TL;DR
This paper extends the Direct Simulation Monte Carlo (DSMC) method to efficiently compute gradients for optimization problems involving the Boltzmann equation with general collision models. By developing an "adjoint DSMC" approach that handles variable collision rates and angle-dependent scattering, the authors enable more robust and accurate parameter calibration for complex kinetic simulations.
What Problem Does This Paper Address?
Optimizing parameters in complex physical models, like those described by the Boltzmann equation, is crucial for accurate simulations and predictions. However, directly computing gradients of objective functions with respect to model parameters can be computationally prohibitive due to the high dimensionality of parameters and the lack of explicit analytical relationships between parameters and model solutions. The Boltzmann equation, widely used in areas from rarefied gas dynamics to semiconductor physics and financial modeling, presents a particular challenge due to its integro-differential nature and complex collision operator. Previous work by the authors successfully applied adjoint DSMC to simpler cases (Maxwell molecules with constant collision rates), but generalizing this to more realistic collision models with variable rates and velocity/angle dependencies remained an open problem.
Methods
The core contribution is the development of a generalized adjoint DSMC method. This approach builds upon the Discretize-Then-Optimize (DTO) framework, where the adjoint of the DSMC algorithm is derived. The key novelty lies in handling the rejection sampling step, which is necessary for collision models where the collision rate varies with particle velocities. This involves:
- Adjoint Jacobian Matrix: Deriving a new matrix that captures the dependence of the collision outcome on particle velocities.
- Score Function Term: Introducing a new term in the adjoint equation that accounts for the variable collision rate.
- Two Derivation Approaches: The adjoint method is derived through both a direct differentiation approach and a Lagrangian multiplier approach, yielding the same results and providing complementary insights into the adjoint variables as "influence functions."
- Algorithm Modifications: The paper presents modified DSMC algorithms (Algorithm 1 and 3) to handle general collision kernels, including those with velocity and angle dependence. Algorithm 3 is specifically optimized for separable collision kernels.
Datasets
Not discussed in the paper. The paper focuses on the theoretical development and numerical validation of the adjoint DSMC method using simulated data generated by the DSMC algorithms themselves.
Key Insights and Ideas
- Rejection Sampling is Key: The primary challenge in generalizing adjoint DSMC is the rejection sampling step required for variable collision rates. The authors successfully incorporate this into the adjoint framework.
- Angle Dependence Matters: For collision kernels that depend on the scattering angle, a specific term (47) in the adjoint matrix is crucial for accurate gradient computation. This term accounts for the dependence of the post-collision velocity on the pre-collision velocity.
- DTO and OTD Hybrid: The method combines the DTO approach (discretizing first, then differentiating) with an OTD step (differentiating the expectation over the rejection sampling decision before sampling). This hybrid approach circumvents difficulties arising from differentiating discontinuous decision functions.
- Efficiency Gains: While Algorithm 1 is more general, Algorithm 3 offers improved efficiency for separable collision kernels by sampling collision parameters only for actual collisions.
Key Results
- The developed adjoint DSMC method accurately computes gradients for spatially homogeneous Boltzmann equations with general collision models, including those with velocity and angle dependence.
- Numerical tests demonstrate that the gradient error decreases with an increasing number of particles (N), as expected, and that the random errors in the adjoint DSMC gradients are significantly smaller than those from finite difference approximations.
- The importance of the additional term (47) in the adjoint matrix for angle-dependent kernels is numerically confirmed, showing that its omission leads to incorrect adjoint gradients. For angle-independent kernels, this term has a negligible effect due to averaging.
Limitations and Future Work
- Spatially Homogeneous Case: The current work focuses on the spatially homogeneous Boltzmann equation. Extending the method to spatially inhomogeneous cases remains a future direction.
- Computational Cost: While more efficient than direct gradient computation, DSMC simulations can still be computationally intensive, especially for large numbers of particles and time steps.
- Future Work: The authors plan to extend the adjoint DSMC method to Coulomb collisions and apply it to large-scale optimization problems constrained by Boltzmann equations. They also suggest exploring the method for other kinetic models.
Summary updated