Domain-Specific Languages for Scientific Computing

14^th June 2022, 13:00 Ashton Lecture Theatre
Dr. Navjot Kukreja
Department of Computer Science, University of Liverpool

Abstract

In this talk, I will give an overview of my research in order to foster discussions and collaborations, and as a way to introduce myself to the Department. I will discuss the simulation of wave propagation using finite difference methods, inverse problems, and how the combination of these two (Full-Waveform Inversion) helps the Oil and Gas Industry find Petroleum. This is a mathematical problem that typically runs distributed on hundreds of nodes of an HPC cluster, and this scale leads to software performance being the primary driver of engineering and design decisions when writing software to implement this. I will then talk about how we addressed this problem by building a Domain-Specific Language (the compiler, really). I will talk specifically about the language features built to enable this DSL to be used for large-scale backpropagation (inverse problems need the same backpropagation as a neural network training, but at a much larger scale) including Automatic Differentiation. Next, I will discuss the upcoming area of scientific machine learning and how it faces the same performance issues as large-scale inverse problems. I will finally discuss my ongoing work in addressing these performance issues.