H. Badorreck
M. Jupé
D. Ristau
A. Wienke

Exploring Machine Learning Interatomic Potentials

International Conference on Frontiers of Optical Coatings (FOC)
22.-26. Oktober
Tianjin
2023
Type: Konferenzbeitrag
Abstract
For coatings and fundamental material studies, simulations are a great tool to investigate details that are nearly impossible to access experimentally. The Laser Zentrum Hannover e.V. and partner institutes therefore developed a multiple scale model, named Virtual Coater®. In this framework, the full coating process can be simulated by combining different methods for the material flow in the coating chamber, the atomistic growth, and the optical properties of the coating. For the atomistic growth molecular dynamics (MD) can be applied. MD is based on the Newtonian equations of motion and therefore allows a very accurate description of the growth process in principle. However, in a recent study we find [1], that the accuracy is limited regarding structural properties. The accuracy can be adjusted by the choice of the interaction potential. Empirical potentials of e.g. the Lennard-Jones or Buckingham type can be computed fast, but their accuracy is rough. An advanced approach is the use of Machine Learning Interatomic Potentials (MLIP). Here, an AI-based training algorithm is used to transform the quantum mechanical interaction potentials into a data-based model, which has a very high accuracy but requires an extremely high computational effort due to the complex physics-based models. MLIP enable MD simulations with nearly quantum mechanical accuracy and only moderate increased computational cost compared to empirical potentials. These can be used to relax grown structures as an additional step or also for the growth simulation itself by massive use of parallel GPUs. Currently, the application of MLIP is still very limited to exemplary studies in the context of university research. The implementation of the Virtual Coater® in the context of the cluster of excellence {\textquotedbl}PhoenixD"allows for the first time the application to real coating processes and the solution of questions from the experimental environment. An important aspect is also the research on structural differences of large area coatings for example with the large-scale IBS coating machine “IBS2000”, which will enable us to coat substrates uniformly up to 2m diameter. We compare the results of structures with MLIP-MD applied with structures from standard MD on the one side and quantum mechanically computed structures on the other side. The radial distribution function and coordination numbers are important benchmarks here. Further, the resulting much higher accuracy of simulated structures, especially for stress or vibrational absorption are shown.