Process parameters are critical in determining the quality and performance of laser deep penetration welding (LDPW) by influencing key factors like breaking tensile strength (BTS). Optimizing these parameters is essential for ensuring consistent weld quality and improving overall efficiency. As the dimensionality of process parameters and sample sizes increase, the relationship between process parameters and BTS becomes increasingly complex, uncertain, and nonlinear. While machine learning (ML) techniques struggle to handle this complexity, deep learning (DL) methods, although effective at capturing nonlinear patterns, are constrained by their lack of interpretability, limiting their wider application in industrial environments. To address these challenges, an intrinsically and post-hoc interpretable approach combining the Kolmogorov-Arnold Network (KAN) and genetic algorithm (GA) is proposed. An LDPW experimental platform is first established, employing an orthogonal design (L81) with five factors at three levels to gather datasets. A lightweight and interpretable KAN with a single hidden layer is then developed to establish the nonlinear relationship between process parameters and BTS. During optimization, the trained KAN serves as a surrogate model to define decision variables, constraints, and objective functions, with interpretability enhanced through graphical and symbolic representations that clarify how process parameters impact the BTS. Furthermore, the post-hoc Shapley additive explanation (SHAP) method is employed to provide global and local insights into the overall and sample-specific influence on BTS. Validation of the optimization results confirms that the proposed method outperforms other state-of-the-art approaches.