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

TL;DR

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.

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.

Figures (20)

• Figure 1: (a) Overview of the workflow used in this work to train GNN models for an end-to-end prediction of the Peierls barrier and potential energy changes. The process begins with the generation of initial and final dislocation structures in random BCC multi-component alloys, which are then minimized to compute the initial and final potential energies. Next, NEB simulations are performed to determine the minimum energy path and calculate the Peierls barrier. The input to the machine learning model is the graph representation of the alloys, with nodes encoding spatial and chemical information, and edges representing bond types. Two supervised GNN models are trained to predict graph-level labels: the Peierls barrier and potential energy change. (b) GNN architecture. The graph input is first passed into an input block to enlarge the dimension of node features. The node embedding, edge features and connectivity of input graphs are then input to the message passing block where the information from the neighbors of each node in the graph is aggregated to update the hidden features of the nodes. The output of the message passing block is then input into a global pooling later which outputs a graph-level embedding by adding node embeddings across the node, and it is connected to a multilayer perceptron (MLP) that returns a predicted graph label, i.e. Peierls barrier or potential energy change.
• Figure 2: Evaluation of the GNN model for Peierls barrier prediction on the test set. Test set contains new compositions that never appeared in the training set. (a) Comparison of machine learning (ML) predictions and NEB results for the Peierls energy barrier. (b) Comparison of the ML results and NEB results for the mean energy barriers of binary and ternary compositions in the test set. The correlation coefficient $R^2$ between the predictions and the ground truth is 0.97.
• Figure 3: Evaluation of the GNN model for the potential energy change prediction on the test set. Test set contains new compositions that never appeared in the training set. (a) Comparison of ML predictions and NEB results for the potential energy change. (b) Comparison of the ML results and NEB results for the solute/screw interaction energy parameter, Eq. \ref{['eg:dep']} of binary and ternary compositions in the test set. The correlation coefficient $R^2$ between the predictions and the ground truth is 0.89.
• Figure 4: Overview of the GNN-powered, LLM-based multi-agent system developed here. The system consists of a human and an AI assistant agent at its core, where the human poses queries, and the AI assistant provides responses, seamlessly steering the problem-solving process with the help of integrated tools. These tools are responsible for various tasks, including planning, coding, and multi-modal analysis, and each incorporates a set of AI agents that dynamically collaborate to solve complex tasks. A key component is the physics tool, which includes newly developed graph neural network (GNN) models to retrieve essential physical parameters (such as the Peierls barrier and potential energy change) as well as physics-based theories (like solute-strengthening theory). The GNN models enable the rapid prediction of fundamental material properties, bypassing the need for costly atomistic simulations. The iterative collaboration between agents within the tools and the seamless interaction between the human and AI assistant allows for efficient resolution of complex materials design challenges.
• Figure 5: Overview of a typical workflow executed by our multi-agent system for accelerated and automated multi-component alloy design and analysis. Upon receiving a query from the user, the process begins with the assistant agent, which calls the planning tool to generate a detailed plan. The assistant agent follows this plan by invoking the relevant tools and providing the necessary input functions. These tools are integrated into our system to equip the agent with advanced capabilities, including material property predictions through physics-based retrieval tools (GNN models), data visualization via code execution (through a code-executor pair of agents), and multi-modal analysis (through a multi-modal agent-user pair) for enhanced visualization and reasoning. The system incorporates a human-in-the-loop design, fostering dynamic collaboration between humans and AI. This enables continuous refinement of ideas and queries, resulting in a more adaptive and efficient problem-solving process.