A Differentiable Neural-Network Force Field for Ionic Liquids.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Accurate modeling of ionic liquid potential energy is crucial for understanding their behavior.
• First-principles calculations are computationally expensive, limiting their application.
• Developing efficient and accurate models is essential for advancing ionic liquid research.

Purpose of the Study:

• To develop NeuralIL, a novel model for predicting the potential energy of ionic liquids.
• To achieve high accuracy comparable to first-principles methods with significantly reduced computational cost.
• To enable efficient prediction of arbitrary derivatives of the potential energy.

Main Methods:

• Implementation of a multilayer perceptron with spherical Bessel descriptors for atomic environments.
• Utilizing fully automatically differentiable programming for training on ab initio forces and energies.
• Employing the Swish-1 activation function for maintaining neural network differentiability.
• Encoding element-specific density within spherical Bessel descriptors.

Main Results:

• Achieved out-of-sample accuracies better than 2 meV atom^-1 for energies and 70 meV Å^-1 for forces on ethylammonium nitrate.
• Demonstrated that encoding element-specific density is key to descriptor accuracy.
• Showed that training on forces drastically reduces data requirements for neural network force fields.
• Found that separate treatment of long-range interactions is unnecessary for dense ionic systems.

Conclusions:

• NeuralIL offers a computationally efficient and accurate method for modeling ionic liquid potential energies.
• The model's ability to train on forces significantly enhances data efficiency.
• Ensemble learning can be used for extrapolation detection with small datasets.
• NeuralIL provides a high-quality representation of the potential energy surface without explicit long-range interaction treatment.