Ventilation flow reconstruction based on mean velocity measurement using physics-informed neural networks

Abstract

This study proposes a method using Physics-Informed Neural Network (PINN) to reconstruct high-resolution ventilation flow fields from sparse measurement data, addressing spatial resolution limitations in traditional experiments. Data from scaled-model aircraft cabin experiments validated the approach, demonstrating effective utilization of mean velocity data to enhance resolution. The results revealed accurate velocity field reconstruction when using the complete dataset, including regions challenging for measurement using particle image velocimetry. Even with reduced measurement data, maintaining approximately 300 or more data points ensured reconstruction accuracy with deviations below 10% of inflow velocity. Using fewer than 100 data points showed localized accuracy decline, yet key flow features remained.