Convergence, Adaptivity, and Applications of Physics-Informed Machine Learning
Abstract
Extensive work in applying deep learning to broader fields of science and engineering have been emerging in recent times, to include materials informatics, thermodynamics, and numerous other fields of computational sciences. Advances in these areas have been of particular excitement as future materials and new and informative laws of nature can be learned from data, even if that data is less than what would typically be required of a deep learning approach. In this work, we focus on the development and democratization of Physics-Informed Deep Learning, a field of science that was proposed before the turn of the century but has recently been gaining rapid popularity among academia and industry alike.
This dissertation is centered around recent work in physics-informed deep learning, as well as other areas of deep learning applications in computational sciences, such as materials informatics. Specifically, we will address recent advances in training stability and convergence of PINN solvers to semi-linear and stiff problems where the baseline PINN fails to converge or train effectively. We will discuss specific applications of PINNs to computational science domains where it could provide a force multiplier to researchers, and work performed in deep learning estimation of phase field modeling. Additionally, we will discuss the open-source package TensorDiffEq, a Python package based on Tensorflow that allows for easy implementation of PINN-based forward, inverse, and data assimilation solvers.
Citation
McClenny, Levi Daniel (2022). Convergence, Adaptivity, and Applications of Physics-Informed Machine Learning. Doctoral dissertation, Texas A&M University. Available electronically from https : / /hdl .handle .net /1969 .1 /197313.