Physics-Informed Neural Networksに基づく３次元弾性体の固有振動数の推定

抄録

We aim at developing a Physics-Informed Neural Network (PINN) to estimate eigenfrequencies of 3D elastic bodies rapidly and accurately. As the first step, we consider a Convolutional Neural Network (CNN), which is based on the VGGnet. The CNN is trained by a dataset of geometries, volumes, and (first-order) eigenfrequencies of elastic bodies. The relative error of the estimated eigenfrequency was about 15% and 30% for the training and validation data, respectively. This result indicates to consider a PINN by introducing eigenequations to the loss function, for example.