AI for Scientific Discovery: A Survey of Methods, Applications, and Open Challenges

Abstract

Artificial intelligence is increasingly used not merely to automate existing scientific workflows but to actively accelerate the process of scientific discovery itself, from proposing candidate drug molecules and novel materials to inferring governing physical laws directly from experimental data. This paper surveys the growing field of AI for scientific discovery, organizing methods into a taxonomy spanning generative molecular design, graph-based property prediction, structure prediction, symbolic regression, and experiment design through active learning and Bayesian optimization. We review applications across drug discovery, materials science, physics, astronomy, climate science, and chemistry, summarize commonly used benchmarks, and present a worked case study illustrating how a generative model, a property predictor, and an active learning loop can be combined into an end-to-end discovery pipeline. We discuss open challenges including data scarcity in specialized scientific domains, the gap between computational prediction and experimental validation, and the difficulty of ensuring that AI-proposed hypotheses respect known physical constraints. We conclude by outlining future directions connecting AI for science with foundation models, automated laboratories, and human-AI collaborative discovery workflows.