Machine learning the electric field response of condensed phase systems using perturbed neural network potentials

Area of Science:

• Computational physics and chemistry
• Materials science
• Condensed matter physics

Background:

• Interactions between condensed phases and electric fields are crucial for natural and technological processes.
• Accurate molecular simulations of these interactions are computationally expensive, limiting current research scope.
• Existing methods like ab-initio molecular dynamics (AIMD) face significant computational challenges.

Purpose of the Study:

• To introduce a novel computational method, Perturbed Neural Network Potential Molecular Dynamics (PNNP MD), for simulating systems under electric fields.
• To extend the accessible time and length scales for molecular dynamics simulations involving electric fields.
• To enable accurate atomistic insights into diverse condensed phase systems interacting with external electric fields.

Main Methods:

• Development and application of Perturbed Neural Network Potential Molecular Dynamics (PNNP MD).
• Machine learning of dielectric properties using two neural networks trained on zero-field molecular dynamics data.
• Validation against ab-initio molecular dynamics (AIMD) for accuracy assessment.

Main Results:

• PNNP MD accurately machine learns dielectric properties of liquid water, including relaxation dynamics, dielectric constant, and field-dependent IR spectrum.
• The method achieves high accuracy even at strong electric field strengths (approx. 0.2 V Å⁻¹).
• Neural networks demonstrate reliable extrapolation capabilities for field response, trained solely on zero-field configurations.

Conclusions:

• PNNP MD offers a computationally efficient and accurate approach for simulating condensed phase systems in electric fields.
• The method overcomes limitations of traditional AIMD, enabling simulations at larger scales.
• PNNP MD is a versatile, modular, and improvable tool for gaining atomistic understanding of electric field interactions across various materials.