WorksPhysics and Astronomy
Thermodynamics of gas giants accreting via shocks: the role of molecular hydrogen dissociation
Abstract
Within the core-accretion framework, gas giants acquire most of their mass through accretion shocks during runaway gas accretion. The shock thermodynamics—governed by radiative cooling, accretion history, and equation of state (EoS)—sets the post-shock gas entropy and initial thermal state, strongly influencing both observables and subsequent evolution. We aim to build a coupled evolutionary model predicting the initial thermodynamic conditions, intrinsic luminosities, and upper limits on magnetic field strengths of gas giants across diverse accretion histories. We couple radiation hydrodynamics and dissociation to long-term planetary interior evolution by training a physics-informed neural network (PINN) predictor on 1000 first-principles simulations resolving the shock physics. This PINN provides post-shock outer boundary conditions for MESA to evolve the planetary envelope beyond runaway accretion. The PINN predicts higher post-shock temperatures than the canonical cold-start model, with the discrepancy widening with planetary mass, while post-shock entropy jumps sharply when dissociates at K. Gas giants evolved with PINN-derived boundary conditions achieve intrinsic luminosities 4–10 times higher than cold-start models. Under optimistic assumptions, the maximum magnetic field strength at reaches a few gauss for a planet, G for , and G for . Incorporating detailed radiation hydrodynamics and realistic thermodynamics into accretion shock models yields higher initial entropies, brighter intrinsic luminosities, and stronger magnetic fields, providing improved constraints on the masses and accretion rates of directly imaged exoplanets.
Introduction
Following the discovery of 51 Pegasi b [59], elucidating the origins of exoplanetary systems has become a central objective in modern astronomy. Nevertheless, the exact mechanisms governing planet formation remain incompletely understood, largely due to the scarcity of observational constraints during the earliest evolutionary stages. In recent years, direct imaging [14] has emerged as a powerful technique for identifying young, forming bodies. These detections rely primarily on high-contrast infrared observations, which capture the high intrinsic thermal luminosity of young, massive planets [57, 58, 48, 52, 18, 37, 10] (van Cappelleen et al. 2025). Furthermore, actively accreting protoplanets can now be probed via H emission (Keppler et al. 2018; Haffert et al. 2019; Zhou et al. 2021; Close et al. 2025a,b; Zhou et al. 2025), offering a diagnostic of magnetospheric accretion—a process that generally requires strong planetary magnetic fields (Batygin & Adams 2025; though see Chen et al. 2026).
Two primary paradigms account for the formation of gas giants. The first is the core-accretion model [61, 76, 56], in which massive solid cores assemble via the coagulation of dust and planetesimals before accreting gas from the surrounding protoplanetary disk (PPD). More recently, this classical picture has been extended by pebble accretion theory [68, 49, 43], which alleviates the slow core-growth timescales and migration bottlenecks inherent to planetesimal accretion [29]. The second paradigm is the gravitational instability model [16, 13, 65], wherein gas giants form directly via the gravitational collapse of localized, self-gravitating clumps within the PPD. Gravitational instability naturally accounts for gas giant formation in low-metallicity environments [64], operates effectively at wide orbital separations (Ni et al. 2025), and preferentially produces high-mass companions [39].
Within the core-accretion framework—the favored formation pathway for Jupiter and Saturn—planets acquire the bulk of their mass during runaway gas accretion. In this regime, low-angular-momentum gas falls onto the protoplanet at near free-fall velocities, driving a strong accretion shock that radiates away a substantial fraction of the kinetic energy [56, 8]. Because most of the planetary envelope passes through this shock front, the fraction of entropy retained in the post-shock gas fundamentally sets the initial thermal state of the young giant. Depending on the efficiency of this radiative cooling, the resulting hot- or warm-start states can amplify convection- and rotation-driven magnetic dipole fields [31], regulate magnetospheric accretion from the circumplanetary disk (CPD) [6], shape satellite system architectures [7, 34], and promote the long-term survival of dilute cores [84, 80, 45].
To characterize this shock physics, Marleau et al. [54]; Marleau et al. [55] employed one-dimensional radiation-hydrodynamic (RHD) simulations to map post-shock thermodynamic states across planetary masses and radii. Chen & Bai [19] extended this approach by incorporating a realistic equation of state (EoS), showing that dissociation strongly modifies the post-shock entropy. Because dissociation and temperature-dependent opacities introduce strong nonlinearities, mapping these thermodynamic regimes requires extensive parameter surveys. However, while one-dimensional RHD models capture the local shock microphysics, their computational expense precludes direct coupling to long-term evolutionary calculations.
In reality, runaway gas accretion is tightly coupled to the secular evolution of the planetary interior. Resolving both processes simultaneously from first principles is computationally prohibitive owing to the extreme separation of timescales between the planet’s Kelvin–Helmholtz cooling timescale () and the local free-fall accretion timescale (). To bridge this gap, we train a physics-informed neural network (PINN) on a suite of RHD shock simulations to serve as a fast surrogate model for post-shock thermodynamic states, establishing a self-consistent pipeline that links short-timescale shock microphysics to long-term interior evolution. From the resulting interior structures, we can self-consistently derive the planet’s intrinsic luminosity [79] and an upper limit on its dynamo-generated magnetic field strength across arbitrary accretion histories.
The remainder of this paper is organized as follows. In Sect. 2, we detail the physical formulation of the accretion shock model and the PINN architecture. We present the RHD simulation results in Sect. 2.2 and evaluate the predictive performance of the PINN in Sect. 2.4. In Sect. 3, we couple the RHD code Guangqi [20] with the Modules for Experiments in Stellar Astrophysics (MESA) via the trained PINN to construct self-consistent evolutionary tracks for young gas giants. Finally, we discuss the broader implications of our findings in Sect. 4 and summarize our conclusions in Sect. 5.
Methods
We first generate an extensive dataset of first-principles RHD simulations using Guangqi, and subsequently use these data to train a fast physics-informed neural network (PINN) predictor for the post-shock thermodynamic state.
