CNN-Based Quantitative Lineshape Analysis in Two-Dimensional Coherent Spectroscopy for Many-Body Interactions and Frequency-Dependent Homogeneous Linewidths

Quantitative lineshape analysis in two-dimensional coherent spectroscopy (2DCS) is essential for extracting parameters such as dephasing rates, inhomogeneity, and many-body interactions [1]. Early analytical approaches [2] established methods to separate homogeneous and inhomogeneous contributions, while later developments incorporated machine learning for linewidth extraction [3]. However, these methods become untractable when the homogeneous response is modified by many-body effects or varies with frequency near a mobility edge in prescence of non-Gaussian inhomogeneity. To address this challenge, we developed convolutional neural networks (CNNs) trained on simulated spectra generated from generalized inhomogeneity using the bivariate spectral distribution function (BSDF) [4]. One model captures many-body interactions, and another learns frequency-dependent homogeneous linewidths. Since the training data encode the nonlinear optical response, the CNNs inherently learn physically meaningful features, enabling advanced lineshape analysis of complex 2D spectra beyond analytical limits.

References

[1] R. Singh et al., Phys. Rev. B 94, 081304(R) (2016).

[2] M. E. Siemens et al., Opt. Express 18, 17699 (2010).

[3] S. Namuduri et al., J. Opt. Soc. Am. B 37, 1587 (2020).

[4] B. De et al., Opt. Lett. 50, 4502 (2025).