Physics-informed neural network framework for fast and accurate nanofluid heat-transfer prediction in microchannel heat sinks

Abstract

Introduction Physics-informed, data-efficient surrogates for microchannel cooling are needed as power densities in electronics and photonics approach thermal limits, yet design workflows still rely on empirical correlations or computationally expensive simulations. This study aims to provide a fast, interpretable, physically consistent, and reliable predictor for heat transfer coefficient, pressure drop, and hotspot temperature in nanofluid microchannels. Methods We developed a hybrid physics-informed neural network trained on more than 10,000 tabulated records curated from a Kaggle repository and treated as compiled heterogeneous operating condition response pairs with limited case-level traceability. The framework uses physics-constrained inputs derived from nanofluid mixture rules and slip-related descriptors, and optimizes a composite loss combining data mismatch, reduced-order conservation-residual penalties, and boundary-condition inconsistency penalties in operating-condition space. To enrich sparse regions, physics-consistent synthetic augmentation was applied by sampling admissible conditions and retaining only feasible samples that satisfied constraint checks. Results The proposed model outperformed strong baselines, achieving a mean absolute error of 11.3, a root mean squared error of 18.4, and a coefficient of determination of 0.96 on the test set. Physics-consistency auditing showed 98.7% feasibility with low residual magnitudes and low boundary-violation rates. Inference profiling yielded approximately 15.7 ms on a GPU and 120 ms on a CPU for single-sample evaluation. Discussion The results support the use of physics-informed learning as an auditable surrogate for rapid design screening in nanofluid microchannel applications. The framework reduces reliance on repeated high-fidelity simulation while reporting residual and boundary metrics alongside predictive accuracy. However, controlled hardware testbed validation is still required before industrial deployment.

More by Galiya Ybytayeva

• Emotionally expressive facial animation driven by EmotionBERT embeddingsScientific Reports · 2026
• Thermal mode modeling using neural network technologies and the finite element methodBulletin of Electrical Engineering and Informatics · 2026
• Precision healthcare in action: A robust MRI framework for breast cancer diagnosis with enhanced feature analysis and boundary refinementBiomedical Signal Processing and Control · 2026
• Explainable multi agent reinforcement learning framework for secure and adaptive communication in UAV swarm based fanetsScientific Reports · 2026 · Citations: 6