Simulation? Machine Learning? Simulation X Machine Learning?: A decision system for research integrating building physic simulation and machine learning methods in the early design stage

Abstract

Researchers have leveraged machine learning technologies and physics-based simulation in predicting daylight and other factors relevant to building energy consumption. However, this is still an emerging research area with comparatively less literature volume than its respectable fields. The less common expertise in building physics and machine learning is one of the significant attributes of the comparable smaller field. Moreover, there is no generalized method outlining the thought process behind integrating simulation and machine learning methods or cost-benefit analysis of choosing to implement simulation, machine learning, or both in the current literature. This thesis proposes a framework that identifies the considerations researchers should ask step-by-step in the simulation and machine learning workflow and analyzes these methods' advantages and drawbacks. The proposed framework is demonstrated with two proof of concept case studies. The first case study used daylight simulation with Climate Studio and Grasshopper to generate synthetic data automatically to train pix2pix, a conditional generative adversarial network (cGAN). The model was hosted on a web interface using p5.js that allows users to create their building designs and provide design feedback simultaneously. The second case study is a thought experiment that employs pedestrian wind comfort with Eddy3D and an artificial neural network (ANN) model in Python.