Toward Accurate PAH IR Spectra Prediction: Handling Charge Effects with Classical and Deep Learning Models.

Area of Science:

• Computational Chemistry
• Astrochemistry
• Environmental Science

Background:

• Polycyclic aromatic hydrocarbons (PAHs) are vital in astrochemistry, environmental studies, and combustion.
• Interpreting their infrared (IR) spectra is difficult due to spectral similarities and the presence of both neutral and charged species.
• First-principle calculations offer accuracy but are computationally expensive, limiting their use.

Purpose of the Study:

• To develop machine learning (ML) models for predicting PAH IR spectra.
• To enable simultaneous prediction for both neutral and ionized PAH molecules.
• To overcome the computational limitations of traditional methods.

Main Methods:

• Developed an XGBoost model using Morgan fingerprints.
• Implemented a graph neural network (GNN) utilizing molecular graph representations.
• Incorporated molecular charge information using one-hot or learnable neural network encodings.

Main Results:

• Both ML models achieved excellent predictive performance for PAH IR spectra.
• Successfully enabled fast and accurate prediction of charged PAHs IR spectra for the first time.
• The XGBoost model achieved state-of-the-art accuracy, while the GNN shows future potential.

Conclusions:

• ML models offer a scalable and efficient approach to predicting PAH IR spectra.
• The developed models can accurately handle both neutral and charged PAHs.
• Future work should address data scarcity for heteroatomic PAHs.