Using atomic clustering based on structural and electronic descriptors that consider surrounding environment to...

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Density Functional Theory (DFT) is crucial for predicting material properties.
• Accurate evaluation of DFT functionals is essential for reliable predictions.
• Existing methods may not capture the nuances of local atomic environments.

Purpose of the Study:

• To develop a detailed method for evaluating DFT functional accuracy.
• To leverage machine learning for atomic clustering based on descriptors.
• To enhance the assessment of 13C NMR chemical shift predictions.

Main Methods:

• Utilized a machine learning-based clustering approach.
• Employed structural and electronic descriptors for atom grouping.
• Analyzed 30,436 carbon atoms from the QM9 dataset, forming 36 intuitive clusters.
• Evaluated 84 DFT functionals for 13C NMR chemical shift calculations.

Main Results:

• Successfully generated 36 distinct atomic clusters reflecting similar chemical environments.
• Demonstrated that grouping atoms by environment reduces prediction errors.
• Showcased the capability of the method to provide detailed accuracy insights.
• Identified variations in DFT functional performance across different atomic clusters.

Conclusions:

• Atomic clustering based on machine learning offers a refined approach to DFT functional evaluation.
• This method provides more granular insights into the accuracy of DFT calculations for specific chemical environments.
• The developed technique enhances the reliability of predicting 13C NMR chemical shifts.