Machine Learning as a Bridge Between Scales for Accelerated Computation
When
Scientific endeavors study a great number of space and time scales, from the subatomic to the intergalactic. Analytic descriptions of physical phenomena are often confined to operate over specific length and time scales, over which the physics is separated from the complexity of larger or smaller phenomena. However, in recent decades, the Computational Science paradigm has emerged, which can treat a great variety of systems by simulating the interaction of their constituents. This approach proves useful for explaining phenomena from the ground up, showing how they emerge from constituent small-scale interactions, but in many situations is intractably expensive.
Recently, Machine Learning (ML) has arisen as an automation paradigm for creating computational models. While the classic domain of ML is to perform human tasks automatically with high throughput, it is also extremely valuable for modeling in computational science, allowing one to amortize the cost of expensive simulations via large-scale data analysis. As such, ML provides highly effective techniques for accelerating computational science. We highlight a spectrum of our work on ML for accelerating computational science by providing a bridge between scales, and the physical considerations needed to build useful models. This provides valuable bridges between electronic scales, molecular scales, mesoscopic scales, and continuum scales. Zoom: https://arizona.zoom.us/j/99138847381