A Forecasting System of Computational Time of DFT/TDDFT Calculations under the Multiverse ansatz via Machine Learning and Cheminformatics

Abstract

A forecasting system for predicting computational time of density-functional theory (DFT) calculation is presented. The forecasting system is established under the many-worlds interpretation of multiverse ansatz, in which the molecules and Kohn-Sham equation are the trunk and every calculating parameters (e.g. basis set, functional) are branch points that generate result's worlds. Every world is constituted by the solved wave functions and the accompanying data (e.g. computational time) after solving. Several machine-learning models, including random forest, long short-term memory, message passing neural network and multilevel graph convolutional network models, are employed for the prediction of computational time of any molecule belonging to a given world. For the molecules that belong to a world without pre-trained models, additional efforts are used for linear combination of models from adjacent world in order to give reasonable predictions. Benchmark results show that the forecasting system can predict proposed times with mean relative error normally around 20% when comparing to these of practical calculations.