基于物理约束的分子光谱预测深度模型

Accurate modeling of mid-infrared vibrational-rotational transition spectra is pivotal for detecting biosignatures and prebiotic molecules in exoplanetary atmospheres. Conventional line-by-line radiative transfer methods encounter prohibitive computational costs when generating high-resolution spectra across extensive parameter spaces, particularly limiting real-time atmospheric retrieval and large-scale exoplanet surveys. To address these challenges, a physics-informed deep learning framework was developed for rapid and precise spectral generation. The architecture incorporates three key components: A parameter-encoding layer establishing global correlations between temperaturepressure conditions and spectral line parameters; Multi-head self-attention mechanisms capturing long-range dependencies in vibrational-rotational features; A physics-constrained decoder incorporating residual modules derived from line profile equations to minimize non-physical deviations. The framework demonstrated successful reconstruction of molecular absorption cross-sections from the HITRAN database at 0.01cm-1 resolution, achieving a 100× acceleration compared to conventional HAPI simulations while maintaining spectral fidelity. The framework accurately preserved fundamental spectroscopic principles, including line intensity scaling and rotational temperature dependencies, across diverse atmospheric conditions. This approach represents the first integration of spectroscopic constraints into neural network-based spectral generation, enabling interpretable temperature-pressure-spectral correlations and compatibility with photochemical network-driven biosignature assessments. The method now provides a computationally efficient solution for next-generation spectral databases, significantly advancing molecular characterization of exoplanetary environments and enhancing biosignature detection systems through photochemical network integration.