First-principles calculations
Physical model
Because the accretion flow evolves on a much longer timescale than the dynamical and shock-cooling timescales, we seek steady-state solutions. To this end, we adopt the radiation-hydrodynamic framework of Chen & Bai [19] and extend it to include the planet’s internal luminosity, as illustrated schematically in Figure 1.
Figure 1. Schematic diagram of the hydrodynamic processes in our model. Gas infalls at free-fall velocity from the outer boundary of the computational domain (defined in Sect. 2.1.3) and forms an accretion shock at the top of the planetary atmosphere, after which the shocked gas cools and settles onto deeper layers. Here, and denote the internal luminosity from the planetary interior and the accretion luminosity, respectively.
Formulated in a spherically symmetric laboratory frame, the governing equations combine hydrodynamics with flux-limited diffusion (FLD) radiation transport and a realistic EoS:
where , , , , , and denote the gas density, radial velocity, pressure, temperature, total energy density, and specific internal energy, respectively, and , , and are the gravitational constant, speed of light, and radiation constant. The gas composition is fixed at 74% hydrogen and 26% helium by mass. Whereas Chen & Bai [19] employed an EoS accounting for both dissociation and hydrogen ionization, they demonstrated that ionization occurs only in very strong shocks. To accelerate our parameter survey, we neglect hydrogen ionization here; further details of this EoS are provided in Appendix A.
We adopt the same opacity tables [78, 53], FLD radiation transport scheme [51], and implicit solver [3, 4] as in Chen & Bai [19]. Consequently, the radiation energy density (), radiation flux (), radiation–gas energy exchange rate (), Planck mean opacity (), and Rosseland mean opacity () are evaluated identically to Chen & Bai [19].
A key physical addition in this work is the planetary internal luminosity, modeled via the volumetric heating source term in Eq. (3):
where is the combined volume of the three innermost spherical shells. Distributing uniformly across this thin inner layer approximates a fixed basal radiative flux while avoiding numerical instabilities.1 Because this heating zone is geometrically thin relative to the atmospheric scale height and resides deep within the optically thick envelope, any error introduced by this treatment is negligible.
Boundary and initial conditions
The boundary conditions are configured to represent runaway gas accretion in the absence of external irradiation or prior assumptions regarding the deep planetary interior. Following Chen & Bai [19], we impose the radiation boundary conditions
where and denote the inner and outer boundary radii, respectively.
At the hydrodynamic outer boundary, gas enters steadily at the local free-fall velocity for a prescribed mass accretion rate :
At the inner boundary, a reflective condition is maintained until the basal gas pressure reaches a specified target value. Thereafter, the primitive variables are held fixed and the ghost-cell velocity is set to zero, allowing gas to exit the domain whenever the mass flux predicted by the approximate HLLC Riemann solver (Chen et al. [21]) is negative. This treatment mimics a quasi-static envelope whose structure remains stationary over a free-fall timescale. Over secular evolutionary timescales, however, the thermodynamic state at evolves, necessitating the coupled shock–interior framework developed in Sect. 3.
Because we consider only steady-state solutions, the final profiles are insensitive to the initial state provided the initial gas mass in the domain is negligible. We therefore initialize all runs with , , and .
Computational grid
To resolve the sharp shock structure, we employ a geometrically stretched radial grid extending from to . The base grid comprises cells with a constant geometric stretching ratio . We apply static mesh refinement (SMR) across the shock region, with each refinement level doubling the spatial resolution. Using five to six refinement levels yields a finest cell width between 12.04 km and 54.20 km. By comparison, the local atmospheric pressure scale height,
is resolved by more than 16 cells near the inner boundary.
A representative simulation from Guangqi
Parameter survey with Guangqi
Figure Figure 2 displays a representative steady-state solution for a accreting protoplanet within a domain spanning at and . The infalling gas accelerates in free fall from the outer boundary (Figure 2a) before decelerating abruptly at an accretion shock near . The flow requires s to build up an upper atmosphere and reach the target inner boundary pressure. We define the post-shock radius as the location downstream of the Zel’dovich spike where the gas and radiation temperatures equilibrate (Figure 2e), marked by open circles in each panel. As shown in Figure 2i, radiative preheating from the shock drives substantial H dissociation in the upstream flow (Chen & Bai [19]), leaving several percent of the hydrogen in atomic form in the post-shock region. Simultaneously, the intense accretion luminosity sublimates silicate grains, reducing the Rosseland mean opacity to (Figure 2d).
Figure 2. Representative steady-state accretion shock solution from Guangqi for a gas giant with , , , and . Panels show radial profiles of (a) gas velocity, (b) density, (c) pressure, (d) Planck mean opacity (blue) and Rosseland mean opacity (red), (e) radiation temperature (blue) and gas temperature (red), (f) reduced radiation flux, (g) luminosity, (h) specific entropy, and (i) H mass fraction. Open circles mark the post-shock state.
We draw 1000 parameter combinations spanning the bulk of the physical regime relevant to runaway gas accretion: planetary mass , radius (Spiegel & Burrows [79]; Hatzes & Rauer [38]; Bashi et al. [5]; Kubyshkina & Fossati [47]; Thomas et al. [82]), mass accretion rate (Mordasini et al. [63]; Lambrechts et al. [50]), and internal luminosity (Marley et al. [56]; Mordasini [62]). Because cases with very low internal luminosity produce nearly indistinguishable post-shock states, we restrict our sample to , where is the accretion luminosity assuming complete conversion of gravitational potential energy into radiation. Below, we examine how each physical parameter governs the post-shock specific entropy, .
Planet mass, radius, accretion rate, and hydrogen dissociation
In the absence of H dissociation and internal molecular excitations, the specific entropy follows the ideal-gas expression:
where is the mean molecular weight, is the adiabatic index, and defines a reference thermodynamic state. Because is set primarily by the inflow ram pressure while is regulated by radiative transport, both quantities follow approximate power-law scalings (Marleau et al. [55], adapted from Eqs. 16 and 34b of):
where and are the post-shock radius and Stefan–Boltzmann constant. In this work, because is at the base of the atmosphere while is measured at the post-shock radius in Guangqi simulations. Without H dissociation, typically remains below 1600 K. Within this regime, silicate dust sublimation drives a steep drop in with increasing temperature [78].2 Adopting a representative scaling together with (consistent with our EoS when H remains intact) yields
where constant terms are absorbed into . Equation (16) predicts that correlates positively with and negatively with both and . Despite the strong nonlinearities of the full RHD system, the Guangqi results in Figure 3 clearly recover two of these three trends.
