Kolmogorov-Arnold Graph Neural Networks Applied to Inorganic Nanomaterials Dataset

TL;DR

The paper tackles property prediction for inorganic nanomaterials using graph neural networks. It introduces and applies Kolmogorov-Arnold Networks (KAGNNs) to the CHILI dataset, including KAEdgeCNN, KAGCN, and KAGIN, with extensive hyperparameter tuning. Results show significant classification improvements and state-of-the-art performance on several CHILI-3K tasks, though regression tasks remain challenging and KAN models come with higher parameter counts. Overall, KAGNNs reduce model imbalance across GNN architectures and extend strong performance to inorganic materials, suggesting broad applicability and potential explainability advantages.

Abstract

The recent development of Kolmogorov-Arnold Networks (KANs) introduced new discoveries in the field of Graph Neural Networks (GNNs), expanding the existing set of models with KAN-based versions of GNNs, which often surpass the accuracy of MultiLayer Perceptron (MLP)-based GNNs. These models were widely tested on the graph datasets consisting of organic molecules; however, those studies disregarded the inorganic nanomaterials datasets. In this work, we close this gap by applying Kolmogorov-Arnold Graph Neural Networks (KAGNNs) to a recently published large inorganic nanomaterials dataset called CHILI. For this, we adapt and test KAGNNs appropriate for this dataset. Our experiments reveal that on the CHILI datasets, particularly on the CHILI-3K, KAGNNs substantially surpass conventional GNNs in classification, achieving state-of-the-art results.