Light radiation, the primary mode of energy dissipation in nuclear explosions, profoundly impacts both the ecological environment and human society. A thorough understanding of its characteristics, propagation dynamics, and energy distribution is therefore essential for evaluating and protecting against nuclear explosion damage effects.

This study introduces the Kolmogorov-Arnold network (KAN) to construct an interpretable model for predicting the area of light radiation damage. The model utilizes multiple optimization algorithms to invert key source term parameters, namely the explosion yield, height of explosion, and explosion location.

Based on the theory of nuclear explosion fireballs, a light radiation model was developed and integrated with ArcGIS Pro software to visualize thermal energy distribution on real-world maps. A dataset correlating explosion yield and height with damage area was then generated, quantified according to established standards for biological burn injuries. The KAN was trained on this dataset, leveraging its unique advantage of providing explicit, interpretable formulas for prediction. To validate its efficacy, the KAN's performance was benchmarked against eight other algorithms, including Gated Recurrent Unit (GRU), Extreme Learning Machine (ELM), and Random Forest (RF). A loss function was constructed for the radiation model to facilitate the inversion of source term parameters via multiple optimization algorithms.

The results demonstrate that the KAN model achieves high prediction accuracy while yielding an interpretable formula for the damage area. Furthermore, both the genetic algorithm and the Hippopotamus optimization algorithm successfully inverted the nuclear source term parameters with high fidelity.

This methodology facilitates both the rapid prediction of damage effects and the accurate inversion of source parameters, thereby enhancing emergency response efficiency and aiding in strategic protective decision-making.