Implicit neural representation and exploration of volumetric Data with Kolmogorov–Arnold networks

Exploring the common signals composing scientific data is essential for understanding the underlying physical phenomena. However, extracting such signals remains challenging for traditional volumetric representations. Implicit neural representations (INRs) have emerged as a powerful framework for representing volumetric data in a memory-efficient and continuous manner. Nonetheless, fitting these signals typically requires deep networks, limiting interpretability. Kolmogorov–Arnold Networks (KANs), recently proposed as an alternative neural architecture, leverage spline-based activation functions to approximate data through adaptive basis functions. We leverage a KAN-based implicit neural representation to model volumetric data and evaluate its performance across multiple scientific datasets. Both quantitative and qualitative results demonstrate our method’s superior representation capability and its potential for capturing complex spatiotemporal structures. We also introduce an explorative interface, allowing users to intuitively identify prominent spline units and examine their relationships with volumetric structures. The interface also supports interactions for visualizing and editing volumetric data, thereby helping users understand how specific features are represented by individual basis functions. Case studies demonstrate that our method provides a flexible strategy for scientific exploration of volumetric data.