Specifically, we show that non-linear dimensionality reduction of atom-centered symmetry functions yields distinct clusters of amorphous ice configurations in the resulting latent space. We show that applying kernel density estimation to these clusters allows us to accurately classify atomic environments as being HDA, LDA, or neither. We also use this latent space to determine physically interpretable collective variables associated with transitions between HDA and LDA. The resulting model is used to quantify thermodynamic properties (i.e., transition temperatures and pressures) associated with the transition.
We show that this method can be used to analyze any set of structurally distinct configurations beyond HDA/LDA ice by considering two additional material systems. First, we use local atomic environments to characterize differences between multiple coarse-grained potentials for small molecules that yield similar macroscopic properties. Second, we determine microscopic mechanisms for (anti-) plasticization behavior in simulations of polymers by characterizing the local environments of water molecules embedded in these polymer materials.
Overall, this work shows how applying machine learning to ensembles of local atomic environments from MD simulations can reveal interpretable physical insights into materials behavior.