Reproduce “A Machine Learning Approach to Generate Residual Stress Distributions using Sparse Characterization Data in Friction-Stir Processed Parts” on OpenResearch

About this paper

Shadab Anwar Shaikh, Kranthi Balusu, Ayoub Soulami

A machine learning-based Residual Stress Generator (RSG) accurately infers full-field residual stress distributions in friction-stir processed parts using only a few sparse experimental measurements. This approach substantially reduces the experimental effort required for comprehensive material characterization, achieving over a 10-fold reduction in data acquisition compared to traditional methods.

Key insights

• Deep learning models, particularly U-Net architectures, can effectively learn the complex, non-linear spatial relationships and "latent structure" within residual stress distributions.
• Combining physics-informed simulation data for training with sparse experimental validation provides a robust method to bridge the gaps between purely computational and purely experimental approaches.
• The characteristic in-plane variations and general patterns of FSP-induced residual stresses (e.g., tensile in the stir zone, compressive at edges) can be reliably reproduced by the ML model.

Summary by alphaXiv.