Innopolis University specialists took second place in the competition for the use of artificial intelligence. According to the press service of the IT university, the Russians lost only to the team from Microsoft Research.

Participants in the Open Catalyst Challenge AI Application Contest developed algorithms to apply deep learning to quantum chemistry and the search for catalysts for renewable energy. The organizers of the competition are Meta AI (Facebook AI Research) and Carnegie Mellon University. It was held as part of the largest annual conference on machine learning and computational neuroscience NeurIPS.

For work in February 2021, the participants received the largest dataset on the quantum-chemical properties of materials and molecules in the world – the dataset contains 1.2 million molecular relaxations with the results of more than 250 million DFT calculations. The main challenge is the transition from time-consuming and resource-intensive quantum-chemical calculations, which take up to 24 hours per structure, to faster predictions based on machine learning – less than 0.1 seconds per prediction. This dataset allows training models to find more active and energy efficient catalysts for converting small molecules into more valuable products and for storing electrical energy.

“Global warming due to excessive concentration of CO2 in the atmosphere, as well as unequal consumption of electricity in large cities, depending on the time of day, has become a challenge for the creation of new ways of storing energy, including from renewable sources. One of these approaches is the use of CO2 as a battery molecule, – explains Ruslan Lukin, Head of the Laboratory for Product Development in the Field of Artificial Intelligence in New Materials, Innopolis University – Through electrochemical processes, carbon dioxide can be converted into more valuable small molecules – methane, ethylene, methanol, ethanol – and is used both for generating electrical energy back, and as a building block for the production of polymers and large-scale chemical products. ”

During the competition, Innopolis University specialists built models based on graph neural networks with information transfer to predict the adsorption energy from the catalyst and reagent structure. With the help of this, it is possible to find catalysts that will make these processes possible, as well as select the most selective and energy efficient catalysts from the point of view of catalytic processes among the vast space of catalysts. In addition, this dataset will allow finding more efficient electrocatalysts for hydrogen energy and the production of fertilizers directly from atmospheric nitrogen, as well as more environmentally friendly automotive catalysts.

A total of 30 solutions from teams from Microsoft Research, Technical University of Denmark, Carnegie Mellon University, Technical University of Munchen, Texas A&M University, KAUST, AIRI, etc. took part in the competition. This number is due to a very high entry threshold: powerful GPU servers are needed to train graph neural networks on data on structures, good expertise in deep learning on chemical structures and methods of working with data of crystal structures.

The groups worked on solutions until October. They were evaluated according to the MAE (Mean Absolute Error) model accuracy metric. “The check was carried out using several test breakdowns, and the proportion of predictions whose error does not exceed 0.01 eV were taken into account, which suggests that machine learning models can predict properties with the same accuracy as quantum-chemical calculations. The accuracy of the solution of the winners from Microsoft Research is 0.547 eV, our accuracy is 0.618 eV, ”explained Ruslan Lukin. Together with him, three more specialists from the Institute of Artificial Intelligence of the Innopolis University worked in the team – Rostislav Grigoriev, Maxim Fadeev and Adel Yarullin.

“This year we began to develop research in the field of finding and improving deep learning architectures for predicting the properties of materials and molecules. The architectures presented by our team in the competition are universal and can be used both to predict the properties of crystalline materials for use in catalysis, to search for materials for neuromorphic computations, as well as to predict the properties of molecules in the search for new drugs. – said the Head of the Institute of Artificial Intelligence of the University of Innopolis Ramil Kuleev – This will significantly reduce the time spent on calculations and experiments, and will also allow in the future to more effectively solve the inverse problem of searching for materials and molecules with desired properties.