Accelerating Simulations using Agentic Workflows for Multiphysics Problems

The focus of this talk is on the development of natural language to workflow generation for multiphysics problems. Such translation of user intent into synthesis of simulation pipeline has the potential to vastly accelerate the use of computational tools for design and modeling of complex systems. We discuss three topics: a) the reasoning process for generating chains of thoughts (COTs), and the mathematical underpinnings of such approaches; b) the code generation process including the bias towards programming languages implicit in large language models (LLM); c) a comprehensive framework for knowledge retention including LLM augmentation using retrieval-augmented generation (RAG) approach. Examples in the automatic generation of simulations for predicting ignition delay time for combustion problems will be presented.