aLLoyM: “Phase Diagram Prediction System”

Abstract:  Phase diagrams are essential thermodynamic tools that describe equilibrium phase stability as functions of temperature and composition in alloy systems. They guide alloy design, heat-treatment optimization, and microstructural control in critical industries. However, experimental determination of phase diagrams is costly and time-intensive. Although computational approaches such as CALPHAD (Calculation of Phase Diagrams) provide reliable thermodynamic modelling through Gibbs free energy minimization, they depend on curated parameter databases and expert assessment, limiting rapid exploration of new material systems.

In this work, we introduce aLLoyM, a domain-adapted Large Language Model (LLM) developed for structured alloy phase diagram prediction. Thermodynamic data generated from CALPHAD assessments in the Computational Phase Diagram Database (CPDDB), covering 389 binary and 38 ternary systems, were systematically sampled to produce over 800,000 equilibrium data points. These data were transformed into multi-task Question–Answer (Q&A) pairs and used to finetune the Mistral-Nemo-Instruct model via Low-Rank Adaptation (LoRA), enabling efficient domain specialization.

The framework supports three thermodynamic reasoning tasks: full phase information prediction, phase name inference, and inverse experimental condition prediction. Performance was evaluated under both interpolation and extrapolation settings to assess generalization. Results show substantial improvement over baseline LLM performance and demonstrate the model’s ability to infer plausible phase behaviour for previously unseen systems.

These findings highlight the potential of integrating

Large Language Models with computational thermodynamics to develop scalable AI-assisted tools for accelerating alloy design and materials discovery.