Accurate rendering of room acoustics requires dense Room Impulse Response (RIR) data across large volumes, a process often constrained by resource-intensive data acquisition. To address this, various soundfield reconstruction techniques have been developed, with physics-guided machine-learning models emerging as the most efficient. This paper presents a method to spatially extrapolate RIRs by applying a recently developed point-neuron learning framework that embeds the fundamental solution of the wave equation directly into its architecture ensuring strict adherence to physical laws. By leveraging prior knowledge of the acoustic environment and generalizable features such as dominant reflection locations and relative frequency response, our approach effectively integrates early reflections and late reverberation into a cohesive model. Comparative evaluations against competing methods demonstrate that the proposed method consistently achieves superior reconstruction performance across diverse acoustic scenarios.
Bastine et al. (2025) studied this question.