A new perspective on the machine learning-based acceleration of the two level topology optimization process
Please login to view abstract download link
Structural Topology Optimization (TO) has increasingly integrated Machine Learning (ML) to mitigate the computational costs of traditional iterative methods like SIMP. While various one-shot architectures [1] have been proposed to accelerate these workflows, their relative performance in generating structurally consistent initial solutions remains a subject of investigation. This work presents a comparative study of different neural network architectures using a methodology based on a distance metric in the parameter space [2] as a performance benchmark. The study integrates these architectures within the 2-Level TO paradigm [3], evaluating their ability to predict coarse-grid topologies that serve as starting points for localized refinements. By utilizing a linear manifold learning strategy [2] to explore the high-dimensional parametric space, we provide a standardized framework to quantify the accuracy and structural soundness of the predictions made by each model. Our analysis focuses on comparing the number of SIMP iterations required for convergence after the initial ML prediction, as well as the overall computational efficiency. This benchmarking approach seeks to identify the strengths and limitations of different DL backbones when coupled with distance-based initialization, aiming for the consistent 75% reduction in computational effort previously observed in data-driven frameworks.