Source-linked AI summary
Generative Monte Carlo Sampling for Constant-Cost Particle Transport
Joseph A. Farmer, Aidan Murray, Johannes Krotz, Ryan G. McClarren
TL;DR
Monte Carlo particle transport is highly faithful but incurs costs that grow with scattering events. GMC reformulates cell transmission as conditional generation, using neural networks to sample exit states directly. It preserves reported transport statistics while maintaining constant per-cell cost in optically thick regimes, with extensions beyond the demonstrated setting left for future work.
Problem
Monte Carlo remains a high-fidelity transport method, but its computational cost is tethered to the number of scattering events and no prior method preserved full statistical fidelity across materials and regimes.
Method
GMC trains a generative model to reproduce conditional particle-exit distributions for cell transmission without simulating scattering histories, using optical coordinates and a model-agnostic generative framework.
Results
GMC preserves scalar flux and energy-deposition profiles with the same 1/sqrt(N) statistical convergence rate as standard Monte Carlo while retaining constant O(1) cost per cell transmission in optically thick limits.
Takeaways & Limitations
GMC aligns particle transport with neural-network inference, allowing a single optically parameterized model to generalize across diverse materials and geometries within its training distribution.
Takeaways & Limitations
The demonstrated work is limited to steady-state, monoenergetic simulations; energy- and time-dependent extensions are deferred to future work.
Abstract
from arXiv · showhide
We present Generative Monte Carlo (GMC), a novel paradigm for particle transport simulation that integrates generative artificial intelligence directly into the stochastic solution of the linear Boltzmann equation. By reformulating the cell-transmission problem as a conditional generation task, we train neural networks using conditional flow matching to sample particle exit states, including position, direction, and path length, without simulating scattering histories. The method employs optical coordinate scaling, enabling a single trained model to generalize across any material. We validate GMC on two canonical benchmarks, namely a heterogeneous lattice problem characteristic of nuclear reactor cores and a linearized hohlraum geometry representative of high-energy density radiative transfer. Results demonstrate that GMC preserves the statistical fidelity of standard Monte Carlo, exhibiting the expected $1/\sqrt{N}$ convergence rate while maintaining accurate scalar flux profiles. While standard Monte Carlo computational cost scales linearly with optical thickness in the diffusive limit, GMC achieves constant $O(1)$ cost per cell transmission, yielding order-of-magnitude speedups in optically thick regimes. This framework strategically aligns particle transport with modern computing architectures optimized for neural network inference, positioning transport codes to leverage ongoing advances in AI hardware and algorithms.
Introduction
Monte Carlo remains highly faithful for particle transport, but its cost is tied to resolving every scattering event. The paper introduces generative models as direct Monte Carlo samplers to reduce this bottleneck while preserving physics fidelity.
- Monte Carlo provides high-fidelity particle-transport modeling across reactor, radiative-transfer, and other streaming-and-scattering applications.
- Hardware and algorithmic advances have improved Monte Carlo implementations, but resolving every scattering interaction remains its fundamental computational bottleneck.
- Generative modeling offers learned mappings from simple base distributions to complex target distributions, enabling efficient sampling in scientific-computing applications.
- GMC trains a neural network to sample conditional particle-exit distributions directly instead of following particle random walks.
- GMC is presented as the first use of generative models as direct samplers inside a Monte Carlo algorithm, with one trained model intended to cover a geometry-class training distribution.
- An ML-based Monte Carlo algorithm can benefit from advances in AI hardware and algorithms, making model accuracy and efficiency central to simulation performance.
- Unlike condensed-history and random-walk methods, GMC requires no analytical or diffusion approximations, with accuracy limited by neural-network representation capacity and training data.
Theoretical Background
The linear Boltzmann equation balances streaming and collisions against scattering gains and independent sources. Standard Monte Carlo estimates transport quantities by simulating particle histories, tracking path lengths, absorption, and scattering events.
- The steady-state, monoenergetic linear Boltzmann equation enforces particle conservation in phase space defined by position and direction.
- The equation balances particle loss from streaming and collisions against particles scattered into a direction and independent production.
- Scalar flux integrates angular flux over directions and represents total particle path length per unit volume per unit time.
- Scalar flux determines reaction-rate densities through the product of scalar flux and the relevant macroscopic cross section.
- Monte Carlo reconstructs macroscopic transport solutions by simulating individual particle histories and aggregating their contributions to ensemble averages.
