Physics-informed federated scientific machine learning for constitutive modeling under heterogeneous and decentralized data conditions
COMPUTERS & STRUCTURES, cilt.331, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 331
- Basım Tarihi: 2026
- Doi Numarası: 10.1016/j.compstruc.2026.108419
- Dergi Adı: COMPUTERS & STRUCTURES
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Aerospace Database, Applied Science & Technology Source, Compendex, INSPEC, zbMATH, Academic Search Ultimate (EBSCO), Engineering Source (EBSCO), Technology Collection (ProQuest)
- Anahtar Kelimeler: Constitutive model, Data heterogeneity, Data privacy, Federated learning, Inverse method, Scientific machine learning
- Bursa Uludağ Üniversitesi Adresli: Evet
Özet
The integration of neural network-based methods into scientific modeling has opened new opportunities for tackling complex challenges in constitutive modeling within computational mechanics. In practical settings, however, challenges arise due to the privacy constraints, distributed nature of data, and the difficulty of transferring large datasets. Federated learning offers an effective solution by enabling decentralized training of a global model while keeping data localized. In this work, we present, for the first time, a physics-guided Federated Scientific Machine Learning (FedSciML) framework for both inverse parameter identification of a constitutive model and direct approximation of polycarbonate material behavior under heterogeneous and decentralized data conditions. To systematically study data heterogeneity, we introduce controlled data generation strategies that produce non-independent and identically distributed (non-IID) datasets and employ the 1-Wasserstein distance to quantify distributional differences. We further examine how varying the number of participating clients influences the level of heterogeneity. The results show that the proposed FedSciML framework effectively calibrates the constitutive model by identifying a unified parameter set that accurately reproduces experimental observations. In addition, the framework demonstrates strong predictive capability in learning material behavior even under severe data scarcity and highly non-IID conditions. Overall, both the global federated model and the local client models achieve accurate extrapolation across unseen regimes, without requiring direct data sharing. This highlights the robustness and practical relevance of the proposed approach.