Integrating parametric design and machine learning for room acoustics prediction in the early design stage of concert halls

Abstract

The room acoustics design of concert halls plays an important role in sound quality, and machine learning has been proposed as an advanced approach for predicting room acoustics. However, previous studies often use geometric (such as length, width, and height) and material parameters as features, resulting in high accuracy but limited generality when applied to different room types. Hence, this study proposed a machine learning based parametric design framework on the Rhino-Grasshopper platform. The approach used distributed equivalent absorption areas as features by considering the importance of the first reflection in concert halls, thus improving generality in the early design stage. Trained on a dataset combining two different types of concert halls, the fivefold cross-validation results demonstrated that the Mean Absolute Percentage Error (MAPE) for Reverberation time (T₃₀) and the Root Mean Square Error (RMSE) for both Clarity 80 (C₈₀) and Sound strength (G) remained within 1 Just Noticeable Difference (1 JND). The high accuracy highlights the advantages of this framework in comparing different types of concert halls and broadens the capability of architects to optimize design proposals in the early design stage.