AI–CFD Framework for Nanofluid Heat Transfer: Surrogates, Reinforcement Learning, and Super-Resolution

Computational Fluid Dynamics (CFD) is indispensable for analysing complex flow
and heat transfer phenomena, yet its high computational cost restricts scalability and real-time
applications. This study introduces a unified AI–CFD framework that integrates deep learning
(DL) surrogate modelling, deep reinforcement learning (DRL) optimisation, and
super-resolution (SR) visualisation to transform CFD into an adaptive, solver-agnostic
framework. Using CFD datasets from nanofluid heat transfer simulations, convolutional neural
networks and transformer architectures were trained to predict velocity and temperature fields
with high fidelity, achieving 25–30× speedup while maintaining accuracy within 5% error
margins. DRL agents, implemented via Proximal Policy Optimisation, dynamically tuned inlet
velocity and nanoparticle fraction, yielding up to 18% improvement in heat transfer coefficient
compared to baseline CFD control strategies. In parallel, generative adversarial networks
upscaled coarse CFD outputs, reconstructing fine-scale turbulence structures with PSNR
> 32 dB and SSIM > 0.93, thereby enhancing interpretability without prohibitive computational
expense. Together, these modules form a synergistic pipeline that accelerates CFD simulations,
adaptively optimises thermal performance, and restores visual fidelity. Unlike prior studies that
investigated surrogates, RL agents, or SR methods separately, this work integrates all three into
a comprehensive framework, bridging fluid mechanics and artificial intelligence for scalable,
real-time thermal system design. The framework was validated through mesh independence,
solver benchmarking, and spectral fidelity checks across Fluent and OpenFOAM, confirming
robustness and scalability across platforms.