Quantitative Prediction of Vertical Ionization Potentials from DFT via a Graph-Network-Based Delta Machine Learning...

Area of Science:

• Computational Chemistry
• Quantum Chemistry
• Machine Learning in Chemistry

Background:

• Accurate wave function theories (e.g., CCSD(T)) are computationally expensive for large systems.
• Density functional theory (DFT) is feasible but often lacks quantitative accuracy for electronic changes.
• There is a need for computationally efficient methods that achieve high accuracy in chemical process modeling.

Purpose of the Study:

• To develop an efficient delta machine learning (ΔML) model for accurate prediction of vertical ionization potentials.
• To integrate molecular fragmentation, error cancellation, and machine learning for enhanced accuracy.
• To improve upon the limitations of DFT in describing electronic changes during chemical processes.

Main Methods:

• Utilized the Connectivity-Based Hierarchy (CBH) scheme for systematic error correction based on molecular fragmentation.
• Developed an automated method for identifying ionization sites using electron population difference maps.
• Employed a graph-based quantum mechanics/machine learning (QM/ML) model incorporating atom-centered features of CBH fragments.
• Integrated electronic descriptors from DFT, specifically electron population difference features.

Main Results:

• The ΔML model achieved coupled cluster accuracy for vertical ionization potentials.
• The graph-based QM/ML model significantly increased prediction accuracy.
• Incorporating DFT-derived electronic descriptors improved model performance beyond chemical accuracy, approaching benchmark accuracy.
• The best models demonstrated robust performance, with reduced dependence on the specific DFT functional used.

Conclusions:

• The developed ΔML model effectively corrects DFT deficiencies for accurate vertical ionization potential calculations.
• The integration of CBH fragmentation and graph-based ML offers a powerful approach for high-accuracy computational chemistry.
• This method provides a computationally feasible alternative to traditional high-accuracy methods for large systems.