A Spectral-Temporal Decoupling Framework for the High-Fidelity Reconstruction of Low-SNR Transient Absorption Spectra.

Transient absorption spectroscopy (TAS) is a cornerstone for investigating dynamical mechanisms in quantum dots, photovoltaics, and photosynthesis. A primary challenge in the field is the development of high-sensitivity techniques capable of probing species in low-signal regimes, such as single-molecule detection, spatially resolved TAS, and the tracking of short-lived intermediates. While advancements in instrumentation have pushed the physical limits of detection, extracting meaningful dynamics from noise-limited data remains a bottleneck. In this work, we propose a residual U-Net framework integrated with a spectral-temporal decoupling module, namely TS-ResUNet for denoising and reconstruction of transient maps. Quantitative evaluations demonstrate that TS-ResUNet consistently outperforms conventional algorithms and standard U-Net architectures in denoising, while maintaining high fidelity in the extracted spectral and kinetic information. Furthermore, sim-to-real transfer learning performed on experimental data sets indicates that effective adaptation is achievable with a small number of paired training data. This framework provides a robust methodology for refining low signal-to-noise ratio measurements and significantly accelerating data acquisition in ultrafast spectroscopy.