betaFeedback

WorksMathematics

A high-order deterministic dynamical low-rank method for proton transport in heterogeneous media

Abstract

Dose calculations in proton therapy require the fast and accurate solution of a high-dimensional transport equation for a large number of (pencil) beams with different energies and directions. Deterministically solving this transport problem at a sufficient resolution can however be prohibitively expensive, especially due to highly forward peaked scattering of the protons. We propose using a model order reduction approach, the dynamical low-rank approximation (DLRA), which evolves the solution on the manifold of low-rank matrices in (pseudo-)time. For this, we compare a collided-uncollided split of the linear Boltzmann equation and its Fokker-Planck approximation. We treat the uncollided part using a ray-tracer and combine high-order phase space discretizations and a mixture model for materials with DLRA for the collided equation. Our method reproduces the results of a full-rank reference code at significantly lower rank, and thus computational cost and memory, and further makes computations feasible at much higher resolutions. At higher resolutions, we also achieve good accuracy with respect to TOPAS MC in homogeneous as well as heterogeneous materials. Finally, we demonstrate that several beam sources with different angles can be computed with little cost increase compared to individual beams.

Introduction

Accurate modeling of charged particle transport, particularly proton transport, is essential for applications such as radiation therapy, where dose deposition must be carefully controlled. The linear Boltzmann equation provides a physically accurate model for charged particle transport. However, solving it numerically is challenging due to multiscale effects, a high-dimensional phase space (space, angle, and energy), and strongly forward-peaked scattering. The latter necessitates a finely resolved phase space discretization or specialized methodological modifications, thus often leading to prohibitive memory requirements and computational costs. These computational requirements limit the feasibility of using the full Boltzmann model, particularly in settings that require many simulations, such as uncertainty quantification and treatment plan optimization in intensity-modulated proton therapy.

For this reason, Monte Carlo methods, which can statistically solve the full physical model at high precision by simulating large numbers of particles, require acceleration techniques and strong computational resources to be practical [15] [48]. Deterministic solutions to the Boltzmann equation remove the statistical noise and slow convergence of Monte Carlo methods and further allow access to full angle- and energy-resolved particle densities, thus better supporting further computations (e.g. uncertainty quantification) [22] [48]. Although the accuracy of such solvers can compete with that of Monte Carlo methods given equivalent physical modeling [21] [48], they are not as wide-spread due to their high algorithmic complexity and challenges in efficient implementation. Rather, simplified models such as pencil beam algorithms, which apply dose kernels precomputed from Fermi-Eyges theory [16] or Gaussian fits to measurements and Monte Carlo simulations, are frequently used in practice [18] [42, 45, 50]. While these models offer a computationally efficient alternative, they are based on simplifying assumptions which limit precision especially in heterogeneous media [46, 50]. Alternatively, deterministic approaches have been developed to efficiently resolve forward-peaked scattering, mainly based on multi-grid [38, 47] and adaptive grid refinement [29] [33]. While such refinement strategies can effectively resolve angular dependencies, they still require a sufficiently fine spatial resolution [32]. Another approach in computational electron therapy [27] reduces the memory footprint and computational costs by leveraging dynamical low-rank approximation (DLRA) [24]. This approach represents and evolves the solution to the linear Boltzmann equation in a low-rank format, thereby significantly reducing the number of degrees of freedom. Since DLRA is a dimensionality reduction method for time-dependent problems, a central idea in [27] is to interpret the energy as a pseudo-time by using a continuous slowing down approximation [31]. This interpretation treats energy loss in the medium as a surrogate for time evolution, enabling the use of time-dependent model reduction techniques such as DLRA.

A central challenge of DLRA is the inherent stiffness of the evolution equations of low-rank factors [23]. Therefore, robust time integrators that are not affected by this stiffness have been developed. The main idea of these integrators is to move only in flat subspaces of the low-rank manifold, thereby avoiding directions with a high curvature that result in stiff evolution equations [35]. Perhaps the most popular class of robust time integrators is the (augmented) basis-update & Galerkin (BUG) integrator [9, 7], which has since its introduction been modified to enable parallel update computations [8] and to achieve high–order accuracy [6, 25]. Robust integrators for DLRA have been used extensively in kinetic theory [13], ranging from applications in plasma physics [14, 10] to neutron criticality [28] [41].

The use of DLRA in such problems relies heavily on the derivation of efficient spatial and temporal discretizations that preserve underlying structures.

While the method proposed in [27] allows for highly efficient numerical simulations, it is currently limited to electron therapy, water-equivalent materials, as well as first-order temporal and spatial discretizations. An extension to proton simulations is challenging, primarily due to highly peaked scattering coefficients and the requirement of straggling terms in the continuous slowing down approximation. The latter makes the representation of energy as a pseudo-time more challenging and thus complicates the use of DLRA as a time-dependent model reduction technique.

In this work, we propose a novel extension of DLRA to proton transport that addresses these challenges through the following contributions: (1) a ray-tracing model for uncollided particles to handle straggling, (2) high-order phase-space discretizations, and (3) a low-rank material decomposition. The ray-tracing model exploits the small number of relevant angular directions of uncollided particles to achieve low computational costs while being able to efficiently incorporate straggling effects. Collided particles are then advanced without straggling, thus enabling the use of dynamical low-rank approximation. Moreover, sharp gradients in proton transport require high-order phase-space discretizations. To mitigate computational costs imposed by limiters, we propose a nodal second-order upwind stencil combined with a fourth-order Runge–Kutta method in pseudo-time, i.e. energy, for numerical stability. We further use a modal spherical harmonics method (PNP_{\mathrm{N}}) to discretize the directional domain. Lastly, to allow the incorporation of different materials, we propose a low-rank material decomposition similar to [28], which enables efficient evaluation of the DLRA evolution equations.

The remaining manuscript is structured as follows: First we give a short introduction to proton transport and the dynamical low-rank approximation in § 2. The former relies on the continuous slowing down approximation of the Boltzmann or Fokker-Planck equation, including a transformation that enables the use of energy as pseudo-time in the presence of material-dependent parameters. Then, the main method is introduced in § 3, including the ray tracer for the uncollided and spatial and angular discretizations as well as DLRA evolution equations based on the rank-adaptive BUG integrator for the time/energy update of the collided part of the equation. Finally, in § 4, we present a variety of numerical experiments and in § 5, we discuss the computational complexity and memory reduction. We provide a discussion and conclusion in § 6 and § 7.

Background

Transport model for proton therapy

We model the transport of particles through a medium using the continuous slowing down approximation of the steady-state linear Boltzmann equation

Ω⋅∇rψ(E,r,Ω)−∂S(E,r)ψ(E,r,Ω)∂E−12∂2T(E,r)ψ(E,r,Ω)∂E2=Γψ(E,r,Ω),(1)\boldsymbol{\Omega} \cdot\nabla_{\mathbf{r}}\psi(E,\mathbf{r},\boldsymbol{\Omega})-\frac{\partial S(E,\mathbf{r})\psi(E,\mathbf{r},\boldsymbol{\Omega})}{\partial E}-\frac{1}{2}\frac{\partial^{2}T(E,\mathbf{r})\psi(E,\mathbf{r},\boldsymbol{\Omega})}{\partial E^{2}}=\Gamma\psi(E,\mathbf{r},\boldsymbol{\Omega}), \tag*{(1)}

where the phase space of the particle density ψ\psi consists of energy E∈[Emin⁡,Emax⁡]⊂R+E \in[E_{\min},E_{\max}] \subset\mathbb{R}_{+}, space r∈R3\mathbf{r} \in\mathbb{R}^{3} and direction of flight Ω∈S2\boldsymbol{\Omega} \in S^{2}. We denote the stopping power by SS and straggling by TT. The collision operator Γ\Gamma on the right-hand side depends on the differential and total scattering cross sections Σs\Sigma_{s} and Σt\Sigma_{t} and is defined as

Γψ=∫S2Σs(E,r,Ω′⋅Ω)ψ(E,r,Ω′)dΩ′−Σt(E,r)ψ(E,r,Ω)(2)\Gamma\psi=\int_{S^{2}}\Sigma_{s}(E,\mathbf{r},\boldsymbol{\Omega}'\cdot\boldsymbol{\Omega})\psi(E,\mathbf{r},\boldsymbol{\Omega}')\mathrm{d}\boldsymbol{\Omega}'-\Sigma_{t}(E,\mathbf{r})\psi(E,\mathbf{r},\boldsymbol{\Omega}) \tag*{(2)}

in case of the Boltzmann operator and as

Γψ=ξ12L(3)\Gamma\psi=\frac{\xi_{1}}{2}L \tag*{(3)}

for the Fokker-Planck operator, where ξ1\xi_{1} is the first transport coefficient defined as

ξ1=2π∫−11(1−μ0)Σs(E,r,μ0)dμ0,μ0=cos⁡(θ)=Ω′⋅Ω(4)\xi_{1}=2\pi\int_{-1}^{1}(1-\mu_{0})\Sigma_{s}(E,\mathbf{r},\mu_{0})\mathrm{d}\mu_{0},\qquad\mu_{0}=\cos(\theta)=\boldsymbol{\Omega}'\cdot\boldsymbol{\Omega} \tag*{(4)}

and LL is the spherical Laplace–Beltrami operator

L=[∂∂μ(1−μ2)∂∂μ+11−μ2∂2∂Φ2],(5)L=\left[\frac{\partial}{\partial\mu}(1-\mu^{2})\frac{\partial}{\partial\mu}+\frac{1}{1-\mu^{2}}\frac{\partial^{2}}{\partial\Phi^{2}}\right], \tag*{(5)}

where the spherical coordinates μ∈[−1,1]\mu\in[-1,1] and Φ∈[0,2π)\Phi\in[0,2\pi) represent the cosine of the polar angle and the azimuthal angle, respectively. Note, that, under the assumption of highly forward-peaked differential cross sections, the Fokker-Planck operator has been shown to be a first-order approximation to the Boltzmann operator [34].

