Predicting S1 TDDFT Energies from ZINDO Calculations Using Message-Passing ΔML with Electronically Informed Descriptors | JoVE Visualize

Area of Science:

Computational Chemistry
Machine Learning
Quantum Chemistry

Background:

Semiempirical methods offer fast excited-state energy calculations but often lack accuracy.
Time-Dependent Density Functional Theory (TDDFT) provides higher accuracy but is computationally expensive.
Bridging this accuracy-computational cost gap is crucial for large-scale molecular screening.

Purpose of the Study:

To develop a machine learning approach (ΔML) that enhances the accuracy of semiempirical excited-state energy calculations.
To achieve accuracy comparable to higher-level methods like TDDFT with minimal additional computational cost.
To enable efficient computational screening of large molecular datasets for properties like excitation energies and oscillator strengths.

Main Methods:

Trained machine learning models using a dataset of 7600 organic π-conjugated molecules with ZINDO and M06-2X/3-21G* TDDFT calculations.
Utilized an AttentiveFP Message-Passing Neural Network incorporating electronic information (e.g., particle-hole densities) from ZINDO calculations.
Investigated molecular descriptors including Morgan fingerprint and a novel molecular orbital-weighted radial distribution function.
Retrained the ΔML framework for different low- and high-level calculation pairs (e.g., ZINDO to ωB97X-D/6-31G*).

Main Results:

The best ΔML-ZINDO model improved the correlation of ZINDO S₁ energy predictions from 0.77 to 0.96 on a test set.
Achieved a negligible additional cost (∼2 ms/molecule) compared to standard ZINDO calculations (∼2 s/molecule).
Demonstrated retraining capability, improving correlation from 0.88 to 0.99 for ZINDO to ωB97X-D/6-31G* energies.
Enhanced S₁ oscillator strength predictions from a correlation of 0.524 to 0.839, enabling identification of emissive molecules.

Conclusions:

The ΔML approach effectively corrects systematic errors in low-level excited-state calculations, achieving high accuracy.
ΔML offers a computationally efficient solution for large-scale molecular screening and property prediction.
The framework is versatile and can be adapted to various computational chemistry methods and molecular descriptors.