Figure 3. Post-shock specific entropy as a function of initial planetary radius across varying accretion rates (colors). Columns from left to right correspond to , and rows correspond to . Filled squares denote simulations in which the post-shock atomic hydrogen mass fraction exceeds . The upper x-axis indicates the critical accretion rate (Eq. (17)) corresponding to each combination of and .
In Figure 3, simulations with significant H dissociation () are marked by filled squares, whereas dissociation-free models are shown as circles. At fixed and within the dissociation-free regime (typically at large ), decreases with increasing , in agreement with Eq. (16). Similarly, at fixed and , increases monotonically with . By contrast, no prominent negative correlation between and is apparent across the parameter grid.
At high accretion luminosities, intense radiative heating raises the gas temperature to the H dissociation threshold, where each dissociating molecule absorbs 4.48 eV, producing a sharp jump in specific entropy. Consequently, the dissociated
models (filled squares in Figure 3) exhibit substantially higher than the undissociated models—a thermal dichotomy first identified by Chen & Bai [19]. Neglecting internal luminosity, the critical accretion rate required to trigger H dissociation is approximately Chen & Bai [19]
assuming a geometric covering fraction (adopted throughout this work), a reduced radiation flux following dust sublimation, and a global shock efficiency (Eq. (23)) prior to significant dissociation. Although a finite lowers by roughly –, we retain Eq. (17) as a reference benchmark along the upper -axis of Figure 6 and as a reference contour in Figure 7.
Figure 6. Comparison between post-shock specific entropy predicted by the PINN (curves) and measured directly from Guangqi simulations (filled circles) for a planet. Curves span (colors) for (solid) and 0.3 (dashed). Simulation points correspond to and . The upper x-axis marks the critical accretion rate for dissociation from Eq. (17).
Figure 7. PINN-predicted post-shock atomic hydrogen mass fraction (top row) and specific entropy (bottom row) across the parameter space for (left to right) at fixed . Top panels show contours of (black) alongside the analytical dissociation threshold from Eq. (17) (white dashed). Bottom panels show contours of . Overlaid circles indicate the global shock efficiency (Eq. (23)}) measured in Guangqi simulations: red (78%–83%), purple (83%–88%), orange (88%–93%), green (93%–98%), and blue (98%–100%).
Internal luminosity
Physically, provides an additional basal heat flux that warms the post-shock region and increases . Across our simulation grid, generally correlates positively with when , , and are held fixed.
In certain extreme regimes—typically involving partially dissociated post-shock gas—a very high induces noticeable expansion of the outer envelope, which can cause to decrease slightly. An example appears in Figure 3 when comparing and for and . However, such conditions are unlikely to arise in nature, because low-mass planets cannot sustain high ratios while accretion luminosity dominates.
Opacity
Dust opacity strongly influences the thermal evolution of young planets during their early formation stages [75]. To quantify its impact on , we run two sets of simulations for a planet at low () and high () accretion rates, comparing our fiducial opacity from Semenov et al. [78] against a reduced opacity scaled by a factor of 0.5 (Figure 4).
Figure 4. Effect of dust opacity on post-shock specific entropy for a gas giant at (left) and (right). Solid lines adopt the fiducial dust opacity of Semenov et al. [78], whereas dashed lines adopt half that opacity with all other parameters fixed. The two curves overlap completely in the right panel.
At the low accretion rate (left panel), higher dust opacity yields a larger , because the elevated optical depth impedes radiative cooling behind the shock. At the high accretion rate (right panel), however, shock heating raises temperatures above the dust sublimation threshold, rendering completely insensitive to the initial dust opacity.
Machine learning algorithm
To enable rapid interpolation across our four-dimensional parameter space (, , , and ), we construct a PINN regression model based on a multilayer perceptron (MLP) architecture (Figure 5). The network outputs the post-shock specific entropy (), temperature (), and pressure (), which are passed directly to MESA as outer boundary conditions for interior evolution calculations (Sect. 3).
Figure 5. Architecture of the physics-informed neural network (PINN). Alongside the primary inputs, the physics-guided feature (orange) is supplied to the input layer. Within the dense hidden layers, the post-shock density (green) is incorporated as a supervised latent variable. A customized physics-informed loss function (red) replaces the standard MSE objective. Depending on the EoS formulation in , the output layer returns either or .
Physical priors and thermodynamic constraints are embedded into the PINN at three levels: the input layer, the hidden layers, and the loss function—drawing on recent advances in physics-guided feature engineering [87, 41], physically interpretable latent variables [69], and physics-informed loss functions [30, 67] (Yu et al. 2020).
First, we augment the four baseline inputs with the composite feature , which scales directly with the specific gravitational potential energy released by the infalling gas and governs the overall accretion energy budget. Second, because the post-shock density is required to evaluate the thermodynamic relations in our loss function but is unknown prior to inference, we embed as a latent variable within a hidden layer. During training, this node is supervised against the simulated values, enabling the network to leverage density-dependent thermodynamic couplings during optimization without requiring post-shock quantities at inference time.
Third, rather than relying solely on a standard mean-squared-error (MSE) objective, we train the network using a composite loss function constrained by the EoS:
where is a vector of scalar weights and
Here, , , , and denote the standard MSE losses,
evaluated over samples between the simulated () and predicted () values of , , (or ), and , respectively. The two physics-informed penalty terms, and , are defined as
In both expressions, and are evaluated analytically via the EoS, the Saha equation, and hydrogen mass conservation (see Appendix A). Although and share a similar functional structure, they serve distinct roles: enforces thermodynamic consistency with the underlying EoS and Saha equilibrium on the training set, whereas enforces structural consistency between the thermodynamic states implied by the simulated and predicted pairs. Embedding physical structure across the inputs, latent representation, and loss function improves physical interpretability and mitigates overfitting on our 1000-model training set.
