We are pleased to announce the publication of our latest research, titled "Developing Machine Learning Models for Accurate Prediction of Radiative Efficiency of Greenhouse Gases," in Journal of the Taiwan Institute of Chemical Engineers. This groundbreaking work leverages machine learning (ML) models to accurately predict the radiative efficiency (RE) of greenhouse gases (GHGs), which play a critical role in global warming.
Our research introduces ML models trained on a comprehensive dataset generated through density functional theory (DFT) and infrared (IR) spectra calculations. These models are designed to efficiently predict the RE of a vast number of halogenated and non-halogenated GHGs, enabling more informed decision-making in climate change mitigation.
Key Findings:
- Large Dataset Using DFT Calculations: We developed a dataset of approximately 82,000 molecules using DFT to compute IR spectra and radiative efficiency values. This dataset enabled us to train machine learning models to predict GHG radiative efficiency, a critical factor in understanding their impact on climate change.
- Top-Performing Machine Learning Models: Among various architectures, the directed message-passing neural network (DMPNN) model emerged as the most accurate, particularly when combined with quantum mechanical features (ml-QM-DMPNN). These models showed impressive performance, achieving root mean square errors (RMSEs) of 0.0582 and 0.0564 Wm−2ppb−1, respectively, when refined using transfer learning with experimental data.
- Transfer Learning for Improved Accuracy: To enhance the model's predictive accuracy, we employed transfer learning techniques. By first training the models on the DFT dataset and then refining them using a smaller set of experimental data, we significantly improved the prediction of experimental REs, demonstrating the model's potential for predicting climate-relevant properties of GHGs.
This study represents a significant advancement in using machine learning to assess the environmental impact of greenhouse gases. By providing a more efficient method for predicting RE, the research offers valuable insights into the design of environmentally-friendly chemicals and materials that contribute to climate change mitigation.
For more details, please access the full publication here.