Modeling Zinc Complexes Using Neural Networks.

Area of Science:

• Computational chemistry
• Materials science
• Drug discovery

Background:

• Accurate modeling of molecular energetics is crucial for chemical and biological studies.
• Deep learning accelerates the development of quantum chemistry models for potential energy surfaces.
• Existing deep learning models primarily focus on organic molecules due to data availability and simpler electronic structures.

Purpose of the Study:

• To develop a deep learning architecture for modeling the energetics of zinc organometallic complexes.
• To address the limitations of traditional sampling methods for complex molecular systems.
• To improve the accuracy of predicting relative energies for zinc conformers.

Main Methods:

• Compiled a diverse dataset of zinc complexes using metadynamics for enhanced sampling.
• Developed a novel deep learning architecture incorporating partial charges to model long-range interactions.
• Utilized neural network potentials for energy calculations.

Main Results:

• The deep learning model accurately predicts the energetics of zinc organometallic complexes.
• Partial charges were identified as critical for modeling long-range interactions in neural network potentials for zinc complexes.
• The developed model achieved a mean absolute error (MAE) of 1.32 kcal/mol compared to the double-hybrid PWPB95 method.

Conclusions:

• Deep learning offers a powerful approach for modeling the energetics of complex organometallic systems.
• The inclusion of partial charges significantly enhances the accuracy of neural network potentials for zinc complexes.
• This work advances the exploration of chemical space for organometallic compounds, outperforming semiempirical methods.