Generative design of physical objects using modular framework

Engineering Applications of Artificial Intelligence. 2023. Vol. 119. Article 105715.

Nikita O. Starodubcev, Nikitin N., Andronova E., Gavaza K., Sidorenko D., Kalyuzhnaya A.

In recent years generative design techniques have become firmly established in numerous applied fields, especially in engineering. These methods are crucial for automating the initial stages of the engineering design of various structures, which reduces the amount of routine work. However, existing approaches are limited by the specificity of the problem under consideration. In addition, they do not provide the desired flexibility in choosing a method for a particular problem. To avoid these issues, we proposed a general approach to an arbitrary generative design problem and implemented a novel open-source framework called GEFEST (Generative Evolution For Encoded STructure) on its basis. This approach is based on three general principles: sampling, estimation, and optimization. This ensures the freedom of method adjustment for the solution of the particular generative design problem and therefore enables the construction of the most suitable one. A series of experimental studies was conducted to confirm the effectiveness of the GEFEST framework. It involved synthetic and real-world cases (coastal engineering, microfluidics, thermodynamics, and oil field planning). The flexible structure of GEFEST makes it possible to obtain results that surpass baseline and state-of-the-art solutions: 12% improvement in the coastal engineering problem; 9% in microfluidics; 8% in thermodynamics and 7% in oil field planning.

Keywords: Deep Learning, Generative design, Optimization problems