Abstract
We propose a novel machine learning algorithm for simulating radiative transfer. Our algorithm is based on physics informed neural networks (PINNs), which are trained by minimizing the residual of the underlying radiative transfer equations. We present extensive experiments and theoretical error estimates to demonstrate that PINNs provide a very easy to implement, fast, robust and accurate method for simulating radiative transfer. We also present a PINN based algorithm for simulating inverse problems for radiative transfer efficiently.
Introduction
Radiative transfer—the study of how energy in the form of electromagnetic radiation moves through a medium—is essential to fields ranging from astrophysics and climate modeling to medical imaging. However, simulating it is notoriously difficult because the fundamental Radiative Transfer Equation (RTE) is a high-dimensional integro-differential equation. In its most general form, radiative intensity depends on up to seven variables: three for space, one for time, two for angle, and one for frequency.
Traditional numerical methods, such as finite elements or finite differences, suffer from the "curse of dimensionality," where computational requirements grow exponentially with the number of dimensions. Furthermore, media with highly varying optical properties (heterogeneity) make designing robust, efficient solvers even more complex. This paper introduces a novel machine learning approach using Physics-Informed Neural Networks (PINNs) to overcome these limitations.
Methods
The proposed approach leverages PINNs as a continuous approximation (an ansatz) for radiative intensity, rather than discretizing space and time into a grid. The network is trained by minimizing a loss function that incorporates the residual of the RTE itself, alongside initial and boundary condition mismatches.
Key technical components include:
- Residual Minimization: The physics of the system is embedded directly into the training process.
- Quadrature for Integrals: To handle the integral term (scattering kernel) within the RTE, Gauss-Legendre quadrature is employed.
- Sobol Sequences: Low-discrepancy Sobol sequences are used to select training points, effectively mitigating the curse of dimensionality compared to uniform grids.
- Inverse Problem Adaptation: Unknown physical parameters, such as absorption coefficients, are treated as trainable variables, allowing the network to solve inverse problems simultaneously.
The architecture typically consists of a standard fully connected deep neural network, as illustrated in the figure below.
Experimental Results
The method was evaluated on several synthetic benchmarks, including 1D monochromatic slabs, 3D monochromatic unit cubes (5D problem), and 3D polychromatic time-dependent cases (6D problem). The authors conducted extensive experiments with various hyperparameter configurations to ensure robustness across these diverse scenarios.
Performance and Accuracy
The PINN approach demonstrated remarkable accuracy across dimensions. In 1D benchmarks, the generalization error was as low as 0.24%. Even in the complex 6D polychromatic case, the error in radial flux remained approximately 2.1%.
Computational efficiency is a significant advantage; training times ranged from 20 minutes for simple cases to 2-3 hours for complex 3D time-dependent problems on a single GPU. This represents a substantial speedup over traditional high-dimensional solvers. Furthermore, the PINN successfully recovered unknown absorption coefficients in inverse problems with less than 3% error.
Discussion and Conclusion
A key insight of this work is the theoretical error estimate showing that the generalization error for PINNs depends only logarithmically on the dimension. This explains their superior performance in high-dimensional spaces where traditional methods fail. Additionally, PINNs offer a unified framework for both forward and inverse problems and can handle sharp gradients and discontinuities despite being smooth approximators.
Future work will address the stability of time-dependent error bounds for long-duration simulations and the challenges posed by purely scattering media. The ultimate goal is the integration of PINN-based radiation modules into comprehensive radiation-hydrodynamics codes for complex physical simulations.
Summary updated