Improved vapor pressure predictions using group contribution-assisted graph convolutional neural networks (GC²NN)
A novel hybrid architecture combining graph convolutional neural networks with group contribution methods to predict vapor pressures of organic molecules with state-of-the-art accuracy.
The vapor pressures (p_vap) of organic molecules play a crucial role in the partitioning of secondary organic aerosol (SOA). Given the vast diversity of atmospheric organic compounds, experimentally determining p_vap of each compound is unfeasible.
The Challenge
Machine learning algorithms allow the prediction of physicochemical properties based on complex representations of molecular structure, but their performance crucially depends on the availability of sufficient training data. Existing methods—including parameterizations and traditional group-contribution approaches—struggle to capture molecular connectivity and the positional influence of functional groups.
Our Approach
We propose a novel approach to predict p_vap using group contribution-assisted graph convolutional neural networks (GC²NN). The architecture combines two complementary components:
- Graph component: A graph convolutional neural network that processes molecular graphs containing atom and bond features as representations of molecular structure
- Group contribution component: A shallow fully-connected neural network that processes numerical molecular descriptors like molar mass
Molecular graphs allow the model to better infer molecular connectivity compared to methods using other, non-structural embeddings. We achieve best results with an adaptive-depth GC²NN, where the number of evaluated graph layers depends on molecular size.
Results
We present two vapor pressure estimation models that achieve strong agreement between predicted and experimentally-determined p_vap.
General Model
A broad-scope model suitable for both organic and inorganic molecules:
- Mean absolute error: 0.69 log-units
- R²: 0.86
Specialized SOA Model
A model specialized on organic compounds with functional groups often encountered in atmospheric SOA:
- Mean absolute error: 0.37 log-units
- R²: 0.94
The adaptive-depth GC²NN models clearly outperform existing methods, including parameterizations and group-contribution methods, demonstrating that graph-based ML techniques are powerful tools for the estimation of physicochemical properties, even when experimental data are scarce.
Published in Geoscientific Model Development Read the full paper →