Benchmarking general neural network potential ANI‐2x on aerosol nucleation molecular clusters

Abstract

AbstractNew particle formation including atmospheric aerosol nucleation and subsequent growth, contributes to about half of cloud condensation nuclei and can grow to form haze under certain conditions, so its role in climate change and air quality is indispensable. However, various kinds of nucleation precursors create vast combinations of molecular clustering, hindering the understanding the detailed picture of nucleation mechanism. The recently appeared general neural network potential, ANI‐2x, covering most of elements composing nucleation clusters, looks promising to be embedded into the nucleation theoretical workflows to improve the nucleation simulation accuracy. Here we benchmarked ANI‐2x's performance on both low and high energy isomers based workflows through comparing it with semi‐empirical (PM7) and DFT (ωB97XD/6‐31++G(d,p)) methods. Results show that ANI‐2x is superior to PM7 in single point energy and geometry optimization calculations when compared against DFT. However, generally ANI‐2x's accuracy on high energy isomers is still far less than that on low energy isomers. Besides, force comparison indicates that PM7's accuracy is better than that of ANI‐2x. After all, accuracy of the workflow that focuses on low energy isomers can be benefitted from ANI‐2x while for high energy isomers based workflow, PM7 is still the better choice than ANI‐2x due to the more important role of force labels than that of energy labels in preparing machine learning dataset. However, due to the linear relation between the ANI‐2x force and DFT force, a scale factor of approximately 0.50 is expected to greatly improve the ANI‐2x force performance.