Hybrid AI model cuts sound field interpolation time 1,000x

Researchers (Hayato Komaba, Gen Sato, Ken Kurata, Yusuke Ikeda) have proposed a hybrid framework that dramatically speeds up sound field interpolation. Traditional physics-informed neural networks (PINNs) are accurate but slow, requiring iterative fine-tuning for each new sound field. Their new method combines PINN pre-training with a physics-informed extreme learning machine (PIELM) tailored for acoustic fields. Instead of re-training the entire network, the approach uses pre-trained hidden-layer weights and only adapts the output layer via closed-form computation—eliminating the need for iterative optimization.

In simplified 1D free-field simulations, the hybrid model achieves interpolation accuracy comparable to standard PINN fine-tuning, but with adaptation time reduced by more than three orders of magnitude (i.e., over 1,000x faster). This breakthrough addresses a key bottleneck for applications requiring real-time processing or online learning, such as spatial audio in AR/VR, active noise control, or adaptive acoustic sensing. The work has been submitted to IEEE for publication, signaling its relevance to the audio and speech processing community.