In order to later use energy as pseudo-time and separate angular and spatial dependency in the collision operator we apply the transformation ψ~(E,r,Ω):=S(E,r)ψ(E,r,Ω)\tilde{\psi}(E,\mathbf{r},\boldsymbol{\Omega}) := S(E,\mathbf{r})\psi(E,\mathbf{r},\boldsymbol{\Omega}) and assume that the tissue at each point in space is composed of a mixture of the same set A={H,C,N,O,Na,Mg,P,S,Cl,Ar,K,Ca}\mathcal{A} = \{\mathrm{H}, \mathrm{C}, \mathrm{N}, \mathrm{O}, \mathrm{Na}, \mathrm{Mg}, \mathrm{P}, \mathrm{S}, \mathrm{Cl}, \mathrm{Ar}, \mathrm{K}, \mathrm{Ca}\} of base materials with differing weights wi(r)w_i(\mathbf{r}), i∈Ai \in\mathcal{A} [2]. Then the stopping power and cross sections can also be computed as weighted mixtures of the respective parameters for the individual elements using the Bragg additivity rule [17]. Note that this model for material composition is much more accurate [17, 20] than assuming water-equivalence as in [27]. Then scattering cross sections Σs\Sigma_s are computed as

Σs(E,r,μ0)=∑i∈Awi(r)Σs;i(E,μ0),(6)\Sigma_{s}(E,\mathbf{r},\mu_{0})=\sum_{i\in\mathcal{A}}w_i(\mathbf{r})\Sigma_{s;i}(E,\mu_{0}), \tag*{(6)}

where Σs;i\Sigma_{s;i} is the scattering cross-section of material ii. The total scattering cross section Σt(E,r)\Sigma_t(E,\mathbf{r}) is computed analogously, where for material ii we have Σt;i(E)=2π∫−11Σs;i(E,μ0) dμ0\Sigma_{t;i}(E)=2\pi\int_{-1}^{1}\Sigma_{s;i}(E,\mu_0)\,d\mu_0. The resulting transport equation then takes the following form

−∂ψ~(E,r,Ω)∂E=−Ω⋅∇rψ~(E,r,Ω)S(t,r)+12∂2T(E,r)∂E2S(E,r)ψ~(E,r,Ω)+∑i∈AwiΓiψ~(E,r,Ω)S(E,r),(7)-\frac{\partial\widetilde{\psi}(E,\mathbf{r},\mathbf{\Omega})}{\partial E}=-\mathbf{\Omega}\cdot\nabla_{\mathbf{r}}\frac{\widetilde{\psi}(E,\mathbf{r},\mathbf{\Omega})}{S(t,\mathbf{r})}+\frac{1}{2}\frac{\partial^2T(E,\mathbf{r})}{\partial E^2S(E,\mathbf{r})}\widetilde{\psi}(E,\mathbf{r},\mathbf{\Omega})+\sum_{i\in A}w_i\Gamma_i\frac{\widetilde{\psi}(E,\mathbf{r},\mathbf{\Omega})}{S(E,\mathbf{r})}, \tag*{(7)}

where for Boltzmann

Γiψ=∫S2Σs;i(E,r,Ω′⋅Ω)ψ(E,r,Ω′) dΩ′−Σt;i(E,r)ψ(E,r,Ω)(8)\Gamma_i\psi=\int_{S^2}\Sigma_{s;i}(E,\mathbf{r},\mathbf{\Omega}'\cdot\mathbf{\Omega})\psi(E,\mathbf{r},\mathbf{\Omega}')\,d\mathbf{\Omega}'-\Sigma_{t;i}(E,\mathbf{r})\psi(E,\mathbf{r},\mathbf{\Omega}) \tag*{(8)}

and for Fokker-Planck

Γiψ=ξ1;i2L,ξ1;i=2π∫−11(1−μ0)Σs;i(E,r,μ0) dμ0.(9)\Gamma_i\psi=\frac{\xi_{1;i}}{2}L,\qquad\xi_{1;i}=2\pi\int_{-1}^{1}(1-\mu_0)\Sigma_{s;i}(E,\mathbf{r},\mu_0)\,d\mu_0. \tag*{(9)}

Modeling of physical quantities

The particle and material dependent quantities in the transport equation include scattering cross sections, stopping power and straggling coefficients. The accuracy of our model strongly depends on a good modeling of these quantities. We determine the values for the set AA containing 12 elements which are used to constitute tissue in the human body and mix these according to their proportions at each point in space. For comparability, we extract the stopping power for each material and different energies from the Monte Carlo code TOPAS MC [40]. Such data is not as easily obtained for straggling and differential cross sections, therefore we determine these based on physics models.

We use William’s model [49] to compute the modified straggling coefficient:

T(r,E)=∑i∈A1(4πϵ0)2Ni(r)4πe4Zi(4Ii3mevp(E)2ln⁡2mevp(E)2Ii),(10)T(\mathbf{r},E)=\sum_{i\in A}\frac{1}{(4\pi\epsilon_0)^2}N_i(\mathbf{r})4\pi e^4Z_i\left(\frac{4I_i}{3m_e v_p(E)^2}\ln\frac{2m_e v_p(E)^2}{I_i}\right), \tag*{(10)}

where ϵ0\epsilon_0 is the vacuum permittivity, Ni(r)N_i(\mathbf{r}) are the atomic densities of atom ii within the material at position r\mathbf{r}, ZiZ_i the respective atomic numbers, IiI_i the ionization energies, mem_e the electron rest mass and vp(E)v_p(E) the velocity of the incident proton.

For the differential elastic scattering, we use Molière’s model [44, 36] which has been found in [3] to achieve good agreement with TOPAS MC and improved results compared to the small-angle first Born approximation. Here, the macroscopic differential cross section of the mixture material is given by

Σs(E,r,μ0)=∑i∈Aτi;lab(μ0)4Ni(r)α2k(E)2(1−μ0+χα;i(E)),(11)\Sigma_s(E,\mathbf{r},\mu_0)=\sum_{i\in A}\tau_{i;\mathrm{lab}}(\mu_0)\frac{4N_i(\mathbf{r})\alpha^2}{k(E)^2(1-\mu_0+\chi_{\alpha;i}(E))}, \tag*{(11)}

where χα;i(E)=χ0;i(E)2(1.13+3.76ai(E)2)\chi_{\alpha;i}(E)=\chi_{0;i}(E)^2(1.13+3.76a_i(E)^2) is the Molière corrected screening parameter according to [44, 36] with χ0;i(E)=1.13αZi1/3mecp(E)\chi_{0;i}(E)=\frac{1.13\alpha Z_i^{1/3}m_ec}{p(E)} and ai=Ziαβ(E)a_i=\frac{Z_i\alpha}{\beta(E)} representing the deviation from the Born approximation [1]. Further, β=vp(E)c\beta= \frac{v_p(E)}{c} is the relative velocity, α\alpha the fine structure constant and k(E)=p(E)ℏk(E) = \frac{p(E)}{\hbar} the reduced wave number with p(E)p(E) the momentum and ℏ\hbar the reduced Planck constant. Note, that equation (11) contains a factor to convert the cross sections from the center of mass to the laboratory frame of reference, which can be computed as

τi;lab(μ0)=(1+2μ0⋅mpmi+(mpmi)2)321+μ0⋅mpmi.\tau_{i;\mathrm{lab}}(\mu_0) = \frac{\left(1+2\mu_0\cdot\frac{m_p}{m_i}+\left(\frac{m_p}{m_i}\right)^2\right)^{\frac{3}{2}}}{1+\mu_0\cdot\frac{m_p}{m_i}} .

where mpm_p is the proton rest mass and mim_i the rest mass of the target nucleus [2]. Note, that since protons and neutrons have almost equal mass mpmi=Ai\frac{m_p}{m_i}=A_i can be used to simplify the computation, where AiA_i is the atomic mass number.

For simplicity, we neglect nuclear scattering (including production of secondary neutrons and recoil nuclei) and absorption as their effect on the deposited energy is relatively small [12] [39] and they do not affect the numerical method presented here.

Dynamical low-rank approximation

Discretizing the transport equations presented in the previous section will result in a large matrix ordinary differential equation

u˙(t)=F(t,u(t)),u(t0)=u0,(12)\dot{\mathbf{u}}(t)=\mathbf{F}(t,\mathbf{u}(t)),\qquad\mathbf{u}(t_0)=\mathbf{u}_0, \tag*{(12)}

where u(t)∈Rn×m\mathbf{u}(t)\in\mathbb{R}^{n\times m} and energy will act as a pseudo-time. Here, nn refers to the number of spatial cells and mm is the number of moments of the modal angular discretization.

A rank rr approximation to u(t)\mathbf{u}(t) is given by u(t)≈X(t)S(t)V(t)T\mathbf{u}(t)\approx\mathbf{X}(t)\mathbf{S}(t)\mathbf{V}(t)^{\mathsf{T}} where X(t)∈Rn×r\mathbf{X}(t)\in\mathbb{R}^{n\times r} and V(t)∈Rm×r\mathbf{V}(t)\in\mathbb{R}^{m\times r} are orthonormal basis matrices and S(t)∈Rr×r\mathbf{S}(t)\in\mathbb{R}^{r\times r} is a not necessarily diagonal coefficient matrix. Here, the rank rr matrices form a smooth manifold, which we denote by M\mathcal{M}. The core idea of dynamical low-rank approximation is to evolve the low-rank factors instead of the full solution in time [24]. This is achieved by projecting the dynamics of (12) onto the tangent space of M\mathcal{M} at the current solution. The projection restricts the dynamics of the solution to the low-rank manifold M\mathcal{M}, thus ensuring a memory-efficient solution representation during the entire simulation. The projected equations then read

u˙r(t)=P(ur(t))F(t,ur(t)),(13)\dot{\mathbf{u}}_r(t)=\mathbf{P}(\mathbf{u}_r(t))\mathbf{F}(t,\mathbf{u}_r(t)), \tag*{(13)}

where for Z=XΣVT\mathbf{Z}=\mathbf{X}\mathbf{\Sigma}\mathbf{V}^{\mathsf{T}}, the projector onto the tangent space of M\mathcal{M} at Z\mathbf{Z} takes the form

P(Z)G=XXTG(I−VVT)+GVVT.\mathbf{P}(\mathbf{Z})\mathbf{G}=\mathbf{X}\mathbf{X}^{\mathsf{T}}\mathbf{G}(\mathbf{I}-\mathbf{V}\mathbf{V}^{\mathsf{T}})+\mathbf{G}\mathbf{V}\mathbf{V}^{\mathsf{T}}.

