Simultaneous learning of static and dynamic charges.

Area of Science:

• Computational chemistry
• Atomistic machine learning
• Quantum mechanics

Background:

• Accurate modeling of condensed-phase systems requires capturing long-range interactions and electric response.
• Static charges (Coulomb interactions) and dynamic charges (atomic polar tensors/Born effective charges) represent these phenomena.
• Efficiently learning both charge types within a single model is a significant challenge.

Purpose of the Study:

• To critically compare different approaches for learning static and dynamic charges simultaneously.
• To evaluate the impact of coupling strategies and dielectric screening on model accuracy and computational cost.
• To determine the most practical modeling approach for condensed-phase and cluster systems.

Main Methods:

• Comparison of three approaches: independent charge learning, coupled learning with global screening, and coupled learning with environment-dependent screening.
• Utilizing bulk water and water clusters as test systems.
• Analyzing accuracy and computational cost for each approach.

Main Results:

• Coupled learning requires dielectric screening correction, which is complex in heterogeneous systems.
• Learned, environment-dependent screening improves dynamic charge accuracy but offers negligible gain over independent predictions.
• Coupled approaches increase computational cost compared to independent models.

Conclusions:

• Despite theoretical links, independent modeling of static and dynamic charges is the more practical and computationally efficient choice.
• This holds true for both condensed-phase and isolated cluster systems.
• Atomistic machine learning models benefit from simpler, independent charge learning strategies.