The Center for Scientific Communication of the Moscow Institute of Physics and Technology reported that a new technology based on machine-learned potentials will help use AI to create high-precision models of materials and production processes, including welding and heat treatment.
Machine-learned potentials are specialized AI systems trained on limited volumes of full quantum chemical calculation results. They can spend three times less time analyzing information than traditional algorithms.
To optimize the performance of these AI systems, an efficient implementation for GPU should be developed, as well as the addition of long-range interactions and improved methodologies for creating databases for modeling. In addition, the integration of experimental data promises further improvement in the accuracy and applicability of machine-learned interatomic potentials.
Scientists from different countries are now wondering how these algorithms can be used, and have identified the pros and cons of using existing machine-learned potentials.
Researchers from Russia presented calculation systems based on moment tensor potentials (MTP). In terms of accuracy, MTP is on par with global counterparts, and even surpasses them in performance. The main difference is that Russian calculation systems can be trained on fragments of complex systems.
So far, these technologies cannot be used on graphics cards (GPUs). But scientists are on the way to discovering this possibility, because GPUs have higher computing power than CPUs (central processing units), where the moment tensor potential method is currently used.
Related materials:
Scientists from Chelyabinsk will create the first quantum infrared microscope in Russia
AI technologies will help develop the infrastructure of 6G networks