Learning physical models that can respect conservation laws (Q6198203)

The authors formulate the problem of learning physical models that can respect conservation laws from the finite volume perspective by writing the governing conservation law in integral form rather than the commonly-used in SciML (scientific machine learning) differential form. This permit allows for instance to incorporate the global integral form of the conservation law as a linear constraint into black-box ML models. This allows in turn to develop a two-step framework that first trains a black-box probabilistic ML model, and then constrains the output using a probabilistic constraint of the linear integral form. A framework called ProbConserv is proposed for incorporating conservation constraints into a generic SciML architecture. One of the main advantages of this framework is that it combines the integral form of a conservation law with a Bayesian update. A complete analysis of ProbConserv on learning with the generalized porous medium equation (GPME) is provided. ProbConserv is effective for easy GPME variants, performing well with state-of-the-art competitors; and for harder GPME variants it outperforms other approaches that do not guarantee volume conservation. ProbConserv enforces physical conservation constraints, maintains probabilistic uncertainty quantification (UQ), and deals well with shocks and heteroscedasticities. In each case, it achieves superior predictive performance on downstream tasks.