Predicting stress–strain relationships of additively manufactured materials under compression and tension using transfer learning and Wasserstein distance-based dataset pruning

Additive manufacturing enables the design of materials and structures with tunable mechanical properties. However, predicting the stress–strain relationships of additively manufactured materials remains a challenge due to complex process-structure–property relationships in additive manufacturing. Wh...

Full description

Saved in:
Bibliographic Details
Main Authors: Chenglong Duan, Dazhong Wu
Format: Article
Language:English
Published: Elsevier 2025-08-01
Series:Materials & Design
Subjects:
Online Access:http://www.sciencedirect.com/science/article/pii/S026412752500629X
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Additive manufacturing enables the design of materials and structures with tunable mechanical properties. However, predicting the stress–strain relationships of additively manufactured materials remains a challenge due to complex process-structure–property relationships in additive manufacturing. While machine learning has been used to predict stress–strain relationships, large volumes of training data are usually required to achieve high predictive performance. In addition, ML models may not be generalizable when process conditions vary. To address these issues, a generic transfer learning (TL) framework that integrates distance-based dataset pruning (DP) is developed. Four distance metrics, including Euclidean, Cosine, Mahalanobis, and Wasserstein distances, are used to prune redundant data in the source domain by determining the distance between the source domain and the training data in the target domain. Only the most representative stress–strain curves in the source domain are retained to pre-train a long short-term memory model. We demonstrate the TL framework on two cases. Case 1 includes a compressive dataset collected from fused filament fabrication (FFF) fabricated polylactic acid (PLA) samples under compression tests. Case 2 includes a tensile dataset collected from laser powder bed fusion (L-PBF) fabricated AlSi10Mg samples under tension tests. Experimental results show that TL integrated with Wasserstein distance-based DP achieves the best predictive performance and computational efficiency, with an average mean absolute percentage error of 15.19 % for Case 1 and 19.51 % for Case 2, respectively.
ISSN:0264-1275