搜索

x
中国物理学会期刊

人工智能赋能的跨尺度热设计:以机器学习为主线的从原子到器件综述

Artificial Intelligence-Enabled Cross-Scale Thermal Design: A Machine-Learning-Centered Review from Atoms to Devices

PDF
导出引用
  • 随着高功率电子器件、宽禁带半导体、异质集成与先进封装技术的快速发展,热输运研究正由单一热物性表征转向贯穿材料、界面、结构与器件的跨尺度协同热设计。本文以人工智能赋能的跨尺度热设计为核心,系统综述机器学习在从原子到器件热传导研究中的应用进展。首先,介绍机器学习辅助的热物性预测与目标导向材料设计,重点讨论热物性数据生成、模型构建、高通量筛选、反向搜索及约束验证;其次,阐述基于机器学习势函数与通用预训练原子势的界面热输运原子级设计,分析异质固固界面、二维范德华界面和固液界面中的热输运机制、适用边界及界面热边界参数桥接;进一步总结机器学习代理模型、多目标优化和拓扑优化在先进封装、热点抑制、冷板/微通道散热器及器件可靠性设计中的应用。最后,讨论该领域在数据质量、物理约束、模型泛化、跨尺度参数传递与实验闭环验证等方面面临的挑战,并展望其在高功率电子、先进封装和智能热管理器件中的发展前景。

     

    With the rapid development of high-power electronics, wide-bandgap semiconductors, heterogeneous integration, and advanced packaging technologies, heat transport research is shifting from the characterization of individual thermophysical properties toward cross-scale thermal design across materials, interfaces, structures, and devices. Heat conduction involves atomic interactions, interfacial thermal transfer, heat-flow regulation, and device-level heat dissipation. Conventional high-fidelity calculations, empirical models, and case-by-case simulations are often limited by high computational cost, restricted design-space exploration, and insufficient parameter transfer across scales. Artificial intelligence, especially machine learning, provides new methodological support for thermal transport research and high-performance thermal management through data mining, surrogate modeling, and optimization.
    This review focuses on artificial-intelligence-enabled cross-scale thermal design, with machine learning as the central thread. Organized around a chain of data, models, parameter transfer, device optimization, and experimental validation, it first clarifies the boundaries of AI/ML methods, physical solvers, and optimization algorithms. It then reviews thermophysical-property prediction and target-oriented materials design, covering data generation, model construction, candidate generation, inverse design, uncertainty quantification, and validation. Atomistic interfacial design using system-specific machine-learning interatomic potentials and universal pretrained atomistic models is discussed for heterogeneous solid–solid interfaces, van der Waals interfaces involving two-dimensional materials, and solid–liquid interfaces. Attention is paid to cross-scale parameter bridging, including rotation of anisotropic thermal-conductivity tensors, interfacial thermal-conductance boundary conditions, combination of intrinsic and process-induced thermal resistances, and propagation of parameter uncertainty to device-level outputs. Applications of surrogate modeling, multi-objective optimization, and machine-learning-assisted topology optimization are reviewed for advanced packaging, hotspot mitigation, cold-plate and microchannel design, manufacturability, and reliability. A GaN/BAs heterostructure serves as an end-to-end case study of transferring material and interfacial parameters into finite-element models with experimental benchmarking.
    Taken together, AI-enabled cross-scale thermal design is moving beyond isolated prediction and optimization toward an integrated framework linking data generation, parameter transfer, device-level decision making, and experimental feedback. However, realizing this vision requires addressing key challenges in data quality and standardization, physical consistency, out-of-distribution generalization of atomistic models, coarse-graining of nonequilibrium phonon information, cross-scale uncertainty propagation, and closed-loop experimental validation. Future progress will depend on the deeper integration of multiscale simulation, physics-informed machine learning, data-driven optimization, additive manufacturing, and digital twins, ultimately advancing heat transport research toward reliable and intelligent thermal design from atoms to devices under realistic operating conditions.

     

    目录

    /

    返回文章
    返回