Physics-based modeling and data-driven modeling are the two major categories of approaches for modeling in science and engineering. Traditionally, computational hydraulics models solve the governing partial differential equations (PDEs) using numerical schemes such as finite volume method. These PDEs are the mathematical description of the underlying physics. Data-driven approach uses the data, which suppose to have the physics embedded, and learn the dynamics of the system. Recently, the combination of both physics-based and data-driven modeling approaches has gained increasing popularity due to the rapid development of machine learning (ML) and artificial intelligence (AI). Among many of the proposed combination strategies, differentiable programming has emerged as a very promising path forward. Models built with differentiable programming are “differentiable”, which means it is possible to accurately and efficiently calculate gradients with respect to model variables and parameters. Differentiability opens the door for many possibilities, such as quick optimization, parameter estimation, inversion, and the discovery of new knowledge embedded in high-dimensional space. This talk will demonstrate some proof of concept and simple differentiable hydraulics models.
