AI model predicts material fatigue with 96.5% accuracy for circular factories

A new paper from researchers at multiple institutions introduces an AI framework that predicts both functional behavior and material fatigue for returned products in circular factories. The system uses a convolutional encoder to extract loading patterns from spindle forces and shaft torque, feeding an LSTM that predicts nine functional variables as Gaussian estimates. In parallel, the same loading history is translated into fatigue information via finite-element stress reconstruction, S-N/Miner damage evaluation with Haibach extension, and Paris-law crack-growth analysis. A streaming replay algorithm consolidates these branches into reliability trajectories.

In held-out tests, the model achieved a mean 2%-tolerance accuracy of 0.9652 across nine outputs, with near-perfect thermal variable predictions. Drive motor current and load speed were the most demanding, with R² values of 0.9750 and 0.9924 respectively. Torque history proved especially important, and conventional LSTM outperformed GRU and xLSTM in short-history settings. The work addresses a critical gap in circular manufacturing, where returned products have heterogeneous degradation states and require instance-specific reliability assessment.