A Multitask Learning Model for Predicting Various Types of Metal–Organic Framework Stability

This study introduced a MOF stability prediction neural network (MOFSNN), a multitask learning architecture to predict various stability metrics of MOFs, including thermal, solvent removal, water, acid, base, and boiling water stabilities. Based on the crystal graph convolutional neural network (CGCNN), MOFSNN incorporates a lattice parameter embedding layer and task-specific attention layers to enhance the prediction accuracy and robustness. Using a data set compiled from literature, the MOFSNN model was demonstrated to outperform the traditional machine learning models and the original CGCNN model, especially in tasks with limited data. Uncertainty analysis using latent space variance (LSV) and latent space entropy (LSE) proved to effectively control prediction reliability. Moreover, the model’s ability to extrapolate to unseen materials further validated its great potential for MOF stability prediction. This study highlights the efficacy of multitask learning in leveraging correlations among different stability metrics, advancing MOF stability prediction and laying a solid foundation for material discovery.