Deep Learning of Activation Energies

Area of Science:

• Computational chemistry
• Machine learning
• Chemical reactivity

Background:

• Quantitative prediction of reaction properties like activation energy is limited by data scarcity.
• Accurate predictions are crucial for computer-assisted reaction mechanism generation and organic synthesis planning.

Purpose of the Study:

• To develop a template-free deep learning model for predicting activation energy.
• To train and validate the model on a novel, diverse dataset of gas-phase quantum chemistry reactions.

Main Methods:

• Developed a template-free deep learning architecture.
• Utilized reactant and product molecular graphs as input.
• Trained the model on a new dataset of quantum chemistry reaction data.

Main Results:

• The deep learning model achieved accurate predictions of activation energy.
• Model predictions align with established chemical reactivity principles.
• Demonstrated the model's capability on a diverse set of reactions.

Conclusions:

• The developed model shows promise for predicting chemical reactivity.
• Availability of quantitative reaction data is key for advancing predictive models.
• Future work will focus on expanding data and refining predictive methods.