MetaPhysiCa: OOD Robustness in Physics-informed Machine Learning

This A fundamental challenge in physics-informed is because the standard ML part of PIML, which tends to machine learning (PIML) is the design of robust learn spurious associations, will perform poorly in our OOD PIML methods for out-of-distribution (OOD) setting. We then propose a promising solution: Combine forecasting tasks. These OOD tasks require meta learning with causal structure discovery to learn an learning-to-learn from observations of the same ODE model that is robust to OOD initial conditions and can (ODE) dynamical system with different unknown adapt to OOD parameters of the dynamical system. In our ODE parameters, and demand accurate forecasts OOD tasks, OOD robustness means that the robustness is even under out-of-support initial conditions and tied to interventions over the initial conditions and unknown out-of-support ODE parameters. In this work we parameters of the system, not on arbitrary interventions as propose a solution for such tasks, which we define the system evolves from the initial state. This is an important as a meta-learning procedure for causal structure distinction. There can be multiple ODE models that discovery (including invariant risk minimization).