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

1 Introduction

• MPEAs offer tailored-property potential across vast composition spaces, while refractory BCC alloys have drawn interest for retaining high-temperature strength beyond current superalloys.MPEAs comprise three or more elements and can exhibit strength, fracture toughness, ductility, and resistance to hydrogen embrittlement beyond pure or dilute counterparts.Source: §1 · p. 2
• BCC plasticity depends on screw-dislocation motion, whose theories require costly atomistic calculations of Peierls barriers and solute-screw interaction energies across random solute environments.The large multicomponent design space and need for many realizations to obtain statistically accurate averages make these calculations particularly challenging.Source: §1 · p. 2
• Because property-specific ML can miss interdisciplinary knowledge, multi-agent systems combine multimodal data and external physics knowledge through LLM-powered agents with distinct roles.The agents dynamically collaborate and divide different aspects of alloy-design problems.Source: §1 · p. 2
• The system combines planning and reasoning LLMs, specialized collaborating agents, and external tools, including a GNN that predicts atomic-level alloy properties.It builds on AtomAgents, whose atomistic simulations became prohibitively costly as alloy-system complexity increased.Source: §1 · p. 2
• Trained on a relatively small portion of ternary Nb-Mo-Ta BCC composition space, the GNN accurately predicts Peierls barriers and solute/screw interaction energies for rapid physics retrieval.The paper presents examples of using the multi-agent approach to explore alloy design and predict macroscopic yield stress.Source: §1 · p. 3

2.1.1 Peierls barrier

• 37 meV MAE for Peierls-barrier predictions and R2 = 0.97 show accurate performance on test compositions unseen during training.Predicted mean barriers also agreed with ground truth across binary and ternary alloys.Source: §2.1.1 · pp. 4–5, Figure 2 · p. 5

2.1.2 Potential energy change predictions

• 60 meV MAE for potential-energy-change predictions and R2 = 0.89 accompany good agreement in solute/screw interaction energy for binary and ternary alloys.The predicted interaction parameter provides an input to solute-strengthening theories.Source: §2.1.2 · pp. 5–6, Figure 3 · p. 6

2.2 Multi-modal multi-agent model

• A human-in-the-loop assistant coordinates specialized agents and physics tools; planner and reviewer agents iteratively refine plans before tools predict properties, organize data, and analyze results.The framework builds on AtomAgents and integrates GNN predictions with physics-based theories such as solute-strengthening theory.Source: §2.2 · pp. 6–7
• The system computed Peierls barriers for 231 Nb–Mo–Ta compositions at 5% intervals within seconds per composition, then mapped trends across the design space.High barriers clustered around intermediate Nb and Mo with low Ta; selected-system plots showed nonlinear variation, including a peak near 50% Nb in Nb–Mo and rising trends in both ternary systems as x increased to 0.5.Source: §2.2 · pp. 8–9, Figure 6 · p. 10
• The agents’ ternary-plot analysis was imperfect: it incorrectly inferred that increasing Nb decreases the Peierls barrier, motivating further refinement.The reported trends instead show increasing barriers with Nb in the selected Nb–Mo and Nb–Ta systems, with a peak near 50% Nb in Nb–Mo.Source: §2.2 · pp. 8–9
• The predicted interaction-energy parameter varied nonlinearly: it peaked near 30% Nb in Nb–Mo, was higher at high Nb in Nb–Ta, and rose with x in both ternary systems.The (NbMo)2xTa1−2x trend eventually stabilized, while (NbTa)2xMo1−2x decreased as Mo became minimal.Source: §2.2 · p. 9, Figure 6 · p. 10
• The system completed yield-stress prediction tasks across temperatures for binary and ternary BCC alloys, using GNN-derived Peierls and interaction parameters for the Maresca-Curtin theory.Other inputs, including lattice constant, kink formation energy, and vacancy/interstitial formation energies, were averaged over pure elements; binary predictions were plotted alongside experimental data.Source: §2.2 · pp. 9, 11, Figure 7 · p. 11
• The agent identifies lattice distortions impeding dislocation motion as a key solid-solution strengthening mechanism.Source: §2.2 · p. 12
• The multi-agent collaboration combines AI-driven physics predictions, established theoretical frameworks, and plotting and visualization reasoning for materials design.The paper presents this multifaceted integration as a pathway that could transform how materials design is approached and executed.Source: §2.2 · p. 12
• Alloy strength typically decreases with temperature; Maresca-Curtin predictions align for Nb81Mo19 but diverge for Nb95Mo5 and Nb75Mo25.The agent suggests composition, microstructure, temperature effects, and experimental variability may explain discrepancies, and proposes incorporating more detailed microstructural and compositional factors. Its analysis covers binary and ternary alloys, with the yield-stress-versus-temperature summary shown in Figure 8.Source: §2.2 · pp. 11–12, Figure 8 · p. 12

Summary and future perspective

• The GNN predicts screw-dislocation energy changes, Peierls barriers, and solute/screw interaction energies accurately in seconds, enabling rapid alloy-space exploration with minimal accuracy loss versus costly atomistic simulations.The method could extend to high-entropy alloys, FCC and HCP structures, and edge dislocations, supporting searches for alloys with higher strength.Source: §3 · pp. 12–13
• The multimodal system combines LLMs, specialized collaborating agents, and a GNN, replacing AtomAgents’ direct-simulation physics retrieval to explore ternary properties and predict BCC yield stress validated against experimental data.Demonstrations cover Peierls barriers and solute/screw interaction energies across the ternary composition space.Source: §3 · p. 13
• The system supports multimodal model and tool integration, iterative error correction through agent feedback, and human-in-the-loop refinement of requests and model parameters.The described system uses LLMs including o1-preview and o1-mini.Source: §3 · p. 13
• Future potential spans other disciplines through evolving or task-customized LLMs, domain-specific model integration, and links with SciAgents for hypothesis generation followed by testing and refinement.The authors anticipate that advances in LLMs could improve system performance.Source: §3 · p. 13

