A machine learning correction for DFT non-covalent interactions based on the S22, S66 and X40 benchmark databases

Area of Science:

• Computational Chemistry
• Supramolecular Chemistry
• Machine Learning in Science

Background:

• Non-covalent interactions (NCIs) are crucial in supramolecular chemistry but challenging to quantify accurately.
• Existing computational methods for NCIs are either inaccurate at low theory levels or computationally expensive at high levels.
• There is a need for efficient and accurate computational approaches to study NCIs.

Purpose of the Study:

• To develop an efficient approach to correct non-covalent interaction calculations.
• To improve the accuracy of low-level theoretical calculations for NCIs.
• To reduce the computational cost associated with accurate NCI predictions.

Main Methods:

• A novel correction method for density functional theory (DFT) calculations using a general regression neural network machine learning approach.
• Investigation of various DFT methods (M06-2X, B3LYP, PBE, etc.) with small basis sets (6-31G*, 6-31+G*).
• Inclusion of solvent effects using the conductor-like polarizable continuum model (water and pentylamine solvent).

Main Results:

• The machine learning correction improved the root mean square errors of DFT calculations by at least 70%.
• Achieved a mean absolute error of 0.33 kcal/mol, comparable to high-level ab initio methods, at a fraction of the computational cost.
• Validation parameters (R², q², etc.) exceeded 0.92, indicating good model stability, robustness, and predictive power.

Conclusions:

• The proposed correction model can be easily applied post-DFT calculations to enhance NCI accuracy.
• The method introduces only one parameter, ensuring ease of application and flexibility.
• This machine learning correction offers a viable and efficient alternative to traditional methods for calculating NCIs.