Discovery of fatigue strength models via feature engineering and automated eXplainable machine learning applied to the welded transverse stiffener

This research introduces an integrated framework that combines Automated Machine Learning (AutoML) with Explainable Artificial Intelligence (XAI) techniques for the prediction of fatigue strength Δ σ c in welded transverse stiffener details. The methodology synergizes expert-guided as well as algorithmic feature engineering to ensure both, high accuracy and interpretability across five model variants ( M 1 - M 5 ). An initial data curation phase was conducted to clean and harmonize an extensive fatigue test database in accordance with EN 1993-1-9 as sound base for comparative AI model assessment. Ensemble learners (e.g. LightGBM, CatBoost) achieved test RMSE ≈ 30 MPa and R Test 2 ≈ 0 . 78 over the full stress range. While M 3 attained the best training fit ( R Train 2 = 0 . 917 ), M 5 showed superior generalization, especially in the 0–150 MPa domain (RMSE ≈ 12.7 MPa, R Test 2 ≈ 0 . 58 ). SHAP interpretability consistently highlighted stress ratio R , stress range Δ σ i , yield strength R e H , and TIG dressing treatment as dominant predictors, with geometric descriptors playing secondary roles. The results suggest that M 5 offers the best balance between performance and interpretability. This work demonstrates a physically grounded, scalable method for AI-augmented fatigue assessment in structural engineering. • Unified AutoML + XAI pipeline for fatigue strength prediction of welded stiffener. • AutoML tuned tree algorithms achieve best R² values on both, training and test set. • Combining expert knowledge with algorithm-generated interactions. • Explainability built-in via SHAP values reveals dominant fatigue drivers. • Δσ , post-treatment, R , R eH are ranked highest by SHAP, affirming fatigue characteristics.