Zhao Wensheng | MSc in Electronic Science and Technology

Adaptive Deep Reinforcement Learning Optimization Design Process for Hybrid Pin-Fin Microchannel Heat Sink Based on Hybrid Neural Network Acceleration

Wed, 01 Jan 2025 00:00:00 +0000

With the growing complexity of integrated circuits and heterogeneous microsystem integration, effective thermal management and intelligent design are crucial for preventing thermal failures. This study proposes an adaptive deep reinforcement learning-based (ADRL) optimization method for pin-fin microchannel heat sinks, enhancing iterative optimization efficiency with a hybrid neural network (HNN). A hybrid pin-fin microchannel model is established, with simulation-generated datasets capture temperature and pressure drop performance under various parameters. The proposed HNN significantly improves the prediction accuracy under small sample conditions. Compared with deep neural network and convolutional neural network, the root mean square error, mean absolute error, relative error, and comprehensive standard deviation are reduced by 92.6%, 93.7%, 92.5%, 97.2% and 96.4%, 93.7%, 90%, 92.3% respectively, effectively mapping structural parameters to performance indicators. Using the proximal policy optimization algorithm in the ADRL framework, this method optimizes the hybrid heat sink layout under non-uniform heat flow conditions, reducing the maximum temperature of high-power heat sources by 5.34%, the maximum temperature of low-power heat sources by 10.01%, and the pressure drop by 53.05%.

Multiobjective Deep Reinforcement Learning Driven Collaborative Optimization of TSV-Based Microchannel and PDN for 3-D ICs

Wed, 01 Jan 2025 00:00:00 +0000

This study introduces a multiobjective deep reinforcement learning (MODRL) framework for the concurrent thermal-hydraulic optimization of the through-silicon-via (TSV) microchannel heat sink (MCHS) embedded in 3-D integrated circuits (3-D ICs) power delivery network (PDN). By exploiting the inherent structural synergy between TSVs and pin-fin MCHS, the proposed method enhances thermal management in high-density 3-D ICs. The framework integrates deep reinforcement learning (RL) with multiobjective optimization and computational fluid dynamics (CFD) simulations, enabling an efficient exploration of the high-dimensional design space to resolve tradeoffs between thermal efficacy and fluidic resistance. Relative to baseline, the optimized design achieves a reduction in maximum chip temperature of up to 3.3% while concurrently lowering the overall pressure drop by 17.2%. Impedance analysis further validates the design’s superiority, showing that the optimized TSV geometry effectively suppresses high-frequency peak impedance. Compared with standard deep reinforcement learning (SDRL) and genetic algorithm (GA), MODRL converges faster by 57.1% and 62.5%, respectively, showing stronger convergence. These results highlight the advantages of the MODRL intelligent optimization framework in design speed and its great potential in driving the development of next-generation 3-D integrated circuits, especially in applications requiring high power density and high reliability.

Smart cooling: Hydrogel-enhanced adaptive jet impingement utilizing through silicon via for integrated microsystems

Wed, 01 Jan 2025 00:00:00 +0000

In three-dimensional integrated circuits, variations in power consumption across different regions lead to uneven temperature distribution, which can compromise system stability and reliability. While microchannels etched on the chip’s backside are commonly used for cooling, traditional designs provide a fixed cooling capacity and are often inefficient in targeting specific hotspots. Moreover, straight microchannels spanning the entire system can result in overcooling in low-power areas and insufficient cooling in high-power regions.

This study presents a novel design of an adaptive jet impingement cooling structure that combines hydrogel, jet impingement heat sink (JIHS) and through silicon via (TSV) technology. The structure features vertical channels and utilizes the thermally induced deformation of hydrogel to achieve adaptive cooling. This design allows the cooler to be strategically placed at hotspots, dynamically adjusting microfluidic injection in response to temperature fluctuations. As a result, overcooling in low-power regions and inadequate cooling in hotspots are mitigated, improving thermal uniformity. Compared to conventional jet impingement heat sinks, the proposed adaptive jet impingement heat sink improves temperature uniformity by 12.21 %, reduces thermal spreading resistance by 13 %, and increases maximum total thermal resistance by only 3.08 %. The maximum pressure drop increases by just 1.28 kPa. Therefore, with the increasingly complex integrated microsystem architecture, the adaptive impingement jet heat sink has better comprehensive heat dissipation performance than the traditional impingement jet heat sink under complex heat distribution.

Hybrid neural network based multi-objective optimal design of hybrid pin-fin microchannel heatsink for integrated microsystems

Mon, 01 Jan 2024 00:00:00 +0000

With the rapid advancement of 2.5D/3D heterogeneous integrated microsystems, the performance and fast intelligent design for thermal management are unprecedentedly required to address the electrical and mechanical reliability issues caused by thermal runaway. In this work, a hybrid neural network, featuring a small dataset requirement, is developed to accelerate the design of the hybrid pin-fin microchannel heatsink. Assisted by the trained surrogate model and the non-dominated sorting genetic algorithm, a powerful heatsink characterizing power-adaptive cooling capacity is designed. Firstly, a hybrid pin-fin microchannel heatsink is modeled and validated with the experimental data. The critical structural parameters correlated with the heat transfer and hydraulic performance are analyzed and identified through numerical simulation. A hybrid neural network serving as a surrogate model, is then developed to map the relationship between key structural parameters and the targeted performance indexes. The hybrid neural network achieves a prediction accuracy of at least 94.33 % and outperforms traditional networks, including DNN and CNN, in RMSE, MAE, and RE. It improves by 93.4 %, 89.5 %, and 87.8 % over DNN, and by 91.7 %, 93.0 %, and 91.9 % over CNN. The non-dominated sorting genetic algorithm is performed to explore the Pareto front where the intelligent design of power-adaptive pin-fin layout under uneven thermal profile is achieved. The performance indexes of the optimized heatsink are validated with that from the computational fluid dynamics. Compared with the original structure, it is found that enhancements of 5.58 %, 10.76 % and 45.73 % are achieved in the maximum temperature of high-power heat source, low-power heat source and the pressure drop of microchannel.