A graphical illustration of dynamical low-rank approximation can be found in Figure 1. Since this manifold can exhibit high curvature, conventional time integration methods require prohibitively small step sizes. Therefore, careful construction of new time integration methods that take the geometry of M\mathcal{M} into account when solving (13) is required [23]. Robust time integration methods include the projector–splitting integrator [35] and basis-update & Galerkin (BUG) integrators [9, 7, 8]. The main idea of these integrators is to move on flat subspaces of the low-rank manifold, thereby avoiding directions with high curvature. The augmented BUG integrator [7], which we use in this work, first updates the basis matrices U\mathbf{U} and V\mathbf{V} in parallel, constraining the dynamics to

Graphical illustration of dynamical low-rank approximation

Figure 1. Graphical illustration of dynamical low-rank approximation. The solution at time t0t_0 lies on the manifold of rank rr matrices M\mathcal{M}. It will not stay on M\mathcal{M} if it moves according to the original dynamics (i.e., into direction F(ur(t0))\mathbf{F}(\mathbf{u}_r(t_0))), but remains of rank rr if the original dynamics are projected onto the tangent space Tur(t)M\mathcal{T}_{\mathbf{u}_r(t)}\mathcal{M} via the projector P(ur(t))\mathbf{P}(\mathbf{u}_r(t)).

{KV(t0)T∣K∈Rn×r}⊂Mand{U(t0)LT∣L∈Rm×r}⊂M.(14)\{\mathbf{K}\mathbf{V}(t_0)^{\mathsf{T}} \mid\mathbf{K} \in\mathbb{R}^{n \times r}\} \subset\mathcal{M} \quad\text{and} \quad\{\mathbf{U}(t_0)\mathbf{L}^{\mathsf{T}} \mid\mathbf{L} \in\mathbb{R}^{m \times r}\} \subset\mathcal{M}. \tag*{(14)}

This leads to the differential equations

K˙(t)=F(t,K(t)V0T)V0,K(t0)=U0S0,(15)\dot{\mathbf{K}}(t) = \mathbf{F}(t,\mathbf{K}(t)\mathbf{V}_0^{\mathsf{T}})\mathbf{V}_0,\qquad\mathbf{K}(t_0) = \mathbf{U}_0\mathbf{S}_0, \tag*{(15)}
L˙(t)=F(t,U0L(t)T)TU0,L(t0)=V0S0T.(16)\dot{\mathbf{L}}(t) = \mathbf{F}(t,\mathbf{U}_0\mathbf{L}(t)^{\mathsf{T}})^{\mathsf{T}}\mathbf{U}_0,\qquad\mathbf{L}(t_0) = \mathbf{V}_0\mathbf{S}_0^{\mathsf{T}}. \tag*{(16)}

The time updated basis matrices at time t1=t0+Δtt_1=t_0+\Delta t, which we denote by U^\widehat{\mathbf{U}} and V^\widehat{\mathbf{V}} can then be retrieved by a QR decomposition of [K(t1),U0]=U^S^K[\mathbf{K}(t_1),\mathbf{U}_0]=\widehat{\mathbf{U}}\widehat{\mathbf{S}}_{\mathbf{K}} and [L(t1),V0]=V^S^L[\mathbf{L}(t_1),\mathbf{V}_0]=\widehat{\mathbf{V}}\widehat{\mathbf{S}}_{\mathbf{L}}, where [A,B]∈Rm×2r[\mathbf{A},\mathbf{B}]\in\mathbb{R}^{m\times2r} denotes the concatenation of two matrices A,B∈Rm×r\mathbf{A},\mathbf{B}\in\mathbb{R}^{m\times r}. Lastly, to update the coefficient matrix S\mathbf{S}, the integrator constrains the dynamics to {U^SV^T∣S∈R2r×2r}⊂M2r\{\widehat{\mathbf{U}}\mathbf{S}\widehat{\mathbf{V}}^{\mathsf{T}}\mid\mathbf{S}\in\mathbb{R}^{2r\times2r}\}\subset\mathcal{M}_{2r} by solving the Galerkin system

S˙(t)=U^TF(t,U^S(t)V^T)V^,S(t0)=(U^TU0)S0(V0TV^).(17)\dot{\mathbf{S}}(t)=\widehat{\mathbf{U}}^{\mathsf{T}}\mathbf{F}(t,\widehat{\mathbf{U}}\mathbf{S}(t)\widehat{\mathbf{V}}^{\mathsf{T}})\widehat{\mathbf{V}},\qquad\mathbf{S}(t_0)=(\widehat{\mathbf{U}}^{\mathsf{T}}\mathbf{U}_0)\mathbf{S}_0(\mathbf{V}_0^{\mathsf{T}}\widehat{\mathbf{V}}). \tag*{(17)}

The time-updated coefficient matrix S^\widehat{\mathbf{S}} is retrieved from S(t1)\mathbf{S}(t_1). Since the time-updated solution Y^=U^S^V^T\widehat{\mathbf{Y}}=\widehat{\mathbf{U}}\widehat{\mathbf{S}}\widehat{\mathbf{V}}^{\mathsf{T}} is of rank 2r2r, a new rank can be determined by a truncated singular value decomposition of S^\widehat{\mathbf{S}}. Given S^=PΣQT\widehat{\mathbf{S}}=\mathbf{P}\mathbf{\Sigma}\mathbf{Q}^{\mathsf{T}}, where Σ=diag⁡(σ1,⋯ ,σ2r)\mathbf{\Sigma}=\operatorname{diag}(\sigma_1,\cdots,\sigma_{2r}) collects the singular values of S^\widehat{\mathbf{S}}, the singular value decomposition is truncated at a new rank r1r_1 such that

∑i=r1+12rσi≤ϑ\sum_{i=r_1+1}^{2r}\sigma_i\leq\vartheta

with a user-determined truncation threshold ϑ\vartheta, see [7]. This yields a new rank r1r_1 factorization Y1=U1S1V1T\mathbf{Y}_1=\mathbf{U}_1\mathbf{S}_1\mathbf{V}_1^{\mathsf{T}}. The process of evolving the solution along (15) with subsequent truncation is repeated until the factorized solution at a given final time tkt_k is reached after kk steps. The equations (15) inherit the stiffness of the original problem, thus, no further stiffness is added from the underlying geometry of M\mathcal{M}. Therefore, classical time integration methods can be used to evolve the factorized solution. For a detailed discussion of BUG integrators, their inherent parallelism, and extensions to higher order, see [9, 7, 8, 6, 25].

Methods

In this section we detail the numerical method used to solve the problem described in § 2.1. We loosely follow the strategy in [27] of splitting the problem in collided and uncollided as well as scattering and streaming parts and using a PN\mathrm{P}_N-based dynamical low-rank approach for the collided equations. In contrast to [27], we however use a ray tracer including energy straggling for the uncollided part and the derivation for the collided equations differs due to the material dependency in cross sections and stopping power as well as the higher order spatial discretization. Together with the efficient incorporation of multiple materials as detailed in § 2.1.1, this enables the derivation of an efficient and accurate numerical solver to simulate proton radiation in a realistic setting.

Collided-uncollided split

We start by splitting equation (7) into a collided and uncollided part using a first collision source method. This will facilitate the inclusion of straggling and avoid challenging boundary conditions in the DLRA part of the method. We now separate the angular flux into collided particles ψc\psi_c and uncollided particles ψu\psi_u such that ψ=ψu+ψc\psi= \psi_u + \psi_c. For collided particles, we assume that straggling can be neglected, that is,

∂2Tψc∂E2≈0.(18)\frac{\partial^2 T\psi_c}{\partial E^2} \approx0. \tag*{(18)}

Then, we obtain

Ω⋅∇ψu=∂Sψu∂E+12∂2Tψu∂E2+∑i∈AwiΓout;iψu(E,r,Ω),(19)\boldsymbol{\Omega} \cdot\nabla\psi_u = \frac{\partial S\psi_u}{\partial E} + \frac{1}{2}\frac{\partial^2 T\psi_u}{\partial E^2} + \sum_{i\in A} w_i\Gamma_{\mathrm{out};i}\psi_u(E,\mathbf{r},\boldsymbol{\Omega}), \tag*{(19)}
−∂Eψ~c(E,r)=−Ω⋅∇ψ~c(E,r)S(E,r)+∑i∈AwiΓiψ~c(E,r)S(E,r)+∑i∈AwiΓin;iψu(E,r,Ω),(20)-\partial_E\widetilde{\psi}_c(E,\mathbf{r}) = -\boldsymbol{\Omega}\cdot\nabla\frac{\widetilde{\psi}_c(E,\mathbf{r})}{S(E,\mathbf{r})} + \sum_{i\in A}w_i\Gamma_i\frac{\widetilde{\psi}_c(E,\mathbf{r})}{S(E,\mathbf{r})} + \sum_{i\in A}w_i\Gamma_{\mathrm{in};i}\psi_u(E,\mathbf{r},\boldsymbol{\Omega}), \tag*{(20)}

where Γout;i\Gamma_{\mathrm{out};i} and Γout;i\Gamma_{\mathrm{out};i} are the in- and outscattering parts of the collision operator, respectively. So,

Γin;iψ=∫S2Σs;i(E,r,Ω′⋅Ω)ψ(E,r,Ω′) dΩ′,(21)\Gamma_{\mathrm{in};i}\psi= \int_{\mathbb{S}^2}\Sigma_{s;i}(E,\mathbf{r},\boldsymbol{\Omega}'\cdot\boldsymbol{\Omega})\psi(E,\mathbf{r},\boldsymbol{\Omega}')\,\mathrm{d}\Omega', \tag*{(21)}
Γout;iψ=Σt;i(E,r)ψ(E,r,Ω)(22)\Gamma_{\mathrm{out};i}\psi= \Sigma_{t;i}(E,\mathbf{r})\psi(E,\mathbf{r},\boldsymbol{\Omega}) \tag*{(22)}

for the Boltzmann operator and

Γin;iψ=Γiψ(E,r,Ω),Γout;iψ=0(23)\Gamma_{\mathrm{in};i}\psi= \Gamma_i\psi(E,\mathbf{r},\mathbf{\Omega}), \qquad\Gamma_{\mathrm{out};i}\psi= 0 \tag*{(23)}

for the Fokker-Planck operator.

