TO-MambaLDM: Multi-physics informed Mamba-enhanced latent diffusion for cooling structure topology optimization

To enable high-precision 3D in vitro fabrication of medical hydrogel scaffolds, improving the stability and temperature-control performance of the cooling system in 3D printing equipment is essential. The channel layout of the liquid-cooling plate is therefore critical. Conventional topology optimization can produce feasible designs, but it typically depends on repeated finite element analyses, leading to high computational cost and slow convergence. Here, we propose TO-MambaLDM, a multiphysics-driven generative topology optimization framework that integrates a latent diffusion model with a state–space architecture. Compared with existing data-driven and diffusion- or Transformer-based generative design methods, TO-MambaLDM explicitly incorporates thermo–fluid operating conditions into the generative process and enhances long-range dependency modeling to improve physical consistency and connectivity. The framework captures thermo–fluid–structural coupling by conditioning the latent space with multi-channel field maps and boundary matrices. Mamba models long-range dependencies, while diffusion generates high-resolution, physically consistent designs. A manufacturability loss promotes channel connectivity and fabrication feasibility. Experiments show that TO-MambaLDM achieves a reconstruction accuracy of IoU = 0.93 and reduces temperature and pressure MSEs to 7.4 and 7.2, respectively, outperforming multiple baseline models, with improvements of 18.0% and 11.1% in temperature and pressure prediction accuracy. Transfer learning further verifies its ability to generate manufacturable, high-performance cooling structures across diverse inlet–outlet configurations. TO-MambaLDM establishes a unified, physically grounded paradigm for rapid cooling structure design in additive manufacturing, electronics cooling, and energy systems.