Machine learning enables long time scale molecular photodynamics simulations.
Julia Westermayr1, Michael Gastegger2, Maximilian F S J Menger1,3
1Institute of Theoretical Chemistry , Faculty of Chemistry , University of Vienna , 1090 Vienna , Austria .
Chemical Science
|December 21, 2019
Summary
Machine learning accelerates simulations of photo-induced processes. This method enables accurate nanosecond-scale photodynamics, overcoming limitations of traditional quantum chemistry calculations.
Area of Science:
- Computational chemistry
- Photochemistry
- Machine learning
Background:
- Simulating photo-induced processes is crucial but computationally expensive.
- Current methods struggle with long timescales due to quantum chemistry costs.
Purpose of the Study:
- To develop a machine learning approach for efficient and accurate photodynamics simulations.
- To enable simulations on nanosecond timescales, previously inaccessible.
Main Methods:
- Utilized deep neural networks to learn the relationship between molecular geometry and electronic properties.
- Integrated machine learning models into molecular dynamics simulations.
- Replaced expensive quantum chemistry calculations with learned potentials.
Main Results:
- Achieved accurate photodynamics simulations on nanosecond timescales.
- Demonstrated superior computational efficiency compared to standard excited-state molecular dynamics.
- Validated the approach using the methylenimmonium cation example.
Conclusions:
- Machine learning offers a powerful solution to the computational bottleneck in photodynamics.
- The developed method provides a viable path for studying complex photo-induced processes over extended timescales.
- This approach significantly enhances the accessibility and efficiency of photodynamics research.
