Teaching a neural network to attach and detach electrons from molecules

Area of Science:

• Computational Chemistry
• Materials Science
• Quantum Mechanics

Background:

• Machine learning (ML) models, specifically Deep-Neural Networks (DNNs), offer quantum mechanical (QM) accuracy for simulations.
• Current DNN potentials are limited to neutral or closed-shell ions due to architectural constraints.
• Simulating open-shell ions is crucial for understanding various chemical processes.

Purpose of the Study:

• To develop an improved ML framework for simulating open-shell anions and cations.
• Introduce the AIMNet-NSE (Neural Spin Equilibration) architecture.
• Enable accurate prediction of molecular energies and properties for charged species.

Main Methods:

• Developed the AIMNet-NSE architecture for predicting molecular energies across various charge and spin states.
• Validated the model against reference quantum mechanical (QM) simulations.
• Utilized learned atomic representations and derived descriptors for reactivity modeling.

Main Results:

• AIMNet-NSE predicts molecular energies with 2-3 kcal/mol error and spin-charges with ~0.01e error for organic molecules.
• The model accurately bypasses QM calculations for deriving ionization potentials and electron affinities.
• Demonstrated the utility of ML-derived descriptors in modeling chemical reactivity, such as electrophilic aromatic substitution regioselectivity.

Conclusions:

• The AIMNet-NSE framework significantly advances ML capabilities for simulating open-shell molecular systems.
• This approach enables efficient and accurate computation of key chemical properties and reactivity descriptors.
• The developed model opens new avenues for large-scale simulations in chemistry and materials science.