The performance of digital receivers used in modern military radar, communication, and surveillance systems is often limited by the performance of the analog-to-digital converter (ADC) used to digitize the received signal. Ultrafast ADCs are critical in military applications such as military software-defined radio, radar, and electronic counter-warfare (ECW) that require high sampling rates and large bandwidths.
Mainly, an electronic ADC must carry out two functions over a time-varying signal. On the one hand, it samples the signal and holds it for a period of time, on the other, quantifies those samples in a certain number of levels. To do so, two important parameters are taken into account: the sampling period and the number of levels. The number of levels is set by the number of bits of the specific ADC. As a general rule, the ADCs behaviour gets worse both for amplitude and timing errors, being the jitter the most important one for timing errors and the thermal noise and non-linearities the most important one for amplitude errors. All these error sources deteriorate the effective number of bits (ENOB) in the ADC
Traditionally, in modern information systems, electronic analog-to-digital conversion methods have supported high-accuracy quantization and operational stability due to the mature manufacturing of electronic components; nevertheless, their bandwidth limitations and high timing jitter hinder the development of electronic methods toward broadband high-accuracy ADCs for next-generation information systems. Aperture jitter i.e. inability of ADCs to sample at precisely defined times has been a major bottleneck on the way towards higher speeds and better accuracy.
Facilitated by photonic technologies, the bottlenecks of bandwidth limitations and timing jitter are elegantly overcome. Photonic ADCs, operating with very low timing jitter, are envisaged to bring ADC performance to new levels.
Photonic ADCs, which perform sampling using ultra-stable optical pulse trains generated by Ultra-stable mode-locked lasers with jitter levels many orders of magnitude lower exist today could improve ADC performance by orders of magnitude.
“This jitter scaling from microwave to optical sources, together with rapid progress in electronic-photonic integration via the silicon photonics technology platform, gives confidence that the orders of magnitude in jitter reduction, possible with mode-locked lasers, can be exploited to bring ADC performance to levels well beyond what is possible today,” write Anatol Khilo, Steven J. Spector and others.
Photonically-assisted ADCs (PADC) can offer significant advantages to the standard all-electronic approach by using photonic sampling with ultra-low jitter pulses for superior performance. Moreover, by optically subsampling the signal a direct downconversion can be performed at the same stage, obtaining substantial improvements in performance by achieving a very precise sampling of the signals thanks to the ultra-low jitter of the pulsed lasers, a signal downconversion without mixing and conversion stages, reducing the spurious and the LO leakage and saving mass, volume and complexity thanks to the use of distributed architectures and fiber optic remote delivery, especially in the proposed scenario in which hundreds of digitalizers are required, write S. Pantoja and others scinetists from Spain.
The concept of a photonic satellite where optical subsystems reduce the microwave and digital electronics complexity on-board communications spacecraft is considered very attractive for its potential to reduce the mass and volume of a payload requiring increased processing bandwidth.
The authors have built the laboratory demonstrator of a photonically-assisted ADC for Broadband regenrative satellite payloads that will have to receive, route and retransmit hundreds of channels and need to be designed so as to meet such requirements of larger bandwidth, system transparency and flexibility. One important device in these new architectures is analog to digital converter (ADC) and its equivalent digital to analog converter (DAC). These will be the in/out interface for the use of digital processing in order to provide flexible beam to beam connectivity and variable bandwidth allocation.
In general, there are four main classes of photonic ADCs: 1) photonic assisted ADC in which a photonic device is added to an electronic ADC to improve performance, 2) photonic sampling and electronic quantizing ADC, 3) electronic sampling and photonic quantizing ADC, and 4) photonic sampling and quantizing ADC.
Deep-learning-powered photonic analog-to-digital conversion
Scientists have used deep learning technologies to improve the performance of photonic analogue-to-digital converters (ADCs), opening the door to advanced information systems. Modern information processing technologies, such as radar, imaging, and communications systems, require ADCs that are high speed, accurate and have broadband capacity. Although electronic ADCs offer high speed and accuracy, their bandwidth is limited, hindering the development of more advanced information technologies.
Facilitated by photonic technologies, the bottlenecks of bandwidth limitations and timing jitter are elegantly overcome. However, since the imperfect properties and setups of photonic components give rise to system defects and can deteriorate the performance of ADCs, designing an advanced ADC architecture remains challenging. Now, Weiwen Zou and colleagues from Shanghai Jiao Tong University in China have overcome this limitation by using deep learning technologies to improve the performance of photonic ADCs.
Recently, deep learning technologies have made substantial advances in a variety of artificial intelligence applications, such as computer vision, medical diagnosis, and gaming. By constructing multiple layers of neurons and applying appropriate training methods, data from images, audio, and video can be automatically extracted with representations to be used in the inference of unknown data.
Data recovery and reconstruction tasks, including speech enhancement, image denoising, and reconstruction, are well accomplished with convolutional neural networks (CNNs, neural networks based on convolutional filters), thereby demonstrating the ability of deep neural networks to learn the model of data contamination and distortion and to output the recovered data. Therefore, it is believed that machine learning technologies, including deep learning, can offer substantial power for photonic applications, write chinese scinetists.
By combining the data recovery capabilities of neural networks with the technical advantages of electronics and photonics technologies, the team created a photonic ADC that overcomes the limitations on bandwidth, laying the foundations for next-generation information systems, including ultra-wideband radars and high-resolution microwave imaging.
Ther deep-learning-powered photonic analog-to-digital conversion (DL-PADC) architecture is composed of three main cascaded parts: a photonic front-end, electronic quantization, and deep learning data recovery. In the photonic front-end, a low-jitter pulsed laser source provides the sampling optical pulse train and the precise quantization clock, thereby ensuring low noise from the source.
An electrooptic modulator (E/O) subsequently provides broadband radio frequency (RF) reception by incorporating the photonic advantage in terms of signal bandwidth. Via optical multichannelization, the sampling speed in each channel is lowered for compatibility with the electronic quantization. Driven by the precise quantization clock from the optical source, electronic quantizers are exploited with their high quantization accuracy.
In practice, the defects in the photonic front-end can pervade the quantized data; hence, the deep learning data recovery realizes distortion elimination of the quantized data, which is essential for overcoming the tradeoff among bandwidth, sampling rate, and accuracy for traditional ADCs.
Deep learning data recovery includes two steps with two functional neural networks: “linearization nets” and “matching nets.” The former executes nonlinearity elimination and the latter interleave data in multiplexed channels with channel mismatch compensation. Unlike the traditional dual-balanced linearization method, the linearization nets require no complicated setups or miscellaneous data processing steps.
In addition, the matching nets accomplish the interleaving via time-domain representations. Because time-domain representations avoid the problems of data length variation and spectrum aliasing, matching nets are more effective than spectral analysis algorithms, which are adopted in state-of-the-art mismatch compensation schemes.
References and Resources also include: