KANs Layer Integration: Benchmarking Deep Learning Architectures for Tornado Prediction

Abstract

Tornado occurrence and detection are well established in mesoscale meteorology, yet the application of deep learning to radar-based tornado detection remains nascent and under-validated. This study benchmarks DL approaches on TorNet, a curated dataset of full-resolution, polarimetric WSR-88D radar volumes. We evaluate three canonical architectures (CNN, VGG19, Xception) under five optimizers and assess the effect of replacing conventional MLP heads with Kolmogorov–Arnold Network (KAN) layers. To address severe class imbalance and label noise, we implement radar-aware preprocessing and augmentation, temporal splits, and recall-sensitive training. KAN-augmented variants generally converge faster and deliver higher rare-event sensitivity and discriminative power than their baselines, with Adam and RMSprop providing the most stable training and Lion showing architecture-dependent gains.

Radar, meets learnable activations

Tornadoes are rare events buried in noisy, imbalanced radar data: exactly the regime where architecture and training choices matter most and are least validated. This study builds a reproducible baseline suite on TorNet, a curated dataset of full-resolution, polarimetric WSR-88D (Weather Surveillance Radar-1988 Doppler) radar volumes, and asks a focused question: what happens when the conventional MLP classification heads of standard deep architectures are replaced with Kolmogorov–Arnold Network (KAN) layers, whose learnable activation functions promise more expressive decision boundaries per parameter?

Setup

Architectures: CNN, VGG19, and Xception, each evaluated with its standard MLP head and a KAN-augmented head.
Optimizers: five, spanning the common production choices (including Adam, RMSprop, and Lion), to separate architecture effects from optimization effects.
Rare-event rigor: radar-aware preprocessing and augmentation, temporal splits (no leakage from future storms into training), and recall-sensitive training to handle severe class imbalance and label noise. Models are compared on accuracy, precision, recall, and ROC-AUC.

Findings

KAN-augmented variants generally converge faster and deliver higher rare-event sensitivity and discriminative power than their MLP-headed baselines.
Adam and RMSprop provide the most stable training; Lion’s gains are architecture-dependent, a practical caution for teams choosing optimizers by default.
The paper contributes a reproducible baseline suite for TorNet, evidence on when KAN integration helps tornado detection, and optimizer–architecture guidance for rare-event forecasting with weather radar.

@article{yang2025kans, title = {KANs Layer Integration: Benchmarking Deep Learning Architectures for Tornado Prediction}, author = {Yang, Shuo (Luna) and Vilataj, Ehsaneh and Raza, Muhammad Faizan and Srinivasan, Satish Mahadevan}, journal = {Big Data and Cognitive Computing}, volume = {9}, number = {12}, pages = {324}, year = {2025}, publisher = {MDPI}, doi = {10.3390/bdcc9120324} }