Uncollided part: Ray tracer

For the uncollided particles, we use an in-house analytical ray tracer implemented in fortran [see also 2, 32].

Having split off the singular part as described above, the PDE along the ray for ψu\psi_u reads

∂ψu(E,z)∂z+∑i∈Awi(z)Γout;iψu(E,z)=∂S∗(E,z)ψu(E,z)∂E+12∂∂ET(E,z)∂ψu(E,z)∂E,(24)\frac{\partial\psi_u(E,z)}{\partial z} + \sum_{i\in A} w_i(z)\Gamma_{\mathrm{out};i}\psi_u(E,z) = \frac{\partial S^*(E,z)\psi_u(E,z)}{\partial E} + \frac{1}{2}\frac{\partial}{\partial E}T(E,z)\frac{\partial\psi_u(E,z)}{\partial E}, \tag*{(24)}

where zz denotes the one-dimensional spatial position, i.e. depth along the ray, and we define S∗=S+12dTdES^* = S + \frac{1}{2}\frac{\mathrm{d}T}{\mathrm{d}E} to obtain the standard diffusion form of straggling. We use a discontinuous Galerkin scheme to discretize the PDE in energy. The energy domain is divided into 128 equally-sized groups. The slowing down term is based on the Lax-Friedrichs scheme, whereas the straggling term utilizes the SIPG method. Polynomials up to second order are used as basis vectors (3 dof per energy element). After discretization in EE, the system can be written as

ME∂ψu(z)∂z+Gψu(z)=0,(25)\mathbf{M}_E\frac{\partial\psi_u(z)}{\partial z} + \mathbf{G}\psi_u(z) = 0, \tag*{(25)}

where ME\mathbf{M}_E is the mass matrix for the energy basis and G\mathbf{G} contains elements corresponding to absorption, slowing down and straggling. Here, ψu(z=0)=f(E)\psi_u(z = 0) = f(E) is the projection of a given energy distribution (Gaussian). We use a Crank-Nicolson scheme to solve this ODE system along the depth. The step size is determined such that it is smaller than the maximum step length (0.01 cm). The unscattered flux distribution is computed at the cell midpoints of the mesh. Based on the spatial and angular distribution of the beam source, a collection of start and end points for rays is generated. The solution is then summed over all ray positions within a beam source and the resulting space and energy dependent flux acts as a source term for the collided equation through inscattering.

Collided part: Dynamical low-rank approximation

We now want to use DLRA to efficiently solve the costly collided part of the equation. Since DLRA is a method designed for time-dependent problems, we now transform from energy EE to pseudo-time t(E)=Emax⁡−Et(E) = E_{\max} - E, resulting in the following equation:

∂tψ~c(t,r)=−Ω⋅∇ψ~c(t,r)S(t,r)+∑i∈AwiΓiψ~c(t,r)S(t,r)+∑i∈AwiΓin;iψu(t,r,Ω).(26)\partial_t\widetilde{\psi}_c(t,\mathbf{r}) = -\mathbf{\Omega}\cdot\nabla\frac{\widetilde{\psi}_c(t,\mathbf{r})}{S(t,\mathbf{r})} + \sum_{i\in A} w_i\Gamma_i\frac{\widetilde{\psi}_c(t,\mathbf{r})}{S(t,\mathbf{r})} + \sum_{i\in A}w_i\Gamma_{\mathrm{in};i}\psi_u(t,\mathbf{r},\mathbf{\Omega}). \tag*{(26)}

Next, we will discretize this equation in angle and space to obtain a (pseudo)time dependent ordinary matrix differential equation.

Angular discretization

We use a modal discretization of the angular variable. To this end, we define the real-valued spherical harmonics basis

mℓk={(−1)k2(Yℓk+(−1)kYℓ−k),k>0,Yℓ0,k=0,−(−1)ki2(Yℓ−k−(−1)kYℓk),k<0.m_{\ell}^{k} = \begin{cases} \frac{(-1)^{k}}{\sqrt{2}}\left(Y_{\ell}^{k} + (-1)^{k}Y_{\ell}^{-k}\right), & k > 0,\\ Y_{\ell}^{0}, & k = 0,\\ -\frac{(-1)^{k}i}{\sqrt{2}}\left(Y_{\ell}^{-k} - (-1)^{k}Y_{\ell}^{k}\right), & k < 0. \end{cases}

which uses the spherical harmonics basis functions

Yℓk(Ω)=2ℓ+14π(ℓ−k)!(ℓ+k)!eikφPℓk(μ),Y_{\ell}^{k}(\Omega) = \sqrt{\frac{2\ell+ 1}{4\pi}\frac{(\ell-k)!}{(\ell+k)!}}e^{ik\varphi}P_{\ell}^{k}(\mu),

where PℓkP_{\ell}^{k} are the associated Legendre polynomials. Collecting all basis functions up to degree NN in a vector

m=(m00,m1−1,m10,m11,⋯ ,mNN)T∈Rm,m=(N+1)2\mathbf{m} = (m_{0}^{0},m_{1}^{-1},m_{1}^{0},m_{1}^{1},\cdots,m_{N}^{N})^{\mathsf{T}} \in\mathbb{R}^{m},\quad m=(N+1)^{2}

gives the modal approximation ψc(t,r,Ω)≈u(t,r)Tm(Ω)\psi_{c}(t,\mathbf{r},\Omega)\approx\mathbf{u}(t,\mathbf{r})^{\mathsf{T}}\mathbf{m}(\Omega). Since we deal with monodirectional beam sources in this work, for a specific source with beam direction Ωin\Omega_{\mathrm{in}}, the uncollided flux ψu(t,r,Ω)=ψu(t,r)δ(Ω−Ωin)\psi_{u}(t,\mathbf{r},\Omega)=\psi_{u}(t,\mathbf{r})\delta(\Omega-\Omega_{\mathrm{in}}) at a specific point in space and time/energy is a scalar and the transformation from nodal to modal representation in angle can be done using the vector TM=m(Ωin)\mathbf{T}_{M}=\mathbf{m}(\Omega_{\mathrm{in}}). For a generalization to beams with a distribution in angle see [27]. Then, the PNP_{N} equations read

∂tu(t,r)=−A⋅∇u(t,r)S(t,r)−∑i∈Awi(r)Σt;i(t)u(t,r)S(t,r)+∑i∈Awi(r)Gi(t)u(t,r)S(t,r)+∑i∈Awi(r)Gi(t)TMψu(t,r),(27)\begin{aligned} \partial_{t}\mathbf{u}(t,\mathbf{r})={}&-\mathbf{A}\cdot\nabla\frac{\mathbf{u}(t,\mathbf{r})}{S(t,\mathbf{r})}-\sum_{i\in\mathcal{A}}w_{i}(\mathbf{r})\Sigma_{t;i}(t)\frac{\mathbf{u}(t,\mathbf{r})}{S(t,\mathbf{r})}\\ &+\sum_{i\in\mathcal{A}}w_{i}(\mathbf{r})\mathbf{G}_{i}(t)\frac{\mathbf{u}(t,\mathbf{r})}{S(t,\mathbf{r})}+\sum_{i\in\mathcal{A}}w_{i}(\mathbf{r})\mathbf{G}_{i}(t)\mathbf{T}_{M}\psi_{u}(t,\mathbf{r}), \tag*{(27)} \end{aligned}

where A⋅∇:=A1∂x+A2∂y+A3∂z\mathbf{A}\cdot\nabla:=A_{1}\partial_{x}+A_{2}\partial_{y}+A_{3}\partial_{z} with Ai:=∫S2mmTΩi dΩA_{i}:=\int_{\mathbb{S}^{2}}\mathbf{m}\mathbf{m}^{\mathsf{T}}\Omega_{i}\,\mathrm{d}\Omega.

The diagonal scattering matrix of material ii is defined as Gi(t)=(Gi;pq)1≤p,q≤m\mathbf{G}_{i}(t)=(G_{i;pq})_{1\leq p,q\leq m} where

Gi;pp(t,r)=2π∫[−1,1]Pk(μ0)Σs;i(t,μ0) dμ0,G_{i;pp}(t,\mathbf{r})=2\pi\int_{[-1,1]}P_{k}(\mu_{0})\Sigma_{s;i}(t,\mu_{0})\,\mathrm{d}\mu_{0},

for Boltzmann and

Gi;pp(t)=−ξ1;i2k⋅(k+1),G_{i;pp}(t)=-\frac{\xi_{1;i}}{2}k\cdot(k+1),

for Fokker-Planck, with p=ℓ2+ℓ+kp=\ell^{2}+\ell+k, 0≤k≤N0\leq k\leq N, ∣ℓ∣≤k|\ell|\leq k. The total cross section is then Σt;i(t)=Gi;11(t)\Sigma_{t;i}(t)=G_{i;11}(t). Note, that we further apply the extended transport correction to the scattering moments in the Boltzmann equation [11] and the correction described in [30] [37] for the Fokker-Planck moments. Both methods aim at a ”delta-function corrected expansion” [30], where the extended transport correction is chosen such that the truncated expansion matches moments 0 to N+1N+1 exactly [11].

Spatial discretization

In this work, we consider a full three-dimensional spatial domain, i.e., the spatial variable is r=(x,y,z)T∈R3\mathbf{r} = (x,y,z)^{T} \in\mathbb{R}^{3}. In order to sufficiently resolve the sharp gradients of the energy deposition near the Bragg peak, a higher-order method is necessary [33]. Here, we use a second order upwind scheme. For this, the spatial domain is split into n=nx⋅ny⋅nzn = n_x \cdot n_y \cdot n_z cells in the form of a structured rectangular grid with uniform one-dimensional grids x1≤x2≤…≤xnxx_1 \leq x_2 \leq\ldots\leq x_{n_x}, y1≤y2≤…≤ynyy_1 \leq y_2 \leq\ldots\leq y_{n_y} and z1≤z2≤…≤znzz_1 \leq z_2 \leq\ldots\leq z_{n_z} with grid sizes Δx,Δy,Δz\Delta_x,\Delta_y,\Delta_z respectively.

