Machine Learning for Predicting Chemical Potentials of Multifunctional Organic Compounds in Atmospherically Relevant Solutions

Abstract

We have trained the Extreme Minimum Learning Machine (EMLM) machine learning model to predict chemical potentials of individual conformers of multifunctional organic compounds containing carbon, hydrogen, and oxygen. The model is able to predict chemical potentials of molecules that are in the size range of the training data with a root-mean-square error (RMSE) of 0.5 kcal/mol. There is also a linear correlation between calculated and predicted chemical potentials of molecules that are larger than those included in the training set. Finding the lowest chemical potential conformers is useful in condensed phase thermodynamic property calculations, in order to reduce the number of computationally demanding density functional theory calculations.

Atmospheric Sulfuric Acid–Multi-Base New Particle Formation Revealed through Quantum Chemistry Enhanced by Machine Learning

“…The quantum machine learning (QML) program employs ML to model molecular quantum chemistry (QC) properties (e.g., electronic energies). Several studies ,, have shown that kernel ridge regression (KRR) can accurately model chemical properties, even when trained on relatively small training databases. In our previous study, we used QML and KRR to predict electronic binding energies (calculated at the ωB97X-D/6-31++G(d,p) level of theory; further referred to as DFT) for the SA–water system, where the binding energy is defined as follows. normalΔ E el = E el cluster − prefix∑ E el monomer We showed that, due to correlation of binding energies between DFT and some low level of theory (in our case semiempirical GFN1-xTB, , further abbreviated as XTB), the low level of theory can be used to improve the accuracy of the ML model.…”

“…The quantum machine learning (QML) program 45 employs ML to model molecular quantum chemistry (QC) properties (e.g., electronic energies). Several studies 42,46,47 have shown that kernel ridge regression (KRR) can accurately model chemical properties, even when trained on relatively small training databases. In our previous study, 42 we used QML and KRR to predict electronic binding energies (calculated at the ωB97X-D/6-31++G(d,p) level of theory; 48 further referred to as DFT) for the SA−water system, where the binding energy is defined as follows.…”

Predicting liquid–liquid phase separation in ternary organic–organic–water mixtures

“…30 In solvent water, the number of intramolecular H-bonds correlates with the pseudo-chemical potential. 37 However, the calculated mixtures contain non-polar organic compounds and the most favorable conformers may not be the same in polar and non-polar solutions. Here, the conformers for COSMO therm calculations were therefore selected using their pseudo-chemical potentials, instead of their intramolecular H-bonding.…”

“…The pseudo-chemical potentials were predicted using a machine learning model. 37,41 200 lowest pseudo-chemical potential conformers in each of the two solvents (at most 400 conformers) were selected for density functional theory calculations using the COSMO conf program. 42,43…”

“…The pseudo-chemical potentials were predicted using a machine learning model. 37,41 200 lowest pseudo-chemical potential conformers in each of the two solvents (at most 400 conformers) were selected for density functional theory calculations using the COSMOconf program. 42,43 The COSMOconf calculations included single-point calculations at the BP/SV(P) level of theory, followed by removal of conformers with similar pseudo-chemical potentials in a set of commonly used solvents.…”