Representation-Dependent Machine Learning of the Isotropic-Nematic Transition in the Lebwohl-Lasher Model

Abstract

Machine-learning detection of phase transitions depends not only on the learning algorithm, but also on whether the input representation preserves the symmetries of the system. We examine this for the weak first-order isotropic--nematic transition of the three-dimensional Lebwohl--Lasher model, whose apolar and continuously degenerate nematic phase makes raw molecular configurations a challenging input for unsupervised learning. Principal component analysis (PCA) and a convolutional autoencoder (CAE) fail to identify the transition from raw configurations because rotationally equivalent nematic states can appear far apart in the input space. When the same configurations are transformed into a rotationally invariant local-correlation representation, both methods recover transition-sensitive signatures and bimodal coexistence distributions. A supervised three-dimensional convolutional neural network (CNN), by contrast, accurately predicts the scalar order parameter from raw configurations when given order-parameter labels. The Lebwohl-Lasher model therefore separates unsupervised phase discovery from supervised order-parameter regression and shows that symmetry-respecting input representations are needed for unsupervised machine learning in orientationally ordered systems.