Physics-inspired machine learning of localized intensive properties.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Machine learning (ML) is widely used for predicting chemical properties, particularly energies and forces in molecules and materials.
• Current atomistic ML models often use a 'local energy' approach for size-extensivity and linear scaling, which can introduce errors for localized electronic properties.

Purpose of the Study:

• To explore strategies for accurately learning intensive and localized properties using machine learning.
• To address the limitations of size-extensive models for predicting properties like excitation or ionization energies.

Main Methods:

• Investigated different pooling functions in atomistic neural networks for molecular property prediction.
• Developed and tested an orbital weighted average (OWA) approach.

Main Results:

• The orbital weighted average (OWA) approach demonstrated accurate prediction of orbital energies.
• The OWA method effectively identified the locations of specific electronic properties within molecules.
• Analysis of pooling functions revealed their impact on predicting localized properties.

Conclusions:

• Standard size-extensive ML models are not optimal for all electronic properties, especially localized ones.
• The proposed orbital weighted average (OWA) method offers a more accurate way to predict localized orbital energies.
• This work provides a new strategy for enhancing machine learning models in computational chemistry for specific property predictions.