Researchers from the Skoltech Materials Center (part of the VEB.RF Group) have proposed a new approach to modelling the mechanical properties of heterogeneous materials, combining machine learning with active learning on local chemical configurations. The method enables calculations for large systems containing tens of thousands of atoms with accuracy comparable to direct quantum-mechanical calculations (DFT), but without requiring large computational resources. The study has been published in the journal Computational Materials Science.
The developed methodology is based on moment tensor potentials (MTPs), which are trained on data obtained from density functional theory calculations. During the simulation, the system automatically identifies local atomic environments for which the potential’s energy predictions become unreliable (extrapolative), extracts them from the large simulation cell, and sends them for additional DFT calculations. These fragments are then added to the training set, and the potential is retrained, progressively expanding its domain of applicability. The authors applied this approach to WC‑Co composites — hard alloys based on tungsten carbide, also known as pobedit, widely used in industry due to their high hardness and fracture toughness.
The approach, based on active learning on local chemical configurations, enables the modelling of large polycrystalline and composite systems, including defects and grain boundaries that are inaccessible to direct DFT calculations, while maintaining the accuracy of quantum-mechanical description. The method is applicable to a wide range of heterogeneous materials — from ceramics and metal ceramics to nanostructured composites.
“Developing new materials with desired properties requires a detailed understanding of the relationship between structure and mechanical characteristics. However, direct modelling of heterogeneous systems using DFT is impossible due to limitations on the number of atoms: these calculations are only applicable to systems of a few hundred atoms, whereas real polycrystalline materials contain tens of thousands of atoms. Our proposed approach solves this problem: we automatically collect information about local atomic configurations encountered in the material and use it to train the potential, maintaining DFT accuracy at scales sufficient for modelling industrial composites. This gives engineers and materials scientists a tool for predictive calculation of material properties without costly experiments,” commented Alexander Kvashnin, the study’s principal investigator, Professor at the Skoltech Materials Center.
“Using WC‑Co as an example, we demonstrated how the method allows quantitative description of the transition from brittle to ductile behavior as the cobalt binder content increases. Such data are important for designing materials with an optimal balance of hardness and fracture toughness. Our approach is universal and can be applied to a wide variety of composite and polycrystalline materials,” noted Faridun Jalolov, the first author of the study, a PhD student in the Materials Science and Engineering program at Skoltech.