The cell of index idx⁡(i,j,k):=(k−1)⋅nx⋅ny+(j−1)⋅nx+i\operatorname{idx}(i,j,k) := (k-1) \cdot n_x \cdot n_y + (j-1) \cdot n_x + i is defined on Ii,j,k:=[xi,xi+1]×[yi,yi+1]×[zi,zi+1]I_{i,j,k} := [x_i,x_{i+1}] \times[y_i,y_{i+1}] \times[z_i,z_{i+1}] with i=1,…,nxi = 1,\ldots,n_x, j=1,…,nyj = 1,\ldots,n_y, k=1,…,nzk = 1,\ldots,n_z. Assuming we want to solve the semi-discrete equation (27) from the previous section, we now collect the numerical solution in a matrix u(t)∈Rn×m\mathbf{u}(t) \in\mathbb{R}^{n \times m} with entries uidx⁡(i,j,k),p(t)u_{\operatorname{idx}(i,j,k),p}(t), p=1,…,mp = 1,\ldots,m. We define the second order upwind difference stencils Dx,y,z+,−∈Rn×n\mathbf{D}_{x,y,z}^{+,-} \in\mathbb{R}^{n \times n} as

Dx;idx⁡(i,j,k),idx⁡(i,j,k)+,−=±32Δx,Dx;idx⁡(i,j,k),idx⁡(i±1,j,k)+,−=∓42Δx,Dx;idx⁡(i,j,k),idx⁡(i±2,j,k)+,−=±12Δx,Dy;idx⁡(i,j,k),idx⁡(i,j,k)+,−=±32Δy,Dy;idx⁡(i,j,k),idx⁡(i,j±1,k)+,−=∓42Δy,Dy;idx⁡(i,j,k),idx⁡(i,j±2,k)+,−=±12Δy,Dz;idx⁡(i,j,k),idx⁡(i,j,k)+,−=±32Δz,Dz;idx⁡(i,j,k),idx⁡(i,j,k±1)+,−=∓42Δz,Dz;idx⁡(i,j,k),idx⁡(i,j,k±2)+,−=±12Δz.\begin{aligned} D_{x;\operatorname{idx}(i,j,k),\operatorname{idx}(i,j,k)}^{+,-} &= \frac{\pm3}{2\Delta_x}, \quad D_{x;\operatorname{idx}(i,j,k),\operatorname{idx}(i\pm1,j,k)}^{+,-} = \frac{\mp4}{2\Delta_x}, \quad D_{x;\operatorname{idx}(i,j,k),\operatorname{idx}(i\pm2,j,k)}^{+,-} = \frac{\pm1}{2\Delta_x}, \\ D_{y;\operatorname{idx}(i,j,k),\operatorname{idx}(i,j,k)}^{+,-} &= \frac{\pm3}{2\Delta_y}, \quad D_{y;\operatorname{idx}(i,j,k),\operatorname{idx}(i,j\pm1,k)}^{+,-} = \frac{\mp4}{2\Delta_y}, \quad D_{y;\operatorname{idx}(i,j,k),\operatorname{idx}(i,j\pm2,k)}^{+,-} = \frac{\pm1}{2\Delta_y}, \\ D_{z;\operatorname{idx}(i,j,k),\operatorname{idx}(i,j,k)}^{+,-} &= \frac{\pm3}{2\Delta_z}, \quad D_{z;\operatorname{idx}(i,j,k),\operatorname{idx}(i,j,k\pm1)}^{+,-} = \frac{\mp4}{2\Delta_z}, \quad D_{z;\operatorname{idx}(i,j,k),\operatorname{idx}(i,j,k\pm2)}^{+,-} = \frac{\pm1}{2\Delta_z}. \end{aligned}

Further, we collect the eigenvectors of Ax,y,z=Vx,y,zΛx,y,zVx,y,zT\mathbf{A}_{x,y,z} = \mathbf{V}_{x,y,z}\boldsymbol{\Lambda}_{x,y,z}\mathbf{V}_{x,y,z}^{T} in Vx,y,z\mathbf{V}_{x,y,z} and split the eigenvalues collected in Λx,y,z=diag⁡(λ1x,y,z,…,λmx,y,z)\boldsymbol{\Lambda}_{x,y,z} = \operatorname{diag}(\lambda_{1}^{x,y,z},\ldots,\lambda_{m}^{x,y,z}) into (Λx,y,z;p,p+)p=1,…,m=max⁡(0,λpx,y,z)(\Lambda_{x,y,z;p,p}^{+})_{p=1,\ldots,m} = \max(0,\lambda_{p}^{x,y,z}) and (Λx,y,z;p,p−)p=1,…,m=min⁡(0,λpx,y,z)(\Lambda_{x,y,z;p,p}^{-})_{p=1,\ldots,m} = \min(0,\lambda_{p}^{x,y,z}). Then we obtain a large matrix differential equation of the form

u˙(t)=FS(u(t))+G(t,u(t)),(28)\dot{\mathbf{u}}(t) = \mathbf{F}_{S}(\mathbf{u}(t)) + \mathbf{G}(t,\mathbf{u}(t)), \tag*{(28)}

where

FS(u(t)):=(Dx+(VxS−1(t)u(t))Λx++Dx−(VxS−1(t)u(t))Λx−)VxT+(Dy+(VyS−1(t)u(t))Λy++Dy−(VyS−1(t)u(t))Λy−)VyT+(Dz+(VzS−1(t)u(t))Λz++Dz−(VzS−1(t)u(t))Λz−)VzT,G(t,u(t)):=∑i∈AwiS−1(t)u(t)Σt;i+∑i∈AwiS−1(t)u(t)Gi+∑i∈Awiψu(t)TMGi,\begin{aligned} \mathbf{F}_{S}(\mathbf{u}(t)) :={}& \left(\mathbf{D}_{x}^{+}\left(\mathbf{V}_{x}\mathbf{S}^{-1}(t)\mathbf{u}(t)\right)\Lambda_{x}^{+}+\mathbf{D}_{x}^{-}\left(\mathbf{V}_{x}\mathbf{S}^{-1}(t)\mathbf{u}(t)\right)\Lambda_{x}^{-}\right)\mathbf{V}_{x}^{T} \\ &+\left(\mathbf{D}_{y}^{+}\left(\mathbf{V}_{y}\mathbf{S}^{-1}(t)\mathbf{u}(t)\right)\Lambda_{y}^{+}+\mathbf{D}_{y}^{-}\left(\mathbf{V}_{y}\mathbf{S}^{-1}(t)\mathbf{u}(t)\right)\Lambda_{y}^{-}\right)\mathbf{V}_{y}^{T} \\ &+\left(\mathbf{D}_{z}^{+}\left(\mathbf{V}_{z}\mathbf{S}^{-1}(t)\mathbf{u}(t)\right)\Lambda_{z}^{+}+\mathbf{D}_{z}^{-}\left(\mathbf{V}_{z}\mathbf{S}^{-1}(t)\mathbf{u}(t)\right)\Lambda_{z}^{-}\right)\mathbf{V}_{z}^{T}, \\ \mathbf{G}(t,\mathbf{u}(t)) :={}& \sum_{i\in A}\mathbf{w}_i\mathbf{S}^{-1}(t)\mathbf{u}(t)\boldsymbol{\Sigma}_{t;i}+\sum_{i\in A}\mathbf{w}_i\mathbf{S}^{-1}(t)\mathbf{u}(t)\mathbf{G}_i \\ &+\sum_{i\in A}\mathbf{w}_i\psi_u(t)\mathbf{T}_{M}\mathbf{G}_i, \end{aligned}

here wi\mathbf{w}_i and S(t)∈R+n×n\mathbf{S}(t) \in\mathbb{R}_{+}^{n \times n} are diagonal matrices containing the material composition weights and stopping power at each spatial grid point, respectively. Here, ψu\psi_u is the uncollided solution from the ray tracer evaluated at the (same) spatial grid points. Since solving this at sufficiently high resolution is extremely costly we use the dynamical low-rank approximation to evolve the solution in (pseudo-)time.

Energy/time update using the dynamical low-rank approximation

For this, we further split equation (25) into its streaming and scattering parts using a first-order Lie splitting, i.e.,

u˙I(t)=FS(uI(t))uI(t0)=u(t0),(29)\dot{\mathbf{u}}_{\mathrm{I}}(t) = \mathbf{F}_{\mathrm{S}}(\mathbf{u}_{\mathrm{I}}(t)) \qquad\mathbf{u}_{\mathrm{I}}(t_0) = \mathbf{u}(t_0), \tag*{(29)}
u˙II(t)=G(t,uII(t))uII(t0)=uI(t1).(30)\dot{\mathbf{u}}_{\mathrm{II}}(t) = \mathbf{G}(t,\mathbf{u}_{\mathrm{II}}(t)) \qquad\mathbf{u}_{\mathrm{II}}(t_0) = \mathbf{u}_{\mathrm{I}}(t_1). \tag*{(30)}

For the streaming update (26a) we use the augmented BUG integrator from section 2.2 to update uI\mathbf{u}_{\mathrm{I}} in factorized form from U0,S0,V0\mathbf{U}^{0}, \mathbf{S}^{0}, \mathbf{V}^{0} to U12,S12,V12\mathbf{U}^{\frac{1}{2}}, \mathbf{S}^{\frac{1}{2}}, \mathbf{V}^{\frac{1}{2}}, by simply plugging in FS\mathbf{F}_{\mathrm{S}} for the right hand side and using a fourth-order Runge–Kutta method for the time updates. Note, that we still do a full time step from t0t_0 to t1=t0+Δtt_1 = t_0 + \Delta t and the 12\frac{1}{2} in the exponent merely denotes that we have only completed the streaming part of the update, i.e., U12S12V12,⊤≈uI(t1)\mathbf{U}^{\frac{1}{2}}\mathbf{S}^{\frac{1}{2}}\mathbf{V}^{\frac{1}{2},\top} \approx\mathbf{u}_{\mathrm{I}}(t_1). The fourth-order Runge–Kutta method is primarily used to stabilize the stiff streaming operators in FS\mathbf{F}_{\mathrm{S}}.

