Towards Agentic Intelligence for Materials Science

TL;DR

This paper champions a pipeline-centric shift in AI for materials science, arguing that true discovery acceleration requires agentic LLMs capable of planning, acting, and learning across the entire discovery loop. It links foundation-model pretraining, domain adaptation, tool integration, and autonomous experimentation into a single end-to-end system, with credit assignment from real-world discovery outcomes guiding upstream design and data curation. By analyzing reactive tasks (prediction, mining, generation, optimization, and verification) and contrasting proxy benchmarks with end-to-end discovery rewards, the authors lay out a practical roadmap toward safe, memory-enabled, long-horizon agents (e.g., self-driving labs and Scientist AI) that can generate, test, and revise hypotheses while coordinating humans and robotic systems. The work emphasizes uncertainty quantification, multimodal data fusion, and explainability as essential components for trustworthy automation, and it calls for governance and safety mechanisms as agentic ecosystems become more capable. Overall, the paper provides a cohesive framework and a staged progression from static models to autonomous, safety-aware AI4MatSci agents aimed at discovering novel, useful materials.

Abstract

The convergence of artificial intelligence and materials science presents a transformative opportunity, but achieving true acceleration in discovery requires moving beyond task-isolated, fine-tuned models toward agentic systems that plan, act, and learn across the full discovery loop. This survey advances a unique pipeline-centric view that spans from corpus curation and pretraining, through domain adaptation and instruction tuning, to goal-conditioned agents interfacing with simulation and experimental platforms. Unlike prior reviews, we treat the entire process as an end-to-end system to be optimized for tangible discovery outcomes rather than proxy benchmarks. This perspective allows us to trace how upstream design choices-such as data curation and training objectives-can be aligned with downstream experimental success through effective credit assignment. To bridge communities and establish a shared frame of reference, we first present an integrated lens that aligns terminology, evaluation, and workflow stages across AI and materials science. We then analyze the field through two focused lenses: From the AI perspective, the survey details LLM strengths in pattern recognition, predictive analytics, and natural language processing for literature mining, materials characterization, and property prediction; from the materials science perspective, it highlights applications in materials design, process optimization, and the acceleration of computational workflows via integration with external tools (e.g., DFT, robotic labs). Finally, we contrast passive, reactive approaches with agentic design, cataloging current contributions while motivating systems that pursue long-horizon goals with autonomy, memory, and tool use. This survey charts a practical roadmap towards autonomous, safety-aware LLM agents aimed at discovering novel and useful materials.

Figures (10)

• Figure 1: An overview of the key sections and an illustrative end-to-end pipeline encompasses key elements like general pre-training tasks & data, foundation language models, domain-specific tasks & data, materials-oriented model adaptation, goal-driven LLM agents, and iterative open-ended materials experimentation. The colored arrows show how feedback signals from open-ended experiments are routed to the agentic LLM, the materials science-tuned LLM, the pre-trained LLM, and their corresponding task modules. Different colors of arrows indicate the sources of the feedbacks. For example, if general-purpose pre-training data does not contribute positively to the ultimate goal of novel and useful materials discovery, adjustments should be made by revising the pre-training tasks and corpus or fine-tuning the pre-trained LLM accordingly to better align model knowledge with the ultimate objective of discovering novel materials. Dashed arrows denote the forward information flows. Note that humans are not only responsible for monitoring the open-ended experiments, but also co-improve with the agentic LLM.
• Figure 2: Technology tree of AI4Material science research. With the emergence of LLMs and agents, research on material science initially focused on domain specific task, primarily concentrating on seperate reactive tasks. Subsequent research has delved deeper, gradually integrating more with agentic systems for material science.
• Figure 3: Taxonomy of recent progress of AI and LLMs.
• Figure 4: The gradual integration of General AI into the materials science workflow. This framework illustrates the progressive evolution of broad, pre-trained foundation models as they are adapted for scientific discovery. Rather than a sudden replacement, the diagram depicts how general capabilities are iteratively refined through domain adaptation and feedback loops, allowing general-purpose AI to be gradually and effectively applied to specialized materials challenges.
• Figure 5: Taxonomy of reactive tasks in materials science from an AI perspective, Part 1.