Machine learning and physics based empirical equations for fatigue life prediction of AM Ti-6Al-4 V

Overview

Fatigue life decides whether an additively manufactured Ti-6Al-4 V part can be qualified, yet fatigue data are scarce and depend on stress, defects and build orientation all at once. Synthetic data can fill the gap, but whether generated data still behave like the real material is rarely checked.

The study benchmarks several data-augmentation strategies under strict fidelity checks, so the stress–defect–life relationships survive augmentation before any model is trained. On the most reliable strategy, four regression models are compared and the best one is interpreted with SHAP to show which factors drive fatigue life.

Two practical results come out of it: an interactive tool that estimates fatigue life from stress, orientation and defect inputs, and orientation-specific, physics-guided Basquin-type equations that can be used on their own.

Key Findings

• —The Residual Bootstrapped Conditional Regression Sampler (RB-CRS) was the most reliable augmentation strategy.
• —A compact artificial neural network gave the best accuracy of four regression models, with a test R² of about 0.98.
• —SHAP analysis: stress amplitude explains 74% of the predicted variation, build orientation 16%, defect size 7%, distance to surface 2% and defect depth 1%.
• —Orientation-specific Basquin-type equations track experiments up to 10⁸ cycles (R² 0.94–0.95).
• —The model is deployed as an interactive tool for rapid fatigue-life screening.

This is our summary of the paper. The full text, with its abstract, figures and data, is on the publisher’s site.