An XGboost enhanced physics-informed machine learning framework for multiaxial fatigue life prediction

Material damage often manifest as multifaceted processes that are highly dependent on loading modes and magnitudes. Consequently, physical models based on a single mechanism may exhibit limited predictive capability when applied across diverse loading scenarios. To address this limitation, this study proposes a load-dependent adaptive multi-physics-informed machine learning framework. The proposed framework primarily consists of two components: a set of PINN models based on different damage mechanisms and an XGBoost adaptive selector. The first component incorporates the core partial differential relationships from different damage mechanisms into each loss function. The second component utilizes an XGBoost multi-classification algorithm to establish an adaptive selector. The proposed XGBoost-PINN framework can adaptively select the most suitable mechanism from the PINN model group for prediction according to specific loading conditions. Additionally, by constructing a transfer learning network, this architecture is further applied to additively manufactured (AM) Ti6Al4V with considering the influence of surface roughness, showcasing its robust generalization ability.