关键词: 机器学习, 内嵌式微通道散热, 多目标优化, 多物理场协同设计, 集成系统热管理

Abstract: With the rapid development of artificial intelligence computing, advanced packaging, and wide bandgap power devices, electronic systems are continuously evolving toward higher power density, higher integration, and stronger multiphysics coupling. Thermal management has become a critical bottleneck that limits device performance, reliability, and energy efficiency. In 2.5D/3D heterogeneous integration systems, the coexistence of multiple hotspots, elongated cross layer heat conduction paths, aggravated thermal crosstalk, and accumulated thermal resistance at package interfaces all substantially increase the difficulty of heat dissipation. For power devices based on silicon carbide (SiC) and gallium nitride (GaN), the high heat flux density near the junction further imposes more stringent requirements on efficient cooling. By moving the cooling location closer to the heat source, embedded microchannel cooling exhibits clear advantages in shortening heat transfer paths, enhancing local heat transfer, and improving system integration, and has gradually evolved into advanced forms such as hierarchical manifold microchannels, pin fin enhanced microchannels, jet impingement cooling, and composite cooling structures embedded in silicon interposers. Meanwhile, as the complexity of microfluidic cooling structures continues to increase, conventional design methods that rely on computational fluid dynamics (CFD) simulations and repeated trial and error can no longer meet the demand for rapid optimization. Machine learning provides a new technical route for rapid modeling, multi-objective optimization, and intelligent design of microfluidic cooling structures, driving the field from direct high fidelity simulations toward surrogate model accelerated evaluation, machine learning driven optimization, and online thermal management under dynamic thermal loads. Focusing on two representative application scenarios, namely 2.5D/3D heterogeneous integration and power devices, this paper systematically reviews the major thermal challenges, key structural forms, representative application progress, and machine learning driven design optimization methods for microfluidic cooling.

Key words: machine learning, embedded microfluidic heat sink, multi-objective optimization, multi-physics collaborative design, integrated system thermal management