Optimal control of partial differential equations in PyTorch using automatic differentiation and neural network surrogates
The take-home message of this paper is that solving optimal control problems can be computationally straightforward, provided that differentiable partial differential equation (PDE) solvers are available. Although this might seem to be a strong limitation and the development of differentiable PDE solvers might seem arduous, for many problems this is not the case. In particular, for linear partial differential equations, they are equivalent to a linear equation system. Therefore it is just sufficient to be able to solve linear equation systems in an automatic differentiation-capable library like PyTorch and be able to differentiate through the linear solver. Using open-source libraries, like torch_sparse_solve, we have easy to use differentiable direct solvers (here: KLU) at hand, which makes solving linear PDEs straightforward. For nonlinear PDEs, the approach above might be not sufficient, since we do not have an equivalent linear equation system for the underlying PDE. In this case, we can then use the library torch-fenics, which enables us to define PDEs using the finite element library FEniCS and then differentiate through this problem in PyTorch. We complement the proposed methodology with various optimal control problems constrained by stationary and instationary PDEs, linear and nonlinear PDEs (from Poisson to fluid-structure interaction), space-time and time-stepping formulations, parameter estimation, right-hand side control, initial condition control and boundary condition control, finite difference and finite element discretizations and neural network surrogates. All developments of this work are accompanied with the respective source codes published on GitHub.