For the scattering update, we use the matrix projector splitting integrator for the collided flux and the augmented BUG integrator on the remainder according to [26]. This reduces the update to a K-, S- and L-step for inscattering from the uncollided flux and just the L-step for self-scattering of the collided flux. Thus, to perform the scattering update (26b), we need to solve the following differential equations

  1. LL-step, collided:

L˙(t)=−∑i∈AU12,⊤wiS−1(t)U12Σt,iL(t)+∑i∈AU12,⊤wiS−1(t)U12L(t)Gi.\dot{\mathbf{L}}(t) = -\sum_{i\in A}\mathbf{U}^{\frac{1}{2},\top}\mathbf{w}_i\mathbf{S}^{-1}(t)\mathbf{U}^{\frac{1}{2}}\Sigma_{t,i}\mathbf{L}(t) + \sum_{i\in A}\mathbf{U}^{\frac{1}{2},\top}\mathbf{w}_i\mathbf{S}^{-1}(t)\mathbf{U}^{\frac{1}{2}}\mathbf{L}(t)\mathbf{G}_i.
L(t0)=S12V12,⊤.\mathbf{L}(t_0) = \mathbf{S}^{\frac{1}{2}}\mathbf{V}^{\frac{1}{2},\top}.

Determine the updated basis using a QR decomposition of L(t1)=S~12V~12\mathbf{L}(t_1) = \widetilde{\mathbf{S}}^{\frac{1}{2}}\widetilde{\mathbf{V}}^{\frac{1}{2}}.

  1. KK-step, uncollided:

K˙(t)=∑i∈AwiS−1ψuTMGiV12,\dot{\mathbf{K}}(t) = \sum_{i\in A}\mathbf{w}_i\mathbf{S}^{-1}\psi_u\mathbf{T}_{M}\mathbf{G}_i\mathbf{V}^{\frac{1}{2}},
K(t0)=U12S12.\mathbf{K}(t_0) = \mathbf{U}^{\frac{1}{2}}\mathbf{S}^{\frac{1}{2}}.

Determine the updated and augmented basis at t1=t0+Δtt_1 = t_0+\Delta t using a QR decomposition of [K(t1),U12]=U^SK[\mathbf{K}(t_1),\mathbf{U}^{\frac{1}{2}}] = \widehat{\mathbf{U}}\mathbf{S}_{K}.

  1. LL-step, uncollided:

L˙(t)=∑i∈AU12,⊤wiS−1ψuTMGi,\dot{\mathbf{L}}(t) = \sum_{i\in A}\mathbf{U}^{\frac{1}{2},\top}\mathbf{w}_i\mathbf{S}^{-1}\psi_u\mathbf{T}_{M}\mathbf{G}_i,
L(t0)=S~12V~12.\mathbf{L}(t_0) = \widetilde{\mathbf{S}}^{\frac{1}{2}}\widetilde{\mathbf{V}}^{\frac{1}{2}}.

Determine the updated and augmented basis at t1=t0+Δtt_1 = t_0+\Delta t using a QR decomposition of [L(t1),V12]=V^SL[\mathbf{L}(t_1),\mathbf{V}^{\frac{1}{2}}] = \widehat{\mathbf{V}}\mathbf{S}_{L}.

4. SS-step, uncollided:

S˙(t)=∑i∈AU^TwiS−1ψuTMGiV^,\dot{S}(t)=\sum_{i\in A}\hat{U}^{\mathsf{T}}w_iS^{-1}\psi_u\mathsf{T}_{M}G_i\hat{V},
S(t0)=(U^TU^12)S12(V~12,TV^)S(t_0)=\left(\hat{U}^{\mathsf{T}}\hat{U}^{\frac{1}{2}}\right)S^{\frac{1}{2}}\left(\widetilde{V}^{\frac{1}{2},\mathsf{T}}\hat{V}\right)

and set S1=S(t1)S^{1}=S(t_1).

Similar to [27], we use an implicit Euler method for the first step to treat stiff scattering terms and an explicit Euler method for the time updates in the remaining steps.

Numerical results

In the following, we present a number of numerical experiments in homogeneous and heterogeneous media to validate and compare our method against an otherwise equivalent full-rank method as well as a state-of-the-art Monte Carlo code (TOPAS MC) [40] with only electromagnetic interactions. We will first consider a single uni-directional beam in different media and later add a second beam with a different incoming angle to investigate the effect this has on the chosen ranks. All beams have a Gaussian spatial distribution in the two-dimensional plane perpendicular to the beam direction with 3mm standard deviation and the intial beam energy follows a Gaussian distribution with standard deviation 1% of the mean beam energy. Since the problem at full angular and spatial resolution is infeasible to solve at full

rank, we will validate the low-rank method against the deterministic full-rank method on a coarser grid and then compare a higher resolution low-rank solution to the Monte Carlo reference. For the spatial resolution we always use a coarser gridsize of 1 mm and a finer one with 0.25 mm. Since initial numerical experiments indicate that more angular modes are required when using the Boltzmann collision operator, we choose N=75N=75 for the coarser and N=115N=115 for the finer resolution here. In case of the Fokker-Planck approximation, which uses a simplified collision operator, N=19N=19 for the coarser and N=75N=75 for the finer resolution have proven sufficient. The truncation tolerance used to adapt the rank is chosen as ϑ=0.01\vartheta=0.01 in all cases.

Homogeneous Case

First, we consider a 90 MeV proton beam in a homogeneous 2×2×72 \times2 \times7 cm box (see Figure 2a). The waterlike material has 0 Hounsfield Units (HU). These values are converted to material composition weights wiw_i and a density of 1.018 gcm3\frac{\mathrm{g}}{\mathrm{cm}^{3}} using the conversion described in [43].

Figure 2. Schematic representations of the domain size, composition and beam directions in the three numerical test cases: (a) homogeneous, water-like 2×2×72 \times2 \times7 cm domain with single beam in z-direction hitting middle of x-y-plane (b) heterogeneous 2×2×72 \times2 \times7 cm domain representing water with a lung insert with a single beam identical to (a) and (c) homogeneous, water-like 2×4×42 \times4 \times4 cm domain with two perpendicular beams pointing in z- and negative y-direction.

Figure 3 shows an overview of two-dimensional slices and profiles of the deposited energy using the Boltzmann collision operator at different resolutions compared to the Monte Carlo reference and a full-rank deterministic solver. It is apparent, that DLRA can approximate the full-rank solution very well and at a much lower rank according to Figure 4 (a). It is also clear that the coarse grid resolution is not sufficient to fully capture the characteristics of the solution. When increasing both spatial and angular resolution we can however see that our model is capable of more closely approximating the Monte Carlo reference as well (see Figure 3). The only significant deviation occurs directly behind the Bragg peak. Here, the solution takes on slightly negative values, which is most likely due to the nature of our angular discretization combined with the steep gradients in this area. Since the full and low-rank solutions match almost exactly, differences to the Monte Carlo reference should however be attributed to the choices of physics models and numerical discretizations.

Surface plots and profiles of deposited energy for Monte Carlo, DLRA, and full-rank methods

Figure 3. Surface plots (a)-(d) and profiles (e)-(g) of deposited energy in the homogeneous test case with a single beam for the Monte Carlo reference compared to DLRA at high and low spatial and angular resolution and the full-rank deterministic method based on the Boltzmann equation. Lines in (a) indicate positions of profiles plotted in (e)-(g).

Ranks chosen by rank-adaptive integrator for different spatial and angular resolutions in the homogeneous and heterogeneous test cases

Figure 4. Ranks chosen by rank-adaptive integrator for different spatial and angular resolutions and the Boltzmann equation in (a) the homogeneous and (b) heterogeneous test case with a single beam.

The Fokker-Planck approximation on the other hand shows smaller differences between the coarse and highly resolved solutions, but tends to underestimate the energy deposition in the region of the Bragg peak and overestimate it slightly in the build-up before the peak as can be seen in Figure 5. This might be attributed to the correction that we use to stabilize the solution [30], which has however proven to be necessary to avoid stronger artifacts near material boundaries during our studies. In Figure 5 (e), the solution also exhibits similar negative areas as previously observed for the Boltzmann solver. Figure 6 further shows that the ranks for the Fokker-Planck approximation are consistently higher than for the Boltzmann equation. Interestingly, in contrast to the Boltzmann solver, the discretizations now affect the rank significantly and finer discretizations are not directly correlated with higher ranks. Here, the ranks are consistently higher for high energies when using a finer angular discretization, but are larger in the case of coarse angular discretizations for small energies. Due to the higher ranks and worse results compared to the Boltzmann solver for high resolutions, using the Fokker-Planck approximation (within the current framework) would only be beneficial for a cheap estimate at a low resolution for which the Boltzmann solver has not converged sufficiently.

Surface plots and profiles of deposited energy

Figure 5. Surface plots (a)–(d) and profiles (e)–(g) of deposited energy in the homogeneous test case with a single beam for the Monte Carlo reference compared to DLRA at high and low spatial and angular resolution and the full-rank deterministic method based on the Fokker–Planck approximation. Lines in (a) indicate positions of profiles plotted in (e)–(g).

Ranks chosen by rank-adaptive integrator for different spatial and angular resolutions and the Fokker-Planck approximation in the homogeneous and heterogeneous test cases with a single beam

Figure 6. Ranks chosen by rank-adaptive integrator for different spatial and angular resolutions and the Fokker-Planck approximation in (a) the homogeneous and (b) heterogeneous test case with a single beam.

