Ultrafast Electronic Coupling Estimators: Neural Networks versus Physics-Based Approaches.

Area of Science:

• Computational Chemistry
• Materials Science
• Biophysics

Background:

• Accurate electronic coupling matrix elements are crucial for simulating charge transfer phenomena.
• Existing physics-based methods like the analytic overlap method (AOM) have limitations.

Purpose of the Study:

• To investigate and compare the performance of neural-network-based coupling estimators against the AOM.
• To evaluate different reference data sampling protocols for machine learning (ML) approaches.

Main Methods:

• Utilized neural network models for estimating electronic coupling.
• Employed various data sampling strategies: random, farthest point, and query by committee.
• Compared ML performance against the established analytic overlap method (AOM).
• Introduced a Δ-ML approach using AOM as a baseline.

Main Results:

• Neural network estimators achieved lower errors, especially maximum errors, compared to AOM.
• ML methods required significantly less training data (hundreds of points) than previous ML studies (thousands).
• The Δ-ML approach demonstrated superior performance with a modest computational overhead (factor of 2).
• Flexible π-conjugated molecules presented challenges for ML due to orbital delocalization.

Conclusions:

• Neural network-based methods show promise for accurate electronic coupling estimation.
• The Δ-ML approach offers an effective balance of accuracy and computational cost.
• Future ML developments should incorporate long-range descriptors to address challenges with flexible molecules.