Data efficient assimilation of multi fidelity information

Abstract

Using deep learning for accelerated material discovery applications has been widely explored by many researches. Generating high fidelity data for material science applications is generally expensive but deep learning methods require a large amount of training data for accurate property predictions. Therefore, material scientists sometimes resort to low cost inaccurate models for structure property prediction. However, it is necessary for some material science problems to predict properties at highest level of accuracy. In this work, we present a proof of concept of a multi fidelity neural network which leverages the low and high fidelity data to predict properties at highest fidelity level. Deep neural nets are necessary to model the non linear cross correlation between high and low fidelity data corresponding to micro structure images at high dimensional space. The use of this method is demonstrated in Organic Solar Cells to predict high fidelity multiple properties of interest.