AI code generation and numerical codes

Project Description

A great deal of work is being done on automated code generation through LLMs, for example, through Github’s work on Copilot. As yet, though, little has been done to consider the implications of AI-assisted code generation for highly numerical code, for example, the parallel simulation codes used to develop digital twins of complex physical phenomena. We propose to study the behaviour of LLM-assisted coding for mathematically sophisticated floating-point software. This may lead to development of new domain specific languages closer to the natural language that would be used by mathematicians to describe a programme’s key characteristics, for example “A 2-d implementation of the wave equation on an adaptive mesh, parallelised with a Halo-Swap