Component-Based Machine Learning for Multi-Element Aggregation and Interaction

Accurately and efficiently predicting airflow, mass transport, and temperature distribution in indoor environments is essential for the design phase, supporting exploration and decision-making with respect to indoor comfort and energy efficiency. However, the high computational cost of computational fluid dynamics (CFD) simulations remains a significant challenge and barrier for application of such methods. The data-driven models render a high-potential fast alternative to replace the CFD simulation to predict the indoor environment. However, generalization and reusability are still limited. Therefore, we propose component-based machine learning (CBML) for spatial flow prediction with better generalization ability than the current machine learning methods. In this paper, we tackle the aggregation of two data-driven prediction components; the success of this aggregation forms a fundamental method of CBML for CFD. The CBML surrogate model for this purpose includes three sub-models, a convolutional autoencoder with residual network (CAER), a multilayer perceptron (MLP), and a convolutional neural network (CNN). Testcase data represent a 2D rectangular room equipped with two inlets on the left and right wall; prediction of the aggregation of two separate predictions of flow caused of left inlet and right inlet forms training and test object in this paper. The CAER serves an order reducer for reducing the dimensionality of spatial data and extracting the features of components. The MLP works as a predictor to map the boundary conditions with component features. Last but not least, the CNN is regarded as an aggregator to discover how two components affect others. Comparison of predictions with CFD simulation results shows a maximum absolute error of less than 0.09 m/s for 95% of the flow field in real-time prediction. Besides, the compressed features of single-inlet and multi-inlet are mapped into latent space by t-SNE, illustrating the correlations between features of different components. These outputs demonstrate that the CBML approach is able to predict the two aggregated flow fields, thus, effectively captures complex feature interactions among multiple components, and the latent features contain useful fluid information.