CAREER: Energy-Efficient and Compact Foundation Model for Universal Virtual Sensing in Dynamic Energy Systems

This NSF CAREER project aims to develop a new class of energy-efficient artificial intelligence (AI) models that can infer the internal behavior of complex energy systems from limited sensor measurements. Many critical infrastructures, including power grids and thermal energy systems, contain regions that cannot be directly instrumented, yet safe and reliable operation requires knowledge of internal conditions such as temperature, voltage, and power balance. Existing AI methods are often computationally intensive, tailored to a single system or component, and challenging to deploy in power-constrained environments such as substations, industrial facilities, and remote monitoring stations. The project will bring transformative change by creating compact foundation models that learn the governing structure of dynamic energy systems and reconstruct unmeasured states in real-time while operating within strict energy budgets on facility-grade or fog-level compute. This will be achieved by integrating physics-aware learning methods, adaptive modeling strategies, and energy-efficient architectures. The intellectual merit of the project includes advancing the theoretical foundations for compact foundation models that generalize across heterogeneous dynamic energy systems and establishing principles for energy-efficient AI. The broader impacts of the project include improving the reliability and resilience of national energy infrastructure, enabling real-time monitoring of critical systems with limited sensing coverage, developing open-access educational curricula on AI for energy systems, and training the next generation of students at the intersection of artificial intelligence, energy engineering, and infrastructure resilience. The project addresses a fundamental challenge in modern energy infrastructure: many systems are governed by differential–algebraic equations, which couple dynamic physical processes with network constraints, as seen in electric power grids where dynamic generator behavior must satisfy algebraic power-flow balance constraints. Traditional monitoring and simulation approaches are often too computationally expensive for real-time deployment and too specialized to generalize across different energy technologies. This research develops a compact neural-operator foundation model that learns mappings between sparse sensor observations and full system state fields while incorporating physical constraints through topology-aware representations and constraint-aware calibration. A neuroscience-inspired spiking neural layer enables event-driven inference that reduces energy consumption relative to conventional neural networks, supporting deployment on embedded edge platforms. The framework will be validated on representative dynamic energy systems, including power-grid state estimation using sparse phasor measurement unit observations, together with hardware-in-the-loop experiments that demonstrate real-time operation under strict power constraints. Together, these CAREER activities will establish energy-efficient foundation models that can be deployed across diverse energy infrastructures without reliance on centralized cloud computing. 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. NSF Award ID: 2543177 | Program: 01002627DB NSF RESEARCH & RELATED ACTIVIT | Principal Investigator: Syed Bahauddin Alam | Institution: University of Illinois at Urbana-Champaign, URBANA, IL | Award Amount: $587,981 View on NSF Award Search: https://www.nsf.gov/awardsearch/show-award/?AWD_ID=2543177 View on Research.gov: https://www.research.gov/awardapi-service/v1/awards/2543177.html