Materials Informatics: Emergence to Autonomous Discovery in the Age of AI

Abstract

ABSTRACT We provide a perspective on the evolution of materials informatics, tracing its conceptual roots to foundational ideas in physics and information theory and its maturation through the integration of machine learning and artificial intelligence (AI). Early contributions from Chelikowsky, Phillips, and Bhadeshia laid the groundwork for what has become a transformative approach to materials discovery. The U.S. Materials Genome Initiative catalyzed a surge in activity, and the period from 2014–2016 marked the first impactful applications of machine learning to materials problems. Since then, the field has seen rapid advances, particularly with the advent of deep learning and transformer‐based large language models (LLMs), which now underpin tools for property prediction, synthesis planning, and inverse design. We present the subject not as a collection of tools, but as an evolving research ecosystem – tracing the historical context, current situation, and future development direction of the field. We review key methodologies—including approaches for sequential design, such as Bayesian Optimization and Reinforcement Learning, and transformers —highlighting their role in accelerating discovery and automating experimentation with growing efforts in building autonomous self‐driving laboratories. We discuss common pitfalls, issues, and solutions associated with LLMs and Bayesian Optimization as applied to the study of specific materials systems. Given the costs of pre‐trained AI models for specific materials, we consider the merits of specialist LLMs versus today's state‐of‐the‐art generalists. Finally, we assess emerging challenges and the potential for AI to evolve from a predictive tool into a collaborative partner in research. We argue that with advances in active learning, uncertainty quantification, and retrieval‐augmented generation, a new era of autonomous materials science is within reach—one in which the human is increasingly taken out of the loop.