BCAM Scientific Seminar: Keeping an eye on the physics within the ML storm: From physics-informed neural networks to making training on physics-based synthetic data work on real data

Link to the session here


Abstract
Neural networks, and the process behind training them, have gained a lot of attention in recent years, resulting in the rise of a new breed of scientists, the “Data scientist”, with their motto: it is all in the data. Some data scientists have even “hypothesized” that they will no longer need our scientific theories and knowledge to solve physical problems.
For decades, we have relied on our knowledge, experience, and developed tools to interpret data and predict the key geo/physical properties behind them. We, specifically, use numerical solutions of our physical laws to invert the data and obtain physical models of the Earth. So, machine learning, which is also a numerical tool, but probably a more intrusive one, has been, fairly, accused of using the data to predict the physical properties, as simple “tasks”, with our physical knowledge buried in the hidden layers (the “black box” phenomenon), nowhere to be found. My task in this presentation is to hopefully convince you that training neural networks, which relies on a mix of statistics and inverse theory, can benefit from our geo/physical laws and a priori knowledge to guide us through the maze of degrees of freedom the neural networks may offer, and specifically to alleviate the many weaknesses/gaps/biases in data. This can be accomplished by instilling our physical laws and their corresponding characteristics into the neural network. It also can be accomplished by utilizing synthetic data for training in which physical laws are honored. However, this step will require that we devise approaches to make our synthetic data training work on Real data and I will share some of those.