Session: 03-03-01: Applications of AI

Paper Number: 160623

160623 - Physics-Guided Machine Learning With Attention Mechanism and Bayesian Inference for Fatigue Life Prediction of Am-Built Materials 

Estimating the accurate fatigue life of additively manufactured components used in the aerospace industry and other critical engineering applications has been extensively investigated over the past decade. As the adoption of additive manufacturing (AM) technologies increases in almost all vital industries because of their advantages over traditional manufacturing methods, there is a growing need for reliable methods to predict the fatigue life of AM-built materials. Although additive manufacturing offers numerous benefits, it also introduces defects such as gas pores, lack of fusion, and surface roughness, which negatively impact the mechanical and fatigue properties of the materials. Several effective methodologies exist to predict fatigue life, each with unique strengths.  Despite the advancements in these methods, achieving precise fatigue life predictions is challenging due to varying material properties. Researchers are proactively employing machine learning techniques to address this challenge to maximize computational efficiency and improve prediction accuracy. Achieving highly accurate, extrapolatable, generalizable, interpretable, and physically consistent results in machine learning-based fatigue prediction remains a challenge.

To address these issues, we proposed two distinct neural networks: 1) A multimodal approach known as the Attention-aware Physics-guided Neural Network (AAUPgNN). This network employs a U-net architecture, physics guidance for the loading parameters, and a gated attention (GA) mechanism to achieve highly accurate predictions, utilizing both the loading parameters and CT scan images. 2) Bayesian Physics-Guided Neural Network (BPgNN), which enhances extrapolation capabilities through a combination of Bayesian principles and physics-based guidance. This approach uses loading parameters as well as surface roughness parameters extracted.

Data for this study were collected from in-house fatigue testing and imaging, as well as from extensive literature sources with surface roughness parameters such as Ra, Ry, Rz, and ¯ρ are the average roughness, peak-to- valley height, 10- points roughness, and average root radius of five dominant valleys. The same parameters are extracted from the high-quality STL files from the CT scan images.  The AAUPgNN architecture consists of a five-layer U-Net, which processes CT images and integrates loading parameters through multilayer perceptrons. Gated attention mechanisms enhance feature representation, while skip connections preserve spatial information. In parallel, a Bayesian neural network was developed to quantify
uncertainty in predictions. This network utilizes extracted surface roughness parameters from STL files and loading parameters, allowing for a comprehensive analysis of the factors influencing fatigue life. While AAUPgNN trained
with the in-house CT scan crack initiation data, the BPgNN trained and tested using the concatenated literature and in-house data with extracted surface roughness parameters.

Due to the unavailability of image data for fatigue prediction, we used in-house data containing 47 crack initiation locations. We employed 80% of these locations, along with the loading parameters, to train the AAUPgNN model. The model was then tested on the remaining data, which demonstrated highly accurate predictions when compared to the traditional neural net-
works, probabilistic neural networks, and the Probabilistic Physics-Guided Neural Network 2.0 (PPgNN 2.0). The attention mechanism implemented in each decoder layer provides valuable insight into the significance of various characteristics of the specimen.

The Bayesian approach effectively trains the network with sparse data and demonstrates robust extrapolation capabilities with unseen data. For example, training data with an average surface roughness (Ra) of less than 0.4 (normalized value) was used, while data with Ra greater than 0.4 was reserved for testing and prediction. While there are discrepancies between the original S-N curve, the results show the potential of Bayesian inference to address data sparsity and enable reliable extrapolation in fatigue life prediction.

This study introduces new methods for predicting the fatigue life of additively manufactured (AM) materials by merging physics-based machine learning approaches with attention mechanisms and Bayesian inference. This combination enhances
the accuracy of fatigue predictions, improves extrapolability, and reduces reliance on manual feature extraction. Future research will focus on expanding our dataset and improving our models.
 

Presenting Author: Yongming Liu Arizona State University

Presenting Author Biography: Yongming Liu is a professor of aerospace and mechanical engineering at the School for Engineering of Matter, Transport, and Energy at Arizona State University. He heads the Prognostic Analysis and Reliability Assessment Laboratory (PARA). He joined ASU in 2012.

His research interests range from fatigue and fracture of engineering materials and structures, probabilistic computational mechanics, risk assessment and management to multi-physics damage modeling and structural durability, multi-scale uncertainty quantification and propagation, imaging-based experimental testing, diagnostics, and prognostics.