Reactive Chemistry at the Unrestricted Coupled Cluster Level: High-Throughput Calculations for Training Machine...

Area of Science:

• Computational chemistry
• Quantum chemistry
• Materials science

Background:

• Accurate atomistic modeling of chemical reactions requires high-level electronic structure theory.
• Common methods like density functional theory (DFT) have limitations in describing bond breaking/formation energetics.
• Generating large datasets for machine learning interatomic potentials (MLIPs) using high-fidelity methods is computationally challenging.

Purpose of the Study:

• To develop novel methods and workflows for automating unrestricted coupled cluster calculations.
• To create a high-quality dataset of energies and forces for gas-phase reactions using unrestricted CCSD(T).
• To develop and evaluate a transferable MLIP trained on unrestricted CCSD(T) data for improved accuracy in chemical reaction modeling.

Main Methods:

• Development of new computational methods and workflows to automate unrestricted coupled cluster calculations.
• Calculation of energies and forces for 3119 organic molecule configurations using unrestricted CCSD(T).
• Training and validation of a machine learning interatomic potential (MLIP) on the generated unrestricted CCSD(T) dataset.

Main Results:

• A comprehensive dataset of energies and forces for gas-phase reactions was generated using the gold-standard unrestricted CCSD(T) level of theory.
• Analysis revealed significant differences between DFT and unrestricted CCSD(T) descriptions of chemical reactions.
• The MLIP trained on unrestricted CCSD(T) data demonstrated superior accuracy in force prediction and activation energy reproduction compared to DFT-trained models.

Conclusions:

• The developed methods overcome the computational challenges of unrestricted coupled cluster calculations, enabling the creation of high-fidelity datasets.
• Transitioning from DFT to unrestricted CCSD(T) data for MLIP training significantly enhances predictive accuracy for chemical reaction energetics.
• This work paves the way for more reliable MLIPs in computational chemistry and materials science.