New deep learning framework predicts gas adsorption in materials with only basic physics data
-
A new generalized framework is proposed to predict gas adsorption properties in materials using only the potential energy surface (PES) as the descriptor. The PES is voxelized into a 3D image that a convolutional neural network (CNN) can process.
-
As a proof of concept, the method is applied to predict CO2 uptake in metal-organic frameworks (MOFs). The CNN model outperforms a model using geometric descriptors and requires 100x less training data.
-
The transferability of the approach is shown by examining CH4 uptake prediction in covalent organic frameworks (COFs). Again the CNN model performs better than using geometric descriptors.
-
The proposed framework is applicable beyond reticular chemistry as it is based on universal interactions captured in the PES. For example, it could be used to predict properties of organic molecules.
-
Refinements like better potentials to construct the PES voxelization and modifications like transfer learning can further improve model performance and data efficiency.