Co-Optimization of Composition in CrabNet
Traditional, data-driven materials discovery involves screening chemical systems with machine learning algorithms and selecting candidates that excel in a target property. The number of candidates to screen grows infinitely large as the number of included elements and the fractional resolution of compositions increases. The computational infeasibility and probability of overlooking a successful candidate grow likewise. Our approach shifts the optimization focus from model parameters to the fractions of each element in a composition. Using a pretrained network, CrabNet, and writing a custom loss function to govern a vector of element fractions, compositions can be optimized such that a predicted property is maximized or minimized withing a chemical system.
More information is available in the peer-reviewed CoCoCrab paper.
This project has been updated since its publication. The code now allows for a dopant threshold to be defined such that elements with fractions below the defined threshold are removed from consideration in property predictions. This allows CoCoCrab to better explore chemical spaces - dropping and adding elements to the composition as needed. The dopant threshold (among other optimization parameters) is defined within the optimize.py file.