SBIR Phase I: Physics-Informed Control and Optimization of Thermochemical Reactor Arrays

The broader/ commercial impact of this Small Business Innovation Research (SBIR) Phase I project is to enable more efficient conversion of secondary organic materials into value-added products by optimizing the monitoring and operation of small-scale industrial reactors. Current secondary organic materials processing systems often rely on sparse, noisy sensor data, necessitating manual adjustments that decrease efficiency and inflate operational costs. This project addresses the challenge of extracting actionable insights from imperfect data in harsh settings by utilizing physics-informed neural networks (PINNs). By advancing fundamental methods for learning from real-world industrial data, this research supports national priorities such as economic development and energy independence. This technology will improve reliability across a broad spectrum of distributed energy and manufacturing systems, providing a scalable solution for durable resource management. It also has the potential to mitigate the risk of catastrophic wildfires. This project will pursue a first-of-a-kind approach to learning reliable physical behavior from sparse and potentially unreliable sensor data from an actual fleet of chemical reactors already deployed in the field, where conventional data-driven models often fail. The central technical risk lies in embedding physical constraints directly into machine learning models so that they can generalize across varying operating conditions while remaining robust to missing or corrupted measurements. The scope of the Phase I research is to establish the feasibility of physics-constrained learning methods for inferring unmeasured internal reactor states and supporting operational decision-making in small-scale thermochemical systems. The intellectual contribution of this work is the development and evaluation of learning frameworks and techniques that combine simplified physical models with data-driven representations, allowing models to respect conservation laws and known process behavior while adapting to real-world variability. The project will investigate methods for conditioning these models on changing boundary and initial conditions, identifying and filtering faulty sensor data, and quantifying uncertainty in inferred states. Performance will be assessed using quantitative criteria tied to predictive accuracy, robustness to noise, and consistency with known physical trends. Successful completion of Phase I will demonstrate that physically informed learning can extract decision-grade information from limited data, laying the groundwork for scalable optimization and control methods in Phase II and beyond. 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.