Machine Learning for Multi-Discipline Parametric Analysis in Architectural Practice

Abstract

Parametric analysis performs building performance analysis by simulating multiple design alternatives compared to the single design alternative analysis of conventional methods. It is an emergent, data-driven approach to architectural design. It has the benefits of enabling the optimization of design spaces, reducing uncertainty in design decision making, and establishing the most sensitive input parameters of the design space. However, it also faces data analytical challenges in practical application. These challenges include the size of design spaces, the speed and accuracy of simulation, and the presence of complex design space conditions. This study investigates machine learning algorithms as a design space reduction strategy for overcoming these data analytical challenges and thereby facilitating the implementation of parametric analysis in architectural practice. It focuses primarily on residential projects designed by Perkins&Will Architects in British Columbia. The study asks if machine learning algorithms can be effective at design space reduction for multi-discipline parametric analysis. In addition, it asks which of the algorithms we study - artificial neural networks, support vector machines and random forests - is the most effective? Are machine learning algorithms robust under complex design space conditions? Further, what is the influence on algorithm performance of the impact factors sample size, sensitivity analysis, feature selection and hyperparameter tuning? The methodology for addressing these questions involves the use of real world case study projects, dynamic simulation of the case study design spaces, and machine learning experiments on the simulated design spaces. A methodological framework referred to as Design Space Construction is used for defining the design spaces, integrating the multi-disciplinary simulation infrastructure and evaluating the analysis outcomes. Machine Learning experiments are performed to assess the effectiveness of the algorithms and the influence of performance impact factors (sample size, sensitivity analysis, feature selection and hyperparameter tuning). The study finds that machine learning algorithms can indeed be effective for multi-discipline parametric analysis. In particular, artificial neural networks are the most effective algorithm. Likewise, hyperparameter tuning is the most influential impact factor affecting algorithm performance. Lastly, machine learning algorithms can be robust under complex design space conditions.