occur along the contour, that Figure 3 identifies as the region of strong gradients in the fluid velocity. One can also notice how in Figure 3b the most intense gradients of the velocity (with the exception of regions affected by numerical floors in the center of the funnel), are localized at the mid-plane around , which matches very well the region of most intense acceleration in Figure 2b. The picture of flow-shear-driven acceleration also removes the need for different acceleration processes acting inside the jet of M9 and in the sheath region of both M9 and M0 discussed earlier in this section, as the behavior of the velocity gradient can explain the difference between these cases. With turbulence as driving mechanism, a cohesive picture emerges of how different accretion regimes affect particle acceleration. In M9, the powerful jet driven by the Blandford-Znajek (BZ) [38] provides the strongest turbulence in the sheath region, where the velocity field accelerates to relativistic speeds all the way to the center of the jet itself. In M0, the BZ process is inactive due to the BH’s zero spin, and the weaker winds from the Blandford-Payne (BP) mechanism [51] act as the source of the velocity gradient. This gradient is not only weaker but also does not extend to the BH’s pole, explaining the reduced acceleration in the innermost funnel region. Consistently, S9 shows weaker acceleration: although the BH has high spin, the available magnetic flux is insufficient to drive a strong jet like that in M9, and lower accretion rates fail to energize BP winds to the same level as in M0, though they still generate mild turbulence in the sheath. The weaker confinement of high-energy particles in S9 further limits acceleration. These insights may also inform our understanding of other accretion regimes exhibiting shear-driven dynamics, such as the hydrodynamic ADIOS model [e.g. 14, 74, 75], where mid-plane inflow transitions to outflows at higher latitudes. However, since the velocity shear layers in ADIOS are expected to be slower and weaker than in the magnetized cases considered here, the underlying acceleration mechanisms might be modified, warranting dedicated investigation.
Finally, as already remarked, using the radial velocity as a proxy for the total velocity (Tab. III justifies this as an order-of-magnitude estimate, at least in the innermost parts of the system), Figure 1b shows that S9 reveals rather uniform and stable in time gradients of the velocity (the arch-like structure that were already highlighted in Sec. IV A 1) through the whole disk. This correlates very well with the behavior of the acceleration shown in Figure 2c, as previously noted. Together, these results strongly suggest that particle dynamics are primarily governed by scattering against turbulence driven by flow shear.
Particle Trapping
Particles occasionally become temporarily trapped in localized regions, where repeated scattering reverses or randomizes their motion over short timescales, effectively confining them for extended periods. This behavior is relevant for assessing how specific system phenomena influence particle dynamics. For example, if a phenomenon lasts but particles remain in any given region for only , its impact is limited. Conversely, if the phenomenon occurs on a timescale , its effect can be significantly amplified for particles trapped longer in the region.
In order to assess this behavior, over the whole trajectory of a particle, its position is mapped to the post-processing grid, and the time each particle spends in a given region of size , is used to compute an average, that allows to observe regions of the disk that are more effective at trapping particles. At the same time, this can be combined with the previous (Sec. IV A 2) diagnostic to get the quantity , that quantifies the efficiency of a given region in accelerating (or decelerating) particles. Not considering azimuthal motion is particularly significant for this diagnostic, as the fluid, far from the BH, moves primarily azimuthally, which means that if advection alone were responsible for particle motion, this could result in very large . This information is also needed to evaluate the nominal diffusion coefficient in the poloidal plane as a function of position (not shown, maximum values are reported in Table IV), as defined in Eq. (10): .
The physical explanation for the trapping will be examined in detail in Sec. IV B, but a trend emerges with the highest average trapping times occurring in the innermost regions (), which points again to scattering by turbulence. Importantly, if the main mechanism impacting the dynamics were magnetic reconnection, there should be an opposite relation between acceleration and trapping, as the reconfiguration of the magnetic field nulls occurring at the reconnection sites are at the same time accelerating and “freeing” particles from their gyration constraint. The localization of strong trapping in high-field regions suggests such zones will play an even greater role in simulations with a corona-like structure, where regions of strong magnetization can form also at lower poloidal angles than the funnel region ([27]).
The absolute maxima of the trapping time reported in Table IV all occurred in the outer parts () of the region of interest. Analysis in the larger domain ( and ) shows the dynamics reaching over narrow portions of the domain (roughly following ) for all simulations, and up to in the MAD cases, though the latter value can be directly linked to the initial conditions of the simulation, as it occurs at which is (in the mid-plane) where the Fishbone-Moncrief torus “ends”, meaning that the magnetic field drops below . Just before decreasing below this threshold, it is almost purely azimuthal, and since the present analysis disregards azimuthal motion for the trapping time, confinement times are “artificially” large. The values of the acceleration rate indicate that the typical particle acceleration process takes place on time scales much smaller than those on which the MHD quantities evolve (see Figure 1a). This implies that the interaction between the particles and the dynamics of the fluid (e.g., the periodic ejections in M9) are going to be limited. The difference between simultaneously evolving the fluid and particles and treating the fluid as a static snapshot is thus expected to be modest.
Individual Particle Dynamics
To better understand the processes experienced by the particles, we focus in this section on a subset of particles that exemplify the larger sample, and study their behavior in detail. Although more particles were analysed, those presented here give an overall picture of the most interesting dynamics occurring in the system. These particles are taken from M9.
Fig. 5a shows, in the left panel, the evolution over time of the factors of selected particles. The right panels show the projection of the trajectories in the (bottom panel) and (top panel) planes.
Inspection of the trajectories in Fig. 5a shows that particles undergoing significant energy changes are those that experience multiple scattering events, and are thus remaining in the domain for rather long times, or those that enter the funnel region. This further supports the interpretation that stochastic acceleration is the main mechanism for particle energization, while magnetic reconnection (or, more generally, acceleration by resistive electric fields) does not seem to play a significant role, at least outside the innermost region () since the currents in the region of interest are not particularly intense or localized and their amplitudes decrease as the distance from the BH increases (Figure 3), whereas the average energy change can occur with almost constant magnitude over much broader regions. At the same time, for higher-energy particles, the largest energy changes take place in between the region of the disk where currents are localized and the region starts (and, as already discussed in Sec. IV A 2, stronger velocity gradients are present), thus it is possible that for these particles electric fields due to grid-scale reconnection may still play a role.
Notably, since, for the most part, particles at such high energies are free-streaming and only occasionally interacting with the fluid, changes in the normal-frame Lorentz factor reflect changes in the metric, that have, roughly, a dependence, same as the magnitude of the magnetic field, ensuring stable resolution of Larmor orbits even for apparently large changes in .
Interaction with structures in the magnetic field
Figure 6 shows 3D snapshots taken in the neighborhood of a particle’s position. The time frame for these snapshots is enclosed in the black square in the right panel of Fig. 5a, which is where the particle is gaining energy through several scattering events. Figure 6 (panels 6a and 6b) show two consecutive scattering events (both scattering events are visible in either panel, the focus is different) occurring in regions where the magnetic field magnitude varies over short distances, both in the parallel and perpendicular directions, and, crucially, where the magnetic field lines are sharply bent [50]. These two panel also show that structures in the magnetic field are advected without being destroyed by reconnection or other processes for significant stretches of time (), and still acceleration takes place to a significant degree, even though it requires several scattering events to have an appreciable effect on the particle’s energy. Panel 6c shows a mirroring event, where the particle interacts with a strong gradient in the direction parallel to the magnetic field, thus leading to the textbook behavior due to the conservation of the adiabatic invariants of motion. Observing this phenomenon in such a complex scenario also allows to point out that the mirroring alone does not have significant impact on the particle energy; it is rather the turbulent environment in which the particle is immersed that leads to overall increase in energy.
Figure 6. Sequence of 3D snapshots of the trajectory of a particle along with magnetic field lines (taken within the black square shown in the left panel of Fig. 5a). The particle undergoes acceleration by being scattered several times on curved magnetic field lines and strong parallel gradients. The trajectory of the particle is shown as discrete diagnostics colored from blue to red with increasing time connected by a continuous black line (particle enters from the right side of the figures). The central (black) point of the trajectory is the particle’s position at the time of the field line snapshot. The field lines are colored (blue to red) based on the base 10-logarithm of the magnitude of the magnetic field.
