Modeling Lightning Activity with Kolmogorov–Arnold Networks based Poisson Regression for Lighting count Prediction

Accurate prediction of lightning activity is critical for understanding convective dynamics, improving severe weather forecasts, and mitigating atmospheric hazards. Traditional regression-based neural approaches produce continuous-valued estimates that are poorly suited for inherently discrete phenomena such as lightning strike counts. This work introduces a Poisson–Kolmogorov–Arnold Network (Poisson-KAN) framework that combines the expressive functional representation of KANs with a Poisson likelihood formulation to model lightning occurrence as a stochastic count process. Using meteorological predictors such as convective available potential energy, humidity, and wind shear, the model captures nonlinear dependencies while maintaining interpretability through its spline-based architecture. Preliminary computational experiments demonstrate improved generalization and physically consistent scaling of predicted lightning rates. This approach highlights how modern machine-learning architectures can be adapted for probabilistic, count-based modeling in atmospheric and geophysical systems.