Leveraging genetic algorithms to maximise the predictive capabilities of the SOAP descriptor

In this collaboration with AstraZeneca, James Darby in Cambridge, and Warwick’s Albert Bartók-Partay, we explore the capabilities of genetic algorithms (GAs) when used to optimise the hyperparameters of the popular SOAP descriptor. SOAPs are a commonly used method to encode atomic local environments, with good performance over a wide range of materials. We show that by using our genetic algorithm package (available at https://github.com/gcsosso/SOAP_GAS) the predictive capabilities of SOAPs are improved for a variety of datasets. There were notable improvements when predicting the solubility of molecules in a prototypical molecular database for drug design, and when predicting the polarizability of the QM7b dataset. We also showcase the improvements on both optimisation speed and quality of our GA when compared to the more commonly used random grid search. Have a look at the paper online or grab a pdf here.