Discovering Low-Viscosity Molecules Using an Integrated Physics-Based Modeling, High-Throughput Screening, and Active Learning Approach (2)— Screening from PubChem Database

Molecules exhibiting lower viscosities than conventional organic solvents are in high demand for applications in electrochemical devices such as lithium-ion batteries and various capacitors. These molecules improve the electric resistance and efficiency of the devices, particularly at low temperatures. To identify low viscous molecules, we have performed screening of 290,000 molecules from the GDB database by applying machine learning (ML) techniques combined with an active learning approach. We identified more than 100 molecules with viscosities less than 0.35 cP. The details of this result will be presented at the part (1) of our talks. As the molecules in the GDB database are limited to those containing C, N, O, S and halogen atoms, we extended our research to the screening of low viscous molecules from the PubChem database [1], which includes molecules that may have atoms other than C, N, O, S and halogens. We built an ML model for viscosity using molecular dynamics (MD) calculated viscosities of 20,465 molecules as a training data set. The resulting ML model was then applied to screen 962,695 molecules from the PubChem database. An active learning approach was applied to the screening where MD calculations were performed on ML-model suggested candidate molecules to iteratively improve the ML model. The screening process resulted in the identification of over 400 molecules with MD-calculated viscosities less than 0.15 cP.

[1] Kim S, Chen J, Cheng T, et al. PubChem 2023 update. Nucleic Acids Res. 2023;51(D1):D1373–D1380.