Submitted:
17 November 2026
Posted:
18 November 2026
You are already at the latest version
Abstract
Accurate temperature measurements are essential in many industrial processes, yet the surface of interest often cannot host a sensor. The temperature must then be inferred from a measurement taken behind a layer of solid material, which amounts to reversing heat diffusion, an ill-posed problem in which noise is amplified. this work presents a physics informed model for temperature reconstruction with a sensor in which the sensitive element sits behind an electrically insulating layer, designed for embedded hardware and small training sets. The specific sensor used for data collection is composed by multiple alumina layers and thermocouples, as explained in detail in the dedicated section. We propose the Regularization-Phase Model (RPM), which keeps the closed form of Tikhonov regularization in Fourier space and learns two real quantities per frequency, a regularization that may be negative and a phase. The regularization accounts for noise and for errors in the amplitude of the forward model, while the phase accounts for delays and for noise correlated with the signal. The optimal parameters are shown to be unique and available in closed form, and the resulting operator coincides with the linear minimum mean square error estimator. Experimental results on synthetic and measured data show that RPM outperforms Tikhonov regularization when the forward model is inexact or the noise is correlated, at the cost of a Fast Fourier Transform, and that sensor delays can be read from the learned phase. Our study shows how physics can prescribe the structure of a learned model, leading to frugal and interpretable algorithms for smart temperature sensing.
Keywords:
inverse heat conduction
; Tikhonov regularization
; physics-informed machine learning
; Ill-posed inverse problems
; LMMSE
; soft sensing
Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.