Development of self-adaptive low-dimension ventilation models using OpenFOAM: Towards the application of AI based on CFD data

Numerous state-of-art CFD (Computational Fluid Dynamics) studies have shown their validity and feasibility in engineering applications but still lack prediction efficiency. It is of great potential to apply artificial intelligence (AI) on the basis of CFD considering their fast development. Thus, the data-dimension reduction of CFD can be very important for the efficiencies of database construction, training and storage. Our previously developed linear low-dimension ventilation model (LLVM) is able to convert high-resolution CFD data into low-dimension grid levels, facilitated the use of fast prediction for ventilation online control. However, limitation still exists considering the dilemma of prediction speed and accuracy, e.g., case of a larger building space. This is due to the neglect of volume contribution ratio from single mesh as well as correlations of cells information when using uniform low-dimension methods. Therefore, we proposed a self-adaptive non-uniform low-dimension model for the data conversion but using lower dimension size with acceptable accuracy. The open-source CFD platform OpenFOAM was used for the package development, called self-adaptive low-dimension tool (LDT), including two modules, i.e., Mon-uniform dividing' and 'self-update'. Error index was defined considering the contribution ratio of individual mesh volume. A series of cases were carried out for demonstration and evaluation. It is found that the proposed model is able to largely improve the data accuracy but with smaller dimension requirement compared to uniform dividing method (e.g., with comparable error index around 16.5% when using zone numbers of 80 for non-uniform and 210 for uniform). Moreover, the self-update module enables users to efficiently and automatically identify the optimal low-dimension zone numbers. This work can be of great importance for the application of CFD-AI techniques.