<p>Structural topology optimization is a key method for designing lightweight and efficient components, particularly in additive manufacturing. In aerospace and automotive engineering applications, where components bear dynamic loads, it is crucial to account for natural frequencies and weight reduction. Traditional optimization methods often require labour-intensive manual setup, especially for tuning input parameters that influence the algorithm’s performance and convergence. Machine learning offers an alternative, streamlining this process and reducing reliance on trial-and-error adjustments. This study introduces a machine learning-based framework to optimize input parameters—such as evolutionary rates and mesh dimensions—while ensuring the creation of lightweight, optimized structures with consistent natural frequencies. The framework uses a neural network trained on a dataset built to describe several configurations of a cantilever beam case study. The primary focus is determining the optimal dataset size for effective neural network training. The analysis aims to minimize dataset size, reducing the time and effort required for data preparation while avoiding overfitting in the neural network. A dissimilarity metric based on problem metadata is used to guide parameter tuning. After training, the artificial neural network is incorporated into the Bayesian Optimization framework as a surrogate model, enabling the new model to rapidly estimate outcomes without running full simulations. This approach substitutes traditional topology optimization algorithms, accelerates time-to-market, and is well-suited for applications with only minor variations in design conditions.</p> Graphical Abstract <p></p>