Exploration via Organic Descriptors in Atomic Layer Processing (EXODIA)

Organic small molecules play a vital role in many atomic layer processes such as area selective atomic layer deposition (AS-ALD), molecular layer deposition (MLD), oxidative MLD (oMLD), and chemical vapor deposition. In these applications, knowledge of physical chemical properties such as vapor pressure (VP), oxidation potential (OP), p K a , melting point (MP), and boiling point (BP), aids in the efficient selection of molecular candidates to be tested. Here, we employ Chemprop, an established message passing graph neural network framework, to develop machine learning (ML) models that predict each of these five physical chemical properties for organic small molecules. Using these ML models, we generated a combined dataset of >43K organic small molecules that includes experimental values where available, and ML predicted values where necessary. The trained ML models achieve a mean absolute error on withheld test sets of 0.2 log10(Psat[torr]) over a range of 10 -4 to 10 3 torr for VP, 0.21V over a range of -1.1V to 2.6V vs. SHE for OP, 0.9 log units over a range of -2.8 to 16.13 for p K a , 22K over a range of 68K - 790.15K for MP, and 12K over a range of 169K - 721K for BP. These models were integrated into two distinct website tools that (1) allow for simple structure to property predictions and (2) allow for the down-selection among the combined dataset of 43K+ molecules based on these five properties. We outline use cases for each tool within the context of small molecule inhibitor selection and oMLD precursor discovery, with a potentially much larger reach within the vapor deposition community.