The highly layered architectures featured in deep learning have paved the way for remarkable recent advances in a wide variety of fields. At the same time, the accuracy of these deep learning models often comes at a cost: it is typically difficult or impossible to determine how a given model works. Although this may not be a problem for some applications, there is a need for machine learning-driven approaches to molecular simulation whereby ML models can not only be called upon to make predictions, but also be interrogated to uncover chemical insights pertaining to why the model is able to make the predictions it makes. For instance, if an ML model, trained on phase space data from chemical simulations, is able to consistently predict the reaction product of a multi-reaction pathway chemical system long before the reaction occurs, then the model must necessarily have learned an approximation to the isocommittor surfaces that separate trajectories bound by distinct product states.  However, obtaining chemical insight from this solution depends on being able to extract information on these surfaces from its often highly indirect representation within the ML model and summarize it in humanly accessible terms.  Kernel learning systems such as support vector machines (SVMs) provide an excellent framework for developing models that not only learn latent predictive patterns in chemical data, and use them to make predictions, but also inform the user as to how the prediction was made. The reason SVMs and related methods excel in this capacity is that the source of the flexibility in their decision boundaries is the expression of the nonlinear target function in terms of a set of known basis functions defined by the choice of kernel and training data, making them susceptible to subsequent analysis.  In particular, every prediction regarding e.g. a transition state between products can be expressed in terms of linear combinations of knowable functions of input features such as atomic coordinates and momenta, facilitating approximation in chemically intelligible terms. This presentation will include a theoretical description of our approach as well as a demonstration whereby the methodology is applied to ab-initio dynamics of the photodissociation of acetaldehyde.

ML models that both learn and teach chemistry via partitioning reactive trajectories by reaction product in phase space using support vector machines

2019 Fall National Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.