ArchOnML; an open-source python package for machine-learning assisted molecular design

The python package ArchOnML ("Archive-On-Machine-Learning") is introduced, which can perform virtual screening projects covering up to millions of structural derivatives through the use of Kernel Ridge Regression models. It supports the full workflow of setting up calculation inputs for external quantum chemistry program packages and post-processesing their outputs for training and predictions. An example screening project for over 1.3 million anthraquinone derivatives is presented, where excitation energies and oscillator strengths of the first two excited singlet and triplet states are predicted from descriptors based on semi-empirical quantum chemistry results. Compared to a non-ML calculation protocol, ArchOnML achieves a speed-up factor of over 400 with mean absolute errors for excitation energies of 0.1 eV. Due to ArchOnML's modular application programming interface, new descriptors, models and interfaces to other external quantum chemistry programs can be added in a straightforward way.