Federated and Physics-Informed AI Models for Real-Time Bio-Nano Digital Twins Using IoBNT
- Publication Type:
- Thesis
- Issue Date:
- 2025
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
Digital twinning in biological systems poses significant challenges due to the highly variable, distributed, and privacy-sensitive nature of biomedical data. Constructing scalable, accurate, and secure digital models that faithfully replicate biological processes demands advanced computational strategies capable of operating across decentralized environments. This thesis addresses the challenge of building a unified and privacy-preserving framework for high-fidelity Digital Twins (DTs) of biological systems, where heterogeneous data and complex physiology require both data-driven and physics-guided modeling, and investigates federated and physics-informed Artificial Intelligence (AI) approaches to achieve this goal. The proposed models are built through the integration of Physics-Informed Neural Networks (PINNs), Federated Learning (FL), Internet of Bio-Nano Things (IoBNT), and advanced deep learning architectures, including Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Fully Connected Neural Networks (FCNNs), and residual neural networks (ResNets). To meet this central challenge, three interrelated frameworks are developed, each extending the previous one to progressively enable scalable, interpretable, and privacy-aware DT construction. The first proposed framework combines FL and CNNs to construct bacterial DTs using data acquired via IoBNT. The second proposed framework employs a hybrid PINN-based architecture, integrated with RNNs, FCNNs, and ResNets, to enable digital twinning of microbial growth in biosystems. Additionally, the third proposed FL-PINN framework is proposed for modeling glucose–insulin dynamics in metabolic regulation using IoBNT-derived signals. Experimental evaluations confirm that the proposed frameworks deliver high accuracy, reliability, and scalability in digital twinning applications. Across all three frameworks, the proposed algorithms reduced data-transfer errors by up to 98%, improved prediction accuracy to a mean absolute error of approximately 0.03, and lowered RMSE to about 0.04, outperforming existing physics-based or purely data-driven methods by margins exceeding 20-30% on benchmark datasets. These gains directly translate to faster convergence and lower communication overhead, which is critical for real-time biomedical DT deployment and clinical decision support. These results ensure consistent predictive stability across clients, reduce communication overhead, and preserve data privacy. Collectively, these integrated contributions establish a unified, future-oriented foundation for developing clinically actionable and privacy-conscious Digital Twins within biotechnology and digital health.
Please use this identifier to cite or link to this item: