Predicting Experimental Heats of Formation via Deep Learning with Limited Experimental Data

Area of Science:

• Computational chemistry
• Machine learning in chemistry

Background:

• Predicting molecular properties is crucial in chemistry.
• Deep learning models require substantial experimental data for training, which is often scarce.
• First-principles quantum mechanical (QM) methods provide accurate but computationally expensive predictions.

Purpose of the Study:

• To develop an efficient and accurate deep learning method for predicting experimental molecular properties.
• To overcome the challenge of limited experimental data in molecular property prediction.
• To leverage the qualitative and semiquantitative accuracy of QM calculations for model calibration.

Main Methods:

• A two-stage deep learning approach was proposed: pretraining a graph neural network (GNN) on QM results.
• Fine-tuning a fully connected neural network (FNN) using limited experimental data.
• The combined model bypasses direct QM calculations for efficiency.

Main Results:

• The combined model achieved high accuracy in predicting experimental heats of formation for organic molecules.
• The model demonstrated a mean absolute error of 1.8 kcal/mol for molecules containing H, C, O, N, or F atoms.
• The approach successfully utilized a small dataset of 405 experimental data points.

Conclusions:

• The proposed pretraining and fine-tuning deep learning strategy effectively predicts molecular properties.
• This method offers a computationally efficient alternative to traditional QM calculations.
• The approach significantly improves prediction accuracy by calibrating QM-informed models with experimental data.