[Paper Review] Rapid and Automated Alloy Design with Graph Neural Network-Powered LLM-Driven Multi-Agent Systems

Alireza Ghafarollahi, Markus J. Buehler|arXiv (Cornell University)|Oct 17, 2024

Manufacturing Process and Optimization4 citations

TL;DR

This paper presents a multi-agent AI system powered by graph neural networks (GNNs) and large language models (LLMs) to accelerate automated alloy design. By integrating physics-informed simulations with LLM-driven agents, the system rapidly predicts key atomic-scale properties—Peierls barrier and solute-screw interaction energy—in NbMoTa-based body-centered cubic alloys, reducing reliance on costly brute-force simulations and enabling efficient exploration of vast compositional spaces with high accuracy.

ABSTRACT

A multi-agent AI model is used to automate the discovery of new metallic alloys, integrating multimodal data and external knowledge including insights from physics via atomistic simulations. Our multi-agent system features three key components: (a) a suite of LLMs responsible for tasks such as reasoning and planning, (b) a group of AI agents with distinct roles and expertise that dynamically collaborate, and (c) a newly developed graph neural network (GNN) model for rapid retrieval of key physical properties. A set of LLM-driven AI agents collaborate to automate the exploration of the vast design space of MPEAs, guided by predictions from the GNN. We focus on the NbMoTa family of body-centered cubic (bcc) alloys, modeled using an ML-based interatomic potential, and target two key properties: the Peierls barrier and solute/screw dislocation interaction energy. Our GNN model accurately predicts these atomic-scale properties, providing a faster alternative to costly brute-force calculations and reducing the computational burden on multi-agent systems for physics retrieval. This AI system revolutionizes materials discovery by reducing reliance on human expertise and overcoming the limitations of direct all-atom simulations. By synergizing the predictive power of GNNs with the dynamic collaboration of LLM-based agents, the system autonomously navigates vast alloy design spaces, identifying trends in atomic-scale material properties and predicting macro-scale mechanical strength, as demonstrated by several computational experiments. This approach accelerates the discovery of advanced alloys and holds promise for broader applications in other complex systems, marking a significant step forward in automated materials design.

Motivation & Objective

• To overcome the computational bottleneck of brute-force atomistic simulations in designing multi-principal element alloys (MPEAs).
• To integrate physics-based insights from atomistic simulations with multimodal data and external knowledge for more accurate and efficient alloy property prediction.
• To automate the exploration of vast compositional spaces in complex alloys like NbMoTa BCC systems by reducing dependence on human expertise.
• To develop a scalable, autonomous AI framework that synergizes GNNs for property prediction with LLM-driven agents for dynamic reasoning and planning.
• To demonstrate the system's capability in predicting macro-scale mechanical strength from atomic-scale phenomena using minimal computational overhead.

Proposed method

• A multi-agent system is implemented using the AutoGen framework, with specialized agents for planning, reasoning, coding, reviewing, and multimodal data handling.
• A GNN model is trained on crystal structure graphs (atoms as nodes, bonds as edges) to predict Peierls barrier and solute-screw interaction energy using PNAConv, GRUCell, BatchNorm, and an MLP head.
• The GNN is trained on 90% of a curated dataset of MPEA compositions with 10% reserved for validation, using Adam optimization and a learning rate scheduler to minimize mean squared error.
• LLM-driven agents use GPT-4 to generate hypotheses, plan exploration strategies, and interpret results, dynamically collaborating to guide the search in alloy design space.
• The system leverages a machine-learned interatomic potential for accurate atomistic simulations and applies Maresca-Curtin screw strengthening theory to link atomic-scale predictions to macroscopic yield strength.
• All tools and functions are implemented as Python functions with defined names, descriptions, and inputs, enabling seamless integration within the agent-based workflow.

Experimental results

Research questions

• RQ1 Can a multi-agent AI system autonomously explore the vast compositional space of complex MPEAs like NbMoTa without human intervention?
• RQ2 How accurately can a GNN model predict critical atomic-scale properties—Peierls barrier and solute-screw interaction energy—compared to brute-force simulations?
• RQ3 To what extent can LLM-driven agents enhance the reasoning and planning capabilities in materials discovery by integrating physics-based knowledge and multimodal data?
• RQ4 Can the synergy between GNNs and LLMs reduce the computational burden of atomistic simulations while maintaining predictive accuracy?
• RQ5 How effectively can this system predict macro-scale mechanical strength from atomic-scale property predictions in BCC alloys?

Key findings

• The GNN model achieved high accuracy in predicting the Peierls barrier and solute-screw interaction energy, significantly reducing the need for expensive all-atom simulations.
• The system reduced computational cost by replacing brute-force simulations with fast, generalizable GNN predictions, enabling rapid screening of alloy compositions.
• The LLM-driven agents successfully coordinated to explore the design space, generating hypotheses and refining search strategies based on predicted properties.
• The integration of physics-based theories, such as the Maresca-Curtin screw strengthening model, enabled accurate prediction of macro-scale yield strength from atomic-scale inputs.
• The multi-agent system demonstrated autonomous, end-to-end operation in alloy design, with convergence observed in training curves and validated performance on unseen compositions.
• The framework is open-source, with data and code available on GitHub, enabling reproducibility and broader adoption in materials science research.