Intelligent material recognition from thermal signatures using hybrid deep learning frameworks

Abstract

Accurate material identification using thermal infrared signatures has significant implications for industrial inspection, autonomous systems, and non-contact diagnostics. In this study, we propose a physics-informed deep learning framework for intelligent material recognition based on transient thermal behavior. A large-scale dataset, ThermalMat-85K, comprising 85,000 long-wave infrared (LWIR) images across ten material classes, was constructed under controlled multi-angle, multi-distance, and multi-environment conditions. Each sample was validated through Fourier transform infrared (FT-IR) spectroscopy and X-ray diffraction (XRD) to ensure chemical and structural integrity prior to thermal acquisition. Seven deep learning architectures were systematically evaluated, including VGG-16, ResNet-50, MobileNet-V3, DenseNet-121, EfficientNet-B4, a CNN+LSTM temporal hybrid, and a proposed Hybrid Net integrating convolutional, transformer, and recurrent modules with cross-modal attention gating. EfficientNet-B4 achieved the highest Top-1 accuracy (93.1%) and macro F1-score (92.8%), demonstrating superior class-wise stability. Residual and dense connectivity architectures also exhibited strong performance, while lightweight and temporal hybrids revealed trade-offs between efficiency and discriminative power. Class-wise confusion analysis indicates that misclassification trends correlate strongly with thermophysical proximity, particularly thermal conductivity and emissivity overlap. These findings confirm that transient thermal decay dynamics encode physically meaningful features suitable for material discrimination. The results establish a robust benchmark for physics-informed thermal material recognition and provide guidance for selecting architectures based on deployment constraints and robustness requirements.