Understanding how a crack grows in metal requires simultaneously calculating the behavior of hundreds of billions of atoms: At the crack tip, where atomic bonds break, and in the surrounding volume of material, where the stress, driving  the crack propagation, is distributed. For modern supercomputers, an accurate modelling of this process remains beyond reach.

Researchers from the Skoltech Artificial Intelligence Center have proposed a solution, published in the journal Computer Physics Communications. Its core idea is to abandon the concept of modeling the entire process atom by atom. The approach could find applications in modeling material properties, creating composites with specified mechanical characteristics, predicting the lifespan of parts and assemblies in mechanical engineering, and developing components with high resistance to mechanical stress in microelectronics.

“We have developed a hybrid approach in which the material is divided into two zones. Where important processes occur, for example, in the contact zone of particles or at the tip of a growing crack, we retain an atomic-level description. The rest of the space is filled with so-called quasi-atoms — coarse-grained particles that can be hundreds or even thousands of times larger than real atoms. Quasi-atoms behave as a cohesive unit and obey the same laws of molecular dynamics,” shared one of the study’s authors, Artem Chuprov, a PhD student in the Computational Systems and Data Analysis in Science and Engineering program.

Artificial intelligence methods made it possible to combine atomic precision with macroscopic scales. The developed algorithm automatically tunes the interaction of quasi-atoms so that the elastic properties of the hybrid model exactly match the reference parameters obtained from full-atom modeling. “The required accuracy of over 99% is achieved in just a few minutes,” specifies another co-author, Egor Nuzhin, a senior research engineer at the center.

To validate the method, the authors simulated the collision of microparticles with radii on the order of fractions of a micrometer — a task that pushes conventional computations to their limits. Copper and silicon were used as test materials. The calculations showed that predictions from continuum theory, which do not account for the atomic structure of the contact zone, require corrections.

“The method opens up new possibilities for modeling friction, fracture, and other processes where we need to simultaneously describe  the atomic mechanism and the overall picture. Our plans include extending the method to new materials and enabling it to tune parameters not only for elasticity but also for viscosity, thermal conductivity, and other properties. The method could also be used for reverse engineering — selecting atomic structures to achieve desired material properties at the macroscale,” added Professor Nikolay Brilliantov, who heads the Multiscale Modeling and Neuromorphic Computing Group at the Skoltech AI Center.