The specialists of the Institute for Problems in Mechanical Engineering of the Russian Academy of Science (IPMash RAS) developed and tested successfully a new approach to modeling crystal growth, which allows to predict crystal properties and defects in fractions of a millisecond. This discovery marks the transition from an expensive and time–consuming experimental approach to creating a technology for synthesizing the materials to its complete end-to-end digital development, from crystal structure prediction to the design of industrial synthesis equipment.

High-quality crystals and thin films are at the heart of modern technology — from microelectronics to laser technology. Although scientists have already learned how to predict the structure of new promising materials using quantum chemistry and artificial intelligence, development of the technology for their industrial synthesis is still the key problem. The existing fundamental theories of crystal growth do not cope well with complex multicomponent systems, that makes the process of optimizing the growth conditions to be a long-term and extremely costly routine which requires thousands of trial experiments.

IPMash RASscientists proposed a solution that dramatically changes the situation.

«We were able to overcome the main limitation — the gap between atomic and macroscopic modeling. Our method unites two worlds: we generate huge amounts of data at the atomic level, and then we “teach” artificial intelligence on them, and it predicts crystal behavior in real growth settings at phenomenal speed. In fact, we have created a „digital twin“ of the epitaxial growth process, which operates with accuracy and speed unattainable by classical theories and numerical methods,» comments Alexey Redkov, a leading researcher at IPMash RAS.

The method is based on a hybrid atomic model implemented on graphics processors (GPU). This made it possible for the first time to simulate a system containing about 600 million atoms over long timescales, which is a record value approaching macroscopic scales. 10 000 virtual experiments were conducted on this model, covering all possible combinations of growth conditions, emerging phenomena and specific features of the growth process.

The data obtained was used to train machine learning models.

The trained AI-model not only revealed the hidden relationships in the growth process, but also learned to predict its result. The critical advantage is speed: the model provides a forecast in fractions of a millisecond, that is significantly faster than direct atomistic modeling on the GPU.

The development represents a transition from analytical theories with their simplifications to data-driven (data-based) methodology. The high prediction rate paves the way for the integration of the AI-model into macroscopic modeling of industrial growth reactors, where it can serve as an accurate boundary condition. This solves the problem of multiscale modeling by linking the atomic and macro levels.

The methodology is applicable to key epitaxy technologies such as MBE, MOCVD, and HVPE. It will significantly accelerate the development and optimization of crystal synthesis for the semiconductor industry, photovoltaics and quantum technologies. In the future, this will be the last link in creating a fully digital circuit for designing the technology and equipment for the industrial synthesis of new promising materials with the necessary properties.

The work was published in Acta Materialia.