Collaborative Research: III: Small: Physics-Informed Probabilistic Forecasting for Risk-Based System Decisions in Small-Data Environments

About This Grant

Many safety-critical engineered systems must remain reliable and safe throughout their intended lifetime to support both public safety and economic stability. System diagnostics plays a vital role in predicting future conditions and guide decisions to prevent unexpected failures or mitigate impacts. However, the existing forecasting methods often depend on large amounts of experimental data or real-world system failure data, which are difficult to obtain, especially for safety-critical systems, where laboratory degradation experiments are time-consuming and field failure events are rare. There are two more key challenges: First, physics-based simulations of how system degrade over time can differ significantly from what actually happens in the real world, especially when a system is severely deteriorated. Second, online monitoring data from a specific system unit may be limited, showing little or no noticeable signs of damage early on. The project aims to develop algorithms and tools that enable reliable system forecasting even in small-data environments; that is, environments that are characterized by imperfect physics, scarce experimental data, and system monitoring data. By probabilistic lifetime prediction, these tools will support rapid, risk-based decisions on quality control and maintenance long before degradation becomes obvious. This project will also promote learning through open-source tools and modules and support STEM education through education and outreach activities. This project will develop a physics-informed probabilistic prognostics platform called Modular Analytics for Prognostics with Small Data, which comprises three main modules. (1) It will derive degradation models from physics-based simulations using sparse symbolic regression and recover the unmodeled degradation physics in the derived models through residual learning under uncertainty. (2) It will enable bi-level degradation model updating at the population and unit levels to achieve unit-specific, probabilistic health forecasting from early-life monitoring data. (3) It will create a degradation-aware policy optimization framework to integrate early-life health forecasts with downstream decision making. Overall, this research lays the foundation for smarter early-life health management strategies that lower life cycle costs and extend service life, and in many cases, promote sustainable practices across industries. 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.