PINN Predictor
Trained on the 1000 Guangqi simulations, the PINN predictor evaluates post-shock thermodynamic states in per query while achieving a coefficient of determination (Appendix B). Figure 6 illustrates its accuracy for a planet by comparing PINN predictions (curves) against direct Guangqi measurements (filled circles) at and . The PINN closely reproduces the simulation benchmarks across both thermodynamic regimes: in the dissociation-free regime, decreases with and increases with , as expected from Sect. 2.2.1, whereas across the dissociation boundary, the network accurately captures the sharp entropy jump.
Because the predictor generalizes continuously across the parameter range defined in Sect. 2.2, we use it to construct high-resolution lookup tables sampled at and
. Figure 7 maps the two most sensitive post-shock diagnostics— (top row) and (bottom row)—across the plane for three accretion rates.
In the top panels of Figure 7, the contours delineate the onset and progression of dissociation. The close alignment between the analytical threshold of Eq. (17) (white dashed curve) and the contour confirms the reliability of this analytical criterion. In the bottom panels, the contours trace the accompanying rise in post-shock entropy, with the and contours closely matching the and levels, respectively.
Finally, we overlay the global shock efficiency measured from Guangqi as colored circles in Figure 7. Defined as the fraction of accretion energy radiated away by the shock ([54, 19]), is given by
where , and , , , and denote the specific internal energy, specific kinetic energy, specific enthalpy compression term, and gravitational potential, respectively, evaluated between the outer boundary and the post-shock radius . Across all 1000 simulations in our survey, spans 78.52% to 99.82%. Lower radiative efficiencies occur systematically when dissociation absorbs the 4.48 eV binding energy per molecule, diverting a substantial fraction of the infalling gravitational potential energy into internal chemical energy rather than emergent radiation.
Application to the primordial states of gas giants with MESA
To assess the impact of dissociation on gas giant formation, we model the accretion and long-term interior evolution of young planets using MESA r10398 (Paxton et al. [70, 71, 72, 73, 74]; Jermyn et al. [42]), following the methodology of Berardo et al. [9].
Initial and boundary conditions
We construct the initial planetary models using the make_planet test suite in MESA. Each simulation starts from a fully convective gaseous sphere of initial mass and radius . Following Paxton et al. [71], we gradually replace the innermost mass of the envelope with an inert core of fixed density ([7]), and subsequently relax the planet for yr to establish thermal and hydrostatic equilibrium.
Runaway gas accretion is implemented via the atmosphere module in run_star_extras.f90, which deposits newly accreted material onto the top of the envelope with a prescribed thermodynamic state. We compare two distinct outer boundary prescriptions:
Photospheric (cold-start baseline): Uses MESA’s default photospheric boundary condition, assigning the accreted gas the local photospheric specific entropy.
Realistic EoS: Replaces the photospheric boundary with the post-shock state (, ) returned by our PINN predictor for the instantaneous values of , , , and (Sects. 2.3 and 2.4), depositing accreted gas at the corresponding post-shock entropy .
The mass accretion rate is parameterized as a Gaussian function of instantaneous planetary mass :
which is symmetric about both the mass midpoint and the temporal midpoint of the accretion phase. As a proof of concept, we model the runaway growth of three representative planets reaching final masses of , using the accretion parameters listed in Table 1. Throughout all simulations, the accreted gas is assumed to have a solar composition of , , and ([2]).
| Saturn-like | Jupiter-like | Super-Jupiter | |
| [M] | 0.3 | 1.0 | 3.0 |
| [Myr] | 0.141 | 0.237 | 0.451 |
| [M yr] | 0.0002 | 0.0003 | 0.0005 |
| [M yr] | 0.002 | 0.005 | 0.01 |
| [M] | 0.126 | 0.2 | 0.2 |
| [M] | 0.063 | 0.1 | 0.1 |
Table 1. Prescribed runaway gas accretion histories for the three representative planetary mass models. The accretion duration is kept short relative to the disk lifetime to represent the rapid runaway accretion phase.
Internal evolution
Figures 8, 9, and 10 illustrate the interior evolution of density, temperature, specific entropy, and molecular hydrogen mass fraction () for the 0.3 , 1.0 , and 3.0 models, respectively. In each figure, the top row shows the “photospheric” boundary case and the bottom row shows the “realistic EoS” boundary case, with vertical grey dashed lines marking the end of accretion (). To highlight key interior transitions, we overlay the Mbar contour (below which metallic hydrogen is expected to form; [40]) on the density and entropy panels, K contours on the temperature panels, on the entropy panels, and contours on the composition panels.
For the lowest-mass case (; Figure 8), the interior profiles remain nearly identical between the two boundary prescriptions. At 1.0 and 3.0 , however, substantial structural differences emerge (Figure 9 and Figure 10). Under the “realistic EoS” boundary condition, the more massive planets develop markedly hotter, higher-entropy interiors, as reflected by the outward shift of the contour. This divergence arises from two effects: first, the realistic post-shock boundary retains higher entropy than the photospheric approximation even prior to dissociation; second, once post-shock temperatures exceed the H dissociation threshold, the specific entropy rises sharply. Because only sufficiently massive planets generate the gravitational potential energy needed to dissociate H at the shock front, high-entropy envelopes naturally emerge at .
Figure 8. Interior structural evolution of the model. Columns from left to right show radial profiles of density, temperature, specific entropy, and molecular hydrogen mass fraction () over time. Top and bottom rows correspond to the “photospheric” and “realistic EoS” outer boundary conditions (Sect. 3.1), respectively. Vertical grey dashed lines mark the end of runaway accretion (). White solid contours in the density and entropy panels indicate Mbar, denoting the approximate threshold for hydrogen metallization. Grey dashed contours in the temperature panels mark K, white dashed contours in the entropy panels trace , and purple solid and dashed contours in the panels denote , respectively.

