Dynamic Ensemble Learning with Transfer Learning for Fatigue Performance Prediction in Ni-Based Superalloys.
Saved in:
| Title: | Dynamic Ensemble Learning with Transfer Learning for Fatigue Performance Prediction in Ni-Based Superalloys. |
|---|---|
| Authors: | Yang, Jiaxing1 (AUTHOR), Du, Fenglou1 (AUTHOR), Lv, Haopeng1 (AUTHOR) lwang199110@163.com, Li, Wang1 (AUTHOR), Wu, Dayong1 (AUTHOR) wudayong_ysu@126.com |
| Source: | Materials (1996-1944). Jun2026, Vol. 19 Issue 11, p2371. 18p. |
| Subjects: | Ensemble learning, Knowledge transfer, Materials science, Heat resistant alloys, Machine learning, Fatigue limit, Tensile tests |
| Abstract: | Accurate prediction of fatigue performance in Ni-based superalloys is hindered by scarce data and poor generalization of conventional machine learning. This study proposes a framework combining dynamic ensemble learning with transfer learning. A tensile prediction model using five base regressors (SVR, RFR, DTR, XGB, MLP) on 1025 tensile samples is first built. A dynamic weighted error feedback ensemble algorithm (DWELA) adjusts base model weights in real-time based on validation errors, improving tensile R2 from 0.90 (best single model) to 0.95. To transfer knowledge to fatigue prediction, a feature alignment transfer learning (FATL) strategy aligns shared features (composition and heat treatment) between source (tensile) and target (fatigue) domains while fine-tuning domain-specific strain features, adapting effectively to a limited fatigue dataset of 622 samples. The resulting ETFPM model evaluated on five independent samples achieves R2 of 0.93 (fatigue stress) and 0.81 (fatigue life), outperforming the best fatigue-trained single model (SVR: R2 = 0.89 and 0.72). Twenty candidate alloys are predicted for screening. The method offers a practical route for fatigue prediction under data-limited conditions. The main novelties are: (i) DWELA's real-time error-driven weight adaptation with hard constraints and early stopping, which improves tensile R2 from 0.90 (best single model) to 0.95; and (ii) FATL's explicit separation of frozen shared features and trainable exclusive features, enabling accurate fatigue prediction (R2 = 0.93 for FS, 0.81 for FL) using only 622 fatigue samples. However, the independent validation is limited to five samples, and the datasets are compiled from the literature with potential heterogeneity in testing protocols and imputation bias for missing values. Further experimental validation is required to confirm broader applicability. [ABSTRACT FROM AUTHOR] |
| Copyright of Materials (1996-1944) is the property of MDPI and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.) | |
| Database: | Engineering Source |
|
Full text is not displayed to guests.
Login for full access.
|
|
| Abstract: | Accurate prediction of fatigue performance in Ni-based superalloys is hindered by scarce data and poor generalization of conventional machine learning. This study proposes a framework combining dynamic ensemble learning with transfer learning. A tensile prediction model using five base regressors (SVR, RFR, DTR, XGB, MLP) on 1025 tensile samples is first built. A dynamic weighted error feedback ensemble algorithm (DWELA) adjusts base model weights in real-time based on validation errors, improving tensile R2 from 0.90 (best single model) to 0.95. To transfer knowledge to fatigue prediction, a feature alignment transfer learning (FATL) strategy aligns shared features (composition and heat treatment) between source (tensile) and target (fatigue) domains while fine-tuning domain-specific strain features, adapting effectively to a limited fatigue dataset of 622 samples. The resulting ETFPM model evaluated on five independent samples achieves R2 of 0.93 (fatigue stress) and 0.81 (fatigue life), outperforming the best fatigue-trained single model (SVR: R2 = 0.89 and 0.72). Twenty candidate alloys are predicted for screening. The method offers a practical route for fatigue prediction under data-limited conditions. The main novelties are: (i) DWELA's real-time error-driven weight adaptation with hard constraints and early stopping, which improves tensile R2 from 0.90 (best single model) to 0.95; and (ii) FATL's explicit separation of frozen shared features and trainable exclusive features, enabling accurate fatigue prediction (R2 = 0.93 for FS, 0.81 for FL) using only 622 fatigue samples. However, the independent validation is limited to five samples, and the datasets are compiled from the literature with potential heterogeneity in testing protocols and imputation bias for missing values. Further experimental validation is required to confirm broader applicability. [ABSTRACT FROM AUTHOR] |
|---|---|
| ISSN: | 19961944 |
| DOI: | 10.3390/ma19112371 |
