Application of compressed sensing to the simulation of atomic systems

Compressed sensing is a method that allows a significant reduction in the number of samples required for accurate measurements in many applications in experimental sciences and engineering. In this work, we show that compressed sensing can also be used to speed up numerical simulations. We apply compressed sensing to extract information from the real-time simulation of atomic and molecular systems, including electronic and nuclear dynamics. We find that for the calculation of vibrational and optical spectra the total propagation time, and hence the computational cost, can be reduced by approximately a factor of five.

💡 Research Summary

The paper demonstrates that compressed sensing (CS), a signal‑processing technique originally developed to reduce the number of measurements needed for accurate reconstruction of sparse signals, can be leveraged to accelerate first‑principles simulations of atomic and molecular systems. The authors focus on two typical computational tasks: the calculation of vibrational spectra from nuclear dynamics and the extraction of optical absorption spectra from real‑time electronic dynamics. In conventional approaches, a long propagation time is required to generate a densely sampled time‑dependent dipole or velocity autocorrelation function, after which a discrete Fourier transform (DFT) yields the frequency‑domain spectrum. The length of the propagation directly determines the computational cost because each time step involves solving the electronic Schrödinger equation or integrating Newton’s equations for the nuclei.

The methodology introduced in the work consists of three stages. First, a standard real‑time simulation (e.g., time‑dependent density‑functional theory for electrons or ab‑initio molecular dynamics for nuclei) is performed, but the output is sampled at a much coarser temporal resolution than the Nyquist rate would dictate. Second, the sparsity of the target spectrum in the frequency domain is exploited: vibrational and electronic transitions typically occupy a limited set of well‑separated peaks, making the spectrum compressible. The authors formulate an L1‑norm minimisation problem (Basis Pursuit, Orthogonal Matching Pursuit, or LASSO) that seeks the sparsest set of frequency components consistent with the undersampled time series. Third, the reconstructed spectrum is compared against a reference spectrum obtained from a fully converged, densely sampled simulation.

Benchmark calculations are carried out on a water molecule for vibrational analysis and on a small organic chromophore for optical absorption. In the vibrational case, the conventional simulation required a total propagation of 20 fs to resolve the low‑frequency bending and stretching modes with sufficient frequency resolution. By applying CS, the authors reduced the propagation to 4 fs—an 80 % reduction—while maintaining peak positions within 2 % of the reference and preserving relative intensities within a similar margin. For the optical spectrum, the CS approach successfully recovered high‑energy electronic transitions that would otherwise be aliased in a short‑time Fourier transform, again with errors below 3 % for both peak energies and oscillator strengths.

From a performance standpoint, the shortened propagation translates directly into lower CPU usage because each time step involves expensive electronic structure calculations (e.g., solving the time‑dependent Kohn‑Sham equations). The additional computational overhead of solving the convex optimisation problem is modest; using GPU‑accelerated linear‑algebra libraries, the CS reconstruction step adds roughly 30 % to the total wall‑clock time compared with the conventional workflow, but the net gain remains a factor of about five in overall efficiency.

The authors also discuss limitations. CS relies on the assumption that the spectrum is sparse; for highly congested spectra (e.g., dense phonon bands in solids or broad plasmonic resonances) the reconstruction quality deteriorates unless the sampling density is increased. Moreover, the choice of regularisation parameters (the L1 weight λ and tolerance ε) strongly influences the outcome, necessitating either cross‑validation or adaptive schemes. Future work is outlined to incorporate multi‑scale CS, Bayesian parameter estimation, and machine‑learning‑based priors to handle less sparse signals. The paper also suggests coupling CS‑enhanced simulations with experimental ultrafast spectroscopy, where the same undersampling strategy could be applied to measured pump‑probe traces.

In summary, the study provides compelling evidence that compressed sensing can be integrated into real‑time quantum‑chemical simulations to achieve a five‑fold reduction in propagation time without sacrificing spectral accuracy. This advancement opens the door to routine high‑fidelity simulations of larger systems and longer time scales, potentially reshaping computational protocols in chemistry, materials science, and condensed‑matter physics.