18.11.2025 BOKU-Met Seminar
Phase transitions, such as the crystallization of a supercooled liquid, involve complex collective rearrangements that also occur in natural systems—for instance, during the freezing of cloud droplets or the formation of ice in aerosols. In all these cases, the full system information lies in a high-dimensional space, and understanding the emergence of order requires dimensionality reduction and pattern recognition. Traditionally, this has relied on physically motivated descriptors, such as symmetry functions or order parameters, to identify structural transitions. Yet such predefined measures can overlook subtle or unexpected features, especially in complex or out-of-equilibrium systems. Machine learning, particularly methods based on data compression such as autoencoders, offers a new route to discover hidden structures directly from raw data. In this work, we explore whether the signatures of phase transitions can be captured without relying on symmetry-based inputs, using particle-resolved data as a model case—illustrating how information-theoretic approaches to pattern recognition may generalize across physical and environmental systems.