Figure 9. Same layout as Figure 8, but for the model. White dotted contours in the panels indicate .

Figure 10. Same layout as Figure 9, but for the model.
In the fourth column of Figure 8–Figure 10, is evaluated using the Saumon–Chabrier–van Horn (SCVH) EoS ([77]). The decline of toward the planetary center reflects pressure-induced ionization and hydrogen metallization at high densities. Because the SCVH EoS is known to overestimate the hydrogen metallization pressure—which occurs around 1 Mbar in young gas giants ([40])—we adopt the Mbar surface as the boundary of the metallic dynamo region in Sect. 3.3. Employing the newer Chabrier–Mazevet–Soubiran (CMS) EoS ([17]) in updated versions of MESA would capture this transition more accurately.
Figure 11a shows the resulting intrinsic luminosity evolution for all three planetary masses, with filled circles marking the end of runaway accretion. Planets evolved with the PINN-predicted “realistic EoS” boundary condition are roughly 4–10 times more luminous than their “photospheric” counterparts. Consistent with the interior structural trends, this luminosity enhancement is strongest for , where post-shock H dissociation injects substantially higher entropy into the accreting envelope ([19]).
Dynamo-induced magnetic field strength
The magnetic field strength of a young gas giant depends primarily on its convective energy flux and rotation rate ([24]). Because our one-dimensional models do not evolve planetary spin, we assume rapid rotation in the regime where dynamo generation is independent of spin, thereby establishing an upper limit on the dipole field strength.
In planetary dynamo theory, efficient magnetic field generation requires high electrical conductivity ([15]). In gas giants, electrical conductivity is dominated by metallic and pressure-ionized hydrogen at pressures Mbar ([32, 40]). Following Fujii et al. [34], we therefore evaluate dynamo generation within the convective region at Mbar using the energy-flux scaling law of Christensen et al. [24]:
where is the vacuum magnetic permeability, is the mean-square magnetic field strength in the dynamo region, is the mean density of that region, is the convective energy flux at the outer boundary of the dynamo zone (, defined by Mbar), and is a geometric factor accounting for radial variations in flux and density. Following Fujii et al. [34], we set the dimensionless efficiency product to .
Assuming a dipolar geometry, the surface-extrapolated equatorial magnetic field strength at is given by
which we compute directly from the evolving interior properties at the Mbar surface.
Figure 11b displays the post-accretion evolution of . For the planet, the magnetic field starts weak ( G) before jumping abruptly to a few gauss at Myr and subsequently decaying. This rapid amplification is triggered when the Mbar layer transitions from a radiative to a convective state. In this low-mass regime, the “photospheric” and “realistic EoS” tracks remain closely aligned.
Figure 11. Evolution of (a) intrinsic luminosity and (b) dipole magnetic field strength at in the equatorial plane over the first 5 Myr for the (dotted), (solid), and (dashed) models. Blue curves represent the default “photospheric” MESA models, whereas red curves show the “realistic EoS” models incorporating post-shock boundary conditions from Guangqi. Filled circles in panel (a) denote the end of runaway accretion, and panel (b) displays the magnetic field evolution following the completion of accretion.
At higher planetary masses ( and ), strong dynamo fields are established immediately after accretion. For the planet, starts at G and declines gradually to G by 10 Myr, exhibiting only minor sensitivity to the outer boundary condition. For the planet, however, the boundary condition exerts a pronounced effect: the “realistic EoS” model generates an initial field of several hundred gauss—more than double the G field of the “photospheric” model—and sustains a systematically stronger field throughout the 10 Myr evolution. This enhanced dynamo output directly reflects the hotter interior and larger convective energy flux produced by shock-driven H dissociation. work, we present the first framework that directly couples first-principles RHD simulations—including post-shock H dissociation—with long-term interior evolution in MESA.
Discussion
Coupling accretion physics and interior evolution
Previous studies of gas giant formation and evolution have typically initialized young planets using constant- analytic prescriptions [8], broad random sampling of initial thermodynamic states [46], or idealized cold- and hot-start assumptions [80]. In this work, we present the first framework that directly couples first-principles RHD simulations—including post-shock H dissociation—with long-term interior evolution in MESA.
By comparing models evolved under default “photospheric” and our “realistic EoS” outer boundary conditions across three representative accretion histories (Sect. 3.2), we quantify how shock microphysics shapes the nascent planet. Overall, the divergence between the two boundary prescriptions grows progressively with planetary mass. While H dissociation begins to influence post-shock entropy around , it becomes a dominant driver of the thermal structure for massive gas giants (). Our interior models incorporate several simplifying assumptions, including an inert core ([8]), a homogeneous envelope composition that neglects primordial composition gradients, and the SCVH EoS ([77]), which overestimates the hydrogen metallization pressure. Although these approximations affect absolute interior profiles, they apply equally to both sets of simulations and therefore do not alter our differential comparison between the “photospheric” and “realistic EoS” boundary conditions. Nevertheless, because a planet’s initial thermodynamic state critically governs both its subsequent internal evolution ([84]) and the architecture of its circumplanetary environment and satellite system ([6, 7]), incorporating self-consistent composition gradients and updated equations of state remains an important avenue for future work.
Implications for observations and early evolution
High-contrast direct imaging has revealed several young protoplanets still embedded within their natal protoplanetary disks, detected via either infrared thermal continuum or H line emission. Disentangling whether the observed infrared flux is dominated by intrinsic planetary emission or ongoing accretion luminosity is crucial for inferring planetary masses and evolutionary states. As shown in Figure 11, our “realistic EoS” boundary conditions yield intrinsic luminosities 4–10 times higher than those predicted by standard cold-start models. Consequently, applying cold-start evolutionary tracks to directly imaged planets significantly underestimates their intrinsic brightness at a given mass—which in turn leads to an overestimate of the inferred planetary mass, particularly in the high-mass regime ().
