Brown University engineering students attended the 2019 American Institute of Chemical Engineers (AIChE) Annual Student Conference in Orlando, with senior Alex Ng '20 taking first place in his division. Ng (chemical engineering) presented in the category of Catalysis and Reaction Engineering VI, and he is advised by associate professor Andrew Peterson.
Ng's project was titled, "Learning on the Job: An Active Learning Calculator for Atomistic Simulations." Neural network potentials have been used to accurately predict the energy and forces on atomic structures, acting as suitable substitutes for traditional electronic structure calculators. However, these predictions often diverge from true values when the neural network encounters images that deviate considerably from those found in the training set. In these cases, it would be desirable to include this image in the training set and re-train the network so it can better understand this region of the potential energy surface, but the manual selection of these images can be tedious and impractical. Ng presented a method whereby a bootstrap ensemble of neural network calculators is used as a hands-free calculator that can identify poorly understood regions on the potential energy surface and retrain the calculator as necessary: for every image, an assessment of prediction certainty is made by evaluating the ensemble width, used as a measure for uncertainty. For images where the ensemble is uncertain, we call the electronic structure calculator and retrain the ensemble calculator on that image. A scheme for accelerated retraining of the neural network is also presented within. We apply this calculator to a 400-image sample of a single water molecule where one of the bond lengths and the bond angle is varied. We show the ability for the ensemble calculator to train itself when it is uncertain about an image, and display how retraining on just a few images allows the calculator to understand the region much better. This active learning calculator should allow for the hands-free training of accurate neural networks to perform electronic structure calculations through sample regions not covered by the training data, considerably reducing computation time and cost for future calculations that retread previously investigated sections of the potential energy surface.
This year's Brown AIChE students that attended the national meeting included: Eric Kwon '20, Ng, and Ryan Bain '21.