OpenTIR-PINN: A Benchmark Dataset of Synthetic Radiative Transfer, Sensor Response Functions, and Quarantined Mineral Endmembers for Physics-Informed Temperature and Emissivity Separation

OpenTIR-PINN is an open-science benchmark dataset designed for training, regularizing, and evaluating Physics-Informed Neural Networks (PINN) and numerical radiative transfer inversion algorithms on the ill-posed thermal infrared (TIR) Temperature and Emissivity Separation (TES) problem. Dataset Contents & Architecture:1. Multi-Temperature Synthetic Training Manifold: 8,134 at-sensor top-of-atmosphere (TOA) radiance spectra paired with surface kinetic temperatures (Ts ~ U(280, 330 K)) and directional-hemispherical emissivities, simulated across 4 standard MODTRAN 6 atmospheric profiles (Tropical, Mid-Latitude Summer, Mid-Latitude Winter, 1976 US Standard) covering Precipitable Water Vapor (PWV) from 0.5 to 4.5 g/cm².2. Quarantined Out-of-Distribution (OOD) Golden Targets: 7 chemically pure, structurally diverse terrestrial endmembers (Quartz, Plagioclase Feldspar/Anorthite, Kaolinite Clay, Calcite, Urban Concrete, Urban Asphalt, and Green Foliage) strictly quarantined prior to data augmentation or temperature-sweep pairing to evaluate true physical generalization without data leakage.3. Sensor Spectral Response Functions (SRFs): Full optical bandpass profiles and numerical calibration matrices for: - HyTES: 256 contiguous channels spanning 7.49 to 12.00 µm (JPL calibrated). - ASTER: 5 discrete thermal infrared bands (Bands 10–14: 8.28, 8.63, 9.08, 10.66, 11.29 µm, RTTOV calibrated).4. Full-Stack Experimental Reproducibility (MODTRAN 6): Complete Fortran input decks (Cards 1–5, tape5) and unformatted radiative transfer spectral tables (tape7) with an automated reproduction runner script.5. Raw Laboratory Spectra Baseline (Pre-Code): Continuous FTIR laboratory reflectance/emissivity spectra (~2,150 spectral points from 1.25 to 15.0 µm) for the 7 Golden Targets and 1,205 base library samples prior to sensor convolution and code execution.6. Differentiable Physics Engine: Vectorized forward radiative transfer operators supporting both NumPy arrays and PyTorch autograd tensors (Planck law, inverse brightness temperature, RTE surface coupling, and sensor filter convolution). Associated Publication:"Physics-Informed Neural Networks with Sobolev Shape Regularization for Ill-Posed Temperature and Emissivity Separation Across Multispectral and Hyperspectral Thermal Infrared Sensors".