Recurrent Neural Network/Machine Learning Predictions of Reactive Channels in H⁺ + C₂H₄ at E_Lab = 30 eV: A Prototype...

Area of Science:

• Computational Chemistry
• Chemical Physics
• Machine Learning in Chemistry

Background:

• Ion cancer therapy (ICT) involves complex chemical reactions.
• Predicting reaction outcomes is crucial for optimizing ICT.
• Current methods may be computationally intensive.

Purpose of the Study:

• To develop and validate a hybrid computational approach for predicting chemical properties in ICT reactions.
• To accelerate the simulation of these reactions using machine learning.
• To assess the accuracy of the proposed method for reaction type and product charge prediction.

Main Methods:

• Developed the Simplest-Level Electron Nuclear Dynamics/Machine Learning (SLEND/ML) approach.
• SLEND models quantum dynamics; ML accelerates predictions using SLEND-generated data.
• Applied SLEND/ML to H+ + C2H4 reactions at 30 eV, a model ICT system.

Main Results:

• Recurrent Neural Network (RNN) and k-nearest neighbor models achieved high accuracy (98.23% and 95.13%) for reaction type prediction.
• RNN demonstrated excellent generalization across frequent and infrequent reaction types.
• RNN achieved low mean absolute errors (0.02-0.07) for predicting product charges.

Conclusions:

• SLEND/ML is an effective and accurate method for predicting chemical properties in ICT reactions.
• The hybrid approach significantly accelerates simulations compared to traditional methods.
• This methodology holds promise for advancing computational studies in ion cancer therapy.