Cutting Through the Noise: Extracting Dynamics from Ultrafast Spectra using Dynamic Mode Decomposition

Coherent multidimensional spectroscopy provides experimental access to molecular structure and sub-picosecond dynamics in solution. Dynamics are typically inferred from the evolution of lineshapes over was a function of waiting time. Numerous spectral analysis methods, such as center/nodal line slope, have been developed to extract these dynamics. However, the extracted dynamics can depend heavily on subjective choices such as the region selected for CLS analysis or the chosen models. In this study, we introduce a novel approach to extracting dynamics from ultrafast two-dimensional infrared (2D IR) spectra using dynamic mode decomposition (DMD). As a data-driven method, DMD directly extracts spatio-temporal structures from the complex 2D IR spectra. We evaluate the performance of DMD in synthetic and experimental spectra containing overlapped peaks. We show that, DMD can retrieve the dynamics of overlapped transitions and cross peaks that are typically challenging to extract with traditional methods. In addition, we demonstrate that combining conditional generative adversarial neural networks (cGANNs) with DMD can recover dynamics even at low signal-to-noise ratios. DMD methods do not require preliminary assumptions and can be readily extended to other multidimensional spectroscopies.