Artificial intelligence enables simulation of millions of atoms under laser excitation

Ultrashort laser pulses with a pulse duration of a few femtoseconds (1 femtosecond = ^10-15 seconds) are an essential tool in many areas of research and industry: In addition to the experimental observation of ultrafast physical phenomena in the laboratory, these lasers are also used in surgery as well as in high-precision material processing in the microelectronics or photovoltaics industry. Extensive, atomic-scale simulations of such processes are essential to develop and support experiments and practical applications. However, physical simulations taking quantum mechanics into account require an enormous computational effort and are limited to a maximum number of about 1000 atoms - too few for the relevant length scales.

Neural networks learn to predict the movements of atoms

Modern machine learning approaches make it possible to learn and reliably predict complicated relationships from large data sets. Results from physical simulations, for example, can be used to train neural networks to predict the motions of atoms based on the positions of their neighboring atoms. Researchers at the University of Kassel in the group of Prof. Dr. Martin Garcia have now succeeded in extending this method to the simulation of laser-excited materials. The crucial point here is the consideration of the electronic temperature, which depends on the strength of the laser pulse. Indeed, the pulse duration is so short that initially only the light electrons are affected by the femtosecond laser, while the heavier atomic hulls remain "cold." Neural networks that include electronic temperature as an additional parameter for predicting atomic motions, in addition to the positions of neighboring atoms, can be used to accurately simulate laser excitation of materials on a large scale and with little computational effort.

Use of AI enables ultralarge-scale simulations of laser processes

In their paper, the researchers show that many important physical properties can be reproduced with high accuracy by artificial intelligence. They use a thin silicon film as an example to demonstrate the potential of the new model: a simulation of more than a hundred thousand atoms shows that surface impurities can be removed using femtosecond lasers. "Due to the high computational efficiency of the model, even simulations with millions of atoms or more are possible, which would be unthinkable with purely physical methods," explains Pascal Plettenberg, one of the authors of the paper. "The method is also easily transferable to other materials and could play an important role in future research into the effects of ultrashort laser pulses."

The third author of the study is Dr. Bernd Bauerhenne. The project was supported by the German Research Foundation (DFG) through grant GA 465/27-1.

More information:

The article Neural network interatomic potential for laser-excited materials has now been published in the journal Communications Materials:

https://www.nature.com/articles/s43246-023-00389-w

Contact:

Prof. Dr. Martin E. Garcia
University of Kassel
Department of Mathematics and Natural Sciences
Solid State and Ultrafast Physics
Phone: +49 561 804-4480
E-mail: garcia[at]physik.uni-kassel[dot]de