Machine learning potential enables multi-state accuracy for ultra-fast photodynamic simulations

Understanding the complex movements of atoms during chemical reactions remains a central challenge in theoretical chemistry, and accurately simulating these processes requires extraordinary computational power. Ivan V. Dudakov, Pavel M. Radzikovitsky and colleagues at Lomonosov Moscow State University and Irkutsk State Research and Technology University are now making significant progress in this field, developing a machine learning interatomic potential that achieves unprecedented accuracy comparable to complex multistate calculations. This breakthrough enables a complete mapping of the potential energy landscape of the molecule, revealing all possible pathways after light absorption, as demonstrated through detailed simulations of the methaniminium cation. Importantly, the team also introduces a new wavepacket oscillation model that couples fundamental quantum calculations with observed reaction rates, providing a transparent framework for examining the importance of quantifying uncertainties in these complex simulations.

Researchers are investigating the potential of transfer learning, especially using advanced computational techniques, to improve the accuracy and efficiency of simulations related to ultrafast photodynamics. This work focuses on developing interatomic potentials that can accurately describe complex molecular systems with limited computational resources, a challenge often encountered in modeling photochemical processes. We introduced a new wave packet oscillation model featuring power-law damping to describe the long-term evolution of the excited state after photoexcitation. This model aims to capture the complex dynamics of energy dissipation and relaxation within molecular systems and provide insight into the mechanisms governing photochemical reactions. The research team demonstrates the application of these combined techniques to simulate the photodynamics of complex molecules, achieving a balance between computational cost and accuracy in describing the underlying physical processes.

Machine learning speeds up photochemical reaction simulations

Accurate simulation of photochemical reactions governed by transitions between electronic states is a central pursuit in theoretical chemistry. Machine learning has emerged as an innovative tool for constructing the required potential energy surfaces, but applying it to excited states is a major challenge due to the computational costs of generating high-level quantum chemical data. Researchers overcame this challenge by developing machine learning interatomic potentials that are trained with limited precision calculations and can accurately predict energies and forces in complex molecular systems. This approach significantly reduces the computational burden associated with simulating excited state dynamics and enables the study of larger and more complex photochemical reactions than previously possible. The resulting potentials exhibit high accuracy in reproducing reference data and good transferability to different molecular configurations and chemical environments.

Molecular photochemistry using quantum mechanical simulations

This research focuses on understanding the photochemistry and dynamics of molecules when exposed to light, with a particular focus on the paths molecules take after absorbing photons. The research team simulated the evolution of molecules on the potential energy plane based on precise quantum chemical calculations. They use methods that determine the electronic structure of a molecule and calculate its potential energy surface. Molecular dynamics simulations track the movement and reactions of molecules over time. Machine learning models predict potential energy surfaces, accelerate dynamics, and quantify prediction uncertainties. The team employs techniques such as Gaussian process regression and deep ensembles to assess the reliability of the results. This combination of high-level quantum chemistry and machine learning enables more accurate and efficient simulations of molecular dynamics, resulting in reliable predictions and insights into photochemical pathways.

Possibility of photodissociation mapping using machine learning

This study provided a major advance in the simulation of photochemical reactions and achieved a detailed understanding of the photodissociation landscape of specific molecules. Researchers have developed a transfer learning protocol for building highly accurate machine learning interatomic potentials, reaching the level of advanced theoretical methods. This enabled for the first time high-level simulations of nonadiabatic dynamics initiated in the excited state and comprehensively mapped competing decay channels, including photoisomerization and hydrogen loss. The research team also introduced a new wavepacket oscillation model, a power-law dynamics framework that directly links quantum transition probabilities to classical rate constants.

This model extracts state-specific lifetimes from first-principles population dynamics and provides a mechanistically transparent and interpretable description of the response. The analysis revealed that incorporating uncertainty corrections from an ensemble of models improves the agreement between different approaches. Importantly, the kinetic fit validates the recently discovered photochemical pathway via a new conical crossing and elucidates the channel-specific lifetime of this minor pathway. Taken together, these contributions provide a robust framework for building accurate machine learning potentials, a complete mapping of photodissociation mechanisms in fundamental model systems, and new kinetics that link quantum transitions to interpretable lifetime models, establishing a generalizable route for simulating ultrafast photochemical processes with both accuracy and clarity.