A robust methodology for dataset preparation and algorithm performance assessment in machine learning prediction of the fatigue life of additive manufactured components

The paper leverages existing literature on fatigue life prediction in additive manufacturing components within the realm of artificial intelligence (AI) and machine learning methods. The first key contribution is a meticulous methodology for dataset preparation. It aims to convert features from natural language to a format suitable for AI algorithms, minimizing over-fitting risks. This process is not limited to AI-specific scalers, but leverages authors’ expertise in the problem’s physics. Additionally, the article evaluates the performance of various machine learning and AI algorithms. This constitutes the second contribution of the paper. Typically, the performance of different algorithms is assessed based on the root mean square comparison of predicted versus true data. However, this widely accepted methodology overlooks the fact that data points are not independent entities; rather, they are grouped into subsets, each associated with a specific fatigue Wohler curve. From a practical standpoint, the engineering entity of interest for prediction is the complete curve, not just the single point belonging to the curve. Therefore, a methodology based on this concept is proposed to reliably assess algorithm performance, serving as a complementary technique to standard performance assessment metrics.