Machine Learning-Guided Discovery of Temperature-Induced Solid-Solid Phase Transitions in Inorganic Materials

TL;DR

This study introduces a machine learning framework that efficiently predicts temperature-induced solid-solid phase transitions in inorganic materials, aiding the discovery of materials for energy, thermal management, and information storage.

Contribution

The paper presents a novel uncertainty-aware machine learning approach combining DFT and neural networks to predict phase transitions at finite temperatures in inorganic crystals.

Findings

Identified over 2,000 potential phase transitions in 50,000 compounds between 300-600 K.

Discovered numerous transitions with large entropy changes suitable for cooling applications.

Found 21 compounds with significant changes in thermal conductivity across phase transitions.

Abstract

Predicting solid-solid phase transitions remains a long-standing challenge in materials science. Solid-solid transformations underpin a wide range of functional properties critical to energy conversion, information storage, and thermal management technologies. However, their prediction is computationally intensive due to the need to account for finite-temperature effects. Here, we present an uncertainty-aware machine-learning-guided framework for high-throughput prediction of temperature-induced polymorphic phase transitions in inorganic crystals. By combining density functional theory calculations with graph-based neural networks trained to estimate vibrational free energies, we screened a curated dataset of approximately 50,000 inorganic compounds and identified over 2,000 potential solid-solid transitions within the technologically relevant temperature interval 300-600 K. Among our key findings, we uncover numerous phase transitions exhibiting large entropy changes (> 300 J K$^{-1}$ kg$^{-1}$), many of which occur near room temperature hence offering strong potential for solid-state cooling applications. We also identify $21$ compounds that exhibit substantial relative changes in lattice thermal conductivity (20-70%) across a phase transition, highlighting them as promising thermal switching materials. Validation against experimental observations and first-principles calculations supports the robustness and predictive power of our approach. Overall, this work establishes a scalable route to discover functional phase-change materials under realistic thermal conditions, and lays the foundation for future high-throughput studies leveraging generative models and expanding open-access materials databases.