Overall, Figure 6 provides a good example of how the interaction of the particles with structure in the magnetic field, and not with reconnection events, can lead to energization over several scattering events. This region of the domain also likely represents how the bursty behavior in MAD disks is going to affect particle dynamics, as it will cause structures in the magnetic field to move in a chaotic manner, enhancing scattering. This is however difficult to disentangle from the overall dynamics of the disk itself when it comes to the acceleration behavior of particles. Furthermore, the time-scales involved in bursts are much longer than those involved in scattering, thus a different type of analysis (left for further work) is needed, requiring to combine, statistically, what was shown in Sec. IV A 3 with the temporal statistics of MAD flows.
All of this is consistent with the statistical analysis presented in IV A 2, where it was shown that the dominance of negative changes in can be explained by the fact that most particles, due to their high energies, are predominantly free-streaming or advected through the domain. Particle acceleration is observed only in regions where scattering events are expected to be more frequent (i.e., where gradients are stronger).
A novel behavior observed in particles propagating along the jet
A remarkable phenomenon occurs frequently for particles traveling within the funnel region of the MAD simulations. Some particles propagate for long distances (from 100 – 300 taking around 100 – 300 , as expected for particles with ) along the jet-sheath but then get reflected back towards the BH. This reflection is particularly strong for particles that move from the jet spine to the jet-sheath, and thus are affected by a gradient that is determined by the numerical floor applied to the density. This is already visible in the trajectories of some of the particles chosen in Fig. 5a. To mitigate this effect, we restrict our analysis to reflections occurring at , i.e., outside the region most affected by numerical floors (see Figure 3). While the numerical strength of the gradient and its localization are affected by the artificial floor, the presence of such a gradient is physically expected, thus this effect will occur regardless of the specific floor implementation. It should be noted, however, that at these distances from the BH, the simulation resolution is modest. The typical Larmor radius for a particle involved in reflections sits at , to be compared with the resolution at of , thus the particle dynamics sit just at the limit of the resolution, but these reflections occur, occasionally, also in well-resolved regions, as will be shown below. The occurrence of these reflections along the jet is not particularly rare, showing up as a significant contribution in the average particle distributions. One can notice the significant tail in the average particle distribution of M9 when we extend the range of the energy axis (using logarithmic spacing) in Fig. 5b.
A possible explanation for this reflection can be found in the drift. A particle traveling in the inner part of the jet is experiencing a helical field with a predominant component. It can be accelerated along its motion, and will eventually have a Larmor radius large enough to access the region of radial magnetic field gradient between the spine and the sheath. The resulting force (drift) from is along , and deflects the particle at a given position in the azimuthal direction, generating a centripetal force that puts the particle in an “orbit” encircling the jet spine. If the region of radial gradient of is sufficiently wide, and the particle has a finite radial velocity, this process can occur again, but in the outer layers of the jet the azimuthal component of the magnetic field becomes stronger, with the drift producing a force directed along . Depending on the relative direction of the azimuthal and vertical fields, the combination of these two deflections can result in an inversion of the direction of motion. If at this stage the particle has left the region of radial gradient of , its motion will have effectively been inverted, and the particle will travel towards the BH.
In several cases, particles were observed traveling along the funnel region several times (see Fig. 5a, bottom right panel), gaining energy until they finally escaped the region that reflects them back and forth. This seems to be a viable mechanism that leads particles that already have very high energies () to being accelerated by a factor , which would be a remarkable figure if independently verified. More interesting is the possibility that lower-energy particles could undergo a similar process, as it would provide a very efficient and almost universal mechanism for the acceleration of particles in these environments. If, as this analysis supports, turbulent flow-shear acceleration is a fundamental process, such particles may undergo acceleration in similar ways, as velocity gradients and shears are expected to persist down to scales below the current numerical resolution. A primary challenge is that, to experience the successive deflections that reverse the direction of motion, a particle must already possess a gyroradius large enough to sample the magnetic field gradient. It is also interesting to consider that particles that have been accelerated over the length of the jet may be “recycled” and contribute to the processes and emission closest to the BH with the energy they have acquired from their acceleration along the jet.
Some of these particles, when getting close to the BH, have large enough energies that the magnetic field does not confine them effectively, and they can get a slingshot assist from the gravitational potential and be flung out of the domain. The reason why a similar effect is not observed in S9 is that in this simulation the fall-off of the magnetic field along the funnel region is faster than in the MAD cases, and thus effects due to magnetic field gradients become quickly less important.
Figure 7 depicts one such particle undergoing the process of reflection in a region where the grid resolution is still . This again underscores that the primary driver of the dynamics, whether it be inversion of the direction of motion or acceleration, are structures in the magnetic field related to shear, turbulence, and curvature, since in Figure 7 a particle was captured undergoing an inversion in the direction of motion, and subsequently acceleration, in a region where the magnetic field does not undergo any reconfiguration. Rather, the magnetic field presents strong shears that result in forces perpendicular to the gradient itself, pointing in the direction the particle ends up moving towards. The intensity of the acceleration increases as the particle moves further along the funnel, where the magnetic field is more ordered still. The direction of the field lines, that was rapidly and chaotically changing in the case of turbulence analyzed in Figure 6, is here instead ordered and smooth (neighboring field lines are parallel to each other and only change direction gradually). This case will resembles more the picture of the particles being accelerated by the shear of the magnetic field, which also results in changes in the particle direction, trapping for rather long times, and increases in the energy of the particle.

Figure 7. Sequence of 3D snapshots of the trajectory of a particle along with magnetic field lines (taken within the black ellipse shown in the left panel of Fig. 5a). The particle undergoes acceleration and inversion of the direction of motion by interacting with the strongly sheared magnetic field. The trajectory of the particle is shown as discrete diagnostics colored from blue to red with increasing time connected by a continuous black line. The central (black) point of the trajectory is the particle’s position at the time of the field line snapshot. The field lines are colored (bright to dark) based on the logarithm of the magnitude.
We note that the trace within the black ellipse in the right panel of Fig. 5a, which corresponds to the snapshots in Figure 7, appears to exhibit a sudden, nearly vertical jump. This is due to the frequency of diagnostic sampling not being able to capture the rapid dynamics, but the actual time-steps of the simulation are orders of magnitude smaller (see Sec. III, and the resolution of the gyration is ensured.
As noted in Sec. IV A 1, particles initialized in the funnel region often have the largest , but in that section, re-entry in the and was not considered. When the analysis is expanded to and both and increase significantly. This is not shown due to the decrease in both the resolution of the simulation and the statistics of particles this far from the BH. However, this behavior now finds a clear explanation in these inversions of the direction of motion, that particles initialized in the funnel are more likely to undergo.
Finally, Fig. 5a also shows particles that stayed in the region for a few transits, with enough energy to not be captured by the gravitational well. It is not uncommon, however, that interaction with the turbulent magnetic field lines, which exhibit rapidly changing magnitudes and chaotic directions, leads a particle to fall into the BH due to a ”sudden” change in direction. This, on the other hand, means that particles initially headed towards the BH might be redirected, allowing them to avoid infall and resulting in rather long trapping times in the innermost region. While one could expect the process to be isotropic, as the scattering is random in nature, and thus roughly half of the particles will be directed towards the BH, while another half will be scattered away from it, this may not necessarily be true. Firstly, in the vicinity of the BH the turbulence is unlikely to be isotropic because of strong spacetime curvature and because of the way the dynamics shape the magnetic field.
A systematic study of turbulence properties in GR environments prospects itself as a natural extension of this work, and is currently undergoing. Secondly, although the scattering itself might be isotropic, its effect on the average dynamics of the particles need not be, due to the imbalance generated by particles falling into the BH. Those that get scattered towards the BH end up crossing the event horizon and are thus lost. This also explains why closest to the BH the largest average energy increases occur (see Figure 2), as particles moving inward have increasing from the changes in the metric, but then never leave in the opposite direction, leading to an imbalance, as was already noted in Sec. IV A 2.
