Machine-learning potential energy surfaces implications in photodissociation process

Multi-state quantum molecular dynamics is one of the most accurate methodologies for predicting rates and yields of different chemical reactions. However, the generation of potential energy surfaces (PES), transition dipoles, and non-adiabatic couplings from ab initio calculations become a challenge, especially because of the exponential growth of computational cost as the number of electrons and molecular modes increases. Thus, machine learning (ML) emerges as a novel technique to compute molecular properties using fewer resources. Yet, the validity of ML methodologies continues in constant development, particularly for high-energy regions where conventional ab initio sampling is reduced. We test the accuracy of the potential energy surfaces interpolated with machine learning (ML) techniques in the solution of the time-dependent Schrodinger equation for the conventional IR+UV bond-breaking process of semi-heavy water. We perform a statistical analysis of the differences in expectation values and dissociation probabilities, which depend on the number of ab initio points selected to generate the machine learning potential energy surface (ML-PES). The energy differences of the electronic excited state modify population transfer from the ground state by driving with a UV laser pulse. We consider as the exact solution the photodynamics implemented with analytical expressions of the electronic ground (X) over tilde (1)A(1) and excited (A) over tilde B-1(1) states. The results of the mean bond distance and dissociation probabilities suggest that ML-PES is suitable for dynamics calculations around the Franck-Condon region, and that standard interpolation methods are more efficient for multistate dynamics that involve dissociative and repulsive energy regions of the electronic states. Our work contributes to the continued inclusion of ML tools in molecular dynamics to obtain accurate predictions of dissociation yields with fewer computational resources and non-written rules to follow in multi-state dynamics calculations.