Dataset generation

• Random ternary NbMoTa compositions are assigned to pristine rectangular cells using the corresponding lattice constant and desired concentrations.Source: Dataset generation · p. 14
• Cells are oriented with glide direction x||[1,1,¯2], glide-plane normal y||[¯110], and dislocation line z||[111], with periodic boundaries along x and z and a free surface along y.Source: Dataset generation · p. 14
• At zero temperature, structures are relaxed with FIRE and cell-dimension relaxation until forces fall below 10e-6 eV/Å and σxx, σxz, and σzz fall below 0.1 MPa; simulations use LAMMPS and an MTP potential [34].Source: Dataset generation · p. 14
• A screw dislocation is introduced at the center of each relaxed cell using the PAD method [16], followed by energy minimization and pressure relaxation to form the initial dislocated configuration.Source: Dataset generation · p. 14
• The final configuration uses the same pristine random structure with the dislocation shifted to the adjacent Peierls valley by distance a; the potential-energy change is final minus initial total energy.Source: Dataset generation · p. 14
• NEB computations [60] in LAMMPS find the minimum-energy path between configurations, and its maximum value is stored as the Peierls barrier.Source: Dataset generation · p. 14
• The procedure is repeated for 200 realizations per composition in the training set and 50 per composition in the test set, as depicted in Figure S1.Source: Dataset generation · p. 14

Graph representation

• The graph contains 705 Mo-atom nodes and 2610 edges within a 16 Å-radius cylinder centered on the screw-dislocation core, connecting atoms less than 2.8 Å apart.The 2.8 Å cutoff is the nearest-neighbor distance in perfect molybdenum.Source: Graph representation · p. 14
• Node features encode solute type as one-hot Nb, Mo, or Ta values and use the pure-Mo displacement z-component to represent the screw-dislocation defect.The defect feature is identical across configurations, so new configurations do not require atomic relaxation during inference.Source: Graph representation · p. 14
• Edge features encode bond type as one-hot representations based on the neighboring nodes.The edge-feature representation is summarized in Table 1.Source: Graph representation · p. 14

Graph neural networks (GNNs)

• The model uses an input block with PNAConv, GRUCell, and BatchNorm to increase node-feature dimensions to 50.It is implemented in PyTorch with PyTorch Geometric.Source: Graph neural networks (GNNs) · p. 15
• Ten message-passing repetitions let nodes exchange connectivity-based messages and update their features using received messages.Source: Graph neural networks (GNNs) · p. 15
• A ReLU-activated MLP with layer sizes 30, 20, 10, and 1 predicts the Peierls barrier or potential energy change.All layers use PyTorch Geometric’s default weight and bias initialization.Source: Graph neural networks (GNNs) · p. 15

Model training and evaluation

• Each dataset was split into 90% training data and 10% validation data.The compositions are shown in Figure S1.Source: Model training and evaluation · p. 15
• Both models were trained for 250 epochs with batch size 32 and Adam, while ReduceLROnPlateau halved the learning rate after 10 epochs without improvement in validation MSE.Training started at 0.0005 on one NVIDIA Tesla V100 with 32GB memory; learning curves in Fig. S4 indicate convergence.Source: Model training and evaluation · p. 15

Solute-strengthening theory in BCC alloys

• Maresca–Curtin theory attributes screw-dislocation strengthening to Peierls-like resistance, kink glide, and cross-kink formation and unpinning, assuming spontaneous kinking at zero load and temperature.The initially straight dislocation becomes kinked to lower its total energy.Source: Solute-strengthening theory in BCC… · p. 15
• The alloy strength is τ(˙ϵ, T) = τxk(˙ϵ, T) + min[τk(˙ϵ, T), τp(˙ϵ, T)], with terms denoting cross-kink unpinning, kink migration, and Peierls strengths, respectively.˙ϵ is the experimental strain rate.Source: Solute-strengthening theory in BCC… · p. 15

• The system combines GPT-4 with dynamic multi-agent collaboration to design AI agents.Source: Agent design · p. 15
• The system is implemented in the open-source AutoGen framework [67].Source: Agent design · p. 15
• UserProxyAgent creates the human user agent, AssistantAgent creates the Assistant, Planner, Reviewer, and coder agents, and MultimodalConversableAgent creates the multimodal agent.Each agent receives a role through a profile description included as its system_message at creation; profiles are shown in Figures S8–S12 in the Supplementary Material.Source: Agent design · p. 15

Function and tool design

• Each tool is a Python function characterized by a name, description, and input properties.The complete tool list and descriptions are available in the corresponding code.Source: Function and tool design · p. 15

Data and code availability

• All data and code are available on GitHub or from the corresponding author upon reasonable request.The repository is https://github.com/lamm-mit/AlloyAgents.Source: Data and code availability · p. 15

Author contributions

• M.J.B. and A.G. conceived the project and developed the GNN and multi-agent system; A.G. curated data, ran tests, analyzed results, and drafted the paper, while M.J.B. supported analysis and finalized it with A.G.Source: Data and code availability · pp. 15–16

• Additional materials are provided as Supplementary Materials.Source: Supplementary Materials · p. 16