The geometric design of liquid-cooled cold plates critically determines their thermal-hydraulic performance, influencing the efficiency and reliability of high-power electronic systems. This study presents a conditional diffusion model-assisted digital twin framework that integrates physics-based multi-objective topology optimization (TO) with generative AI to accelerate high-performance thermal management design. The study first employs multi-objective topology optimization to generate a diverse set of cold plate geometries, together with the corresponding thermal resistance ( \(R_{th}\) ) and pressure drop ( \(\Delta p\) ) under asymmetric thermal loading. A conditional diffusion model is then trained to capture the underlying distribution of cold plate geometries under specified physical conditions, enabling rapid and diverse generation of new designs consistent with user-defined parameters. The new designs generated by diffusion model are evaluated using a surrogate model to predict their thermal-hydraulic performance, facilitating the quick selection of viable candidates without exhaustive full-order simulations. The final designs are validated through high-fidelity finite-element thermal-fluid simulations, confirming strong agreement with reference results. By integrating the predictive rigor of physics-based modeling with the speed and adaptability of generative AI, this work brings physics-informed predictive design off the supercomputer and into the operational world, establishing a scalable and intelligent digital twin paradigm for the design of next-generation thermal management systems.