Discussion
Particle Transport and Acceleration
After considering the points raised above, we now attempt to assemble a coherent picture of acceleration and confinement from these simulations. When it comes to acceleration, our results in Sec. IV A 2 suggest that scattering, whether off turbulent structures or magnetic field curvature, is the primary mechanisms for particle acceleration across all cases analyzed. The localization of the acceleration region (see Figure 2) is, in principle, compatible with both the regions of intense velocity gradients and turbulent fluctuations (see Figure 3 and Figure 4) and regions of intense currents (see Figure 3). However, particles at higher energies lose energy (on average) by interaction with the disk, while at lower energies there is a more significant energy gain. This behavior makes a direct explanation based on the current less likely, since the effect of the electric field should be the same (albeit with different magnitudes) regardless of the particle energy for particle that are overall moving in the same direction. There is also a better match in the amplitude of the fluctuations in regions that provide similar acceleration across the MAD simulations (see Figure 4) than there is for the current, as in Figure 3a the current is much more localized in the mid-plane than in Figure 3b, but the acceleration regions in the disk have roughly the same extension across the two cases, as shown in Figure 2. Considering these point (and the additional arguments presented in Sec. IV A 2 and Sec. IV B) turbulent scattering appears the most plausible explanation for the acceleration observed. A contribution by magnetic reconnection cannot be ruled out, but our analysis suggests its effect is marginal, as particles seldom intersect a current sheet at the appropriate moment. On the other hand, turbulence and curvature are present at all times in regions of strong gradients, and thus provide a more robust explanation.

Figure 5. (a) Trajectories (right panels) and evolution over time (left panel) of for selected particles. The right panels also shows the contour as dash-dotted black line and as dotted black lines. In the plane the contours are shown for . The black square and ellipse represent the particles and time-frames that are expanded upon with the 3D snapshots in Figure 6 and Figure 7 respectively. (b) Average particle distribution in the region of interested for our analysis (, ). The energy axis has been extended with respect to the previous analysis and made with logarithmic spacing. Two different power-law fits for the high-energy tail are shown.
Figure 5 (PDF p. 20). Figure highlighting the behavior of individual particles and how they can impact the average properties of the system in M9. The average particle distribution shows a very extended tail into high-energies, due to particles being accelerated along the jet and then reflected back into the region of interest.
One potential observational consequence of the identification of these acceleration regions is a localized enhancement of the jet brightness along these regions (e.g., [e.g. 76]). Particle acceleration in the jet-sheath could produce limb-brightening even in the absence of significant Doppler boosting [73]. Section IV A 2 shows that both SANE and MAD disks can accelerate particles. In the MAD runs the acceleration is concentrated along the jet, particularly in the sheath, whereas in the SANE runs the efficiency is comparable in the disk and the jet, which can be linked to the peculiar structure of the radial flows in these disks (see Figure 1b). These results support a picture in which turbulent scattering in regions of strong gradients, rather than magnetic reconnection, dominates particle energisation.
In principle, this kind of study can help to constrain the transport properties in the diffusion–advection equation for the phase-space distribution of energetic particles with spatial diffusion coefficient and source term (e.g., Parker [77], Earl et al. [78], Park and Petrosian [79])
For weakly varying spatial dependencies, for example, this equation approximately describes momentum-space diffusion with an added escape term , representing spatial transport out of the region of interest. Since cannot be derived from first principles, but influences the spectral shape, an estimate in the innermost region is of particular interest. The practical challenge is to achieve sufficient numerical resolution that allows one to fully cover the relevant particle-energy range. While the escape times found here are indicative, they are not necessarily representative for modeling neutrino fluxes from different AGN (Abbasi et al. [80]). Still, it is worth noting that the dominant acceleration mechanism is compatible with turbulent scattering and with present turbulent-corona models (Padovani et al. [7], Lemoine and Rieger 2025, Le Bihan et al. [81]).
It is instructive to extract a mono-energetic particle subset from the simulations (ensuring a sufficiently high particle count for statistical robustness) to compare the observed dynamics with theoretical expectations (e.g., Rieger and Duffy [82]) and previous studies (e.g., Kimura et al. [61]). If particle propagation were properly diffusive, the resulting distribution would approach a log-normal form as in simple stochastic or shear acceleration models without escape. As a test, we compare the time evolution of the distribution to Eq. (14) in Rieger and Duffy [82], which captures both the temporal evolution and power-law behavior. While the qualitative trend matches the model, with an initial Gaussian-like phase followed by a broadening tail, fitting a log-normal distribution proves challenging. The initial phase tends to require extreme parameters (), and at later times, the low-energy tail is under-estimated while the high-energy tail is over-estimated. The best-fit power-law index is , and no single parameter set fits distributions at multiple times. This suggests that the scattering dynamics evolve over time and/or differ from local diffusive (no escape) model as considered in Rieger and Duffy [82]. This is not surprising: the simulations are global, include particle loss and potentially capture non-diffusive transport, while the reference model is local and assumes conservation of particles. The discrepancy may thus reflect the complexity of real accretion flows, where global dynamics, irreversible losses and diverse trajectories complicate a universal parameterization. A comparison with previous work offer valuable context, though methodological differences limit direct quantitative comparison. While the dynamics considered here are fully general relativistic, simulations in Kimura et al. [61] neglect gravitational effects on the particles and use Newtonian potentials for fluid evolutions. The present study also considers particles with significantly higher energies (and Larmor radii), required to resolve gyration even near the BH, where magnetic fields can increase by an order of magnitude compared to their values at , a predominantly GR effect, particularly prominent in MAD regimes. The closest simulation for comparison in this work is S9, that, however, for the high-energy particles considered here, exhibits poor confinement, complicating direct comparison. Nevertheless, in reference Kimura et al. [61], the evolving distributions are fitted with Gaussians, whereas the present analysis finds this approach less suitable beyond the initial phases, i.e., well before the reported in Figure 8 of that work, where .

Figure 8. Validation of the particle pusher used in this study done by verifying that particles remain on stable orbits in regions of strong spacetime curvature. The main figure shows two particles moving in the mid-plane of a system with a Kerr BH () at its center. The orbits precess and cover large radial extents (from up to , but the particles neither fall into the BH nor escape at infinity even after . The insets show the projection in the poloidal plane of the chaotic particle trajectories that result from the setup of a static dipole in the Kerr metric. Colored lines represent magnetic field lines, with color indicating the log of the magnetic field magnitude. The purple trace corresponds to a metric with , a particle with energy and angular momentum initialized at . The red trace corresponds to a metric with , a particle with energy , and the same and as the previous one. Crucially, the bounce points remain stable over time, and the radial extent of the orbit remains confined even in such extreme conditions and for .
Other relevant studies include Effenberger and Petrosian [57], which explores the relation between scattering, crossing, and escape times in idealized isotropic turbulence for various injection schemes, including ”beam injection,” analogous to our flow-aligned setup, and ”pancake injection,” similar to Kimura et al. [61]. Similarly, Effenberger and Petrosian [57] describes super-diffusive transport mediated by Lévy flights, which bears resemblance to our observations: particles typically follow weakly-scattering trajectories, dominated by advection, but occasionally undergo large deviations, particularly in the funnel region (Fig. 5a), coinciding with bursts of energy gain. While these studies are local and lack dynamical and general relativistic effects, they highlight universal features of particle transport, such as the interplay between coherent motion and rare, large-scale scattering events, that are also present in our simulations. This suggests that while quantitative comparisons are limited, the underlying physical mechanisms may share commonalities across different modeling frameworks.
Astrophysical context and comparison
The simulations presented here probe particles with Larmor radii satisfying at injection, ensuring adequate orbit resolution throughout the simulation domain up to , where an artificial sink, acting as the horizon, is placed. The chosen range of Larmor radii is also a consequence of the desire to have the same energy distribution at all locations where injection occurs.
