Δ² machine learning for reaction property prediction.

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Quantum Chemistry

Background:

• Δ-learning models accelerate high-level energy calculations but cannot predict reaction properties.
• Reaction properties like activation energies require both high-level geometry and energy evaluations.

Purpose of the Study:

• Introduce a novel Δ²-learning model to predict high-level activation energies from low-level critical-point geometries.
• Enable accurate and efficient characterization of chemical reactions.

Main Methods:

• Developed a Δ²-learning model utilizing atom-wise featurization.
• Trained the model on a dataset of ~167,000 reactions using GFN2-xTB and B3LYP-D3/TZVP energies.
• Validated transferability on external datasets and fine-tuned with Gaussian-4 calculations.

Main Results:

• The Δ² model accurately predicts high-level activation energies from low-level geometries.
• Demonstrated implicit learning of geometric deviations between low-level and high-level structures.
• Achieved near chemical accuracy on unseen reactions and external test sets.

Conclusions:

• The Δ²-learning model offers an efficient strategy for accelerating chemical reaction characterization.
• Combines machine learning with semi-empirical quantum chemistry for high accuracy.
• Fine-tuning improved accuracy by 35% over DFT predictions at a low computational cost.