Simulation-Based Machine Learning: The Emerging Frontier of AI Research

Abstract

Both traditional model-driven approaches and data-driven approaches (machine learning) have their respective advantages and disadvantages. Recognizing the need for a new approach that complements the strengths of both, we have undertaken research in simulation-based machine learning. Our main achievements to date include: (1) replacing simulations with deep learning to accelerate computational processes, (2) developing deep learning methods faithful to physical laws (AI for Science), and (3) solving complex mathematical models using deep learning, transitioning from numerical (discrete) solutions to continuous ones. This article presents an overview of these efforts. While deep learning has already been applied to various fields through end-to-end learning, as exemplified by generative AI, further advancements are anticipated—especially in leveraging automatic differentiation in backpropagation for learning and solving ODEs and PDEs.