Accelerated topology optimization via a deep quantized projector

Topology optimization (TO) often requires prolonged iterative evolution to clearly define the material distribution, incurring substantial computational cost from repeated physical analyses. This paper proposes a modular topology optimization framework via a deep quantized projector (DQP). The proposed method embeds a pretrained vector quantization model into an existing topology optimization loop as a modular learning-based projector. During the offline stage, the framework uses pseudo-topology samples rather than optimization-derived samples to train the model to extract codebook-guided structural features without relying on expensive TO samples. A weighted codebook projection mechanism is introduced to replace hard nearest-code assignment and relax the abrupt switching between code vectors associated with standard vector quantization. The resulting latent features are constructed through a temperature-controlled weighted combination of learned code vectors, enabling continuous transitions among the learned representations. During the online evolution stage, the frozen pretrained projector is employed to reconstruct the current candidate density field, and its output is progressively blended with the current design to avoid an abrupt modification of the optimization trajectory. Numerical experiments in 2-D and 3-D design spaces using different optimization methods show that the DQP-enhanced formulations require fewer optimization iterations and lower wall-clock time across the investigated cases. Representative 2-D trajectories further show that the compliance objective evolves and stabilizes over a shorter iteration range after DQP activation, while faster density separation occurs concurrently along the modified topology-evolution trajectory. Limitations and future research directions are discussed based on the current investigations.