A Physics-Informed MobileNetV2 Framework for Thermal Infrared Gas Classification: Combining Thermodynamic Input Features and Loss-Level Physics Penalization

<title>Abstract</title> <p>Deep-learning models for thermal-infrared gas classification have thus far treated thermal images as ordinary photographs, discarding the fact that every pixel contains information on physically calibrated temperature governed by fixed thermodynamic law. This paper presents a Physics-Informed Machine Learning (PIML) system that classifies gas environments from thermal-infrared images using both physics-grounded input features and a physics-penalized training objective. Rather than utilizing the three-channel RGB representation used in all prior work on the MultimodalGasData thermal-image subset, in this work, each image is converted into a six-channel tensor comprising temperature-normalized RGB channels, horizontal and vertical thermal-gradient maps derived from Fourier's law of heat conduction via the Sobel operator, and a local temperature-anomaly map obtained by neighborhood-mean subtraction. A MobileNetV2 backbone with a custom input-projection block is trained in two phases; frozen-backbone feature adaptation followed by fine-tuning of the top thirty backbone layers — under a physics-informed loss function that penalizes confident gas predictions in thermally flat, physically implausible image regions. The resulting model attains 94.80% test accuracy and 96.67% best validation accuracy across four classes (No Gas, Perfume, Smoke, Mixture), with a Mixture-class F1-score of 99.58% and zero false Mixture alarms across 722 non-Mixture test images. Occlusion-based saliency analysis shows model attention concentrated in thermodynamically meaningful regions (the base of rising smoke plumes, diffuse upper-region warmth for aerosol dispersion, and multi-region activation for simultaneous emissions) thus providing the first physical-plausibility validation reported for this dataset. The model improves by 1.80 percentage points over a thermal-only CNN baseline lacking physics-informed features and closes to within 1.2 percentage points of a full multimodal sensor-fusion result, while trailing a purely accuracy-optimized MobileNetV2 baseline on standard preprocessing by 4.96 percentage points. This study provides a quantified interpretable trade-off between raw accuracy and physical grounding which is argued to be favorable for safety-critical deployment.</p>