Configuration-aware representation for generalizable indoor airflow field surrogates under ceiling fan reconfiguration

Yiting Zhang, Wei Liang, Yue Lei, Roman Buckle, Kian Wee Chen, Ippei Izuhara, Shohei Miyata, Eikichi Ono, Adrian Chong
Building and Environment, Available online 24 August 2026, 115167
https://doi.org/10.1016/j.buildenv.2026.115167

Abstract
Ceiling fans deliver elevated air speed to enhance convective cooling. Coupled with air conditioning, they allow higher cooling setpoints while maintaining thermal comfort, yielding substantial energy savings. Data-driven surrogate models have supported parametric inference for design exploration. To improve generalization, recent studies have incorporated prior knowledge into model inputs, architectures, or training objectives to address incomplete boundary conditions and geometric variations. However, existing surrogate models are often evaluated under fixed operating scenarios, with performance sensitive to the coverage of configurations in the training data. This study develops a configuration-aware representation for indoor air speed field prediction across ceiling fan configurations. The location, operating status, and fan-induced momentum of each fan are encoded as differentiable source fields defined on the target grid and paired with multiple model architectures. Using measured air speed fields from eight fan configurations in a controlled chamber, we evaluated the models under held-out configurations. Compared with a geometry-only baseline, the proposed representation reduces mean absolute error by 35% and increases the spatial correlation with ground truth to 0.815. Across held-out configurations, Fourier neural operator models with the proposed representation achieve the highest accuracy and halve the standard deviation of mean absolute error compared to convolutional neural networks, which struggle to recover local flow features where the test configuration departs from training coverage. The proposed framework enables rapid design-stage evaluation of ceiling-fan configurations within the tested chamber and fan-array setting. With further validation, the representation principle can provide a physically interpretable basis for broader transfer.