Machine learning(ML)-based AI systems are often black box, making it challenging to understand and interpret their decision-making processes. Surrogate models are essential for providing transparency, enabling the extraction of interpretable models that approximate the behavior of these black-box models. This includes querying the model with inputs and using the responses to infer information about the model’s structure and parameters. In this paper, we propose a combinatorial testing approach for surrogate model construction, aiming to efficiently capture the significant interactions between features that drive the original model’s predictions. Our approach leverages t-way testing to generate diverse data points, focusing on interactions potentially influencing specific prediction classes. We iteratively refine the data set based on the pre-trained model’s predictions, identifying crucial patterns and generating additional targeted test cases. This iterative process efficiently constructs a surrogate model that closely mirrors the original model’s behavior, achieving high accuracy with a minimal number of test cases. We evaluate our approach on 4 datasets and 12 ML models and compare the results with state-of-the-art (STOA) approaches. Our experimental results suggest that the proposed approach can successfully construct a surrogate ML model, and in most cases, performs better than STOA approaches in terms of efficiency and accuracy.