In realistic SMBH systems, such as NGC 1068 (, ), the particles relevant for the neutrino production have . The explicit ratio depends on the assumed magnetic field strength, which may reach strength up to in the mid-plane near the horizon and drop to in the corona and farther out (e.g., Lemoine and Rieger 2025, Yuan and Narayan 2014, Inoue et al. 2020). The proton energy required to produce the IceCube neutrinos (Collaboration et al. [5] is (), corresponding to Larmor radii , i.e., a ratio . Resolving such a small ratio, e.g. , would require a grid spacing , which is difficult to achieve with current numerical capabilities. An additional limitation arises from the fact that ADAF-like flows simulated with MAD and SANE tori are not expected to remain physically self-consistent at the high accretion rates inferred for NGC 1068, where radiative cooling can no longer be neglected. In terms of the normalized AthenaK units, the Eddington accretion rate for the latter corresponds to , several orders of magnitude above the rates simulated here (see Figure 1a).
Nevertheless, we can attempt an extrapolation that might prove informative, highlighting the relevance of the cases considered here relative to known sources, despite the noted limitations. NGC 1068 is not a strongly jetted galaxy, and instead thought to host a compact corona-like region (Padovani et al. [7]). Thus the case that seems most relevant among those analyzed in this paper is the S9 simulation, that also has a particle energy range somewhat closer to the mentioned , though still significantly higher. However, as discussed in Sec. IV A 1), SANE disks provide only limited confinement, as the magnetic field strength decreases rapidly with distance from the black hole. This reflects the absence of strongly magnetized regions beyond the innermost () region of the disk. Strongly magnetized regions with strong gradients also appear, from the simulations shown, to be much more efficient at accelerating particles, which may in turn enhance neutrino production in hadronic scenarios. This leaves open the possibility that particles with lower energies than those considered here could be confined by the disk for longer times, since their smaller gyro-radius could prevent (or at least delay) their access to regions of significantly lower magnetic field. This would allow more time for interactions with the fluid, potentially leading to interactions similar to those observed in the MAD cases. However, verifying such effects requires simulations with higher dynamic range than employed here, and thus the true confinement properties of SANE disks at lower particle energies remain somewhat uncertain.
For efficient neutrino production the disk must contain an extended region of strong, preferably turbulent, magnetic field. Our MAD simulations provide such conditions, but they not reproduce the assumed properties of NGC 1068. None of our simulations reveal a clear, magnetized corona that is believed to be one of the defining features of neutrino emitters like NGC1068 [7]. We remind the reader that, in this paper, we adopt a loose definition of “corona” as a region of strongly magnetized plasma that is gravitationally bound (as determined by, e.g., the Bernoulli parameter criterion) and lacks wind-like outflows [42]. Features resembling coronas have, however, been observed in recent simulations that include radiative transport and achieve higher accretion rates [27].
That said, extrapolating our results to lower particle energies suggests an interesting scenario. The magnetic field in the SANE case may be sufficient to confine lower-energy particles long enough for efficient acceleration, while particles above a certain energy threshold escape with minimal deflections. This could render SANE disks more effective at letting cosmic-ray particles escape. Notably, a relatively hard () low-energy proton spectrum with a very soft high-energy tail appears needed to model the neutrino emission in NGC 1068 (e.g., Lemoine and Rieger [12], Murase [53], Le Bihan et al. [81]). Figure 5b suggests that, provided particle feedback on the turbulence can be neglected, efficient acceleration of low-energy particles can yield a hard particle spectrum, while subsequent efficient escape at higher energies may lead to strong spectral softening.
Conclusions
In this work, we have studied the dynamics of energetic particles in advection-dominated accretion flows (ADAF), observing acceleration factors up to in funnel regions through interaction with turbulence. This acceleration is likely driven by shear-type mechanisms [66], or more generally, by interactions with flow-driven magnetic field structures (e.g., [65, 12]), such as turbulent, curved, or sheared field lines. Confinement times of particles in the wider disk region can significantly deviate from diffusive and free-streaming estimates (see Tabs. II and III), sometimes by up to an order of magnitude. We also identify new dynamics specific to the global, GR nature of our simulations. In particular, the global nature of the background GRMHD simulations allows to study particle propagation across regions with significantly varying parameters, that are self-consistently connected by the fluid simulations. This reveals dynamics with features reminiscent of Lévy flights, with small, random scattering events that confine particles in limited regions, occasionally interrupted by large, sudden jumps in both energy and location. The observed, complex particle dynamics at higher energies indicate that using spatially-constant (energy) diffusion coefficients or advection dynamics to model particle transport may not be valid beyond the low-energy regime and thus lead to notable discrepancies. The (GR) nature of the simulations also enables the study of black hole spin effects on particle dynamics in ways that have no Newtonian equivalent. In particular, the presence of strongly magnetized outflows and funnel regions [38, 51] allowed us to observe, for the first time, a reflection process that particles can undergo during propagation along the jets formed in the high-spin MAD case. Since the most intense acceleration region occurs in the funnel-sheath region, these reflections allow particle energization to last longer than would otherwise be estimated.
Overall, the picture that emerges is that strongly sheared velocity gradients drive turbulence, which in turn generates and sustains the magnetic field structures that interact with particles. This is particularly important because such dynamics can cascade self-similarly across the inertial range of the turbulent spectrum, implying that the same acceleration mechanism operates across a wide range of scales, including for lower-energy particles, that are of particular interest for astrophysical applications. Indeed, these observations prove relevant for understanding the origin of ultra-high-energy cosmic rays in AGN jets (e.g., [85, 86]), particularly if particles with lower injection energies also undergo similar processes. This is likely the result of two compounding processes. When the gyro-motion of the particle occurs on scales much smaller than the velocity gradient, the particle interacts with the small-scale turbulent structures generated by the cascade triggered by the large-scale velocity gradient. Once the gyro-radius is large enough, the particle samples the velocity shear directly, enabling efficient acceleration. Tentative extrapolations of these results to neutrino-emitting Seyfert galaxies such as NGC 1068 suggest that efficient acceleration of low-energy particles in the accretion flow near the black hole can yield a hard particle spectrum, while subsequent efficient escape at higher energies may lead to strong spectral softening. It must be stressed, however, that particles in our simulations are purely passive tracers and do not feedback on the fluid.
The results presented here underscore the need for future developments for bridging the large gap between particles involved in neutrino production or non-thermal emission and the scales of the system of interest. Resolving the Larmor orbits of such particles is computationally prohibitive, owing both to the spatial resolution required and to the high Larmor frequencies involved. A possible way to overcome these limitations is to extend AthenaK to evolve particles within the gyro-center approximation [46, 87] (Zou et al. 2024). While this approach would enable studies of particle interactions with the large-scale structures resolved by the flow, capturing the transport and acceleration of low-energy particles may require advanced subgrid models that account for interactions with plasma structures unresolved by the GRMHD simulation. Incorporating radiative cooling, or even radiative transport [26, 27], will be important to accurately model the flow dynamics and corona formation in black hole systems at high accretion rates. Similarly, exploring additional flow configurations, such as ADIOS or other RIAFs, will be of interest for understanding a broader range or accretion (e.g., Abramowicz and Fragile [14], Blandford and Begelman [74], Begelman [75]). Extending the dynamic range of the fluid simulation, currently underway, will offer new insights, while including magnetic reconnection (Ripperda et al. [63]), retrograde and tilted tori (McKinney et al. [88], Das and Porth 2024), and turbulence-driven acceleration mechanisms will further deepen our understanding of particle energization. Lastly, including more physics to model the GRMHD flows, other than the already mentioned radiative transport, such as resistive effects (Ripperda et al. [63]) or even extending the approach to two-fluid models (Most et al. [90]), would represent important developments to the modeling, as the former improves the fidelity of the magnetic field modeling by explicitly having physically motivated reconnection mechanisms, rather than the grid-scale dissipation of the ideal models, and the latter would allow for different turbulent modes to develop in the plasma, thus potentially altering the interaction between particles and fluid.
Acknowledgments
The authors would like to thank A. Bohdan, S. Cerri, M. Lemoine, S. Aerdker and J. Wang for interesting discussions on the topic. The anonymous referee is also thanked for their constructive and detailed feedback. This work was partially funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany´s Excellence Strategy – EXC 2094 – 390783311.
