Machine learning sorts a simulated polymer into ordered and disordered beads

The work

Technical · from the record

Coarse-grained molecular dynamics simulations of 10^6 monomers (1000 chains of 1000 beads) of a poly(vinyl alcohol) model were cooled from reduced temperature 0.9 to 0.75, and each monomer was described by a 17-component fingerprint combining the alignment of its surrounding bond vectors in concentric shells with the conformation of its local chain contour. An autoencoder compressed these fingerprints to a two-dimensional latent space and a two-component Gaussian mixture model separated the latent points into ordered and disordered groups, while hierarchical clustering was applied directly to the uncompressed fingerprints; a multilayer perceptron trained on the clustering labels transferred the decision boundary to all monomers at all time steps. The resulting labels overlapped substantially with two manually thresholded classifiers based on stem length and on average local orientation, with the stem-length threshold raised to 17 consecutive monomers to align the crystalline fractions after the transition. Tracking individual monomers between the crystalline and amorphous classes, the authors report a local maximum in the fraction of crystalline monomers conserved until the end of the run, and a drop in the fraction of newly crystalline monomers that return to the amorphous phase, both before the transition appears in the specific volume.

How AI was used

Machine learning supplied the per-monomer crystalline/amorphous classification used throughout the analysis. For every monomer in the simulation trajectory, a 17-dimensional structural fingerprint was constructed by hand from shell-resolved squared cosine alignments of neighbouring bond vectors within a 3.5 sigma radius and from the pairwise squared cosines among five consecutive bond vectors along the contour. Fingerprints of roughly 10^4 monomers from one slice of the simulation box at the final, lowest-temperature step formed the training set, with inputs remapped through the second Legendre polynomial. A self-supervised autoencoder was trained on this set with a mean-squared reconstruction loss, a learning rate reduced by half after 25 epochs without validation-loss improvement, compressing each fingerprint to two bottleneck neurons. A Gaussian mixture model with full covariance and two components was fitted to the latent vectors to produce AE labels, and agglomerative hierarchical clustering with Ward linkage was applied to the uncompressed fingerprints to produce SF labels. To extend the hierarchical-clustering boundary to configurations at other temperatures, a multilayer perceptron with 17 input neurons was trained on the fingerprint matrix against those labels, with its sigmoid output thresholded at one half, and the mixture model's predict function was used for the latent vectors of all other time steps. Principal component analysis of the conformation fingerprint was used as a one-dimensional order parameter for interpreting the latent map. Two non-learned classifiers, a stem-length criterion and an average-orientation criterion with manually placed thresholds, were computed on the same trajectory for comparison, and identified crystal domains were checked by projecting them perpendicular to the dominant eigenvector of their orientation tensor and Fourier transforming the projection.