Heterogeneous Case

Next, we consider a heterogeneous case mimicking a beam encountering strong heterogeneities along its path as might be the case for example in the head and neck region. For this, a lower density box with -400 HU, roughly corresponding to lung tissue, is placed inside the domain (see Figure 2b). A proton beam with 80 MeV is directed at the box such that the center of the beam hits the interface between the materials. Again, the Hounsfield units are converted to material compositions and densities using the method in [43]. As in the homogeneous case, the full-rank solution is well-approximated by the dynamical low-rank approximation for both the Boltzmann and the Fokker–Planck solvers. This can be seen both in the 2D surface plots in Figure 7c and (d) (Boltzmann) and Figure 8 (c) and (d) (Fokker–Planck) and the one-dimensional solution profiles in Figure 7e and (g). The Boltzmann solver also approximates the Monte Carlo reference reasonably well, although more deviations can be seen than in the homogeneous test case. For example, the surface plots in Figure 7a and (b) show that the highest dose values in the Bragg peaks are slightly underestimated and depth profiles in Figure 7e reveal differences in the local minimum behind the first peak. The ranks required for this more complex test case are expectedly higher than in the homogeneous case (see Figure 4 (b)) and increase for increasingly fine spatial and angular resolutions. All chosen ranks are however still significantly smaller than the number of spatial cells or angular modes and lead to a large reduction of the system size (see also § 5).

Surface plots and profiles of deposited energy in the heterogeneous test case

Figure 8. Surface plots (a)-(d) and profiles (e)-(g) of deposited energy in the heterogeneous test case with a single beam for the Monte Carlo reference compared to DLRA at high and low spatial and angular resolution and the full-rank deterministic method based on the Fokker-Planck approximation. Lines in (a) indicate positions of profiles plotted in (e)-(g).

When comparing the Fokker-Planck approximation to the Monte Carlo reference in Figure 8 (a) and (Figure 8b) we observe similar results as in the homogeneous case as well. Again, the solution tends to underestimate the highest energy depositions and there are small negative areas behind the Bragg peak. In this case, the sharp material boundary however tended to cause artifacts, which can still be slightly seen in Figure 8 (b). While a stronger correction of the angular moments according to [30] [37] reduces the artifacts, it also has a smoothening effect and dampens the height of the Bragg peak, creating a tradeoff between stability and preserving solution characteristics. Figure 6 (b) further shows that ranks for different discretizations have a similar structure as in the homogeneous case, but are consistently higher, as already observed for the Boltzmann solver.

Homogeneous with two beams

Finally, we consider a case with two 50 MeV beams entering the domain at different angles (see Figure 2c). A treatment plan in proton therapy typically consists of a larger number of layered beamlets with different initial energies, positions and angles. Thus, this is a good test case to gauge the impact of a more complex beam source on rank and therefore computational costs and memory. Figure 9 and Figure 10 show a comparison of Monte Carlo to DLRA as well as the ranks chosen using the same tolerance as in the previous test cases. The low-rank method manages to capture the basic dose characteristics well, but appears to slightly smoothing in the sense of losing local “hot-spots” in the solution and corresponding sharp dose gradients. For the Boltzmann solver, the chosen ranks are on average roughly 2 points higher than in the homogeneous single beam test case for the finer angular resolution and even less when comparing the coarser resolutions (compare Figure 4 (a) and Figure 10 (a)). Interestingly, comparing Figure 6 (b) and Figure 10 (b), we see that the ranks are even lower than those of the single beam test case for the Fokker-Planck approximation. This difference could also be due to the lower beam energy in this test case. While numerical experiments have shown that for a tolerance of ϑ=0.01\vartheta= 0.01 the initial beam energy does not affect the

Surface plots and profiles of deposited energy for two perpendicular beams

Figure 9. Surface plots (a)-(c) and profiles (d)-(e) of deposited energy for two perpendicular beams in a homogeneous material for the Monte Carlo reference compared to DLRA based on the Boltzmann equation and the Fokker-Planck approximation. Lines in (a) indicate positions of profiles plotted in (d)-(e).

Ranks chosen by the rank-adaptive integrator

Figure 10. Ranks chosen by rank-adaptive integrator using (a) Boltzmann and (b) the Fokker-Planck approximation in the homogeneous test case with two beams.

Monte Carlo, $10^{8}$ histories

(a) Monte Carlo, 10810^{8} histories

DLRA, $P_{109}$, $\Delta_x = 0.25$ mm

(b) DLRA, P109P_{109}, Δx=0.25\Delta_x = 0.25 mm

DLRA, $P_{75}$, $\Delta_x = 1$ mm

(c) DLRA, P75P_{75}, Δx=1\Delta_x = 1 mm

full rank $P_{75}$, $\Delta_x = 1$ mm

(d) full rank P75P_{75}, Δx=1\Delta_x = 1 mm

Depth

(e) Depth.

Lateral, MC vs DLRA

(f) Lateral, MC vs DLRA.

Lateral, DLRA vs full rank

(g) Lateral, DLRA vs full rank.

Figure 7. Surface plots (a)-(d) and profiles (e)-(g) of deposited energy in the heterogeneous test case with a single beam for the Monte Carlo reference compared to DLRA at high and low spatial and angular resolution and the full-rank deterministic method based on the Boltzmann equation. Lines in (a) indicate positions of profiles plotted in (e)-(g).

rank for the Boltzmann solver, we have seen that the ranks for the Fokker-Planck solver decreased for decreasing beam energies. Further, for both approaches the ranks are mostly affected by the angular discretization, which is somewhat intuitive considering we effectively doubled the amount of relevant directions. Figure 11 also shows that for both Boltzmann and Fokker-Planck, the first two modes of the angular basis indeed represent the directions of the two beams, thus capturing the forward movement of particles with some angular spread due to scattering.

First four dominant angular modes for Boltzmann and Fokker-Planck

Figure 11. First four dominant angular modes after one fifth of the time/energy steps.

Since in theory the rank quadratically impacts the computational costs (see § 5), further studies are necessary to see when it is more efficient to treat beam sources separately or jointly. Our results however suggest that at the very least a relatively narrow spread in angle, which is typically used for beamlets originating from the same beam direction, can be incorporated without a significant increase in computational costs or memory. This would already reduce the amount of individual computations to only the 2-5 beam angles, compared to potentially thousands of pencil beams [5].

Memory, computational complexity and implementation

The dynamical low-rank approximation reduces the computational complexity from O(n⋅m)\mathcal{O}(n \cdot m) to O(r2⋅(n+m))\mathcal{O}(r^{2} \cdot(n + m)) and O(r⋅(n+m))\mathcal{O}(r \cdot(n + m)) with respect to runtime and memory, respectively [27]. Thus, for r≪m,nr \ll m, n we can expect significant cost reductions. In order to conduct our numerical experiments, the described method is implemented in julia with GPU acceleration using CUDA.jl. The code to reproduce our numerical experiments is openly available on GitHub. Since the full-rank method with a meaningful discretization does not fit on the GPU, we further implement a full-rank CPU version using the same numerical methods and physical models. All experiments are run on a workstation PC1 which is comparable to what can be expected in a clinical setting.

Table 1 shows a comparison of the full-rank and DLRA Boltzmann and Fokker-Planck solvers with respect to their theoretical as well as measured costs2 for the homogeneous test case with coarser spatial and angular resolution. We report the minimum execution time out of 20 runs, since the execution path is completely deterministic and thus variations are due to system overhead and background processes on the shared workstation. The theoretical values are computed using the formulas for computational complexity and memory from above and the average rank over time/energy steps is used for DLRA, which is 4.77 and 19.99 for Boltzmann and Fokker-Planck, respectively. We see that in case of the Boltzmann solver, DLRA reduces the costs to less than 0.5% in all cases and consistently achieves an even higher efficiency gain in practice, which demonstrates the combined effects of GPU implementation and complexity reduction. As predicted theoretically, the reductions are larger in terms of memory consumption.

full rank (CPU)DLRA (GPU)% of full rank
Boltzmann
complexity161 728 000769 597.89920.4759%
system size161 728 000161 226.35840.0997%
runtime12h20.4s0.0472%
memory41.4TiB1.69GiB0.0039%
Fokker-Planck
complexity11 200 00011 348 642.8400101.3272%
system size11 200 000567 716.00005.0689%
runtime2684s32.9s1.2258%
memory2.84TiB2.71GiB0.0932%

Table 1. Theoretical and measured time and memory costs of DLRA vs. full rank.

Since for the Fokker-Planck approximation we both use lower orders of angular discretization and the average rank is higher than for Boltzmann, there are no theoretical complexity improvements in the case considered here. In practice, we can however see almost two orders of magnitude speed-up due to the GPU implementation. Thus, even for slightly higher rank and less high-dimensional problems, DLRA can be worthwhile by reducing the memory. Further, we see in Figure 6 that the ranks for Fokker-Planck do not necessarily increase for finer spatial discretizations, so for more highly resolved problems we would also see larger theoretical gains with respect to runtime. Note, that a fair comparison to

the costs of the TOPAS MC reference computation is not as straightforward [4]. For 10810^{8} histories (primary particles) the computation time on the same workstation using up to 48 nodes is still in the ballpark of several hours. However, due to the different numerical discretization methods and included physics, the theoretical complexities are difficult to compare at a constant error. Further, the actual runtime can be almost arbitrarily reduced with more computational power due to the embarrassingly parallel nature of Monte Carlo methods. Thus, we give these values merely as a frame of reference.

Discussion

Our numerical results have demonstrated that the dynamical low-rank approximation can reduce the computational costs and memory associated with three-dimensional dose calculations in proton therapy by several orders of magnitude, both in theory and practice. In particular, we see that protons exhibit very low ranks even compared to other charged particles [27]. This can likely be attributed to the structure of their energy deposition curve, i.e., the dynamics are mostly dominated by advection, followed by close to diffusive behavior for low energies. While ranks expectedly increase in heterogeneous materials, the fact that they do not increase significantly for several beam sources is promising for a use in treatment planning.