References
Validation of the particle pushers
Similarly to the tests done in Bacchini et al. [44, 46], the geodesic terms in the implicit pusher can be tested by initializing particles with values of the angular momentum that result in stable (precessing) orbits (Levin and Perez-Giz [91], Grossman et al. [92]) when positioned close to the BH (not necessarily only on the mid-plane). In Figure 8 two such orbits are shown, integrated for , demonstrating the ability of the pusher to conserve the energy of the particles. Similarly, the full particle push, including electromagnetic terms, can be tested for stability by setting up a static dipole field in the Kerr metric (Takahashi and Koyama [93]) and verifying that the bounce positions of the particles initialized with appropriate energies and angular momentum do not change over time. Figure 8 shows the projection of the particle trajectories that undergo such repeated bounces in one poloidal plane, and one can notice how the bounce points do not change over time (this test was run for ) and the particle trajectory is well confined radially. The precision of this pusher depends mostly on the time-step used to evolve the simulation and on the tolerance used for the iterative procedure.
Impact of particles on code performance
Regarding the specifics of the implementation in AthenaK ([24]), particles are handled at the level of the meshblockpack, as they can easily move across regions and the meshblockpack contains all the information about the fields needed to carry out the particle push.
The average number of iterations in realistic applications, where electromagnetic fields are present and turbulent, is usually around 10, though this also depends on the chosen tolerance, and some particles can take up to 50 iterations. The rate of failure for the convergence of the iterative method is very small, roughly 1 particle every fails to converge within the upper limit of iterations (set at 200). In this case the fallback strategy is an explicit mid-point update. The failure rate improves significantly over time (to fewer than 1 particle in in the later stages) as particles leave the most challenging region of the system closest to the BH.
We run sets of simulations using 0.01, 0.1 and 1 particles per cell (“ppc”), varying from 1 to 4 GPUs (NVIDIA A100 GPUs) and compare them also with a simulation running exclusively the GRMHD computations (the “ref” points in Figure 9). The particle parameters are chosen such that the time-step is not further constrained by particle motion. Since the particle solver is iterative, the performance impact presented here is only indicative, and challenging systems might require more average iteration numbers than presented here. The simulations are set up in such a way that each meshblockpack contains the same number of particles. Indeed, load balancing is an outstanding issue in the particle implementation of AthenaK, currently being worked on.

Figure 9. Summary of performance impact of particles on simulations performed with AthenaK. Points of different colors indicate different particle number (particle per cell “ppc”) as a function of the varying number of GPUs (NVIDIA A100) employed. The top plot shows the total time taken by simulations. The black crosses show the reference case with no particles and only GRMHD computations. The blue line shows the ideal scaling where duration of a simulation scales exactly like . The bottom plot shows the number of particle per cpu-second (“pps”) updates. Here the blue line shows the performance scaling expected expected naively by scaling like .
We note that there is a slight degradation in performance going from to due to the introduction of communication in the execution of the program and the small size of the mesh chosen as reference. In other words, at low particle numbers, the reference problem we chose for this indicative example of performance scaling is not ideal for multi-GPU execution as it is communication/memory-bound. This is particularly visible in the bottom panel of Figure 9, where for the particle-per-cpu-second (“pps”) updates are degrading with , due to the inefficiency of the communication, while at the problem becomes computationally bound, and the expected scaling is recovered, with communication playing a minor role. The top panel of Figure 9 shows that at lower particle densities the impact on performance is negligible, and thus the more in-depth scaling analysis presented in Stone et al. [24] applies, but at higher particle density the cost of particles can quickly become dominant. Indeed, at one can notice how the scalings with follows very closely the expected scaling, with minor inefficiencies introduced by communication. Thus makes the problem computationally bound no matter the size of the original GRMHD problem.
Aside from the increased computational cost of the simulations, there is the additional cost of the communication of the arrays that contain particle information, that comes with an additional memory demand. Since individual particles are tracked, at each timestep, the code must check in each meshblockpack which particles has crossed the boundaries and either send it to the neighboring region of the domain or destroy it if it has left the domain altogether. Thus there are two additional arrays that need to be allocated to contain the information of these particles. In principle, though this is an extreme corner case, all particles within a rank might be sent at the same time, or all might be destroyed, and there is no way to know this a priori, thus the arrays have to be allocated with sufficient size to accommodate for this eventuality. Concretely, allocating arrays with 1/4 of the total size of the total array is sufficient for the present usecases.
Load balancing can be straightforwardly adapted to consider particles by assigning a default weight to each particle, such that meshblocks with more particles might be better distributed across computational resources.
Coarse-graining Analysis
Coarse-graining is used as a technique to analyze the turbulent properties of a medium when the conditions for the application of Fourier analysis (homogeneous background, periodic directions etc...) are not satisfied. In this appendix we describe the implementation of the coarse-graining process as applied in this paper.
As a preliminary step, we construct, starting from the data stored on the multi-refinement mesh, an array with uniform resolution, by taking the highest refinement level available and applying it to the whole domain (larger domains can run into memory issues on some smaller machines, often requiring more than 64 GB), and filling in the ”newly created” points when refining cells by performing tri-linear interpolation of the values in the cell itself and its neighboring ones in post-processing. This concretely means taking the most refined cell in the simulation data as the reference scale , and defining, at each point of the domain, a relative scale . Then the value of the current cell is placed at coordinate of a cell that covers the domain , and the values from the neighboring cells will be placed at positions and , where , as the neighboring cell could be in a region with different refinement. These three values are then used to interpolate the values cell created from the refinement, that will be located at for . Currently this approach does not include values from the cells neighboring any given one at the corners or edges, only through faces.
The uniform array thus created is used to perform coarse-graining (aka filtering) analysis [71]. In short, this consists in defining a coarse grained field at scale by convolving the original field with a scaled kernel , that should satisfy the condition of being third-rank:
where is a scaled Gaussian kernel. Since the system we’re considering is three-dimensional, we consider the product of 3 1-dimensional kernels to perform the convolution.
When doing the convolution, we consider points up to a distance to ensure that a large enough area around each point is used for the convolution (i.e. to limit ”boundary effects” from the kernel being applied to discrete points).
From this coarse-grained (low-pass filtered) field, one can reconstruct a fluctuation spectrum that can be shown [71] to be equivalent to the usual Fourier spectrum in situations where both can be applied:
as the domain for the averaging we choose . After the scale of the convolution domain crosses 40 points (in all directions, since we’re doing the convolution isotropically), the domain of the convolution is incremented by 2 points at each further step. This is mostly to save on post-processing time while only affecting a region of the spectrum that in logarithmic plots already appears very compressed.
Furthermore, it has been shown [71] that using the high-pass-filtered spectrum allows to replicate spectra with steeper index (compared to for low-pass filter), where is the rank of the kernel used to perform the coarse-graining. The high-pass-filtered spectrum is obtained from the coarse-grained field simply by replacing in the previous equations with . Examples of the low-pass and high-pass filtered fields are shown in Figure 10. In these figures one can also notice that the outermost regions of the domain don’t undergo the coarse-graining procedure. This is done to simplify the operations on the multi-level grid, avoiding having to look for neighboring meshblocks outside the area of interest, since in any case the analysis is carried out in the innermost regions of the coarse-grained domain.

Figure 10. (a) M9. Low-pass filter (). (b) S9 High-pass filter (). Heatmaps showing the result of the coarse-graining procedure on the density field.
The coarse-graining procedure also allows to estimate the magnitude of the magnetic field fluctuations locally, as shown in Figure 11 for two different locations of interest for the analysis in the main paper. The average value of the magnetic field is computed within a certain sub-domain over which then the coarse-graining procedure is performed, and from the high-pass filter spectrum one can derive the distribution of the fluctuations. As visible, in the jet-sheath region, where the magnetic field is strong and with a stable structure, the fluctuations are smaller when compared to the disk/outflow region, where turbulence is more predominant and doesn’t allow the formation of stable structures. In the disk, in fact, most of the fluctuations have amplitude comparable to the average magnetic field, meaning that the region is extremely turbulent. Getting closer to the BH (not shown), reduces the amplitude of the fluctuations, as the magnetic field is strengthened and its structure more strongly defined by interaction with the boundary conditions of the BH.

Figure 11. (a) Location (30.0, 0.0, 20.0). (b) Location (15.0, 0.0, 20.0). (c) Location (60.0, 0.0, 15.0). Distribution of the magnetic field amplitude fluctuations at locations of interest for the analysis in the main body of the paper from simulation M9.
