Reviewed by Lexie CornerMar 19 2025
Researchers at Nanjing University, led by Ningmu Zou, have developed a new approach to address the major challenges associated with distributed acoustic sensing (DAS). To process data from DAS systems, they created a new Time-Wavelength Multiplexed Photonic Neural Network Accelerator (TWM-PNNA).
By directly leveraging light signals received from distributed acoustic sensing systems, the proposed photonic neural network architecture provides massive gains in accuracy and efficiency over conventional electronic computations. Image Credit: N. Zou (Nanjing University).
Distributed acoustic sensing (DAS) systems are advanced tools used in infrastructure monitoring. They can detect even the smallest vibrations across tens of kilometers of fiber optic cables. These systems are highly effective for applications such as railway monitoring, submarine cable surveillance, oil exploration, and earthquake detection. However, the large volumes of data produced by DAS systems create a significant processing speed bottleneck, limiting their utility in real-time applications that require quick responses.
Neural networks and other machine learning techniques have emerged as viable solutions to process DAS data more efficiently. Despite improvements in traditional electronic computing, using CPUs and GPUs, their speed and energy efficiency still face fundamental limitations.
Photonic neural networks offer a promising alternative. They use light instead of electricity for computations, which can lead to much faster processing speeds while using significantly less power.
However, integrating these optical computing systems with DAS technologies presents several challenges, particularly in managing complex data structures and ensuring precise signal processing.
This groundbreaking work represents the first successful integration of photonic neural networks with DAS systems that can handle real-time data processing.
Ningmu Zou, Nanjing University
To convert traditional electronic neural network operations into optical processes, the researchers developed a system architecture. In this system, the convolution kernels of the neural network, which act as mathematical filters to extract features from input data, are represented by multiple tunable lasers emitting light at different wavelengths.
To achieve this, the team initially used the Mach-Zehnder modulator to convert the two-dimensional data from the DAS systems into one-dimensional vectors that could then be encoded onto optical signals.
By utilizing a wavelength-selective switch, the researchers successfully implemented convolution operations using light signals instead of electronic computations. This switch allocates specific weights to different wavelength channels.
The researchers focused on two main technical challenges: minimizing the effect of modulation chirp (frequency fluctuations) on optical convolutions and developing reliable methods for optical full-connection operations.
Through extensive testing, the team identified that a critical factor in assessing performance is the ratio of wavelength shift caused by modulation chirp to the wavelength spacing between adjacent laser channels.
They found that recognition accuracy is significantly reduced when this ratio exceeds 0.1. By lowering this ratio or implementing a technique called push-pull modulation, the researchers were able to reduce the effect of chirp and achieve classification accuracy above 90 %, comparable to the 98.3 % accuracy of traditional electronic systems.
The researchers also found that the system could maintain a classification accuracy above 90 % as long as at least 60 % of the complete connection parameters were retained after pruning.
This discovery allows for further reductions in computational load and model size without compromising performance. This could lower costs and simplify the production of these optical systems.
The proposed TWM-PNNA system demonstrated impressive computational power, achieving 1.6 trillion operations per second (TOPS) with an energy efficiency of 0.87 TOPS per watt. The system could theoretically reach 81 TOPS with an energy efficiency of 21.02 TOPS per watt, significantly outperforming similar electrical GPUs.
TWM-PNNA enables the all-optical integration of DAS with high-speed computational systems, providing a unique computational framework for DAS applications. This represents a critical step toward next-generation infrastructure monitoring technology capable of processing vast amounts of sensor data in real-time.
If fully realized, DAS systems have the potential to revolutionize applications in seismic monitoring, transportation safety, and the protection of critical infrastructure.
Journal Reference:
Yu. F., et al. (2025) Time-wavelength multiplexed photonic neural network accelerator for distributed acoustic sensing systems. Advanced Photonics. doi.org/10.1117/1.AP.7.2.026008