XRayPro: A self-supervised multimodal model for MOF application recommendations using PXRD and precursors
Submission Track: Short Paper
Submission Category: AI-Guided Design + Automated Material Characterization
Keywords: Metal-organic frameworks (MOFs), Transformer, Crystal structures, Multimodality, Self-supervised learning, Application recommendation, Carbon capture, Gas storage
TL;DR: This work presents XRayPro - a multimodal model for MOFs that uses PXRDs and precursors to predict geometric and chemical properties, while incorporating a recommendation system and self-supervised pretraining with a CGCNN for improved accuracy.
Abstract: In crystal structures, retrieving properties following synthesis is a time-consuming process. As crystal synthesis is often followed by a crystallinity assessment through the calculation of its powder x-ray diffraction (PXRD) pattern, this information (alongside its precursors) can be leveraged to directly predict the properties of these structures. To address this, we developed XRayPro, a model specifically tailored for metal-organic frameworks (MOFs), which can not only directly predict material properties, but also incorporates a recommendation system to suggest new applications - all done with only a PXRD and the MOF precursors. Additionally, self-supervised learning was done against a crystal graph convolutional neural network (CGCNN) to pretrain our multimodal model, leading to a significant improvement in the data efficiency of our model and enhancing its ability to learn chemistry-reliant and quantum-chemical properties. Our multimodal model not only predicts geometric, chemistry-reliant, and quantum-chemical properties, but the recommendation system has also shown potential in discovering new applications for certain MOFs, particularly in carbon capture and methane storage.
Submission Number: 14
Loading