Retrieving genuine nonlinear Raman responses in ultrafast spectroscopy via deep learning

Noise manifests ubiquitously in nonlinear spectroscopy, where multiple sources contribute to experimental signals generating interrelated unwanted components, from random point-wise fluctuations to structured baseline signals. Mitigating strategies are usually heuristic, depending on subjective biases such as the setting of parameters in data analysis algorithms and the removal order of the unwanted components. We propose a data-driven frequency-domain denoiser based on a convolutional neural network to extract authentic vibrational features from a nonlinear background in noisy spectroscopic raw data. The different spectral scales in the problem are treated in parallel by means of filters with multiple kernel sizes, which allow the receptive field of the network to adapt to the informative features in the spectra. We test our approach by retrieving asymmetric peaks in stimulated Raman spectroscopy, an ideal test-bed due to its intrinsic complex spectral features combined with a strong background signal. By using a theoretical perturbative toolbox, we efficiently train the network with simulated datasets resembling the statistical properties and lineshapes of the experimental spectra. The developed algorithm is successfully applied to experimental data to obtain noise- and background-free stimulated Raman spectra of organic molecules and prototypical heme proteins.