PyITA: A Taylor Expansion-Based Data Augmentation Program for ANN Potentials Applied to TiO2
by Zhengbo Xiang
Abstract – The efficiency of Artificial Neural Network (ANN) potentials enables the modeling of materials at scales that are too computationally expensive for conventional first-principles approaches. However, the force and energy-prediction accuracy of ANNs are generally limited by the availability of training data and training hours. The enlistment of more efficient training methods can partially mitigate this limitation. In this paper, I demonstrate the capabilities of my new Python program, PyITA, which executes a recently demonstrated Taylor expansion-based data augmentation technique11 . Using my program, I was able to evaluate the powerful methodology by constructing and comparing ANN potentials for the chemical species titania (TiO2). I compared the error distributions of the augmented potentials with that of the non-augmented potentials. Potentials were trained on both a large (7815 structures) and a small (500 structures) dataset. I ultimately found insufficient evidence to confirm that the data augmentation method is effective for increasing the accuracy of either force predictions or energy predictions on either dataset.
Video not yet available.