Machine learned calibrations to high-throughput molecular excited state calculations

Area of Science:

• Computational chemistry
• Photochemistry
• Machine learning

Background:

• Understanding molecular excited state properties is crucial for designing light-interacting compounds.
• Experimental methods for chemical discovery are time- and resource-intensive.
• High-throughput virtual screening (HTVS) uses computational chemistry to screen large databases, but initial low-accuracy methods can lead to high false positive/negative rates.

Purpose of the Study:

• To calibrate a high-throughput computational method (xTB-sTDA) against a higher accuracy method (TD-DFT) using machine learning.
• To reduce errors in HTVS for excited state property prediction.
• To enable cost-effective and accurate large-scale excited state screening for accelerated molecular discovery.

Main Methods:

• Machine learning (ML) was employed to calibrate the eXtended Tight Binding based simplified Tamm-Dancoff approximation (xTB-sTDA) against time-dependent density functional theory (TD-DFT).
• The calibrated model was tested for in-domain and out-of-domain accuracy.
• The ML-calibrated model was applied to screen a 250,000-molecule database and validated by calibrating against a higher-level method (CC2).

Main Results:

• The ML calibration achieved an approximately sixfold decrease in in-domain error and a threefold decrease in out-of-domain error.
• The resulting mean absolute error of ~0.14 eV is competitive with existing ML calibrations and superior to linear calibration of xTB-sTDA.
• Application to a large database revealed inaccuracies of xTB-sTDA across chemical space, and the workflow demonstrated generalizability with CC2 calibration.

Conclusions:

• Machine learning provides a cost-effective and accurate approach for large-scale excited state screening.
• This ML-driven calibration enhances the reliability of HTVS, reducing false positives/negatives.
• The developed workflow accelerates molecular discovery for photochemical applications and other disciplines reliant on excited state properties.