{"title":"Interpretable machine learning enabled multi-objective design optimization of fractal-like microchannel heat sinks","authors":"Jingbei Bai, Yaojun Li, Pincheng Xiang, Min Chen","doi":"10.1016/j.applthermaleng.2026.133804","DOIUrl":null,"url":null,"abstract":"<div><div>Fractal-like microchannel heat sinks are promising for high-heat-flux thermal management, but their optimization remains challenging because geometry, hierarchical flow resistance, and heat-transfer performance are strongly coupled under fixed-volume constraints. This study examines whether a physically interpretable resistance scale can be combined with machine learning to avoid imposing a prescribed successive-length rule and support multi-objective optimization. A Poiseuille-based network-resistance descriptor was derived, and 3839 training simulations and 200 independently generated test simulations were constructed. Symbolic regression identified explicit low-complexity relationships for five thermo-hydraulic responses, while a physics-informed multi-task learning model, denoted PI-MTL, learned the residuals between the symbolic predictions and numerical results. With the full training set, PI-MTL achieved a mean coefficient of determination (<em>R</em><sup><em>2</em></sup>) above 0.987. With 20% of the training data and 5% added noise, the mean <em>R</em><sup><em>2</em></sup> remained 0.967, compared with 0.954 for conventional MTL and 0.886 for SimpleMLP. SHAP analysis identified the level-0 channel length as the most influential variable. Relative to two Murray-law-based baselines, the balanced design reduced pressure drop by 5.8–23.0% and bottom-surface temperature difference by 44.2–52.1%, while increasing Nusselt number by 10.6–39.3% and coefficient of performance by 8.2–28.2%. The disk-shaped case revealed non-monotonic hydraulic behavior and a localized Pareto trade-off. Favorable designs used a shorter level-0 channel than the level-1 channel. Smaller aspect ratios improved thermal performance at the cost of increased pressure loss, whereas larger successive width ratios reduced the hydraulic penalty when combined with an appropriate channel-length allocation. The framework provides an interpretable, data-efficient, and transferable approach to multi-objective optimization of hierarchical microchannel heat sinks.</div></div>","PeriodicalId":8201,"journal":{"name":"Applied Thermal Engineering","volume":"309 ","pages":"Article 133804"},"PeriodicalIF":7.5000,"publicationDate":"2026-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Applied Thermal Engineering","FirstCategoryId":"5","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S135943112604113X","RegionNum":2,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"2026/10/28 0:00:00","PubModel":"Epub","JCR":"Q2","JCRName":"ENERGY & FUELS","Score":null,"Total":0}
引用次数: 0
Abstract
Fractal-like microchannel heat sinks are promising for high-heat-flux thermal management, but their optimization remains challenging because geometry, hierarchical flow resistance, and heat-transfer performance are strongly coupled under fixed-volume constraints. This study examines whether a physically interpretable resistance scale can be combined with machine learning to avoid imposing a prescribed successive-length rule and support multi-objective optimization. A Poiseuille-based network-resistance descriptor was derived, and 3839 training simulations and 200 independently generated test simulations were constructed. Symbolic regression identified explicit low-complexity relationships for five thermo-hydraulic responses, while a physics-informed multi-task learning model, denoted PI-MTL, learned the residuals between the symbolic predictions and numerical results. With the full training set, PI-MTL achieved a mean coefficient of determination (R2) above 0.987. With 20% of the training data and 5% added noise, the mean R2 remained 0.967, compared with 0.954 for conventional MTL and 0.886 for SimpleMLP. SHAP analysis identified the level-0 channel length as the most influential variable. Relative to two Murray-law-based baselines, the balanced design reduced pressure drop by 5.8–23.0% and bottom-surface temperature difference by 44.2–52.1%, while increasing Nusselt number by 10.6–39.3% and coefficient of performance by 8.2–28.2%. The disk-shaped case revealed non-monotonic hydraulic behavior and a localized Pareto trade-off. Favorable designs used a shorter level-0 channel than the level-1 channel. Smaller aspect ratios improved thermal performance at the cost of increased pressure loss, whereas larger successive width ratios reduced the hydraulic penalty when combined with an appropriate channel-length allocation. The framework provides an interpretable, data-efficient, and transferable approach to multi-objective optimization of hierarchical microchannel heat sinks.
期刊介绍:
Applied Thermal Engineering disseminates novel research related to the design, development and demonstration of components, devices, equipment, technologies and systems involving thermal processes for the production, storage, utilization and conservation of energy, with a focus on engineering application.
The journal publishes high-quality and high-impact Original Research Articles, Review Articles, Short Communications and Letters to the Editor on cutting-edge innovations in research, and recent advances or issues of interest to the thermal engineering community.