Seating absorption modeling in concert hall geometrical-acoustics simulations: Validation and data-driven post-processing

This is the authors' abstract. We don't add key points for this paper.

Accurately representing the absorption of audience seating remains an uncertainty in geometrical-acoustics (GA) room acoustic simulations, because equivalent absorption area measured in a reverberation room does not uniquely translate to band-dependent absorption coefficients in a hall model. This study evaluates practical seat-absorption modeling cases and tests data-driven post-processing to improve prediction accuracy. A 394-seat rectangular hall with a single-level audience area and stepped seating was measured in two conditions—without seats and with seats—at 16 receiver positions over octave bands (125–4000 Hz). Seat absorption was measured following ISO 354, and an as-built ODEON model was calibrated under the without-seats condition; the seating cases were then applied to predict the with-seats condition. Accuracy for EDT, T 20 , T 30 , C 80 , and D 50 was evaluated using RMSE and JND-normalized standardized RMSE. Although one case showed the most balanced overall performance, the best case was not unique: the optimal modeling assumption depended on frequency band and acoustic parameter. Interaction analysis confirmed significant case-by-band effects for reverberation-related parameters ( T 20 , T 30 ), whereas clarity parameters ( C 80 , D 50 ) showed stable dominance patterns. To move beyond single-case selection, two post-processing approaches were assessed with leave-one-out cross-validation (LOOCV): a frequency × parameter-dependent weighted ensemble and a residual-learning bias correction. The ensemble yielded modest but consistent RMSE reductions (up to 14.6 % for EDT), while residual learning reduced errors for reverberation-related parameters and mitigated systematic bias. The proposed framework supports conditional use (or combination) of seating cases by frequency and parameter, improving room acoustic predictions.