Progress in automatic control of thermal processes and real-time estimation of heat penetration into live tissue has long been limited by the difficulty of obtaining high-fidelity thermodynamic models. Traditionally, in complex thermodynamic systems, it is often infeasible to estimate the thermophysical parameters of spatiotemporally varying processes, forcing the adoption of model-free control architectures. This comes at the cost of losing any robustness guarantees, and implies a need for extensive real-life testing. In recent years, however, infrared cameras and other thermographic equipment have become readily applicable to these processes, allowing for a real-time, non-invasive means of sensing the thermal state of a process. In this work, we present a novel physics-based approach to learning a thermal process's dynamics directly from such real-time thermographic data, while focusing attention on regions with high thermal activity. We call this process, which applies to any higher-dimensional scalar field, attention-based noise robust averaging (ANRA). Given a partial-differential equation model structure, we show that our approach is robust against noise, and can be used to initialize optimization routines to further refine parameter estimates. We demonstrate our method on several simulation examples, as well as by applying it to electrosurgical thermal response data on in vivo porcine skin tissue.