Speaker
Description
Calculating excited state spectra of large systems is often prohibitively expensive with standard frequency-domain methods such as the Casida equations, the Bethe-Salpeter Equation (BSE), or Equation-of-Motion Coupled Cluster (EOM-CC). Real-time methods provide an alternative, as all modes are excited simultaneously. However, long simulation times are required to resolve narrow spectral features with traditional Fourier signal analysis, significantly limiting system size. Super-resolution methods such as Compressed Sensing promise high-resolution spectra from much shorter signals but assume the spectrum to be sparse, an assumption which breaks down in larger systems where sharp features are embedded in a quasi-continuum of smaller nearby peaks. To overcome this, we combine newly designed highly noise-tolerant super-resolution techniques with physically motivated filtering. Using approximate frequency information, we identify the brightest transitions and extract only these important modes from the time propagation. By doing so, we effectively precondition the problem for super-resolution, reducing the number of required time steps for signal reconstruction to a minimum. We demonstrate our approach on systems containing several hundred heavy atoms, achieving up to 20-fold speedups while maintaining spectral accuracy even for signals dominated by large continua.