BRICKS: Compositional Neural Markov Kernels for Zero-Shot Radiation-Matter Simulation

2026/05/07 by Richard Hildebrandt, Evangelos Kourlitis, Baran Hashemi +7 · 1 voice
Computer Science · Physics and Astronomy · #cs.LG #hep-ph

paper · pdf · doi:10.48550/arxiv.2605.06591

Abstract

We introduce a new strategy for compositional neural surrogates for radiation-matter interactions, a key task spanning domains from particle physics through nuclear and space engineering to medical physics. Exploiting the locality and the Markov nature of particle interactions, we create a next-particle prediction kernel using hybrid discrete-continuous transformer models based on Riemannian Flow Matching on product manifolds. The model generates variable-sized typed sets of particles and radiation side effects that are the result of the interaction of an incident particle with a material volume. The resulting kernel can be composed to simulate unseen large-scale material distributions in a zero-shot manner. Unlike mechanistic simulators, our model is designed to be differentiable, provides tractable likelihoods for future downstream applications. A significant computational speed-up on GPU compared to CPU-bound mechanistic simulation is observed for single-kernel execution. We evaluate the model at the kernel level and demonstrate predictive stability over multi-round autoregressive rollouts. We additionally release a novel 20M-event radiation-matter interaction dataset for further research.

Citations

• SoFlow: Solution Flow Models for One-Step Generative Modeling
• Generative Monte Carlo Sampling for Constant-Cost Particle Transport
• CaloClouds3: Ultra-Fast Geometry-Independent Highly-Granular Calorimeter Simulation
• ParaFlow: fast calorimeter simulations parameterized in upstream material configurations
• APEBench: A Benchmark for Autoregressive Neural Emulators of PDEs
• CaloChallenge 2022: A Community Challenge for Fast Calorimeter Simulation
• Deep Generative Models for Ultra-High Granularity Particle Physics Detector Simulation: A Voyage From Emulation to Extrapolation
• Denoising diffusion models with geometry adaptation for high fidelity calorimeter simulation
• New Angles on Fast Calorimeter Shower Simulation
• Efficiently Scaling Transformer Inference
• Message Passing Neural PDE Solvers
• Learning to Simulate Complex Physics with Graph Networks
• Understanding and Improving Layer Normalization
• Mining gold from implicit models to improve likelihood-free inference

Discussions

• This was a true team effort with a very cool student group at TUM and also friends around the world. Can't wait to see where this goes. Paper Link: arxiv.org/pdf/2605.06591 [bsky, 2 points, 0 comments]

Related

• Neural Network Surrogate Models for Absorptivity and Emissivity Spectra of Multiple Elements
• Differentiable Physics-Neural Models enable Learning of Non-Markovian Closures for Accelerated Coarse-Grained Physics Simulations
• Accelerating Science with Generative Adversarial Networks: An Application to 3D Particle Showers in Multi-Layer Calorimeters
• Generative Monte Carlo Sampling for Constant-Cost Particle Transport
• Physics-consistent machine learning: output projection onto physical manifolds