Learning Order: Can Neural Networks Discover Phase Transitions Without Symmetry Functions?

Phase transitions in soft matter systems — from crystallization to gelation — arise from collective particle rearrangements that are challenging to capture in full microscopic detail. Conventional approaches rely on physically inspired order parameters or symmetry functions to characterize emerging structures, but such descriptors may overlook crucial features in the often complex organisation of biolgical materials or synthetic super-structures. Here we investigate whether machine learning can uncover these hidden features directly from raw particle configurations. Using variational autoencoders trained on simulated trajectories of serveral soft matter systems, we show that the latent space encodes clear signatures of structural transitions without the need for handcrafted inputs. Our results suggest that neural networks can serve as unbiased tools to detect and interpret phase behavior in complex soft matter systems, revealing patterns that elude traditional symmetry-based analysis.