Reproduce “Physics-Informed Neural Networks for Thermophysical Property Retrieval” on OpenResearch

About this paper

Ali Waseem, Malcolm Mielle

An iterative Physics-Informed Neural Network (PINN-it) framework estimates the thermal conductivity of building walls from non-invasive thermographic and environmental data. This method provides accurate thermal conductivity estimations, even with limited thermographs over short periods, addressing limitations of current time-intensive or invasive techniques for in-situ building assessments.

Key insights

• Physics-Informed Neural Networks can act as accurate surrogate forward models for transient heat transfer in building facades, consistently modeling temperature profiles under diverse seasonal and environmental conditions.
• An iterative PINN-based approach can reliably estimate unknown thermophysical properties (thermal conductivity) from sparse, non-invasive thermographic data, reducing the need for lengthy measurement campaigns.
• The framework demonstrates resilience to deviations from ideal initial conditions, suggesting practical applicability in real-world scenarios where perfect steady-state initialization is often not met.

Summary by alphaXiv.