Soft Contrastive Learning for Unsupervised Discovery of Phases of Matter

Abstract

A machine learning (ML) framework for phase-diagram construction in condensed matter systems with unknown phase structure is presented. The method combines physically motivated descriptors with contrastive representation learning and a two-stage data-generation workflow. First, a sparse set of configurations is generated by Monte Carlo simulated annealing in order to train a contrastive neural network that produces a low-dimensional embedding. Consequently, representative phase configurations are used as initial conditions for gradient-based optimization on a dense parameter grid, enabling efficient generation of refined datasets for high-resolution phase-diagram construction. In order to optimize the workflow, a variant of the method, in which derived descriptors are used as soft labels in the contrastive objective, is examined as well. The framework is demonstrated on a PbZrO3-inspired model potential formulated in terms of coupled order-parameter fields and used as a physically motivated testbed for exploring configuration diversity and phase behavior. The proposed approach provides a physically informed route to ML-assisted phase-diagram construction that integrates data analysis with adaptive dataset refinement.