Scientists Have Discovered a New Characteristic of Artificial Intelligence

13:09, December 24, 2021

Discovery expands the possibility of creating materials with predetermined properties

Scientists at Ural Federal University, the Institute of Metallurgy, Ural Branch of the Russian Academy of Sciences, and the Institute of High Pressure Physics, Russian Academy of Sciences have discovered an unexpected characteristic of artificial intelligence: to extend the "skills" of describing the characteristics of certain metal alloys to others with significantly different compositions and characteristics. This ability greatly facilitates the task of modeling new materials with given characteristics. Colleagues gave the discovered phenomenon an original definition: compositional tolerance. The article about the studies and the description of the results was published in the Journal of Molecular Liquids.

The discovery was made during the "training" of machine learning potentials based on deep neural networks. Machine Learning Potentials, Deep Neural Network Potentials, DNNPs, are a type of Machine Learning Interatomic Potentials, MLIPs, mathematical functions; they allow describing interaction of atoms in the system, characterizing their movement and spatial position. Deep neural networks - high-performance computing systems based on the principles of brain organization and activity and consisting of several layers; used for object recognition, speech, machine translation.

"Our task was to establish which parameters - the number of neural network layers, the number of neurons in each layer, etc. - have the most significant influence on DNNP, on potential learning conditions, such as speed and stability. The knowledge obtained as a result of the research makes it possible to increase the accuracy and computational performance of Deep Neural Network Potentials," comments Roman Ryltsev, professor at the Training and Research Center "Information Security" of UrFU, head of the Laboratory of Disordered Systems of the Institute of Metallurgy, Ural Branch of Russian Academy of Sciences, research leader and co-author of the article.

The scientists used melts of aluminum, copper, and nickel as a "test-bed". These melts were chosen as a chemically complex, multicomponent, yet well-studied system, suitable and convenient for rigorous potential testing. In addition, these alloys are particularly corrosion-resistant and heat-resistant and are therefore of great practical importance.

Researchers have "tested" and "tuned" on ten compositions of aluminum, copper, and nickel melts several dozens of potentials with different parameter values. Based on the results of numerical experiments, the priority parameters that determine the effectiveness of Deep Neural Network Potentials were identified, and their software embodiment, the most optimal in terms of accuracy and speed of calculations and predictions, was determined. It has been revealed that the compression procedure of the selected Deep Neural Network Potentials model, reminiscent of file archiving, increases the computational performance of the potentials by six times, while maintaining the same accuracy of calculations and forecasts.

Further research by scientists aims to find out whether compositional tolerance is a universal characteristic of Deep Neural Network Potentials, and whether they are able to "recognize," describe and model systems with more components than aluminum, copper and nickel melts, including solid-phase ones.

"But the data already obtained also open up prospects for the development of universal potentials for high-precision, high-performance and cost-effective prediction, programming and creation of a wide range of new materials with a given structure and properties (hardness, toughness, conductivity, etc.), specifically - multicomponent metal alloys," summarizes Roman Ryltsev.

The significance of the research conducted is demonstrated by the support of grants of the Russian Science Foundation (№№ 18-12-00438, 19-73-20053) and use of computing resources of the largest computing centers of Russia at the "Kurchatov Institute" Research Center, the Interagency Supercomputer Center of the Russian Academy of Sciences and the Institute of Mathematics and Mechanics, Ural Branch of the Russian Academy of Sciences.

Reference

The rapid development of software, machine learning models and, in particular, neural networks in the last few years has given rise to a technological revolution in computer atomistic modeling of new materials. It involves predicting, "constructing" and explaining material characteristics at the atomic level, inaccessible in real experiments.

In comparison with "classical" quantum-mechanical methods, the advantage of neural networks is that they are able to describe and predict properties of complex systems, consisting not of hundreds, but of millions of interacting particles on time scales up to hundreds of nanoseconds with high accuracy. At the same time, the computational time required is less than in the case of quantum-mechanical methods.

Until recently, calculations of such scales could only be performed using approximate empirical potentials, which, compared to quantum mechanical methods, gave much lower accuracy, insufficient for solving problems of modern materials science. The problem was solved by combining quantum-mechanical calculations and machine learning methods to efficiently generate Machine Learning Interatomic Potentials, MLIP.