Conversely, H emission is widely interpreted as a tracer of magnetospheric accretion and used to estimate mass accretion rates. Taking the well-studied protoplanet PDS 70c as an example, its observed H luminosity of ([36, 26, 89]) corresponds to a total accretion luminosity of under the assumption of pure magnetospheric accretion ([1]). However, this value is lower than the combined near- to mid-infrared photometric luminosity measured across the , , , , and bands ([60, 85, 86, 11] (Christiaens et al. 2024)), and the true bolometric luminosity of PDS 70c must be higher still. A substantial fraction of PDS 70c’s emission may therefore originate from the planet’s intrinsic cooling luminosity and its circumplanetary disk ([22]). Accurately predicting intrinsic planetary luminosities is thus essential for interpreting multi-wavelength observations of forming planets.
Finally, the planetary magnetic field is generated by dynamo action within the deep metallic hydrogen envelope. Applying the energy-flux scaling relation of Christensen et al. [24] following Fujii et al. [34], our predicted field strengths (Figure 11) represent theoretical upper limits, as the scaling assumes magnetic energy equipartition in the rapid-rotation regime. Even under this optimistic assumption, a Jupiter-mass planet’s equatorial dipole field at declines from G to G over the first 10 Myr. Such field strengths may be insufficient to magnetically truncate the circumplanetary disk around young Jupiter-mass planets ([7]), underscoring the need to explore a wider range of accretion histories with coupled evolutionary models.
Conclusions
In this work, we systematically investigated how H dissociation during runaway gas accretion shapes the initial thermal and magnetic states of young gas giants. Using a radiation-hydrodynamic framework ([19]) coupled with a realistic equation of state incorporating H dissociation via the Saha equation, we conducted 1000 steady-state accretion shock simulations with Guo & Y. [20] across a broad grid of planetary masses (), radii (), accretion rates (), and internal luminosities (). From this dataset, we trained a physics-informed neural network (PINN; Appendix B) to rapidly and accurately predict post-shock thermodynamic states. Coupling this predictor to MESA enabled us to self-consistently co-evolve the planetary interior structure and post-shock outer boundary conditions across prescribed accretion histories.
Our main findings are summarized as follows:
The post-shock specific entropy predicted by the PINN substantially exceeds the photospheric entropy typically adopted in cold-start models. For massive protoplanets () undergoing rapid accretion, partial H dissociation occurs naturally at the accretion shock (see Figure 10). Consequently, newly formed massive gas giants () attain intrinsic luminosities of during their first 1 Myr (Sect. 3)—roughly an order of magnitude brighter than their cold-start counterparts.
Using the convective energy flux within the metallic hydrogen dynamo region ( Mbar; [24]), we derived upper limits on the equatorial dipole magnetic field strength at (; Sect. 3.3) under fiducial accretion histories:
A Saturn-mass planet () generates a weak dipole field of only a few gauss, which is unlikely to truncate its circumplanetary disk (CPD) and points toward a distinct satellite formation pathway ([33, 34]).
A Jupiter-mass planet () begins with G, decaying to G over 10 Myr. Because these values likely represent theoretical upper limits, magnetospheric accretion from the CPD likely operates only once the disk surface density drops to low values ([35], e.g., g cm), at which point the low-mass CPD cannot efficiently spin down the planet via magnetic braking ([6]).
A super-Jupiter () remains intrinsically luminous (–) over its first 10 Myr and begins with a strong dipole field of G, decaying to G over 10 Myr. Such a field is capable, in principle, of driving magnetospheric accretion ([81]), though the resulting accretion rate depends sensitively on the CPD ionization fraction and warrants further study.
The elevated initial luminosities of massive gas giants significantly enhance their detectability via high-contrast direct imaging. By retaining substantial shock-generated heat, young giants remain bright at near- and mid-infrared wavelengths for hundreds of millions of years ([57]), creating strong synergies with current and upcoming direct-imaging surveys ([23, 28, 91]). Expanding the demographic sample of directly imaged young exoplanets will place stringent observational constraints on post-formation entropy and directly test the accretion shock physics modeled here.
Acknowledgements. This work is financially supported by the Tsinghua University Dushi Program and National Natural Science Foundation of China under grants No. 12342501. The Center of High performance computing at Tsinghua University provided the computational resources. We also acknowledge the Chinese Center for Advanced Science and Technology for hosting the Protoplanetary Disk and Planet Formation Summer School in 2024, which inspired this work.
Physics-informed loss function derivation
The specific entropy per unit mass of an ideal gas mixture is written as
where , , and denote the Boltzmann constant, gas temperature, and mass density, respectively, while and represent the number density and partition function of species .
Neglecting nuclear spin, electronic excitation, and molecular rotational–vibrational terms, the translational partition functions reduce to
where is the Planck constant and is the dissociation energy per H molecule.
Considering only the dissociation reaction , the number densities and are governed by hydrogen atom conservation and the Saha equation:
where is the hydrogen mass fraction adopted in our simulations and is the atomic hydrogen mass. Substituting Eqs. (A.2)–(A.6) into Eq. (A.1) simplifies the specific entropy to
Combining Eqs. (A.5) and (A.6) yields the equilibrium number densities for each species:
Physics-informed loss function derivation
To improve numerical stability and prevent vanishing or exploding gradients during optimization, all input features are standardized prior to training. Because standardization rescales the physical quantities, the original dimensional constants cannot be applied directly. We therefore introduce three global trainable coefficients—, , and —that are updated automatically alongside the network weights:
In terms of , , and , the partition functions become
and the species number densities are given by
Omitting the multiplicative constant during PINN training, the normalized specific entropy and pressure evaluated in the physics-informed loss terms are
and
Comparison between MLP and PINN
We evaluate model accuracy on a held-out test set of simulations using the coefficient of determination () and the mean absolute error (MAE):
where , and denote the predicted and simulated values for sample , and is the sample mean of the simulated target.
Figure B.1 compares the test-set predictions of the trained PINN against those of a baseline MLP across all three simultaneously predicted post-shock variables, while Fig. B.2 shows the corresponding relative residuals. The PINN achieves close agreement with the Guangqi simulations, exceeding across all three variables and reducing the MAE from , , and (MLP) to , , and (PINN), respectively. As shown in Fig. B.2, the baseline MLP struggles at the extremes of the parameter space: it systematically overestimates small values and underestimates large values of and , and exhibits large errors in below 2000 K. Incorporating physical constraints in the PINN effectively eliminates these systematic biases.
