A Neural Network-Based Performance Assessment Approach for Heat-Dissipating Microchannels in Packages

To address the severe computational bottleneck imposed by three-dimensional conjugate heat transfer (CHT) simulations in the thermal design of micro-channel heat sinks (MCHSs), this study develops a high-throughput automated CFD pipeline that integrates geometric feasibility constraints with a cryptographic SHA256-based data-freezing mechanism. From 791 randomly generated designs with extreme geometric characteristics, 638 valid high-fidelity structured samples were retained after strict numerical and geometric screening. On this physically consistent dataset, a physics-informed machine learning (PIML) surrogate was established to map core geometric and derived features to multi-objective thermo-hydraulic responses. For the most challenging target, pressure drop, fluid-mechanics priors including v2 and 1/Dh were explicitly injected into the learning space to improve robustness under highly nonlinear operating conditions. Quantitative results show that the surrogate achieved R2 >.094 on key heat-transfer-related targets such as Nu, maintained a global average MAPE of approximately 567. %, and constrained the local error in the extreme high-pressure-drop regime to 907. %. By replacing kilosecond-level CFD iterations with millisecond-level forward inference, the proposed framework delivered an acceleration of up to O(106) relative to conventional CFD, enabling real-time design-space exploration for next-generation chip thermal management.