RuAviation

The Moscow Aviation Institute has developed software to speed up complex engineering calculations in the aviation industry. MAI researchers have replaced single-type aerodynamic calculations with methods using artificial intelligence. The created neural network library will allow to perform with acceptable accuracy on an ordinary home PC the tasks that today require the use of the most powerful supercomputers, the MAI said.

The project is carried out on the basis of the Laboratory of Artificial Intelligence and Mathematical Modelling of the Institute No. 8 “Computer Science and Applied Mathematics”. This work was started within the framework of a separate decree of the Russian Government, in which the university plays the role of the main developer of new domestic software for complex multiphysics calculations in the field of aviation.

When designing a new type of aviation and space equipment at the stage of selecting the optimal configuration of the sample, it is required to perform many calculations on supercomputers. In particular, researchers are engaged in detailed modelling of aerodynamic processes, which is one of the longest and most time-consuming parts of the work. It takes months of supercomputer work to go through as many variants of aircraft layout as possible, or to calculate a few seconds of flight using traditional computational aerodynamics methods.

The library of the software system was written in Python and C++, and neural networks were based on the PyTorch framework. In their work, MAI specialists used the most modern architectures: physically informed neural networks (PINN) and graph neural networks (GNN).

The developers built into their library a module of data generation for almost any, even the most complex case of body flow. Having received a set of generated data, neural networks are trained on them and then perform fast calculations without using supercomputers. The data generation module is based on software for modelling fluid and gas dynamics processes, namely CFD solvers Logos and OpenFOAM.

The main industrial customer of MAI is the Ministry of Industry and Trade of the Russian Federation, which subsidises the development for implementation at enterprises of the aviation industry. At this stage, the specialists are testing the first version of the library and plan to get the second version with wider functionality and better quality of trained neural networks by the end of the year. The development is also planned to be patented in the near future.

According to the MAI, the module allows to reduce the time for aerodynamic calculations from 2 to 1000 times, now the software library is at the stage of testing and refinement. Within the framework of the whole complex it will start to be tested in December, and according to the plan the beginning of development of the technique at real enterprises will take place next year. Sukhoi Design Bureau, KTRV, State Research and Production Enterprise Region and MKB Vympel will be the first to start testing and using the MAI development.