Figure B.1. Predicted versus simulated post-shock quantities (, , and ) on the test set, with the corresponding values indicated in each panel. Blue and grey markers denote predictions from the PINN and baseline MLP models, respectively. Across , , and , the MAE decreases from 0.2522 , 77.1961 K, and bar (MLP) to 0.1004 , 23.9201 K, and bar (PINN), respectively.
Figure B.2. Relative prediction errors for (a) post-shock specific entropy, (b) post-shock temperature, and (c) post-shock pressure across the test set. Black dashed lines denote zero residual.
Training the PINN converges within 20–30 min, with both training and validation losses decreasing rapidly before plateauing smoothly, validating our early-stopping criterion. Once trained, the network predicts post-shock thermodynamic states in s per query—compared to approximately 20 CPU hours required for a single full Guangqi simulation.
Several methodological aspects warrant further exploration. Whereas generalization in standard neural networks is typically assessed by comparing training and validation losses, rigorously evaluating out-of-domain generalization in PINNs is more complex ([12]) and lies beyond the scope of this work. Confirming reliable extrapolation would reduce dependence on large parameter surveys, allowing the PINN to predict post-shock states beyond the parameter boundaries sampled by Guangqi. Furthermore, because our architecture maps multiple inputs to multiple outputs while jointly optimizing three global physical coefficients, our 1000-model dataset represents a relatively modest sample size. Although PINNs are well suited to data-scarce regimes ([12]), expanding the training survey to Guangqi simulations in future work will further improve predictive robustness across the full parameter space.
References
- [1]Aoyama, Y. & Ikoma, M. 2019, ApJ, 885, L29
- [2]Asplund, M., Amarsi, A. M., & Grevesse, N. 2021, A&A, 653, A141
- [3]Balay, S., Abhyankar, S., Adams, M. F., et al. 2019, PETSc Users Manual, Tech. Rep. ANL-95/11 - Revision 3.11, Argonne National Laboratory
- [4]Balay, S., Gropp, W. D., McInnes, L. C., & Smith, B. F. 1997, in Modern Software Tools in Scientific Computing, ed. E. Arge, A. M. Bruaset, & H. P. Langtangen (Birkhäuser Press), 163–202
- [5]Bashi, D., Helled, R., Zucker, S., & Mordasini, C. 2017, Astronomy & Astrophysics, 604, A83
- [6]Batygin, K. 2018, AJ, 155, 178
- [7]Batygin, K. & Adams, F. C. 2025, Nature Astronomy, 9, 835
- [8]Berardo, D. & Cumming, A. 2017, ApJ, 846, L17
- [9]Berardo, D., Cumming, A., & Marleau, G.-D. 2017, ApJ, 834, 149
- [10]Bernardi, A., Zurlo, A., Cieza, L. A., et al. 2026, ApJ, 1009, L3
- [11]Blakely, D., Johnstone, D., Cugno, G., et al. 2025, AJ, 169, 137
- [12]Bonfanti, A., Santana, R., Ellero, M., & Gholami, B. 2024, Neural Computing and Applications, 36, 22677
- [13]Boss, A. P. 1997, Science, 276, 1836
- [14]Bowler, B. P. 2016, PASP, 128, 102001
- [15]Brandenburg, A. & Subramanian, K. 2005, Phys. Rep., 417, 1
- [16]Cameron, A. G. W. 1978, Moon and Planets, 18, 5
- [17]Chabrier, G., Mazevet, S., & Soubiran, F. 2019, ApJ, 872, 51
- [18]Chauvin, G., Desidera, S., Lagrange, A.-M., et al. 2017, A&A, 605, L9
- [19]Chen, Z. & Bai, X. 2022, Astrophysical Journal Letters, 925, L14
- [20]Chen, Z. & Bai, X.-N. 2026, ApJS, 283, 48
- [21]Chen, Z., Coleman, M. S. B., Blackman, E. G., & Frank, A. 2019, Journal of Computational Physics, 388, 490
- [22]Chen, Z., Yang, J., & Bai, X.-N. 2026, arXiv e-prints, arXiv:2610.21284arxiv.org/abs/2610.21284
- [23]Chilcote, J., Konopacky, Q., De Rosa, R. J., et al. 2020, in Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series, Vol. 11447, Ground-based and Airborne Instrumentation for Astronomy VIII, ed. C. J. Evans, J. J. Bryant, & K. Motohara, 114471S
- [24]Christensen, U. R., Holzwarth, V., & Reiners, A. 2009, Nature, 457, 167
- [26]Close, L. M., Males, J. R., Li, J., et al. 2025a, AJ, 169, 35
- [28]Desidera, S., Chauvin, G., Bonavita, M., et al. 2021, A&A, 651, A70
- [29]Drążkowska, J., Bitsch, B., Lambrechts, M., et al. 2023, in Astronomical Society of the Pacific Conference Series, Vol. 534, Protostars and Planets VII, ed. S. Inutsuka, Y. Aikawa, T. Muto, K. Tomida, & M. Tamura, 717
- [30]Ebert-Uphoff, I., Lagerquist, R., Hilburn, K., et al. 2021, arXiv preprint arXiv:2106.09757arxiv.org/abs/2106.09757
