Predicting phosphorescence energies and inferring wavefunction localization with machine learning.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Phosphorescence is crucial for optoelectronic devices like LEDs and photovoltaics.
• Accurate prediction of singlet-triplet energy gaps is key to discovering new phosphorescent materials.
• Current machine learning models struggle with predicting these gaps due to neglecting spin transition spatial locality.

Purpose of the Study:

• To develop an improved machine learning approach for predicting singlet-triplet energy gaps.
• To address the limitations of standard models in capturing the spatial nature of spin transitions.
• To enhance the discovery of phosphorescent compounds with targeted emission energies.

Main Methods:

• Introduction of novel 'localization layers' into neural network architectures.
• Weighting atomic contributions to accurately model energy gaps.
• Training the model on singlet-triplet energy gaps of organic molecules.
• Applying the enhanced model to predict phosphorescence energies of larger compounds.

Main Results:

• The new model significantly improves the prediction accuracy of phosphorescence energies.
• Localization layers effectively identify critical chemical environments influencing spin transitions.
• Inferred localization weights correlate strongly with ab initio spin density, revealing transition localities without explicit electronic input.

Conclusions:

• Localization layers offer a powerful method for modeling localized phenomena in materials science.
• This approach enhances the predictive accuracy of machine learning models for phosphorescent materials.
• The technique is adaptable to various atom-centered neural network models for broader applications.