Granger causality detection with Kolmogorov-Arnold networks

Abstract

Identifying causal relationships between variables is fundamental to scientific inquiry, and the Granger causality (GC) framework is widely used to provide indications of causal relationships in time series analysis. While the original GC approach was limited to linear interactions, it was later extended to nonlinear cases using neural networks, particularly multilayer perceptrons (MLPs). In this work, we introduce—for the first time—the use of Kolmogorov-Arnold Networks (KANs) for detecting nonlinear Granger causality (GC-KAN), leveraging their functional decomposition properties for enhanced interpretability and performance. We develop a tailored training approach for KAN-based causality inference and evaluate GC-KAN on Vector Autoregressive (VAR) models and chaotic Lorenz-96 systems where the underlying causal relationships are known. Results are compared against component-wise MLP (cMLP), the first and most widely used neural network approach for GC detection. Our results demonstrate the potential of KANs to outperform MLPs in identifying interpretable Granger causal relationships, particularly in detecting sparse causality patterns in nonlinear scenarios. More broadly, our findings highlight the potential of AI-driven methods in uncovering causal structures underlying dynamical laws in nonlinear physical systems.

The authors' abstract, as published at the source. Physical Review Research, 2026 · DOI ↗