Physics-informed Hierarchical Learning for Defect Characterization and Fatigue Life Prediction in Metal Additive Manufacturing

Abstract

Understanding and accurately predicting the fatigue life of additive manufactured (AM) metal parts remains a pressing challenge for their reliable adoption in safety-critical applications such as aerospace, energy, and biomedical systems. In laser powder bed fusion (L-PBF), process-induced volumetric defects are one of the primary factors governing crack initiation and fatigue failure. Traditional fatigue assessment relies on destructive and time-consuming fatigue testing, resulting in fatigue data that is expensive and scarce. Consequently, individual organizations often possess insufficient data to develop robust predictive models. While collaborative learning across organizations provides a potential solution, concerns about heterogeneous inspection capabilities and data privacy hinder effective data sharing and joint model development. Moreover, most existing studies focus on uniaxial loading conditions, whereas engineering components are usually experienced complex multiaxial stress states. These challenges span various levels of fatigue assessment, ranging from defect-level characterization and specimen-level fatigue life prediction to organization-level collaborative learning and multiple-defect level multiaxial fatigue behavior understanding.

To address these challenges, this research develops a physics-informed hierarchical learning framework for defect characterization, fatigue life prediction, and collaborative learning in AM. The framework progressively enhances defect-fatigue understanding and prediction across multiple hierarchical levels through a series of physics-informed data-driven methodologies. Specifically, four methodologies are introduced:

(1) An integrated data-driven analytical framework using kernel support vector regression and interpretable model-agnostic methods for defect criticality analysis and fatigue life prediction of L-PBF 17-4 PH stainless steel parts. As the foundation of the hierarchical framework, this defect-level methodology enhances the understanding of defect-fatigue relationships in a data-driven manner and quantitatively analyzes the importance of defect features to fatigue life.

(2) A multimodal transfer learning (MMTL) framework using process-defect-loading data for defect classification and nondestructive fatigue life prediction of L-PBF Ti-6Al-4V parts. At the specimen level, this framework leverages knowledge learned from a pre-trained model with abundant process and defect data in the source task to predict fatigue life nondestructively with limited fatigue test data in the target task.

(3) A personalized federated transfer learning framework using conditional optimal transport (FedCOT) for collaborative predictive modeling across organizations with heterogeneous inspection capabilities. At the organization level, this framework enables ``target" organizations with limited features to benefit from ``source" organizations with sufficient features in terms of prediction performance, while preserving data privacy through a central server.

(4) A physics-informed multiple instance learning (PhysMIL) framework for multiaxial fatigue life prediction of AM metal parts. At the multiple-defect level, this framework integrates the critical plane mechanics with defect representations from multiple defects within each specimen and learns their contributions to fatigue failure via attention-based aggregation and physics-informed regularization on defect characteristics and fatigue damage mechanisms. It improves both prediction accuracy and interpretability under complex multiaxial loading conditions.

These methodologies establish a unified physics-informed hierarchical learning framework that advances the understanding of process-defect-fatigue relationships, enables nondestructive and data-efficient fatigue life prediction, supports privacy-preserving collaboration across organizations, and extends predictive modeling to realistic multiaxial loading scenarios. They provide practical tools for improving the reliability and qualification of AM components and can be utilized in other engineering domains involving limited, heterogeneous, and privacy-sensitive data.