Abstract

Over the last couple decades, machine learning (ML) and advanced computational strategies have been successful when employed in building physics problems. ML models have known applications in the operations of buildings, including real-time energy monitoring and automated mechanical systems. They can also be used to inform design decisions for complex problems during conceptualization. For acousticians and building designers, trained ML models could provide better insight on expected performance without requiring extensive calculations or room acoustic simulations. This presentation describes a computational method for predicting the design performance (i.e., mid-frequency reverberation time) of various concert halls. More specifically, a random forest (RF) regression model is trained on architectural and measured acoustical data to predict the reverberation time. The RF models are tuned and act as a surrogate model for predicting the mid-frequency reverberation time of concert halls that were not used in model training. The estimated reverberation times are then compared to calculations from conventional equations (i.e., Sabine and Norris-Eyring equations) and room acoustic simulations using ray-tracing methods. This work shows how acousticians and architects can employ ML strategies in the conceptual design phase to improve estimated performance accuracy without expending significant computational resources.