Table of symbols
| Quantity | Symbol | Definition | Use |
| Lapse function | Time dilation between observers. | ||
| Shift vector | Spatial shift in 3+1 decomposition. | ||
| Spacetime metric | – | Full metric tensor. | |
| Spatial metric | Spatial part of | Defines spatial distances in 3+1. | |
| FIDO 4–velocity | 4–velocity of fiducial observer. | ||
| Lorentz factor (FIDO frame) | Measured by fiducial observer. | ||
| Time component of particle 4–velocity | In coordinate basis | ||
| Particle 4–velocity | – | Covariant/contravariant components. | |
| Coordinate 3–velocity | Spatial velocity measured at infinity. | ||
| Proper time | – | Particle’s proper time. | |
| Rest mass density | – | Proper rest mass density of fluid. | |
| Fluid 4–velocity | – | Fluid 4–velocity. | |
| Magnetic field (coordinate basis) | – | Eulerian magnetic field. | |
| Magnetic field (fluid rest frame) | Magnetic field in fluid rest frame. | ||
| Magnetization | Magnetic over rest–mass energy density. | ||
| Internal energy | – | Internal energy density. | |
| Adiabatic index | Ratio of specific heats for relativistic gas. | ||
| Enthalpy | Relativistic enthalpy. | ||
| Alfvén velocity | Characteristic speed of magnetic waves. | ||
| Bernoulli parameter | Measure of fluid binding energy. | ||
| Electromagnetic tensor | – | Faraday tensor. | |
| Dual EM tensor | – | Dual of electromagnetic tensor. | |
| Energy-momentum tensor | – | Stress-energy tensor. | |
| Electric displacement | Electric field measured by FIDO. | ||
| Levi–Civita tensor | Pseudo-tensor for cross products. | ||
| Charge | – | Particle electric charge. | |
| Mass | – | Particle rest mass. | |
| Charge–to–mass ratio | For protons. | ||
| Gravitational radius | Normalization for lengths. | ||
| Larmor radius | Radius of the gyro-orbit. | ||
| Larmor frequency | Frequency of gyration. | ||
| Energy at infinity | Particle energy measured at infinity. | ||
| Spin parameter | – | Dimensionless BH spin (). | |
| Gravitational time | Normalization for time. | ||
| Escape time | – | Time for particle to leave outer boundary. | |
| Infall time | – | Time for particle to fall into sink. | |
| Trapping time | – | Time particle spends in a cell. | |
| Diffusion coefficient | Characterizes diffusive transport. | ||
| Power–law index | – | Index used to fit power–law distributions. | |
| Cylindrical radius | – | Radial coordinate in cylindrical system. | |
| Azimuthal angle | – | Angular coordinate in cylindrical system. | |
| Vertical coordinate | – | Height coordinate in cylindrical system. | |
| Wavenumber | Wavenumber at scale in coarse–graining. | ||
| Scale | – | Coarse–graining scale. | |
| Black hole mass | – | Mass of central black hole. |
Table V. Summary of all symbols and quantities appearing in the paper and their use.
References
- [2]Frank M. Rieger and Amir Levinson. Radio galaxies at VHE energies. Galaxies, 6(4):116, November 2018. doi: 10.3390/galaxies6040116.DOI
- [3]João RT de Mello Neto. Physics and astrophysics of ultra-high energy cosmic rays: recent results from the pierre auger observatory. Physics of PARTICLEs and Nuclei, 53(2):224–232, 2022.arxiv.org/abs/2012.12943
- [4]Noémie Globus and Roger Blandford. Ultra high energy cosmic rays. arXiv e-prints, art. arXiv:2505.21846, May 2025. doi: 10.48550/arXiv.2505.21846.DOI
- [5]IceCube Collaboration, R Abbasi, M Ackermann, J Adams, JA Aguilar, M Ahlers, M Ahrens, JM Alameddine, C Alispach, AA Alves Jr, et al. Evidence for neutrino emission from the nearby active galaxy ngc 1068. Science, 378(6619):538–543, 2022.DOI
- [6]J Bland-Hawthorn, JF Gallimore, LJ Tacconi, E Brinks, SA Baum, RRJ Antonucci, and GN Cecil. The ringberg standards for ngc 1068. Astrophysics and Space Science, 248:9–19, 1997.
- [7]P Padovani, E Resconi, M Ajello, C Bellenghi, S Bianchi, P Blasi, K-Y Huang, S Gabici, V Gómez Rosas, H Niederhausen, et al. High-energy neutrinos from the vicinity of the supermassive black hole in ngc 1068. Nature Astronomy, 8(9):1077–1087, 2024.arxiv.org/abs/2405.20146
- [9]Björn Eichmann, Foteini Oikonomou, Silvia Salvatore, Ralf-Jürgen Dettmar, and Julia Becker Tjus. Solving the multimessenger puzzle of the agn-starburst composite galaxy ngc 1068. The Astrophysical Journal, 939(1):43, 2022.DOI
- [10]Damiano F. G. Fiorillo, Maria Petropoulou, Luca Comisso, Enrico Peretti, and Lorenzo Sironi. TeV neutrinos and hard X-rays from relativistic reconnection in the corona of NGC 1068. The Astrophysical Journal Letters, 961(1):L14, January 2024. doi:10.3847/2041-8213/ad192b.DOI
- [12]Martin Lemoine and Frank Rieger. Neutrinos from stochastic acceleration in black hole environments. Astronomy & Astrophysics, 697:A124, 2025.DOI
- [13]Ramesh Narayan, Aleksander Sadowski, Robert F Penna, and Akshay K Kulkarni. Grmhd simulations of magnetized advection-dominated accretion on a non-spinning black hole: role of outflows. Monthly Notices of the Royal Astronomical Society, 426(4):3241–3259, 2012.DOI
- [14]Marek A Abramowicz and P Chris Fragile. Foundations of black hole accretion disk theory. Living Reviews in Relativity, 16:1–88, 2013.
- [15]Feng Yuan and Ramesh Narayan. Hot accretion flows around black holes. Annual Review of Astronomy and Astrophysics, 52(1):529–588, 2014.DOI
- [16]Ramesh Narayan, Igor V Igumenshchev, and Marek A Abramowicz. Magnetically arrested disk: an energetically efficient accretion flow. Publications of the Astronomical Society of Japan, 55(6):L69–L72, 2003.DOI
- [17]Masaru Shibata and Yu-ichirou Sekiguchi. Magnetohydrodynamics in full general relativity: formulation and tests. Physical Review D, 72(4):044014, 2005.DOI
- [18]Yosuke Mizuno and Luciano Rezzolla. General-relativistic magnetohydrodynamic equations: the bare essential. arXiv preprint arXiv:2404.13824, 2024.arxiv.org/abs/2404.13824
- [20]Masanori Nakamura, Keiichi Asada, Kazuhiro Hada, Hung-Yi Pu, Scott Noble, Chihyin Tseng, Kenji Toma, Motoki Kino, Hiroshi Nagai, Kazuya Takahashi, et al. Parabolic jets from the spinning black hole in m87. The Astrophysical Journal, 868(2):146, 2018.
- [21]Yosuke Mizuno. Grmhd simulations and modeling for jet formation and acceleration region in agns. Universe, 8(2):85, 2022.DOI
- [22]James M Stone, Kengo Tomida, Christopher J White, and Kyle G Felker. Athena++: adaptive mesh refinement framework: design and magnetohydrodynamic solvers. The Astrophysical Journal Supplement Series, 249(1):4, 2020.
- [23]Christian R. Trott, Damien Lebrun-Grandié, Daniel Arndt, Jan Ciesko, Vinh Dang, Nathan Ellingwood, RahulKumar Gayatri, Evan Harvey, Daisy S. Hollman, Dan Ibanez, Nevin Liber, Jonathan Madsen, Jeff Miles, David Poliakoff, Amy Powell, Sivansankaran Rajamanickam, Mikael Simberg, Dan Sunderland, Bruno Turcksin, and Jeremiah Wilke. Kokkos 3: programming model extensions for the exascale era. IEEE Transactions on Parallel and Distributed Systems, 33(4):805–817, 2022. doi:10.1109/TPDS.2021.3097283.
