We present a fully general computational framework which combines state-of-the-art machine learning techniques ( database selection via further point sampling, feature selection via CUR, regularized feed forward neural networks potentials, on-the-fly committee uncertainty estimate ) and molecular mechanics methods ( multiple time-stepping integration, replica-exchange Molecular Dynamics, Metadynamics, Path Integral Molecular Dynamics) to achieve accurate and reliable sampling at affordable computational costs.<br/><br/>The proposed framework is applied in the study of complex mixtures containing strong acids and hydrogen peroxide in phenol. We here characterize the stability of protonated species, dominant non- covalent interactions in solutions, and probe the energetics of possible reactive pathways leading to phenol hydroxylation. These statistics are gathered with first-principles accuracy but without compromising on an explicit description of solvation and on extensive sampling.<br/><br/>The study of acid catalysed phenol hydroxylation is chosen because of its relevance in the industrial production of hydroquinone and catechol, key molecular building blocks in the making of drugs, food, and cosmetics. At the same time, it represents a challenging problem for existing computational methods due to the ambiphilic nature of phenol, giving rise to subtle non-covalent interactions (hydrogen and CH-π bonds).

Machine learning and enhanced sampling to probe the solvation, acidity, and reactions in complex mixtures: The case of acid-catalysed phenol hydroxylation

ACS Fall 2020 Virtual Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.