Differentiability in unrolled training of neural physics simulators on transient dynamics

Unrolling training trajectories over time strongly influences the inference accuracy of neural network-augmented physics simulators. We analyze these effects by studying three variants of training neural networks on discrete ground truth trajectories. In addition to commonly used one-step setups and fully differentiable unrolling, we include a third, less widely used variant: unrolling without temporal gradients. Comparing networks trained with these three modalities makes it possible to disentangle the two dominant effects of unrolling, training distribution shift and long-term gradients. We present a detailed study across physical systems, network sizes, network architectures, training setups, and test scenarios. It also encompasses two modes of computing the simulation trajectories. In prediction setups, we rely solely on neural networks to compute a trajectory. In contrast, correction setups include a numerical solver that is supported by a neural network. Spanning all these variations, our study provides the empirical basis for our main findings: A non-differentiable but unrolled training setup supported by a numerical solver in a correction setup can yield substantial improvements over a fully differentiable prediction setup not utilizing this solver. We also quantify a difference in the accuracy of models trained in a fully differentiable setup compared to their non-differentiable counterparts. Differentiable setups perform best in a direct comparison of correction networks, and the same is observed when comparing prediction setups among each other. In both cases, the accuracy of unrolling without temporal gradients comes relatively close. Furthermore, we empirically show that these behaviors are invariant to changes in the underlying physical system, the network architecture and size, and the numerical scheme. These results motivate integrating non-differentiable numerical simulators into training setups even if full differentiability is unavailable. We also observe that the convergence rate of common neural architectures is low compared to numerical algorithms. This encourages the use of correction approaches combining neural and numerical algorithms to utilize the benefits of both.