- Particles are initialized from a source distribution with an energy weight, while continuous absorption attenuates their weight along trajectories.
- Standard transport advances particles by ray tracing, sampling scattering distances, updating directions at collisions, and crossing cell boundaries otherwise.
- The scalar-flux estimator depends on total track length, while individual scattering events determine the distribution of that total length and the exit state.
Generative Monte Carlo Sampler
GMC replaces collision-by-collision cell transport with conditional neural generation of particle exit states. Conditional flow matching, optical coordinates, and separate internal and boundary models support efficient sampling while preserving transport distributions.
- Generative Monte Carlo Sampler: GMC learns a conditional distribution p(y|c) that maps particle entry states and cell geometry to exit states without simulating intermediate scattering events.Conditional flow matching trains a velocity field that transforms Gaussian noise into learned transport samples.
- Generative Monte Carlo Sampler: Separate Internal and Boundary Models handle volumetric-source particles and particles streaming between adjacent cells, respectively.Boundary particles are rotated into a canonical left-entry frame before inference, reducing the need to learn separate entry-face representations.
- Generative Monte Carlo Sampler: Optical coordinate scaling lets one trained model apply across physical cell dimensions and cross sections when optical thicknesses remain within the training distribution.The optical dimensions are measured in mean free paths using ˜W = Wσs and ˜H = Hσs.
- Generative Monte Carlo Sampler: The inference loop samples a cell exit state, applies analytic absorption and flux tallies, then iterates boundary transport for surviving particles.Absorption attenuates particle weight by exp(−σa∆s), with deposited energy accumulated in the cell tally.
- Generative Monte Carlo Sampler: The output y contains exit position, exit direction, and path length, while the condition c contains optical dimensions, entry position, and direction.The perimeter coordinate pexit identifies the exit face and position, and s records the in-cell path length.
- Generative Monte Carlo Sampler: GMC flow samples need not resemble individual scattering random walks because their ensemble exit distributions are statistically equivalent to standard Monte Carlo.The learned model reproduces correlations between exit direction and path length that arise from transport physics.
Results
GMC reproduces standard Monte Carlo scalar-flux behavior on heterogeneous lattice and linearized hohlraum benchmarks. It also matches Monte Carlo’s statistical convergence while avoiding the optical-thickness-dependent runtime growth of standard transport.
- Results: GMC reproduces the principal scalar-flux features of the lattice benchmark, including absorber shadows and steep gradients from the source.Both MC and GMC use 10^6 particles in the reported flux maps.
- Results: GMC lineouts closely track the MC reference across orders-of-magnitude flux attenuation through absorbing regions.The comparison uses horizontal and vertical cuts through the source center.
- Results: GMC models ray-like streaming, back-scattering from optically thick walls, and flux gradients across material interfaces in the linearized hohlraum benchmark.The geometry is a 1.3×1.3 cm domain discretized on a 112×112 mesh with a boundary source at the left wall.
- Convergence and Performance: Both standard MC and GMC exhibit the expected N^-1/2 convergence rate, with GMC curves coincident with the corresponding MC curves.The study uses five independent runs for each tested particle count.
- Convergence and Performance: GMC runtime is effectively O(1) with respect to optical thickness, whereas standard MC cost increases with the many scattering events in diffusive random walks.The runtime comparison varies optical cell dimensions in a purely scattering medium.
Conclusions
GMC reproduces transport exit distributions and preserves scalar-flux fidelity and Monte Carlo statistical convergence across heterogeneous geometries. Its constant-cost scaling in optically thick regimes also supports execution on modern inference-oriented hardware, while the presented results remain limited to single-energy, steady calculations.
- GMC accurately reproduces complex, multi-dimensional particle exit distributions without simulating scattering histories.
- Optical coordinate scaling enables one trained model to generalize across heterogeneous reactor-like lattices and streaming-dominated hohlraums.
- GMC preserves scalar flux and energy deposition profiles while matching standard Monte Carlo’s 1/sqrt(N) statistical convergence rate.
- O(1) GMC cost contrasts with standard Monte Carlo’s linear cost increase as materials become optically thick in the diffusive limit.GMC uses fixed-step ODE integration to effectively teleport particles across diffusive regions.
- GMC shifts transport computation toward contiguous vector operations, aligning physically accurate simulations with heterogeneous platforms emphasizing tensor operations and neural inference acceleration.
- The presented results focus on single-energy, steady calculations; energy- and time-dependent extensions are described as future work.