Physics Informed Machine Learning of SPH: Machine Learning Lagrangian Turbulence
The emerging and revolutionary field of Scientific Machine Learning (SciML) has made great progress in recent years; with many successful applications in fluid dynamics, turbulence and data-driven discovery of reduced order models. A promising approach within this field is to encode appropriate physical constraints within machine learning algorithms, offering data-driven tools for the automatic discovery of physics-informed models. This talk will show progress made so far in developing such Physics Informed Machine Learning (PIML) techniques in order to discover reduced Lagrangian models for compressible turbulence applications. We will illustrate and analyze working PIML methods capable of learning a Lagrangian based model from Smooth Particle Hydrodynamics (SPH) data. This Lagrangian based PIML method combines modern machine learning tools, such as Automatic Differentiation, Differentiable Programming, Neural Networks, and optimization tools, along with classical (forward and adjoint) based Sensitivity Analysis methods in order to learn (estimate) parameters from data. Furthermore, our hierarchy of models gradually introduces more physical structure, which we show improves interpretability, generalizability (over larger ranges of time scales and Reynolds numbers), preservation of physical symmetries, and requires less training data.
Place: Math, 402 or