We investigated two variations of the transport model: the Boltzmann equation and a Fokker-Planck approximation. Overall, we observe slightly better results using the Boltzmann solver, both in terms of agreement with Monte Carlo, stability and ranks. Our comparative study between the Boltzmann and Fokker-Planck models reveals that, despite the Fokker-Planck model’s simpler collision operator, the Boltzmann model provides more accurate results with lower computational ranks in our high-resolution framework. This suggests that the additional physical fidelity of the Boltzmann operator is crucial and is not a bottleneck when combined with DLRA. For both approaches, the low-rank solution shows near perfect agreement with the equivalent full-rank solution. Comparisons to the Monte Carlo reference show a need for high spatial and angular resolution, as solution quality deteriorates with coarser discretization. Yet, being able to relatively cheaply compute the dose distribution on such fine grids speaks for the efficacy of the dynamical low-rank approximation. The choice of spatial and angular discretizations however needs to be reevaluated, also with view to “unphysical” behavior, i.e., negative regions in the solution. Further, the Fokker-Planck method heavily relies on the implicit scattering update and the correction described in [30] [37] to avoid instabilities and artifacts at the material boundary, respectively. The correction however tends to overestimate the scattering moments [11] which could be a cause of the differences to Monte Carlo that we are seeing. Thus, future work could investigate whether the use of other, higher-order or locally refined spatial and angular discretizations can reduce the required expansion order and eliminate oscillations. Further, since protons undergo more physical interactions at low energies and thus energy loss is non-linear, a proportional, non-uniform energy/pseudo-time discretization could also be of interest to better capture rapid flux changes.

The currently used model for particle interactions still makes several simplifying assumptions. While the use of the continuous slowing down approximation and material composition model is common also in state-of-the-art Monte Carlo codes [19], incorporating nuclear interactions as well as absorption is necessary in order to achieve clinical precision [39]. Note, that we do not see the full impact of neglecting these effects in this work as we compare to a Monte Carlo method which is also limited to electromagnetic interactions. Fortunately, adding nuclear interactions or absorption does not majorly affect the numerical method described here and is more a matter of accurate physical modeling or data (which can be challenging in itself). Based on the good agreement we achieve with the Monte Carlo reference and similar observations made in literature [32], neglecting straggling in the collided part of the equation appears to be a reasonable approximation. However, future work could also investigate strategies for handling a straggling term within the dynamical low-rank approximation of the collided equation.

Conclusion

In this work, we demonstrated that the computational costs and memory associated with deterministic solvers for a proton transport problem can be significantly reduced using the dynamical low-rank method. The chosen ranks are low, making a much higher spatial and angular resolution computationally feasible at little additional costs. We show very good agreement with a full-rank solver based on the same physical models and discretizations. Further, our results agree reasonably well with a TOPAS MC reference. Future work could explore the benefits of higher-order or refined discretizations, include more accurate physics models and extend the framework to full radiation therapy treatment plans.

ACKNOWLEDGEMENTS

Pia Stammer received funding from the German National Academy of Sciences Leopoldina for the project underlying this article, under grant number LPDS 2024-03.

CRediT author statement

Pia Stammer: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Writing - Original Draft, Writing - Review & Editing, Visualization

Niklas Wahl: Software, Writing - Review & Editing

Jonas Kusch: Conceptualization, Methodology, Software, Writing - Original Draft, Writing - Review & Editing

Danny Lathouwers: Methodology, Resources, Writing - Original Draft, Writing - Review & Editing

References

  1. [1]H. A. Bethe. Molière’s Theory of Multiple Scattering. Physical Review, 89(6):1256–1266, Mar. 1953.DOI
  2. [2]T. Burlacu, D. Lathouwers, and Z. Perkó. A Deterministic Adjoint-Based Semi-Analytical Algorithm for Fast Response Change Computations in Proton Therapy. Journal of Computational and Theoretical Transport, 52(1):1–41, Jan. 2023.arxiv.org/abs/2208.11927
  3. [3]T. Burlacu, D. Lathouwers, and Z. Perkó. Yet anOther Dose Algorithm (YODA) for independent computations of dose and dose changes due to anatomical changes. Physics in Medicine and Biology, 69(16), July 2024.arxiv.org/abs/2404.00976
  4. [4]C. Börgers. Complexity of Monte Carlo and deterministic dose-calculation methods. Physics in Medicine & Biology, 43(3):517, Mar. 1998.pubmed.ncbi.nlm.nih.gov/9533131
  5. [5]W. Cao, R. Rocha, R. Mohan, G. Lim, H. M. Goudarzi, B. C. Ferreira, and J. M. Dias. Reflections on beam configuration optimization for intensity-modulated proton therapy. Physics in Medicine & Biology, 67(13):13TR01, June 2022.DOI
  6. [6]G. Ceruti, L. Einkemmer, J. Kusch, and C. Lubich. A robust second-order low-rank BUG integrator based on the midpoint rule. BIT Numerical Mathematics, 64(3):30, July 2024.arxiv.org/abs/2402.08607
  7. [7]G. Ceruti, J. Kusch, and C. Lubich. A rank-adaptive robust integrator for dynamical low-rank approximation. BIT Numerical Mathematics, 62(4):1149–1174, Dec. 2022.arxiv.org/abs/2104.05247
  8. [8]G. Ceruti, J. Kusch, and C. Lubich. A Parallel Rank-Adaptive Integrator for Dynamical Low-Rank Approximation. SIAM Journal on Scientific Computing, 46(3):B205–B228, June 2024.arxiv.org/abs/2304.05660
  9. [9]G. Ceruti and C. Lubich. An unconventional robust integrator for dynamical low-rank approximation. BIT Numerical Mathematics, 62(1):23–44, Mar. 2022.arxiv.org/abs/2010.02022
  10. [10]J. Coughlin, J. Hu, and U. Shumlak. Robust and conservative dynamical low-rank methods for the Vlasov equation via a novel macro-micro decomposition. Journal of Computational Physics, 509:113055, July 2024.arxiv.org/abs/2311.09425
  11. [11]C. R. Drumm, F. , Wesley C., L. , Leonard, and J. Liscum-Powell. An Analysis of the Extended-Transport Correction with Application to Electron Beam Transport. Nuclear Science and Engineering, 155(3):355–366, Mar. 2007.
  12. [12]M. Durante and H. Paganetti. Nuclear physics in particle therapy: a review. Reports on Progress in Physics, 79(9):096702, Aug. 2016.
  13. [13]L. Einkemmer, K. Kormann, J. Kusch, R. G. McClarren, and J.-M. Qiu. A review of low-rank methods for time-dependent kinetic simulations. Journal of Computational Physics, 538:114191, Oct. 2025.arxiv.org/abs/2412.05912
  14. [14]L. Einkemmer and A. Möriggl. Semi-Lagrangian 4d, 5d, and 6d kinetic plasma simulation on large-scale GPU-equipped supercomputers. The International Journal of High Performance Computing Applications, 37(2):180–196, 2023.DOI
  15. [15]Y. Elcim, B. Dirican, and O. Yavas. Dosimetric comparison of pencil beam and Monte Carlo algorithms in conformal lung radiotherapy. Journal of Applied Clinical Medical Physics, 19(5):616–624, Sept. 2018.DOI
  16. [16]L. Eyges. Multiple Scattering with Energy Loss. Physical Review, 74(10):1534–1535, Nov. 1948.DOI
  17. [17]B. Gottschalk. Radiotherapy Proton Interactions in Matter, Mar. 2018. arXiv:1804.00022 [physics].arxiv.org/abs/1804.00022
  18. [18]L. Hong, M. Goitein, M. Buccolini, R. Comiskey, B. Gottschalk, S. Rosenthal, C. Serago, and M. Urie. A pencil beam algorithm for proton dose calculations. Physics in Medicine & Biology, 41(8):1305, 1996.pubmed.ncbi.nlm.nih.gov/8858722
  19. [19]S. Huang, M. Kang, K. Souris, C. Ainsley, T. D. Solberg, J. E. McDonough, C. B. Simone, and L. Lin. Validation and clinical implementation of an accurate Monte Carlo code for pencil beam scanning proton therapy. Journal of Applied Clinical Medical Physics, 19(5):558–572, July 2018.DOI
  20. [20]J. F. Janni. Proton Range-Energy Tables, 1 keV-10 GeV, Energy Loss, Range, Path Length, Time-of-Flight, Straggling, Multiple Scattering, and Nuclear Interaction Probability. Part I. For 63 Compounds. Atomic Data and Nuclear Data Tables, 27:147, Jan. 1982.
  21. [21]M. W. K. Kan, P. K. N. Yu, and L. H. T. Leung. A Review on the Use of Grid-Based Boltzmann Equation Solvers for Dose Calculation in External Photon Beam Treatment Planning. BioMed Research International, 2013(1):692874, 2013.DOI
  22. [22]I. Kawrakow. The effect of Monte Carlo statistical uncertainties on the evaluation of dose distributions in radiation treatment planning. Physics in Medicine and Biology, 49(8):1549–1556, Apr. 2004.
  23. [23]E. Kieri, C. Lubich, and H. Walach. Discretized Dynamical Low-Rank Approximation in the Presence of Small Singular Values. SIAM Journal on Numerical Analysis, 54(2):1020–1038, Jan. 2016.DOI
  24. [24]O. Koch and C. Lubich. Dynamical Low-Rank Approximation. SIAM Journal on Matrix Analysis and Applications, 29(2):434–454, Jan. 2007.DOI
  25. [25]J. Kusch. Second-order robust parallel integrators for dynamical low-rank approximation. BIT Numerical Mathematics, 65(3):31, June 2025.arxiv.org/abs/2403.02834
  26. [26]J. Kusch, L. Einkemmer, and G. Ceruti. On the Stability of Robust Dynamical Low-Rank Approximations for Hyperbolic Problems. SIAM Journal on Scientific Computing, 45(1):A1–A24, Feb. 2023.arxiv.org/abs/2107.07282
  27. [27]J. Kusch and P. Stammer. A robust collision source method for rank adaptive dynamical low-rank approximation in radiation therapy. ESAIM: Mathematical Modelling and Numerical Analysis, 57(2):865–891, Mar. 2023.
  28. [28]J. Kusch, B. Whewell, R. McClarren, and M. Frank. A low-rank power iteration scheme for neutron transport criticality problems. Journal of Computational Physics, 470:111587, 2022.DOI
  29. [29]J. Kópházi and D. Lathouwers. A space–angle DGFEM approach for the Boltzmann radiation transport equation with local angular refinement. Journal of Computational Physics, 297:637–668, Sept. 2015.DOI
  30. [30]M. Landesman and J. E. Morel. Angular Fokker-Planck Decomposition and Representation Techniques. Nuclear Science and Engineering, 103(1):1–11, Sept. 1989.DOI

Paper details

vlnc:W_4ddmjs4x
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents