Synthetic data-driven machine learning for multi-dimensional spectroscopy of molecules

In previous work [1, 2] we have trained deep neural networks using synthetic Hamilton matrices (SHM) with quasi-random matrix elements. The matrices represent atomic and molecular systems (which could exist but do not necessarily exist), dipole-coupled to short and noisy laser pulses. This allows us to generate millions of spectra solving the dynamics required for successful training. The trained network is capable of "purifying" noisy nonlinear photo-electron spectra as if they had been generated by a Fourier-limited pulse.

Here, we investigate, to which extent it is possible to construct the dynamics of a system from measured spectra. In order to identify the network performance, range, and class of systems covered by SHM, here, we consider the setting of multi-dimensional spectroscopy for molecules. To this end, we reconstruct the SHM from the nonlinear spectra using a convolutional neural network where the training is performed with the spectra from synthetic systems and address the transferability of the trained network for the application in real molecular systems through the calculation of dynamics not known to the network, e.g., the ringing of its time-dependent dipole.

[1] Giri, Saalmann and Rost, PRL 124, 113201 (2020).

[2] Giri, Saalmann and Rost, Faraday Discuss. 228, 502 (2021)