- [24]James M Stone, Patrick D Mullen, Drummond Fielding, Philipp Grete, Minghao Guo, Philipp Kempski, Elias R Most, Christopher J White, and George N Wong. Athenak: a performance-portable version of the athena++ amr framework. arXiv preprint arXiv:2409.16053, 2024.
- [25]Aleksander Sadowski, Maciek Wielgus, Ramesh Narayan, David Abarca, Jonathan C McKinney, and Andrew Chael. Radiative, two-temperature simulations of low-luminosity black hole accretion flows in general relativity. Monthly Notices of the Royal Astronomical Society, 466(1):705–725, 2017.
- [26]Yan-Fei Jiang, Omer Blaes, James M Stone, and Shane W Davis. Global radiation magnetohydrodynamic simulations of sub-Eddington accretion disks around supermassive black holes. The Astrophysical Journal, 885(2):144, 2019.
- [27]Lizhong Zhang, James M Stone, Patrick D Mullen, Shane W Davis, Yan-Fei Jiang, and Christopher J White. Radiation grmhd models of accretion onto stellar-mass black holes: I. survey of eddington ratios. arXiv preprint arXiv:2506.02289, 2025.
- [28]Kip S. Thorne, Richard H. Price, and Douglas A. MacDonald. Black holes: the membrane paradigm. Yale University Press, 1986.
- [29]Luciano Rezzolla and Olindo Zanotti. Relativistic hydrodynamics. Oxford University Press, 2013.DOI
- [30]Bart Ripperda, Fabio Bacchini, Oliver Porth, Elias R Most, Hector Olivares, Antonios Nathanail, Luciano Rezzolla, Jannis Teunissen, and Rony Keppens. General-relativistic resistive magnetohydrodynamics with robust primitive-variable recovery for accretion disk simulations. The Astrophysical Journal Supplement Series, 244(1):10, 2019.
- [31]Patrick Chi-Kit Cheong, Lap-Ming Lin, and Tjonnie Guang Feng Li. Gmunu: toward multigrid based einstein field equations solver for general-relativistic hydrodynamics simulations. Classical and Quantum Gravity, 37(14):145015, 2020.
- [32]Charles R Evans and John F Hawley. Simulation of magnetohydrodynamic flows: a constrained transport method. Astrophysical Journal, Part 1 (ISSN 0004-637X), vol. 332, Sept. 15, 1988, p. 659-677., 332:659–677, 1988.DOI
- [33]E. F. Toro. Riemann solvers and numerical methods for fluid dynamics. Springer, 2009.
- [34]Leslie G Fishbone and Vincent Moncrief. Relativistic fluid disks in orbit around kerr black holes. The Astrophysical Journal, 207:962–976, 1976.
- [35]Steven A Balbus and John F Hawley. Instability, turbulence, and enhanced transport in accretion disks. Reviews of modern physics, 70(1):1, 1998.
- [36]Nikolai I Shakura and Rashid A Sunyaev. Black holes in binary systems: Observational appearances. In Symposium-International Astronomical Union, volume 55, pages 155–164. Cambridge University Press, 1973.
- [37]NI Shakura and RA Sunyaev. A theory of the instability of disk accretion on to black holes and the variability of binary x-ray sources, galactic nuclei and quasars. Monthly Notices of the Royal Astronomical Society, 175(3):613–632, 1976.
- [38]Roger D Blandford and Roman L Znajek. Electromagnetic extraction of energy from kerr black holes. Monthly Notices of the Royal Astronomical Society, 179(3):433–456, 1977.
- [39]M Livio, GI Ogilvie, and JE Pringle. Extracting energy from black holes: the relative importance of the blandford-znajek mechanism. The Astrophysical Journal, 512(1):100, 1999.
- [40]Kenji Toma and Fumio Takahara. Electromotive force in the blandford–znajek process. Monthly Notices of the Royal Astronomical Society, 442(4):2855–2866, 2014.
- [41]Roger Blandford, David Meier, and Anthony Readhead. Relativistic jets from active galactic nuclei. Annual Review of Astronomy and Astrophysics, 57:467–509, 2019.DOI
- [42]John F Hawley and Jean-Pierre De Villiers. General relativistic magnetohydrodynamic simulations of black hole accretion disks. Progress of Theoretical Physics Supplement, 155:132–139, 2004.
- [43]Oliver Porth, Koushik Chatterjee, Ramesh Narayan, Charles F Gammie, Yosuke Mizuno, Peter Anninos, John G Baker, Matteo Bugli, Chi-kwan Chan, Jordy Davelaar, et al. The event horizon general relativistic magnetohydrodynamic code comparison project. The Astrophysical Journal Supplement Series, 243(2):26, 2019.
- [44]Fabio Bacchini, Bart Ripperda, Alexander Yuran Chen, and Lorenzo Sironi. Generalized, energy-conserving numerical simulations of particles in general relativity. i. time-like and null geodesics. The Astrophysical Journal Supplement Series, 237(1):6, 2018.
- [45]B. Ripperda, F. Bacchini, J. Teunissen, C. Xia, O. Porth, L. Sironi, G. Lapenta, and R. Keppens. A comprehensive comparison of relativistic particle integrators. The Astrophysical Journal Supplement Series, 235(1):21, mar 2018. doi:10.3847/1538- 4365/aab114. URL https://dx.doi.org/10.3847/1538-4365/aab114.
- [46]Fabio Bacchini, Bart Ripperda, Oliver Porth, and Lorenzo Sironi. Generalized, energy-conserving numerical simulations of particles in general relativity. ii. test particles in electromagnetic fields and grmhd. The Astrophysical Journal Supplement Series, 240(2):40, 2019.DOI
- [48]Y. Saad and M. H. Schultz. GMRES: a generalized minimal residual algorithm for solving nonsymmetric linear systems. SIAM Journal on Scientific and Statistical Computing, 7:856–869, 1986. doi:10.1137/0907058.DOI
- [49]Anthony M Hillas. The origin of ultra-high-energy cosmic rays. IN: Annual review of astronomy and astrophysics. Volume 22. Palo Alto, CA, Annual Reviews, Inc., 1984, p. 425-444, 22:425–444, 1984.DOI
- [50]Martin Lemoine. Particle transport through localized interactions with sharp magnetic field bends in mhd turbulence. Journal of Plasma Physics, 89(5):175890501, 2023.DOI
- [51]Roger D Blandford and DG Payne. Hydromagnetic flows from accretion discs and the production of radio jets. Monthly Notices of the Royal Astronomical Society, 199(4):883–903, 1982.DOI
- [52]EPT Liang. On the hard x-ray emission mechanism of active galactic nuclei sources. Astrophysical Journal, Part 2-Letters to the Editor, vol. 231, Aug. 1, 1979, p. L111-L114., 231:L111–L114, 1979.DOI
- [54]Riccardo Middei, Emanuele Nardini, Chris Done, Elisabeta Lusso, Fausto Vagnetti, Guido Risaliti, Enrico Piconcelli, Stefano Bianchi, Gabriele Matzeu, A Trindade Falcão, et al. Directly tracking the rebrightening of a supermassive black hole accretion disk. The Astrophysical Journal, 1004(2):200, 2026.DOI
- [55]Priyadarshee P Dash, Prantik Nandi, Sachindra Naik, Narendranath Layek, and Sandip K Chakrabarti. Accretion dynamics and coronal geometry in mrk 530: insights from 24 years of x-ray monitoring. arXiv preprint arXiv:2603.06103, 2026.arxiv.org/abs/2603.06103
- [56]Ruchika Dhaka, Ranjeev Misra, and Suraj Kumar Chaurasia. Accretion disc evolution in gx 339-4 across spectral states using nustar, nicer, and insight-hxmt observations. Monthly Notices of the Royal Astronomical Society, 547(4):stag487, 2026.DOI
- [57]Frederic Effenberger and Vahé Petrosian. The relation between escape and scattering times of energetic particles in a turbulent magnetized plasma: application to solar flares. The Astrophysical Journal Letters, 868(2):L28, 2018.
