Machine-Learned Electronically Excited States with the MolOrbImage Generated from the Molecular Ground State

Area of Science:

• Computational Chemistry
• Quantum Mechanics
• Machine Learning

Background:

• Accurate prediction of molecular electronic state properties is crucial for understanding chemical phenomena.
• Traditional methods often face challenges in computational cost and accuracy, especially for excited states and oscillator strengths.
• Deep learning has shown promise in various scientific domains, including chemistry.

Purpose of the Study:

• To introduce a general machine learning framework for probing electronic state properties.
• To develop a novel quantum descriptor, MolOrbImage, for representing molecular orbital information.
• To implement and evaluate a deep convolutional neural network (MO-NN) model for predicting molecular properties.

Main Methods:

• Utilized a novel quantum descriptor, MolOrbImage, capturing quantum information from molecular orbital (MO) states.
• Developed a convolutional neural network (MO-NN) model inspired by computer vision architectures.
• Trained and tested the MO-NN model using orbital energy and electron repulsion integral MolOrbImages against ADC(2)/cc-pVTZ references.

Main Results:

• Achieved high prediction accuracy for transition energies to singlet (MAE < 0.16 eV) and triplet (MAE < 0.14 eV) states.
• Demonstrated significant improvement in predicting oscillator strengths, a previously challenging property.
• Showcased remarkable extrapolation capacity of the MO-NN model to systems outside the training dataset.

Conclusions:

• The MolOrbImage descriptor and MO-NN model provide a powerful and general machine learning framework for electronic structure calculations.
• The developed model offers accurate and efficient predictions of key molecular properties, including challenging ones like oscillator strengths.
• The model's transferability suggests its broad applicability to diverse chemical systems.