High-throughput approaches for PFAS: Generating fast, reliable data for ML investigations

Machine learning (ML) techniques have demonstrated their ability to both interpolate and extrapolate useful information from datasets of known values. Unsurprisingly, this strength has led to an increasing trend in employing ML techniques for accelerating chemical discoveries. However, the availability of large quantities of property-specific data can be a significant challenge. Some experimental data may be costly or difficult to obtain, or, in the case of more novel compounds, may simply not exist. Computational methods are commonly employed to fill in such gaps. However, computed structural and property data can suffer similar detriments to availability at a large scale, and potentially more if differences in computational methodology are taken into account. This potential lack of availability presents problems for ML techniques which require a set of similar data on a large, diverse set of molecules for accurate interpretations and predictions. This presentation will demonstrate our framework methodology for rapidly generating optimized 3D coordinates de novo for a large number of fluorinated compounds and their subsequent use in building a database of computed physical and chemical properties.