- [31]Elias-López, A., Del Sordo, F., Viganò, D., et al. 2025, A&A, 696, A161
- [32]French, M., Becker, A., Lorenzen, W., et al. 2012, ApJS, 202, 5
- [33]Fujii, Y. I. & Ogihara, M. 2020, A&A, 635, L4
- [34]Fujii, Y. I., Ogihara, M., & Hori, Y. 2026, Nature Astronomy
- [35]Fujii, Y. I., Okuzumi, S., Tanigawa, T., & Inutsuka, S.-i. 2014, ApJ, 785, 101
- [36]Haffert, S. Y., Bohn, A. J., de Boer, J., et al. 2019, Nature Astronomy, 3, 749
- [37]Hammond, I., Christiansen, V., Price, D. J., et al. 2023, MNRAS, 522, L51
- [38]Hatzes, A. P. & Rauer, H. 2015, Astrophysical Journal Letters, 810, L25
- [39]Helled, R., Bodenheimer, P., Podolak, M., et al. 2014, in Protostars and Planets VI, ed. H. Beuther, R. S. Klessen, C. P. Dullemond, & T. Henning, 643–665
- [40]Hinz, J., Karasiev, V. V., Hu, S. X., et al. 2020, Physical Review Research, 2, 032065
- [41]Huang, B. & Wang, J. 2022, IEEE Transactions on Power Systems, 38, 572
- [42]Jermyn, A. S., Bauer, E. B., Schwab, J., et al. 2023, ApJS, 265, 15
- [43]Johansen, A. & Lambrechts, M. 2017, Annual Review of Earth and Planetary Sciences, 45, 359
- [45]Knierim, H., Batygin, K., Helled, R., Morf, L., & Adams, F. C. 2026, A&A, 706, A51
- [46]Knierim, H. & Helled, R. 2025, A&A, 698, L1
- [47]Kubyshkina, D. & Fossati, L. 2022, Astronomy & Astrophysics, 668, A178
- [48]Lagrange, A.-M., Bonnefoy, M., Chauvin, G., et al. 2010, Science, 329, 57
- [49]Lambrechts, M. & Johansen, A. 2012, A&A, 544, A32
- [50]Lambrechts, M., Lega, E., Nelson, R. P., Crida, A., & Morbidelli, A. 2019, Astronomy & Astrophysics, 630, A82
- [51]Levermore, C. D. & Pomraning, G. C. 1981, ApJ, 248, 321
- [52]Macintosh, B., Graham, J. R., Barman, T., et al. 2015, Science, 350, 64
- [53]Malygin, M. G., Kuiper, R., Klahr, H., Dullemond, C. P., & Henning, T. 2014, A&A, 568, A91
- [54]Marleau, G.-D., Klahr, H., Kuiper, R., & Mordasini, C. 2017, ApJ, 836, 221
- [55]Marleau, G.-D., Mordasini, C., & Kuiper, R. 2019, The Astrophysical Journal, 881, 144
- [56]Marley, M. S., Fortney, J. J., Hubickyj, O., Bodenheimer, P., & Lissauer, J. J. 2007, The Astrophysical Journal, 655, 541
- [57]Marois, C., Macintosh, B., Barman, T., et al. 2008, Science, 322, 1348
- [58]Marois, C., Zuckerman, B., Konopacky, Q. M., Macintosh, B., & Barman, T. 2010, Nature, 468, 1080
- [59]Mayor, M. & Queloz, D. 1995, Nature, 378, 355
- [60]Mesa, D., Keppler, M., Cantalloube, F., et al. 2019, A&A, 632, A25
- [61]Mizuno, H. 1980, Progress of Theoretical Physics, 64, 544
- [62]Mordasini, C. 2013, Astronomy & Astrophysics, 558, A113
- [63]Mordasini, C., Marleau, G.-D., & Mollière, P. 2017, Astronomy & Astrophysics, 608, A72
- [64]Müller, S., Helled, R., & Mayer, L. 2018, ApJ, 854, 112
- [65]Nayakshin, S. 2010, MNRAS, 408, L36
- [67]Nugroho, A. A., Ashat, A., & et al. 2025, IOP Conference Series: Earth and Environmental Science, 1456, 012008
- [68]Ormel, C. W. & Klahr, H. H. 2010, A&A, 520, A43
- [69]Pawar, S., San, O., Aksoy l u, R., Rasheed, A., & Kvamsdal, T. 2021, Physics of Fluids, 33
- [70]Paxton, B., Bildsten, L., Dotter, A., et al. 2011, ApJS, 192, 3
- [71]Paxton, B., Cantiello, M., Arras, P., et al. 2013, ApJS, 208, 4
- [72]Paxton, B., Marchant, P., Schwab, J., et al. 2015, ApJS, 220, 15
- [73]Paxton, B., Schwab, J., Bauer, E. B., et al. 2018, ApJS, 234, 34
- [74]Paxton, B., Smolec, R., Schwab, J., et al. 2019, ApJS, 243, 10
- [75]Piso, A.-M. A., Youdin, A. N., & Murray-Clay, R. A. 2015, ApJ, 800, 82
- [76]Pollack, J. B., Hubickyj, O., Bodenheimer, P., et al. 1996, Icarus, 124, 62
- [77]Saumon, D., Chabrier, G., & van Horn, H. M. 1995, ApJS, 99, 713
- [78]Semenov, D., Henning, T., Helling, C., Ilgner, M., & Sedlmayr, E. 2003, A&A, 410, 611
- [79]Spiegel, D. S. & Burrows, A. 2012, ApJ, 745, 174
- [80]Sur, A., Burrows, A., Arevalo, R. T., & Su, Y. 2025, ApJ, 994, 186
- [81]Taylor, A. G. & Adams, F. C. 2024, Icarus, 415, 116044
- [82]Thomas, L., Nielsen, L. D., Acuña-Aguirre, L., & Cridland, A. 2025, A&A, 703, A221
- [84]Vazan, A., Helled, R., & Guillot, T. 2018, A&A, 610, L14
- [85]Wang, J. J., Ginzburg, S., Ren, B., et al. 2020, AJ, 159, 263
- [86]Wang, J. J., Vigan, A., Lacour, S., et al. 2021, AJ, 161, 148
- [87]Willard, J. L., Jia, X., Xu, S., Steinbach, M., & Kumar, V. 2022, ACM Comput. Surv., 55
- [89]Zhou, Y., Bowler, B. P., Sanghi, A., et al. 2025, ApJ, 980, L39
- [91]Zhu, Y., Zhao, G., Zhang, X., et al. 2026, arXiv e-prints, arXiv:2608.16215DOI










