Graph network for learning bi-directional physics
This work addresses the need for more generalizable physics-informed neural networks by eliminating problem-specific engineering, which could benefit researchers in computational physics and machine learning.
The authors tackled the problem of learning both forward and inverse models of particle-based physics with a single graph network, achieving at least an order of magnitude higher accuracy in forward dynamics prediction compared to related methods and recovering multi-modal probability distributions for unknown parameters.
In this work, we propose an end-to-end graph network that learns forward and inverse models of particle-based physics using interpretable inductive biases. Physics-informed neural networks are often engineered to solve specific problems through problem-specific regularization and loss functions. Such explicit learning biases the network to learn data specific patterns and may require a change in the loss function or neural network architecture hereby limiting their generalizabiliy. Our graph network is implicitly biased by learning to solve several tasks, thereby sharing representations between tasks in order to learn the forward dynamics as well as infer the probability distribution of unknown particle specific properties. We evaluate our approach on one-step next state prediction tasks across diverse datasets. Our comparison against related data-driven physics learning approaches reveals that our model is able to predict the forward dynamics with at least an order of magnitude higher accuracy. We also show that our approach is able to recover multi-modal probability distributions of unknown physical parameters.