Using Differentiable Physics for Self-Supervised Assimilation of Chaotic Dynamical Systems
Published in Workshop on Differentiable Vision, Graphics, and Physics in Machine Learning at NeurIPS 2020, 2020
Recommended citation: Michael McCabe, Jed Brown. (2020). "Using Differentiable Physics for Self-Supervised Assimilation of Chaotic Dynamical Systems." Workshop on Differentiable Vision, Graphics, and Physics in Machine Learning at NeurIPS. https://montrealrobotics.ca/diffcvgp/assets/papers/16.pdf
We propose a deep learning based data assimilation framework which we call \textit{Amortized Assimilation} for state estimation in high-dimensional chaotic dynamical systems. Amortized assimilators utilize differentiable simulation of physics-derived system dynamics to enable end-to-end physics-aware gradient based training of denoising neural networks which update a simulated system state based on noisy observations. These hybrid models are able to learn to assimilate complex input distributions while maintaining a computable test-time update step in an entirely self-supervised manner using only sequences of noisy observations without loss of accuracy over training with ground truth targets. Numerical experiments demonstrate that amortized assimilators compare favorably with widely used data assimilation methods across common benchmark tasks.
Recommended citation: Michael McCabe, Jed Brown. (2020). “Using Differentiable Physics for Self-Supervised Assimilation of Chaotic Dynamical Systems.” Workshop on Differentiable Vision, Graphics, and Physics in Machine Learning at NeurIPS.