Electronic Descriptors for Supervised Spectroscopic Predictions.

Area of Science:

• Computational Chemistry
• Quantum Chemistry
• Machine Learning Applications

Background:

• Molecular spectroscopic properties are crucial for understanding responses to UV/vis radiation.
• Traditional methods like ab initio and TDDFT are computationally expensive.
• Existing machine learning approaches using only geometric descriptors lack accuracy.

Purpose of the Study:

• To develop a supervised machine learning model for predicting organic molecule absorption spectra.
• To identify effective electronic descriptors for accurate spectral prediction.
• To achieve results comparable to traditional quantum chemistry methods.

Main Methods:

• Utilized supervised machine learning algorithms including Kernel Ridge Regression (KRR), Multiperceptron Neural Networks (MLP), and Convolutional Neural Networks.
• Employed electronic descriptors derived from low-cost DFT calculations: orbital energy differences, transition dipole moments, and charge-transfer character.
• Tested the efficacy of geometric-atomic number descriptors (e.g., Coulomb Matrix) and found them insufficient.

Main Results:

• Neural networks combined with electronic descriptors accurately predict excited state densities and absorption spectra.
• The model provides a very good estimation of charge-transfer character in electronic excited states.
• Achieved results close to chemical accuracy (approximately 2 kcal/mol or 0.1 eV).

Conclusions:

• Machine learning, utilizing DFT-derived electronic descriptors, offers an efficient alternative to computationally expensive methods for predicting molecular spectra.
• The proposed method accurately models absorption spectra and charge-transfer properties of organic molecules.
• This approach holds significant potential for advancing computational spectroscopy.