Machine learning for predicting fatigue properties of additively manufactured materials

Min Yi, Ming Xue, Peihong Cong, Yang Song, Haiyang Zhang, Lingfeng Wang, Liucheng Zhou, Yinghong Li, Wanlin GuoChinese 2023 · DOI

TL;DR

This paper provides a comprehensive review of how machine learning (ML) is being used to predict the fatigue properties of additively manufactured (AM) materials. It highlights various ML approaches, their applications, and the challenges that remain in this rapidly evolving field.

What Problem Does This Paper Address?

Additive manufacturing (AM) offers incredible design freedom for creating complex parts, but the fatigue properties of these materials are notoriously difficult to predict. This is because fatigue performance is influenced by a multitude of factors, including AM processing parameters (like laser power and scan speed), intricate microstructures, residual stresses, surface finish, and post-processing treatments. Traditionally, determining these properties involves expensive and time-consuming experimental testing. This paper addresses the need for more efficient and accurate methods to predict fatigue life and crack growth rates in AM materials, leveraging the power of machine learning to navigate this complex interplay of factors.

Methods

The paper reviews several prominent ML strategies applied to AM fatigue prediction:

  • Feedforward Neural Networks (FNNs): These models learn patterns from input data to predict fatigue life, often trained using experimental data or data generated from physics-based models (like Continuum Damage Mechanics - CDM).
  • Convolutional Neural Networks (CNNs): Primarily used for image analysis, CNNs are applied here to analyze in-situ sensor data from the AM process to correlate it with fatigue performance.
  • Adaptive Network-Based Fuzzy Systems (ANFS): These systems combine fuzzy logic with neural networks to model complex relationships, using "if-then" rules to predict fatigue life based on processing or material properties.
  • Physics-Informed Neural Networks (PINNs): A more advanced approach that integrates physical laws (like PDEs) directly into the neural network's training process, aiming for more robust and interpretable predictions.
  • Support Vector Machines (SVM): These algorithms map data to a higher-dimensional space to find optimal separating hyperplanes, used here for regression tasks to predict fatigue life based on various input features.
  • Random Forests (RF): An ensemble method that builds multiple decision trees and averages their outputs, proving effective for both fatigue life prediction and fatigue crack growth rate prediction.

The paper emphasizes that while these ML models are similar to those used in other fields, they are specifically adapted to incorporate AM-specific features.

Datasets

The paper discusses various datasets used in the reviewed studies, often derived from experimental testing of AM materials. These include:

  • Data from AM processed 300M-AerMet100 steel, SS316L, Ti6Al4V, and AlSi10Mg alloys.
  • Data from LPBF-AM 316L stainless steel, LC4 aluminum alloys, and Inconel 718.
  • Data on Ti6.5Al2ZrMoV titanium manufactured by laser DED.
  • Data on LPBF-AM 17-4 PH stainless steel specimens.
  • Data on LPBF-AM AlSi10Mg alloy.

The paper notes that generating comprehensive datasets for AM fatigue properties is challenging due to the high cost and time involved in experimental characterization. Some studies also leverage data from finite-element simulations to augment experimental datasets.

Key Insights and Ideas

  • AM Complexity Demands ML: The sheer number of interacting variables in AM makes traditional physics-based modeling insufficient. ML offers a powerful way to capture these complex, non-linear relationships.
  • Feature Engineering is Crucial: Identifying and incorporating the right AM-specific features (processing parameters, microstructural characteristics, defect types) is key to successful ML model performance.
  • Physics-Informed ML is Promising: PINNs are emerging as a way to bridge the gap between purely data-driven models and physics-based approaches, potentially leading to more reliable and interpretable predictions.
  • Data Scarcity is a Major Hurdle: The limited availability of high-quality fatigue data for AM materials is a significant bottleneck for training robust ML models.

Key Results

The paper summarizes that various ML models have demonstrated promising results in predicting fatigue properties:

  • FNNs have shown good performance, with accuracy improving with more training data and neurons.
  • CNNs have successfully correlated in-situ sensor data with fatigue life, achieving high ROC curve areas and validation accuracies (e.g., 99.7% and 97.4% respectively).
  • ANFS models have shown capability in predicting fatigue life, with both process-based and property-based models fitting experimental data within a two-error band.
  • SVM models have been used to predict fatigue life, with accuracy varying based on the input features, highlighting the importance of pore analysis.
  • RF models have shown high accuracy in predicting fatigue life and fatigue crack growth rates, often outperforming other ML models. For instance, RF models have achieved R² values of 0.875 for AM TC4 and TC11.

Limitations and Future Work

The authors identify several key challenges and suggest future directions:

  • Small Datasets: The limited availability of experimental fatigue data for AM materials is a major challenge. Future work should focus on data augmentation techniques (e.g., generative models, physics-based simulations) and creating larger, more diverse datasets.
  • Multifarious Features: Selecting the most relevant features from a vast array of processing, microstructural, and post-processing parameters remains difficult. Developing unified feature selection rules and leveraging ML algorithms that can assess feature importance are crucial.
  • Overfitting: Models trained on specific datasets may not generalize well to unseen data or different AM materials. Techniques like regularization, early stopping, and dropout are recommended.
  • Low Interpretability: Many ML models act as "black boxes," making it hard to understand the underlying fatigue mechanisms. Future research should explore methods to improve the interpretability of these models, possibly by integrating more physics knowledge (e.g., through PINNs).
  • Extension to Structure Life: Current ML models primarily predict fatigue properties of standard specimens. A significant challenge is extending this knowledge to predict the fatigue life and damage tolerance of actual AM structures under cyclic loading. This will require integrating ML with structural analysis, numerical simulations, and crack propagation methodologies.

Summary updated

Abstract

Fatigue properties of materials by Additive Manufacturing (AM) depend on many factors such as AM processing parameter, microstructure, residual stress, surface roughness, porosities, post-treatments, etc. Their evaluation inevitably requires these factors combined as many as possible, thus resulting in low efficiency and high cost. In recent years, their assessment by leveraging the power of Machine Learning (ML) has gained increasing attentions. A comprehensive overview on the state-of-the-art progress of applying ML strategies to predict fatigue properties of AM materials, as well as their dependence on AM processing and post-processing parameters such as laser power, scanning speed, layer height, hatch distance, built direction, post-heat temperature, etc., were presented. A few attempts in employing Feedforward Neural Network (FNN), Convolutional Neural Network (CNN), Adaptive Network-Based Fuzzy System (ANFS), Support Vector Machine (SVM) and Random Forest (RF) to predict fatigue life and RF to predict fatigue crack growth rate are summarized. The ML models for predicting AM materials’ fatigue properties are found intrinsically similar to the commonly used ones, but are modified to involve AM features. Finally, an outlook for challenges (i.e., small dataset, multifarious features, overfitting, low interpretability, and unable extension from AM material data to structure life) and potential solutions for the ML prediction of AM materials’ fatigue properties is provided.
BibTeX
@inproceedings{yi2023machine,
  author = {Min Yi and Ming Xue and Peihong Cong and Yang Song and Haiyang ZHANG and Lingfeng WANG and Liucheng ZHOU and Yinghong Li and Wanlin Guo},
  title = {{Machine learning for predicting fatigue properties of additively manufactured materials}},
  booktitle = {Chinese},
  year = {2023},
  doi = {10.1016/j.cja.2023.11.001},
}
Figure