EPFL's ThermoField decodes hidden heat properties from thermal video

Researchers at EPFL have developed ThermoField, a novel framework that goes beyond standard thermal imaging to infer hidden thermophysical properties from time-resolved thermal observations. Traditional thermal scene reconstruction can map temperature fields in complex 3D environments but fails to identify the underlying material properties that govern heat transfer. Inverse methods, on the other hand, estimate physical parameters but require simplified geometries and controlled conditions. ThermoField bridges this gap by integrating differentiable heat-transfer simulation with neural scene representations.

ThermoField represents spatially varying thermophysical properties—such as thermal diffusivity—as continuous neural fields. It constrains these fields using scene geometry, governing heat-transfer physics, and temporal thermal observations. The framework jointly reconstructs 3D geometry, estimates spatially varying thermal diffusivity, and predicts thermal evolution under previously unseen environmental conditions. This enables physically interpretable parameter inference in complex real-world scenes, opening applications in digital twins, infrastructure monitoring, robotics, and scientific imaging. The results establish a unified approach for geometry reconstruction, property estimation, and predictive thermal simulation.