Group Contribution Gaussian Process Hybrid Models for Thermophysical Property Prediction

This repository accompanies the research paper:

“Enhanced Thermophysical Property Prediction with Uncertainty Quantification using Group Contribution-Gaussian Process Regression”

---

Repository Contents

This repository contains all data, scripts, and results related to the study. Below is a breakdown of the key folders and files:

1. Data Files

• Data_for_Model_Building folder containing all curated and preprocessed datasets (*_fcl.csv) for model building for all properties.
• Hvap_data_test_fluorinated_molecules.csv: Test set used to evaluate GCGP ΔHvap model performance on highly fluorinated molecules.

2. Final Results

• Located in the Final_Results folder.
• Includes:
  • Parity plots
  • Numerical model predictions
  • Model performance metrics
  • Model training outputs
  • Results from different random seeds for train/test splits
  • Results (numerical data) from timing tests
  • lml_values.csv: A summary of collected LML values from kernel architecture experiments.

3. Kernel Sweep Experiments

• Found in kernel_sweep_code_and_results folder.
• Contains results of testing multiple kernel designs and model architectures as detailed in the paper.

4. Data Visualization, Data Quality, and Outlier Analysis

• Data_Visualization_Pretraining and Data_Quality_and_Outlier_Checks folders include:
  • Data analysis plots
  • Visualization figures
  • Outlier detection and data quality assessment

5. White Noise Kernel Tests

• Tm_whitenoise_tests folder contains results analyzing the impact of various white noise kernel settings on normal melting temperature (Tm) predictions.

6. Log-Marginal Likelihood (LML) Analysis

• LML_plots folder contains LML plots for various model and kernel combinations across all properties.
• • lml_values.csv in Final_Results folder contains data used in making the plots in LML_plots as well as additional data.

7. Python Scripts

• Code.py folder contains all Python code (*.py) used to analyze data, train models, evaluate model performance, and generate figures.
• Scripts are descriptively named for easy navigation and use.

---

Citation

If you use this code or data, please cite the corresponding paper.

---

Feel free to open an issue or pull request if you have questions, suggestions, or contributions!