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 tranfer 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.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is to design a fast, accurate and easy - to - implement numerical method to simulate the radiative transfer equation in the context of high - dimensionality and optical heterogeneity. Specifically, traditional numerical methods such as the finite element method or the finite difference method face huge computational resource requirements and the curse of dimensionality problem when dealing with the radiative transfer equation. In addition, when dealing with media with different - scale absorption and scattering coefficients and emission terms, it is very challenging to design robust and efficient algorithms for these methods. Therefore, the paper proposes a new algorithm based on Physics Informed Neural Networks (PINNs), aiming to overcome these problems and provide an efficient and accurate solution.
### Main contributions of the paper:
1. **Propose a new algorithm**: The paper proposes a new algorithm based on Physics Informed Neural Networks (PINNs) to approximate the radiative transfer equation (1.1).
2. **Error analysis**: A strict error analysis of the proposed algorithm is carried out, proving the generalization error estimate of PINN, indicating that as long as PINN is well - trained, it can approximate the solution of the radiative transfer equation with high accuracy.
3. **Numerical experiments**: The accuracy and efficiency of the algorithm are demonstrated through a series of numerical experiments.
4. **Inverse problems**: An important advantage of PINNs is that it can approximate inverse problems with the same complexity as forward problems. Therefore, the paper also modifies PINNs to approximate the inverse problem of radiative transfer, that is, to determine the unknown absorption or scattering coefficients from the measured values of radiation intensity.
### Specific problem description:
The radiative transfer equation is a linear partial - integral - differential equation in the following form:
\[ \frac{1}{c} \frac{\partial u}{\partial t} + \omega \cdot \nabla_x u + k u + \sigma \left( u - \frac{1}{s_d} \int_{\Lambda} \int_{S} \Phi(\omega, \omega', \nu, \nu') u(t, x, \omega', \nu') d\omega' d\nu' \right) = f, \]
where:
- \( t \in [0, T] \) is the time variable,
- \( x \in D \subset \mathbb{R}^d \) is the space variable,
- \( \omega \in S = S^{d - 1} \) is the angle variable,
- \( \nu \in \Lambda \subset \mathbb{R} \) is the frequency (or energy level) variable,
- \( c \) is the speed of light,
- \( s_d \) is the surface area of the \( d \)-dimensional unit sphere,
- \( u: D_T \times S \times \Lambda \to \mathbb{R} \) is the radiation intensity to be found,
- \( k: D \times \Lambda \to \mathbb{R}^+ \) is the absorption coefficient,
- \( \sigma: D \times \Lambda \to \mathbb{R}^+ \) is the scattering coefficient,
- \( \Phi: S \times S \times \Lambda \times \Lambda \to \mathbb{R} \) is the scattering kernel,
- \( f: D \times \Lambda \to \mathbb{R} \) is the source (emission) term.
### Solution:
The method proposed in the paper is based on PINNs and trains the neural network by minimizing the residual of the radiative transfer equation. The specific steps include:
1. **Select the training set**: Select internal training points, time - boundary training points and space - boundary training points.
2. **Define the loss function**: Define a loss function that includes PDE residual, initial - condition residual and boundary - condition residual.
3. **Optimization algorithm**: Use a gradient - descent method (such as ADAM or LBFGS) to minimize the loss function and find the optimal weight parameters.