PDaSP Track 3: Privacy-Preserving Dairy-Digitalization with Federated Learning

Modern dairy farming generates enormous amounts of valuable information about milk production, animal health, and farm operations, but this data remains scattered across individual farms and difficult to share due to privacy concerns and technical barriers. This fragmentation prevents farmers from benefiting from advanced computer models that could help them test different management strategies and improve their operations without disrupting actual farm activities or animal welfare. Currently, farmers, especially those with smaller operations, lack access to sophisticated data analysis tools that could help them make better decisions about their herds. This project addresses this challenge by developing a secure digital platform that allows farmers to collaboratively benefit from advanced artificial intelligence without ever sharing their private farm data. The platform will include user-friendly computer assistants that can understand everyday language and provide personalized, data-driven recommendations to help farmers improve their operations. This work serves the national interest by promoting scientific progress in agriculture, advancing national prosperity and welfare through more efficient and sustainable food production, strengthening agricultural data security to protect critical food systems, and supporting the competitiveness of American agriculture in global markets. This project develops a comprehensive, privacy-preserving digital testbed to optimize decision-making in dairy farming through federated-learning empowered digital twins integrated with fine-tuned large language models and retrieval-augmented generation for both centralized policy-making and localized end-user decision support. The research activities include four key components. First, the team will design new or adopt existing ontologies to support interoperability across heterogeneous data sources, facilitating integration of on-farm data with historical datasets and existing artificial intelligence frameworks. Second, the project will extend existing digital twin models to cover additional phenotypes using multimodal data including genomics, sensor streams, and cow history, adapting them for modern artificial intelligence architectures. Third, the research will leverage an established network of dairy farms across Europe and the United States to implement federated training of digital twins, comparing parallel and sequential federated learning schemes against a reference standard model, with evaluation metrics including training time, convergence rates, and communication overhead. During federated training, advanced aggregation techniques will be evaluated and employed, and anomaly detection will be integrated to enhance system security. Fourth, the project will fine-tune large language models using federated instruction tuning and value alignment to support both centralized and localized applications, using parameter-efficient tuning techniques to reduce computational burden. The fine-tuned models will be integrated into retrieval-augmented generation pipelines, enabling language model agents to retrieve and reason over both local farm databases and centralized scientific repositories. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.