상세 보기
Machine learning prediction of bond strength between GFRP bars and UHPC: The critical role of group-aware validation and domain-adversarial learning
- Phoeuk, Menghay;
- Kim, Sangwoo;
- Kwon, Minho
WEB OF SCIENCE
0SCOPUS
0초록
Reliable prediction of bond strength between glass fiber-reinforced polymer (GFRP) bars and ultra-highperformance concrete (UHPC) is required for durable design, yet code provisions and empirical models calibrated for normal-strength concrete or limited datasets give inconsistent predictions. Machine learning (ML) models have predicted bond strength more accurately than closed-form equations, but this accuracy is usually reported under random specimen-level splitting, which ignores that the tests are grouped by experimental program. This study quantifies how such group-level data leakage inflates machine-learned bond-strength predictions and evaluates domain-adversarial learning as a mitigation, using 674 tests from 20 experimental programs. Because the between-program variance proportion is 0.812 (81.2% of the bond-strength variation lies between programs), a random specimen-level split places close analogues of the test specimens in the training set and rewards models for recognizing the source program. Four tree-based ensembles and an artificial neural network reached R2 = 0.946-0.967 under random specimen-level splitting but only 0.631-0.746 under group-level splitting, with mean absolute errors (MAE) 2.2-3.5 times larger. A domain-adversarial neural network (DANN), which penalizes representations that reveal the source program and selects its stopping epoch on program-disjoint splits of the training data, reached R2 = 0.911 (MAE = 2.77 +/- 0.72 MPa) on five unseen programs and R2 = 0.888 (MAE = 2.98 +/- 1.53 MPa) under leave-one-group-out (LOGO) cross-validation. Without group-aware early stopping, the LOGO R2 fell to 0.541; without the adversarial term, R2 fell to 0.887 (hold-out) and 0.878 (LOGO) and the per-program errors became more dispersed (LOGO MAE = 3.35 +/- 1.69 MPa). SHAP attributions are consistent with established bond mechanics: embedment-length ratio and bar surface condition dominate, and the water-to-binder ratio leads the matrix variables. Group-aware validation is recommended as the reporting standard for multi bond databases, with random-level results treated as an reference.
키워드
- 제목
- Machine learning prediction of bond strength between GFRP bars and UHPC: The critical role of group-aware validation and domain-adversarial learning
- 저자
- Phoeuk, Menghay; Kim, Sangwoo; Kwon, Minho
- 발행일
- 2026-12
- 유형
- Article
- 권
- 32
- 언어
- ENG
- 출판사
- Elsevier
- 발행국가
- 네덜란드
- ISSN
- E 2590-1230
P 2590-1230