A research team from Skoltech, AIRI, Tomsk Polytechnic University, and Sber has proposed and tested an approach to predicting the modification of material properties. Artificial intelligence models that were pre-trained on a small amount of data enabled a significant increase in the calculation of the formation energies in possible configurations of higher tungsten boride doped with other metals. The method, which is also applicable to other substances, is presented in the npj Computational Materials journal.

Materials scientists continue to search for new materials for civil and industrial applications. The advancements in computational methods enable the prediction of their crystal structure and properties, whereas traditional experimental searches for chemical modifications are time-consuming and not always efficient. The use of modeling approaches directly leads to challenges because of the many possible realizations of the material’s crystal structure, particularly if they are not ordered.

Machine learning comes to the aid of scientists, which allows predicting target properties of various materials using limited sets of training data. Recently, Graph Neural Networks have proven to be a valuable tool in this area, offering the opportunity for pre-training on all available data in theoretical materials science and subsequent fine-tuning with the use of a limited set of training data.

In the new work, the researchers have proposed an approach to predicting material properties that implements such fine-tuning, but requires only a small number of additional calculations using electron density functional theory, thanks to intelligent selection of additional samples. The study is designed to address the issue of incomplete datasets for structures with chemical composition modifications and enhance thermodynamic stability estimation in the pursuit of functional materials by utilizing hybrid machine learning approaches. The team tested the new approach on finding an optimal dopant (substituent metal) for tungsten pentaboride.

“Previously, we have already developed a method to produce tungsten pentaboride in powder form, an important analog of expensive compounds for heat-resistant ceramic products, drilling equipment in the oil and gas industry. In the new work, we decided to test the new approach on the samples of this compound. First, we chose which metals can complement its structure and form a triple doped compound to improve its mechanical properties. Then we realized that we can only consider a few possible dopant concentrations experimentally, and calculating all configurations is very time-consuming. Based on our small data set, we trained a model that rather quickly predicted the formation energies of all possible configurations of doping with eight transition metals,” shared study co-author Professor Alexander Kvashnin from the Skoltech Energy Transition Center.

In total, the scientists predicted the thermodynamic properties of about 375 thousand structural configurations on a sample of only 200 results of quantum-mechanical calculations. The approach revealed the most promising compounds with improved mechanical properties, which is pentaboride tungsten, doped with tantalum in the percentage of 20 to 60%. The authors showed that modern artificial intelligence models can determine correlations between the composition, properties, and structure of materials. This opens up prospects for extending the proposed approach.

“In our case, the direct use of quantum mechanical calculations could have taken years. Rather than attempting all variants, we devised a strategy to sequentially include only those structures in the training of the graph neural network where it made the most errors. This reduced the combinatorial complexity of the problem, allowing us to achieve acceptable prediction quality for 200 training structures. The trained model enabled the analysis of all dopants in just a few days and the selection of the most promising ones from the perspective of experimental validation. Importantly, although the developed approach was applied to higher borides, it is not limited by construction to any class of compounds and can be used to search for new representatives in any other class of functional materials,” said Roman Eremin, a leading research scientist in the “New Materials Design” group at the AIRI Institute.

The samples were synthesized by the vacuum-free arc method at Tomsk Polytechnic University. A series of experiments were carried out under different synthesis conditions to obtain the predicted structures. The synthesized materials were studied using modern analytical methods.

“The vacuum-free arc method and specialized atmospheric plasma reactor developed by Tomsk Polytechnic University can be operated easily and at a low cost. The equipment is suited for testing quickly for hypotheses about the possibility of a specific predicted compound, such as a higher tungsten boride doped with tantalum. Modern analytical equipment enables the creation of a proof base and the investigation of the structure, morphology, and other features of synthetic materials,” said Alexander Pak, the head of the Laboratory for Advanced Materials in the Energy Industry and a professor in the Department of Electric Power Engineering and Electrical Engineering at Tomsk Polytechnic University.

“The project clearly demonstrates the possibilities of modern neural network architectures for solving applied research problems, in particular, the search for new functional materials. The development of compounds with improved mechanical properties opens up prospects for many industries. Thus, there are opportunities to take further steps in terms of creating experimental samples and testing them in real production processes. We expect that the results will be in demand in the real and other sectors of the economy,” commented Semen Budennyy, Head of Advanced AI Technology Development Division at Sber and a scientific advisor at the AIRI Institute.