Machine-learning potential energy surfaces implications in photodissociation process.

Area of Science:

• Computational Chemistry
• Quantum Dynamics
• Machine Learning Applications

Background:

• Multi-state quantum molecular dynamics offers high accuracy for chemical reaction prediction.
• Ab initio calculations for potential energy surfaces (PES) are computationally expensive, especially for larger systems.
• Machine learning (ML) presents a resource-efficient alternative for computing molecular properties.

Purpose of the Study:

• To evaluate the accuracy of ML-interpolated PES in solving the time-dependent Schrödinger equation for water photodynamics.
• To statistically analyze the impact of ab initio data set size on ML-PES accuracy.
• To compare ML-PES performance against analytical solutions for IR+UV bond-breaking processes.

Main Methods:

• Quantum molecular dynamics simulations using ML-interpolated potential energy surfaces.
• Statistical analysis of expectation values and dissociation probabilities.
• Comparison with exact photodynamics solutions derived from analytical expressions for ground and excited electronic states.

Main Results:

• ML-PES demonstrated suitability for dynamics calculations in the Franck-Condon region.
• Standard interpolation methods proved more efficient for dissociative and repulsive electronic state regions.
• The number of ab initio points significantly influences ML-PES accuracy for dynamics.

Conclusions:

• ML-PES can be effectively integrated into molecular dynamics for accurate dissociation yield predictions with reduced computational cost.
• Careful consideration of ML model limitations in high-energy regions is necessary.
• This work advances the application of ML in multi-state dynamics calculations.