Sign in Subscribe
Artificial intelligence

Decentralized Federated Learning Meets Physics-informed Neural Networks

Strategic Insights

— 1 min read

Research findings on Networks discuss the integration of domain knowledge into the artificial intelligence learning process, with a focus on decentralized settings. The University of Calabria researchers propose the Physics-Informed DFL (PIDFL) architecture to integrate domain knowledge expressed as differential equations. The proposed framework achieves on average over 40% lower test loss compared to the baseline DFLA and outperforms benchmark approaches.

Key Takeaways:

  • The integration of domain knowledge into decentralized federated learning has received significant attention, with most approaches focusing on centralized machine learning scenarios.
  • The proposed PIDFL architecture integrates domain knowledge expressed as differential equations, providing a promising avenue for improving decentralized learning.
  • The serverless data aggregation algorithm for PIDFL is proposed and its convergence is proven, with a discussion of its computational complexity.
  • Comprehensive experiments across various datasets demonstrate the effectiveness of PIDFL, with an average of over 40% lower test loss compared to the baseline DFLA.
  • The PIDFL framework outperforms benchmark approaches (FedAvg, SegGos, and Scaffold) across a variety of datasets.
  • The research highlights the potential of PIDFL in improving decentralized learning through domain knowledge integration.
  • The authors include Reza Shahbazian, Gianvincenzo Alfano, Sergio Greco, Domenico Mandaglio, Francesco Parisi, and Irina Trubitsyna from the University of Calabria.
  • The financial supporters for this research include PNRR MUR projects FAIR and PRIN MUR project S-PIC4CHU.

Statistics:

  • 40%: The average reduction in test loss achieved by the proposed PIDFL framework compared to the baseline DFLA.
  • 323: The issue number of the Knowledge-Based Systems journal where the research is published.
  • 770: The page number of the Journal of Engineering where the news of the research findings is published.
  • 2025: The year the research was conducted and published.

Sources:

  • Decentralized Federated Learning Meets Physics-informed Neural Networks. Knowledge-Based Systems, 2025;323.
  • University of Calabria. Knowledge-Based Systems can be contacted at: Elsevier, Radarweg 29, 1043 Nx Amsterdam, Netherlands. (Elsevier - www.elsevier.com; Knowledge-Based Systems - www.journals.elsevier.com/knowledge-based-systems/)
  • Reza Shahbazian, University of Calabria, Dept Informat Modeling Elect & Syst Engn Dimes, I-87036 Arcavacata Di Rende, Cs, Italy.
  • NewsRx LLC. Findings from University of Calabria Broaden Understanding of Networks (Decentralized Federated Learning Meets Physics-informed Neural Networks). Journal of Engineering. July 21, 2025; p 770.

Read more