Photoisomerization in a Glassy Matrix: Predicting a Broad Distribution of Dynamics with Machine Learning

The response of azo-containing molecules undergoing a trans → cis photoisomerization transition has been primarily studied through simulation and experiment in solution or vacuum. The response of these photoactive molecules in glassy solids, where barriers to motion are significantly higher, is poorly characterized. Results from molecular dynamics simulations show that the dynamics of photoactivated molecules in glassy solids depends critically on local density features. A characteristic power-law wait time for photoisomerization occurs in samples for densities that vary with photoactive molecule and glass-matrix material. Dynamic behavior is driven by difficult-to-identify local density features, which suggests an opportunity for a machine-learning approach. We apply methods from structural analysis of an ideal disordered solid [1] to an all-atom molecular system for the first time, predicting a propensity to isomerize as an analog to “softness.” Our results not only demonstrate that simple machine learning methods can be applied to complex, all-atom molecular systems, but also highlight predictive features of local environments far beyond simple scalar quantities or visually identified features.
[1] Schoenholz, Samuel S., et al., Nature Physics 12, 469 (2016)