물리정보 신경망 (PINN)을 활용한 외팔보의 1D 및 2D 구조해석 기법 비교 연구

Abstract

This study compares one-dimensional (1D) and two-dimensional (2D) structural analysis techniques for cantilever beams using Physics-Informed Neural Networks (PINNs). Structural analysis is a fundamental technology for predicting the safety and performance of structures in various engineering fields, and recently, PINNs utilizing machine learning have emerged as a promising analytical methods. We evaluated the performance of PINNs for 1D Timoshenko beam models and 2D plane stress models. The results demonstrated that the 1D model achieved high accuracy with a maximum deflection error of 0.01% and rotation error of 0.14%, along with excellent computational efficiency, however, it had limitations in capturing stress concentration phenomena. In contrast, the 2D model showed improved accuracy in displacement prediction compared to the 1D model, with 32% improvement in the x-direction and 11% improvement in the y-direction, while accurately capturing stress concentration phenomena around the free end. These results confirmed that the 1D model is suitable for simple displacement prediction or applications where computational efficiency is critical, whereas the 2D model is appropriate for stress concentration analysis or complex stress state analysis. This study is significant in that it provides model selection criteria for effectively utilizing PINNs across various engineering fields, including agricultural engineering.