[1I3-GS-10f] AI application

2:55 PM - 3:10 PM JST(5:55 AM - 6:10 AM UTC)

[1I3-GS-10f-06] A Deep Learning-Based Surrogate Model Capable of Representing Variability in Manufacturing

〇Wataru Arakizeki¹, Naoya Kato¹, Nobuo Hara¹, Akira Kumagawa¹, Kazuki Fujiwara¹, Yoshiro Kitamura¹, Toshihiko Wada¹, Tsubasa Endo², Toshio Otsu², Hiroharu Tamaru², Yohei Kobayashi² (1. Panasonic Holdings Corporation, 2. The University of Tokyo)

Keywords:

Deep Learning,Surrogate Model,Automatic Experiment

Estimating the distribution of variability is crucial for quality control and design optimization in manufacturing.This study proposes a deep learning method that assigns seed embeddings to each training data point and guarantees input-independence through adversarial learning.In comprehensive experiments with 672 conditions on synthetic data, the proposed method achieved the best performance in 35.1% of conditions, comparable to Gaussian process regression (32.1%) and diffusion models (32.7%), while normalizing flows showed 0%.No statistically significant difference was found among the top three methods.However, the proposed method showed consistent advantages for sample sizes of 500 or more.In experiments with laser welding data, the proposed method achieved the highest accuracy when data was limited to 10%, while normalizing flows performed best with sufficient data, indicating the importance of method selection based on data characteristics.