Data Driven Acceleration of Coupled-Cluster Calculations using Machine Learning, Multitask Learning and Physics Imposed Learning

Data-driven coupled-cluster singles and doubles (DDCCSD) method developed by Townsend and Vogiatzis aims at predicting the coupled-cluster t₂ amplitudes using MP2-level electronic structure data with machine learning. In this work we address limitations of the DDCCSD method to expand the applicability and increase the accuracy. First, we implement localized molecular orbitals (LMO) to the DDCCSD method. There is a ten-fold increase in accuracy when the LMO implementation is used compared to the canonical molecular orbital implementation. Next, we introduced five data selection techniques to select data for testing and training. Here we were able to achieve accuracies less than the all-amplitude model using only 10-15% of total datapoints. Then developed three versions of neural networks that introduce multitask learning to DDCC. With these models, we were able to predict both t1 and t2 amplitudes simultaneously. This coupled with physics-imposed learning allowed us to achieve greater accuracies than the conventional DDCCSD methods and due to the architecture of the models we were able to expand the accuracy towards larger chemical systems. Finally we show that DDCC can be expanded further by introducing DDCCSD perturbative (DDCCSD(T)) triples method and DDCCSDT exact triples method.