- [58]Mayank Pathak and Banibrata Mukhopadhyay. Simulating ulxs and blazars as grmhd accretion flows around a black hole. The Astrophysical Journal, 981(2):162, 2025.arxiv.org/abs/2502.03538
- [59]A. Mignone and Jonathan C McKinney. Equation of state in relativistic magnetohydrodynamics: variable versus constant adiabatic index. Monthly Notices of the Royal Astronomical Society, 378(3):1118–1130, 2007.DOI
- [60]Shigeo S. Kimura, Kohta Murase, and Kenji Toma. Neutrino and cosmic-ray emission and cumulative background from radiatively inefficient accretion flows in low-luminosity active galactic nuclei. The Astrophysical Journal, 806(2):159, jun 2015. doi:10.1088/0004-637X/806/2/159. URL https://dx.doi.org/10.1088/0004-637X/806/2/159.
- [61]Shigeo S Kimura, Kengo Tomida, and Kohta Murase. Acceleration and escape processes of high-energy particles in turbulence inside hot accretion flows. Monthly Notices of the Royal Astronomical Society, 485(1):163–178, 02 2019. ISSN 0035-8711. doi:10.1093/mnras/stz329. URL https://doi.org/10.1093/mnras/stz329.
- [62]Hao Zhang, Lorenzo Sironi, and Dimitrios Giannios. Fast particle acceleration in three-dimensional relativistic reconnection. The Astrophysical Journal, 922(2):261, 2021.DOI
- [63]Bart Ripperda, Matthew Liska, Koushik Chatterjee, Gibwa Musoke, Alexander A Philippov, Sera B Markoff, Alexander Tchekhovskoy, and Ziri Younsi. Black hole flares: ejection of accreted magnetic flux through 3d plasmoid-mediated reconnection. The Astrophysical Journal Letters, 924(2):L32, 2022.DOI
- [64]Indu K Dihingia, Yosuke Mizuno, Christian M Fromm, and Ziri Younsi. Impact of radiative cooling on the magnetised geometrically thin accretion disc around kerr black hole. Journal of Cosmology and AstroPARTICLE Physics, 2025(01):152, 2025.DOI
- [65]Martin Lemoine. Generalized fermi acceleration. Phys. Rev. D, 99:083006, Apr 2019. doi:10.1103/PhysRevD.99.083006. URL https://link.aps.org/doi/10.1103/PhysRevD.99.083006.DOI
- [66]Frank M. Rieger. An introduction to pARTICLE acceleration in shearing flows. Galaxies, 7(3):78, September 2019. doi: 10.3390/galaxies7030078.DOI
- [67]Jie-Shuang Wang, Brian Reville, Ruo-Yu Liu, Frank M Rieger, and Felix A Aharonian. Particle acceleration in shearing flows: the case for large-scale jets. Monthly Notices of the Royal Astronomical Society, 505(1):1334–1341, 2021.DOI
- [68]Amir Jafari and Ethan Vishniac. Topology and stochasticity of turbulent magnetic fields. Physical Review E, 100(1):013201, 2019.DOI
- [69]Lukas Bentkamp, Cristian C Lalescu, and Michael Wilczek. Persistent accelerations disentangle lagrangian turbulence. Nature Communications, 10(1):3550, 2019.DOI
- [70]Jie-Shuang Wang, Brian Reville, Yosuke Mizuno, Frank M. Rieger, and Felix A. Aharonian. PARTICLE acceleration in shearing flows: the self-generation of turbulent spine-sheath structures in relativistic magnetohydrodynamic jet simulations. Monthly Notices of the Royal Astronomical Society, 519(2):1872–1880, February 2023. doi:10.1093/mnras/stac3616.DOI
- [71]Dongxiao Zhao and Hussein Aluie. Calculating spectra by sequential filtering. Journal of Renewable and Sustainable Energy, 17(1), 2025.DOI
- [72]Indu K Dihingia, Yosuke Mizuno, Christian M Fromm, and Luciano Rezzolla. Temperature properties in magnetized and radiatively cooled two-temperature accretion flows to a black hole. Monthly Notices of the Royal Astronomical Society, 518(1):405–417, 2023.DOI
- [73]F. M. Rieger and P. Duffy. Shear acceleration in relativistic astrophysical jets. The Astrophysical Journal, 617(1):155–161, 2004. doi:10.1086/425167.DOI
- [74]Roger D Blandford and Mitchell C Begelman. On the fate of gas accreting at a low rate on to a black hole. Monthly Notices of the Royal Astronomical Society, 303(1):L1–L5, 1999.DOI
- [75]Mitchell C Begelman. Radiatively inefficient accretion: breezes, winds and hyperaccretion. Monthly Notices of the Royal Astronomical Society, 420(4):2912–2923, 2012.DOI
- [76]Christian M Fromm, Alejandro Cruz-Osorio, Yosuke Mizuno, Antonios Nathanail, Ziri Younsi, Oliver Porth, Hector Olivares, Jordy Davelaar, Heino Falcke, Michael Kramer, et al. Impact of non-thermal particles on the spectral and structural properties of m87. Astronomy & Astrophysics, 660:A107, 2022.DOI
- [77]E. N. Parker. The passage of energetic charged pARTICLES through interplanetary space. Planetary and Space Science, 13(1):9–49, January 1965. doi:10.1016/0032-0633(65)90131-5.DOI
- [78]J. A. Earl, J. R. Jokipii, and G. Morfill. Cosmic-ray viscosity. The Astrophysical Journal Letters, 331:L91, August 1988. doi:10.1086/185242.DOI
- [79]Brian T. Park and Vahe Petrosian. Fokker-Planck equations of stochastic acceleration: Green’s functions and boundary conditions. The Astrophysical Journal, 446:699, June 1995. doi:10.1086/175828.DOI
- [80]R. Abbasi, M. Ackermann, J. Adams, et al. IceCube search for neutrino emission from X-ray bright Seyfert galaxies. The Astrophysical Journal, 988(1):141, 2025. doi:10.3847/1538-4357/addd05. URL: https://doi.org/10.3847/1538-4357/addd05.
- [81]S. Le Bihan, M. Moisine, and F. Rieger. Neutrino production in turbulent AGN coronae and jets: a self-consistent approach. arXiv e-prints, 2026.
- [82]Frank M. Rieger and Peter Duffy. A microscopic analysis of shear acceleration. The Astrophysical Journal, 652(2):1044–1049, December 2006. doi:10.1086/508056.DOI
- [85]Frank M. Rieger. Active galactic nuclei as potential sources of ultra-high energy cosmic rays. Universe, 8(11):607, November 2022. doi:10.3390/universe8110607.DOI
- [86]Jie-Shuang Wang, Brian Reville, Frank M. Rieger, and Felix A. Aharonian. Acceleration of ultra-high-energy cosmic rays in the kiloparsec-scale jets of nearby radio galaxies. The Astrophysical Journal Letters, 977(1):L20, December 2024. doi:10.3847/2041-8213/ad9589.DOI
- [87]Theodore G Northrop. The guiding center approximation to charged particle motion. Annals of Physics, 15(1):79–101, 1961.DOI
- [88]Jonathan C McKinney, Alexander Tchekhovskoy, and Roger D Blandford. Alignment of magnetized accretion disks and relativistic jets with spinning black holes. Science, 339(6115):49–52, 2013.DOI
- [90]Elias R Most, Jorge Noronha, and Alexander A Philippov. Modelling general-relativistic plasmas with collisionless moments and dissipative two-fluid magnetohydrodynamics. Monthly Notices of the Royal Astronomical Society, 514(4):4989–5003, 2022.DOI
- [91]Janna Levin and Gabe Perez-Giz. A periodic table for black hole orbits. Physical Review D, 77(10):103005, 2008.DOI
- [92]Rebecca Grossman, Janna Levin, and Gabe Perez-Giz. Harmonic structure of generic kerr orbits. Physical Review D, 85(2):023012, 2012.DOI
- [93]Masaaki Takahashi and Hiroko Koyama. Chaotic motion of charged particles in an electromagnetic field surrounding a rotating black hole. The Astrophysical Journal, 693(1):472, 2009.DOI


