Abstract Distributed Acoustic Sensing (DAS) has emerged as a transformative technology for large-scale seismic and CO2 storage monitoring, but field data are frequently contaminated by complex, nonstationary noise that obscures weak subsurface signals. Conventional Machine learning/Deep learning based denoising approaches including convolutional neural networks, autoencoders, and transform-domain filters often require large paired noisy–clean datasets and risk overfitting to empirical patterns, making them unreliable in ground-truth-free deployments. Physics-Informed Neural Networks (PINNs) overcome this limitation by embedding the governing partial differential equations directly into the learning process, enabling noise suppression that remains consistent with physical wave propagation. In this work, we develop and optimize PINN architecture for DAS denoising without any clean dataset using the damped acoustic wave equation as a physics prior. Through staged tuning of network depth, neuron count, learning rate, collocation density, and supervised-to-physics loss weighting, the best configuration achieved a 97% increase in low-frequency energy concentration, 39% reduction in spectral entropy, 43% narrowing of −3 dB bandwidth, and 55% gains in both spatial and temporal smoothness, while preserving overall signal energy. The results demonstrate the potential of PINNs as a robust, physically grounded alternative to purely data-driven DAS denoising methods.
Rasool et al. (Mon,) studied this question.