PriM: Principle-Inspired Material Discovery through Multi-Agent Collaboration

TL;DR

PriM introduces a principle-guided, language-driven multi-agent system for automated materials discovery that embeds physicochemical principles within hypothesis generation and experimental validation. By coordinating Literature and Hypothesis Agents with physics-informed Experiment, Virtual Lab, and Monte Carlo Tree Search–based optimization, PriM delivers transparent reasoning pathways and improved exploration efficiency. In a nanohelix case study, PriM demonstrated substantial improvements in g-factor optimization and interpretability, achieving a maximum $g$-factor of $0.974$ and a 133% improvement over the initial configuration. Overall, PriM bridges data-driven search with scientific principles, enabling faster, more reliable, and more interpretable AI-assisted materials discovery with broad applicability to rational design.

Abstract

Complex chemical space and limited knowledge scope with biases holds immense challenge for human scientists, yet in automated materials discovery. Existing intelligent methods relies more on numerical computation, leading to inefficient exploration and results with hard-interpretability. To bridge this gap, we introduce a principles-guided material discovery system powered by language inferential multi-agent system (MAS), namely PriM. Our framework integrates automated hypothesis generation with experimental validation in a roundtable system of MAS, enabling systematic exploration while maintaining scientific rigor. Based on our framework, the case study of nano helix demonstrates higher materials exploration rate and property value while providing transparent reasoning pathways. This approach develops an automated-and-transparent paradigm for material discovery, with broad implications for rational design of functional materials. Code is publicly available at our \href{https://github.com/amair-lab/PriM}{GitHub}.

Figures (4)

• Figure 1: Overview of PriM. Two phases: hypothesis generation and experimental validation. The Planner initiates a hypothesis loop involving a Literature Agent (gathering prior knowledge) and Hypothesis Agent (formulating testable hypotheses). Validated hypotheses then undergo experimental testing: the Experiment Agent designs conditions, while the Optimizer Agent employs method, i.e., Monte Carlo Tree Search (MCTS), to optimize outcomes. The process iterates through reasoning and roundtable discussions to refine experiments. Crucially, hypotheses are grounded in physicochemical principles, balancing exploration and exploitation to drive discovery.
• Figure 2: Comparison of nanohelices discovery progress with g-factor optimization. PriM achieves a high g-factor value in significantly fewer steps compared to baseline methods, highlighting the benefits of its physicochemical grounding.
• Figure 3: Evolution of Experiment Values Across Iterations by Optimizer Agent.
• Figure 4: Step-by-Step Principle Evolution. Each step records the principles behind the hypothesis, the changes of parameter values, and the achieved g-factor, highlighting key improvements and showcases PriM's ability to balance exploration and exploitation.