Ultra‐fast spectroscopy for <scp>high‐throughput</scp> and interactive quantum chemistry

Francesco Bosia, Thomas Weymuth, Markus ReiherInternational Journal of Quantum Chemistry 2022 · DOI

TL;DR

This paper introduces ultra-fast computational methods for calculating infrared (IR) and ultraviolet-visible (UV/Vis) spectra, enabling real-time spectroscopic feedback during molecular exploration. By developing efficient algorithms for vibrational and excited-state calculations, the authors aim to accelerate high-throughput screening and interactive quantum chemistry.

What Problem Does This Paper Address?

Calculating spectroscopic data (like IR and UV/Vis spectra) for molecules is crucial for understanding their structure, function, and reactivity. However, traditional methods are computationally expensive, making them impractical for high-throughput screening, exploring large molecular ensembles, or enabling interactive, real-time chemical simulations. This "accuracy-time dilemma" limits the speed at which chemists can explore molecular landscapes. The authors aim to bridge this gap by developing methods that provide sufficiently accurate spectral information at a significantly reduced computational cost, enabling interactive and automated chemical exploration.

Methods

The paper presents several key methodological advancements:

  • Ultra-fast IR Spectroscopy:
    • Approximate Structure Optimization: Instead of full, rigorous geometry optimizations, the authors explore using looser convergence criteria. This significantly reduces computational time while maintaining acceptable accuracy for spectral features, especially for high-frequency modes.
  • Partial Hessian Approach: This method reduces the computational cost of calculating the Hessian matrix (which determines vibrational frequencies) by only re-evaluating elements corresponding to molecular fragments that have significantly changed between successive local minima on the potential energy surface. This is controlled by an RMSD threshold.
  • Ultra-fast UV/Vis Spectroscopy:
    • Non-orthogonal Davidson Algorithm: An adaptation of the Davidson algorithm for solving the eigenvalue problem in excited-state calculations, which is more robust for potentially singular overlap matrices.
  • Subspace Diagonalization and Determinant Selection: The authors employ a strategy to limit the size of the excited-state basis by selecting determinants that couple strongly with the ground state. This significantly reduces the computational burden of excited-state calculations.

Datasets

Not explicitly discussed in the paper as training datasets. The authors use molecular structures from simulated trajectories (T1, T2, and MD) for evaluating their methods. These trajectories represent chemical processes like keto-enol tautomerism and molecular dynamics simulations.

Key Insights and Ideas

  • Spectra as Fingerprints: The core idea is that characteristic spectral signals act as "fingerprints" for identifying and characterizing molecular structures, even if absolute accuracy in peak position and intensity isn't always required for initial exploration.
  • Accuracy-Time Trade-off Management: The methods are designed to manage the accuracy-time trade-off by allowing for approximate calculations that can be refined later if necessary. Uncertainty quantification is mentioned as a way to flag potentially erroneous predictions.
  • Leveraging Structural Similarity: The partial Hessian approach effectively exploits the similarity between successive molecular structures in a trajectory to avoid redundant calculations.
  • Basis Set Pruning for Efficiency: For UV/Vis spectroscopy, reducing the size of the excited-state basis set significantly accelerates calculations without a drastic loss in accuracy, especially for lower-energy transitions.

Key Results

  • Significant Speedups: The methods achieve substantial computational speedups for both IR and UV/Vis spectroscopy. For example, approximate structure optimization can lead to speedups of up to an order of magnitude for high-frequency spectral bands. The partial Hessian approach reduced Hessian calculation time from 2341 ms to 511 ms for a specific case.
  • Qualitative Recovery of Spectra: DFTB3, a semi-empirical method, was found to be a good candidate for IR spectroscopy, qualitatively recovering spectra calculated with more accurate DFT methods.
  • UV/Vis Spectra Comparison: The TD-DFTB3 method, with basis set pruning, showed good agreement with more accurate SCS-CC2 calculations for UV/Vis spectra, with a mean absolute deviation in normalized intensity of 0.35.
  • Robustness of Methods: The non-orthogonal Davidson algorithm and basis set pruning were shown to be effective for accelerating excited-state calculations.

Limitations and Future Work

  • Accuracy of Intensities: While DFTB3 qualitatively recovers IR spectra, it can overestimate intensities by factors of two or more for certain modes.
  • Global Rearrangements: The partial Hessian approach is less effective when local minima are connected by global structural rearrangements, as it relies on identifying similar fragments.
  • Initial Guess for Diagonalization: While strategies were explored to improve the initial guess for excited-state calculations, they did not consistently decrease the time to solution.
  • Future Work: The authors suggest extending their approaches to modern semi-empirical models like GFNn-xTB and orthogonalization-corrected methods. They also propose exploring improvements to the iterative alignment algorithm for the partial Hessian approach, such as partitioning the molecule into local fragments for independent alignment. Further investigation into the impact of basis set pruning on higher-energy excited states is also implied.

Summary updated

Abstract

Abstract We present ultra‐fast quantum chemical methods for the calculation of infrared and ultraviolet–visible spectra designed to provide fingerprint information during autonomous and interactive explorations of molecular structures. Characteristic spectral signals can serve as diagnostic probes for the identification and characterization of molecular structures. These features often do not require ultimate accuracy with respect to peak position and intensity, which alleviates the accuracy–time dilemma in ultra‐fast electronic structure methods. If approximate ultra‐fast algorithms are supplemented with an uncertainty quantification scheme for the detection of potentially large prediction errors in signal position and intensity, an offline refinement will always be possible to confirm or discard the predictions of the ultra‐fast approach. Here, we present ultra‐fast electronic structure methods for such a protocol in order to obtain ground‐ and excited‐state electronic energies, dipole moments, and their derivatives for real‐time applications in vibrational spectroscopy and photophysics. As part of this endeavor, we devise an information‐inheritance partial Hessian approach for vibrational spectroscopy, a tailored subspace diagonalization approach and a determinant‐selection scheme for excited‐state calculations.
BibTeX
@article{bosia2022ultrafast,
  author = {Francesco Bosia and Thomas Weymuth and Markus Reiher},
  title = {{Ultra‐fast spectroscopy for <scp>high‐throughput</scp> and interactive quantum chemistry}},
  journal = {International Journal of Quantum Chemistry},
  year = {2022},
  doi = {10.1002/qua.26966},
}
Figure