Chalcogenide Phase Change Materials (PCMs) have emerged as the cornerstone for advanced data storage and computing architectures, particularly Phase Change Random Access Memory (PCRAM) and neuro-inspired computing systems, due to their rapid and reversible switching between high-resistance amorphous and low-resistance crystalline states. However, the traditional trial-and-error research paradigm is increasingly constrained by long development cycles and high costs. Furthermore, conventional theoretical tools like Density Functional Theory (DFT) are limited by small spatial scales and short simulation times, making it difficult to capture complex phase transition dynamics. This review comprehensively elaborates on how the integration of big data analytics and Machine Learning (ML) has revolutionized PCMs research, transitioning it from empirical observation to precise prediction and directional design.

The significant research progress is detailed in three main aspects. High-throughput calculation serves as a data production engine for material property exploration. By automating First-Principles calculations, researchers can rapidly traverse vast compositional spaces. For instance, in the optimization of Sb₂Te₃, High-throughput calculation successfully screened Sc and Y as optimal dopants from transition metals. Calculations revealed that Sc reduces lattice mismatch stress, while Y significantly lowers thermal conductivity, thereby enhancing thermal stability and reducing power consumption. Additionally, high-throughput calculation has accelerated the discovery of novel metastable cubic phases in V-VI and IV-V-VI material systems, establishing a precise correlation between microscopic structure and macroscopic performance. Machine Learning Potentials (MLPs) have enabled large-scale molecular dynamics simulations. By training on high-precision DFT datasets, MLPs achieve ab initio accuracy with the computational efficiency of classical force fields. This breakthrough allows for simulations involving hundreds of thousands of atoms over nanosecond timescales, providing unprecedented insights into crystallization kinetics. Key findings include the observation of the breakdown of the Stokes-Einstein relation in supercooled liquid GeTe, where viscosity and diffusion decouple. Furthermore, ML-driven simulations have clarified the distinct crystallization mechanisms of different alloys—identifying Sb₂Te as growth-dominated and Sb₂Te₃ as nucleation-dominated. In terms of thermal transport, the combination of MLPs and Non-Equilibrium Molecular Dynamics (NEMD) has successfully quantified the thermal conductivity of amorphous phases and revealed how local structural distortions (e.g., reversible octahedral-heptahedral transformations in Sc-doped PCMs) suppress phonon transport.

ML facilitates device-level simulation and inverse design. At the atomic scale, ML models have simulated entire memory cells (exceeding 500 000 atoms), capturing critical phenomena such as element segregation and interfacial diffusion between electrodes and PCMs layers under thermal cycling. Beyond simulation, ML algorithms are applied to the inverse design of functional devices. In neuromorphic computing, ML optimizes synaptic weight updates by leveraging the resistance drift characteristics of PCMs. In photonics, Deep Learning models enable the rapid inverse design of metasurfaces, accurately predicting geometric parameters for desired optical responses (e.g., specific colors or transmission spectra), thereby overcoming the non-uniqueness problem in electromagnetic simulations.