Building Ventilation and Thermal Comfort Prediction Using CFD + ML

A student-laptop project that predicts how well a room is ventilated and how comfortable its occupants are, and then designs the air distribution for the lowest HVAC power. A 2-D finite-volume CFD solver written for the project (incompressible Navier-Stokes + energy with buoyancy, staggered MAC grid, projection method, Chen-Xu zero-equation indoor turbulence model, age-of-air solver) was verified against the Ghia lid-driven cavity and the de Vahl Davis heated cavity (Nusselt number within 0.94 % on 64 x 64) and ran 360 cases of a parametric office section with mixing, ceiling-diffuser and displacement ventilation (3.97 CPU-hours). ISO 7730 / ASHRAE 55 comfort indices (PMV, PPD, draught rate, head-ankle temperature difference, ADPI, age of air, air-change effectiveness) were computed for every case. Machine-learning surrogates (Gaussian process selected for 8 of 11 targets, test R2 0.982 for the occupied-zone temperature, 0.980 for PMV) and a POD + Gaussian-process field model predict comfort and full flow fields in milliseconds. A CFD-verified optimiser finds designs that meet comfort category B with 50 % (summer office) and 47 % (classroom) less HVAC power than a conventional 6-ACH mixing design, which itself fails category B in CFD. A FastAPI web application with three roles and five database options makes everything usable.

Category: Mechanical Engineering / HVAC and Building Services / Computational Fluid Dynamics / Thermal Comfort / Machine Learning / Design Optimisation

Mechanical Engineering Project M07 from ProTechSC Academy, with graphical abstract, video demonstration and project guidance.