Facilitating ab initio QM/MM free energy simulations by Gaussian process regression with derivative observations

Abstract

In this machine-learning-facilitated method, Gaussian process regression (GPR) is used to predict energy and force corrections for a semiempirical QM/MM level to match with ab initio QM/MM results during MD-based free energy simulations.

Minimum Free Energy Pathways of Reactive Processes with Nudged Elastic Bands

“…This way, our results are consistent with previous observations that the AM1/TIP3P level of theory overestimates the reaction barrier (and underestimates the reaction exergonicity), in comparison with the closely related reaction between NH 3 and MeI. It has been shown that further corrections of the Menshutkin’s reaction path and the associated free energies can be performed with a machine learning-based refinement of the MFEP computed at the AM1/TIP3P level of theory . However, in this work, we will focus the following discussion on the comparison with precedents at AM1 level of theory.…”

“…Both FENEB and US produce practically the same trajectory in regions 1 and 3, being the most important differences located in the region of the transition ensemble. In the case of the FENEB path, the variation in the middle zone is quite linear, in good agreement with previous reports. , On the other hand, the path obtained with US differs mainly in the d C–Cl branch. Nevertheless, both methods obtain, in this case, almost the same coordinates for the maximum free energy value, which can be noted by the cross marks.…”

“…In the case of the related MeI species, its reaction with NH 3 exhibits an experimental free energy barrier and change of 23.5 and −34 kcal/mol, respectively. 82 This way, our results are consistent with previous observations 81 that the AM1/TIP3P level of theory overestimates the reaction barrier (and underestimates the reaction exergonicity), in comparison with the closely related reaction between NH 3 and MeI. It has been shown that further corrections of the Menshutkin's reaction path and the associated free energies can be performed with a machine learning-based refinement of the MFEP computed at the AM1/TIP3P level of theory.…”

“…It has been shown that further corrections of the Menshutkin's reaction path and the associated free energies can be performed with a machine learning-based refinement of the MFEP computed at the AM1/TIP3P level of theory. 81 However, in this work, we will focus the following discussion on the comparison with precedents at AM1 level of theory. Panel C of Figure 12 presents the estimated MFEP by our FENEB scheme (black line), projected on the d N−C -d Cl−C plane together with the FES A(d N−C , d Cl−C ) obtained by using multidimensional umbrella integration.…”

“…In our case, however, it served as a case study for the performance of the FENEB method presented herein for modeling a reactive process in complex environments, and its comparison with another popular free energy method (US). The free energy barrier retrieved from the MFEP optimized with FENEB is in good agreement with the previous estimations of 26.3, 79 28.8, 81 27.1, 80 and 26.8 80 kcal/mol, obtained with QM/MM frameworks at the AM1/TIP3P level, 80 using free energy methods based on the exploration of a given FES. This stands for the capability of our FENEB optimization for estimation MFEP and free energy barriers for an actual reactive process, requiring less chemical knowledge of the reaction of interest, since non-distinguished coordinate needs to be specified as in the case of US.…”