Structural topology optimization approaches are widely used to create lightweight and efficient components through additive manufacturing. The race to lightweight components should also account for natural frequencies when designing components and structures subjected to dynamic loads, as in aerospace and automotive engineering. These optimization tools require extensive manual setup, mainly when tuning input parameters that govern algorithmic functions and convergence. In this context, machine learning approaches can circumvent the trial-and-error process associated with the manual setup of simulation factors. This study presents a method that utilizes machine learning to suggest optimal input parameters, such as evolutionary rate and mesh dimensions, for creating lightweight and optimized structures while maintaining a consistent natural frequency. The framework incorporates a neural network trained on a collection of previously solved, analogous problems. A dissimilarity metric derived from problem metadata is used to determine tuning parameters. This approach can be applied to analyze and optimize product configurations where only marginal conditions may change; a Bayesian optimizer based on data coming from the neural network is used to improve the structure, substituting the topology optimization algorithm and reducing the time-to-market of a specific product.