Heat Exchanger Optimization Using CFD + Machine Learning

A design system for the air side of a cross-flow tube-bank heat exchanger that combines its own computational fluid dynamics (CFD) solver, machine-learning surrogates and multi-objective optimisation, and checks every optimised design again with CFD. A two-dimensional lattice-Boltzmann solver (D2Q9 flow, D2Q5 temperature, Numba on the CPU and CUDA through CuPy on a laptop GPU) is validated against the exact laminar channel solution (Nu 7.528 vs 7.541) and published single-cylinder data, and then simulates 320 tube-bank designs (4 rows, Re 40-160, transverse and longitudinal pitch, round or elliptical tubes, inline or staggered) in 32 minutes. A Gaussian-process surrogate predicts the Nusselt number with a 0.77 % error and the pressure-drop (Euler) number with 3.13 % on 40 unseen CFD cases (the Zukauskas textbook correlation: 29.3 %). NSGA-II finds the Pareto front of heat transfer against pressure drop at Re 60, 100 and 140; CFD confirms compact staggered banks with 14-21 % more heat transfer at 19-28 % more pressure drop, and at Re 60 a design with +6.1 % heat transfer and -5.9 % pressure drop against the textbook baseline. A role-based web application exposes prediction, CFD checks, optimisation and the full CFD data base.

Category: Mechanical Engineering / Thermal-Fluid Design / Computational Fluid Dynamics / Machine Learning Surrogate Modelling / Multi-Objective Optimisation

Mechanical Engineering Project M02 from ProTechSC Academy, with graphical abstract, video demonstration and project guidance.