Simulating Solvation and Acidity in Complex Mixtures with First-Principles Accuracy: The Case of CH3SO3H and H2O2 in Phenol

Kevin Rossi, Veronika Jurásková, Raphaël Wischert, Laurent Garel, Clémence Corminbœuf, Michele CeriottiJournal of Chemical Theory and Computation 2020 · DOI · arXiv

Abstract: We present a generally applicable computational framework for the efficient and accurate characterization of molecular structural patterns and acid properties in an explicit solvent using H<sub>2</sub>O<sub>2</sub> and CH<sub>3</sub>SO<sub>3</sub>H in phenol as an example. To address the challenges posed by the complexity of the problem, we resort to a set of data-driven methods and enhanced sampling algorithms. The synergistic application of these techniques makes the first-principle estimation of the chemical properties feasible without renouncing to the use of explicit solvation, involving extensive statistical sampling. Ensembles of neural network (NN) potentials are trained on a set of configurations carefully selected out of preliminary simulations performed at a low-cost density functional tight-binding (DFTB) level. The energy and forces of these configurations are then recomputed at the hybrid density functional theory (DFT) level and used to train the neural networks. The stability of the NN model is enhanced by using DFTB energetics as a baseline, but the efficiency of the direct NN (i.e., baseline-free) is exploited via a multiple-time-step integrator. The neural network potentials are combined with enhanced sampling techniques, such as replica exchange and metadynamics, and used to characterize the relevant protonated species and dominant noncovalent interactions in the mixture, also considering nuclear quantum effects.

Introduction

Simulating chemical reactions in "explicit solvent"—where every solvent molecule is modeled—is a massive computational challenge. To obtain accurate results for acidity and molecular structure, high-level hybrid Density Functional Theory (DFT) is typically required. However, the computational cost of hybrid DFT limits simulations to a few picoseconds, which is insufficient to observe rare events like proton transfer or to achieve statistical convergence.

Researchers often resort to faster, semi-empirical methods like Density Functional Tight-Binding (DFTB), but these can be qualitatively inaccurate for complex systems. This study focuses on a mixture of methanesulfonic acid (CH<sub>3</sub>SO<sub>3</sub>H) and hydrogen peroxide (H<sub>2</sub>O<sub>2</sub>) in phenol, an industrially relevant system for the production of catechol and hydroquinone.

Industrial context: Phenol hydroxylation to hydroquinone and catechol.

Phenol's "ambiphilic" nature—possessing both a polar hydroxyl group and a non-polar aromatic ring—creates a complex network of hydrogen bonds and $\pi$-interactions that are difficult to predict without high-level accuracy.

Methodology

To bridge the gap between speed and accuracy, the authors developed a multi-stage workflow that achieves a $10^4$ speedup over pure hybrid DFT simulations.

The multi-stage computational workflow integrating data-driven selection, ML potentials, and enhanced sampling.

The workflow consists of several key components:

  1. Data-Driven Selection: Farthest Point Sampling (FPS) was used to select 3,304 diverse configurations from preliminary cheap simulations, ensuring broad coverage of the chemical space.
  2. Machine Learning Potentials (MLP): An ensemble of Neural Networks (NN) was trained using a "baselining" (Delta-learning) approach. The NN learns the correction needed to elevate a cheap DFTB calculation to hybrid DFT (PBE0-D3BJ) accuracy.
  3. Multiple Time-Stepping (MTS): A "direct" NN handles fast inner steps, while the more expensive "baselined" NN is used for slower outer steps to optimize efficiency.
  4. Enhanced Sampling: Metadynamics and Replica Exchange were employed to explore rare states, such as acid deprotonation.
  5. Nuclear Quantum Effects (NQE): Path Integral Molecular Dynamics (PIMD) accounted for the quantum nature of hydrogen atoms, which is critical for accurate hydrogen bonding.

Key Results and Insights

The study revealed that standard DFTB qualitatively fails for this system, incorrectly predicting that methanesulfonic acid readily deprotonates in phenol. In contrast, the high-level ML model shows the acid remains protonated, highlighting that high accuracy is essential for basic chemical characterization.

Furthermore, H<sub>2</sub>O<sub>2</sub> was found to be a significant "solvation disruptor," binding strongly to the acid's hydroxyl group and almost completely displacing phenol from the primary solvation shell when quantum effects are included. The framework achieved an energy error of only ~2 meV/atom, effectively reaching "DFT-quality" at a fraction of the cost.

Conclusion and Future Work

This framework enables the first-principles characterization of complex mixtures with unprecedented efficiency. While currently limited by the training set size and a finite cutoff for long-range interactions, it provides a robust foundation for future studies. The next logical step is to apply this methodology to the transition states of the actual chemical reactions, such as the hydroxylation of phenol shown in Figure 1.

Summary updated

Abstract

We present a generally applicable computational framework for the efficient and accurate characterization of molecular structural patterns and acid properties in an explicit solvent using H<sub>2</sub>O<sub>2</sub> and CH<sub>3</sub>SO<sub>3</sub>H in phenol as an example. To address the challenges posed by the complexity of the problem, we resort to a set of data-driven methods and enhanced sampling algorithms. The synergistic application of these techniques makes the first-principle estimation of the chemical properties feasible without renouncing to the use of explicit solvation, involving extensive statistical sampling. Ensembles of neural network (NN) potentials are trained on a set of configurations carefully selected out of preliminary simulations performed at a low-cost density functional tight-binding (DFTB) level. The energy and forces of these configurations are then recomputed at the hybrid density functional theory (DFT) level and used to train the neural networks. The stability of the NN model is enhanced by using DFTB energetics as a baseline, but the efficiency of the direct NN (<i>i.e</i>., baseline-free) is exploited via a multiple-time-step integrator. The neural network potentials are combined with enhanced sampling techniques, such as replica exchange and metadynamics, and used to characterize the relevant protonated species and dominant noncovalent interactions in the mixture, also considering nuclear quantum effects.
BibTeX
@article{rossi2020simulating,
  author = {Kevin Rossi and Veronika Jurásková and Raphaël Wischert and Laurent Garel and Clémence Corminbœuf and Michele Ceriotti},
  title = {{Simulating Solvation and Acidity in Complex Mixtures with First-Principles Accuracy: The Case of CH<sub>3</sub>SO<sub>3</sub>H and H<sub>2</sub>O<sub>2</sub> in Phenol}},
  journal = {Journal of Chemical Theory and Computation},
  year = {2020},
  doi = {10.1021/acs.jctc.0c00362},
}
Figure