MOF-KAN: Kolmogorov–Arnold Networks for Digital Discovery of Metal–Organic Frameworks

Abstract

Digital discovery of functional materials, such as metal–organic frameworks (MOFs), entails accurate and data-efficient approaches to navigate complex chemical and structural space. Based on an innovative deep learning approach, namely, Kolmogorov–Arnold Networks (KANs), we introduce MOF-KAN, a state-of-the-art architecture as the first application of KANs to digital discovery of MOFs. Through meticulous fine-tuning of network architecture, we demonstrate that MOF-KAN outperforms standard multilayer perceptrons (MLPs) in predicting diverse properties for MOFs, including gas separation, electronic band gap, and thermal expansion. Furthermore, MOF-KAN excels in low-data regimes, facilitating robust performance in challenging prediction scenarios. Feature importance analysis reveals that MOF-KAN accurately captures critical features of MOFs relevant to targeted properties. MOF-KAN not only serves as a transformative tool for the rational design of functional materials but also holds broad applicability across various domains in physical sciences.