A research team at SUSU, supported by a V.B. Khristenko Grant, is developing a software package aimed to simplify the engineering design process. In conventional computational fluid dynamics (CFD), solutions are obtained iteratively using computational meshes, which can result in complex simulations taking hours or even days to complete.
According to Ruslan Peshkov, Candidate of Sciences (Engineering), Associate Professor, and Dean of the SUSU Faculty of Aerospace Engineering, Head of the Department of Aircraft Engines, the solver under development is built around a neural network trained to satisfy the fundamental laws of mass, momentum, and energy conservation. During the training stage, the network relies on classical CFD techniques. Once trained, however, it functions as a mathematical model that directly predicts velocity, pressure, and temperature from the coordinates and input parameters of a problem, eliminating the need for a large number of iterative calculations. This significantly accelerates simulations during the early stages of multidisciplinary design optimization, which is particularly valuable for the aerospace industry.
The project is currently focused on developing a neural network training environment. The researchers are investigating various network architectures for applications involving external aerodynamics, internal flow, and heat transfer. The team has already prepared several conference papers for the 8th International Science-to-Practice Conference "Civil Aviation: History and Modernity" and the 18th International Youth Scientific and Technical Conference "Youth. Technology. Space". In addition, a paper has been submitted to a journal included in the State Commission for Academic Degrees and Titles list, and an application for intellectual property registration is being prepared for the software developed as part of the training environment.
A key feature of the project is its focus on end users rather than software developers. The graphical interface will resemble that of standard CFD software packages, allowing engineers to perform simulations without any knowledge of neural network programming. At the same time, adapting and retraining the model for highly specialized applications will require additional expertise. Although training the neural network demands high-performance workstations, routine simulations can later be carried out on a standard personal computer.
Looking ahead, the researchers plan to expand the solver's capabilities to include separated flows, multiphase flows, shock-wave phenomena, and aeroacoustics. They also intend to combine Physics-Informed Neural Networks (PINNs) with Generative Adversarial Networks (GANs). In this approach, one neural network would generate new geometric designs—such as airfoil profiles—while the other would instantly evaluate their performance, including aerodynamic characteristics. Such a combination could enable fully automated optimization of engineering designs.
"The future lies in hybrid approaches," says project leader Ruslan Peshkov. "Neural network algorithms will enable rapid evaluation of multiple design options, while numerical methods will continue to provide final high-accuracy calculations and solve unique problems involving new physical effects."
The researchers emphasize that a PINN-based neural network solver is intended to complement, rather than replace, traditional CFD methods—at least in the foreseeable future.
The project is being carried out as part of the V.B. Khristenko "Step into the Future" Grants Program, an annual funding initiative supporting the university's development strategy within the Priority 2030 program. The grants support promising research projects, the formation of personnel reserve, and the implementation of innovative educational programs that shape the strategic development of SUSU and the Chelyabinsk Region.