
- Photonics Research
- Vol. 13, Issue 6, 1654 (2025)
Abstract
1. INTRODUCTION
The current development of the next generation of mobile communication is aimed at achieving higher data rates, lower latency, and greater capacity, while the available spectrum resources for wireless communication are significantly congested. To establish a high-capacity three-dimensional communication network that integrates air, land, and sea, it is essential to extend wireless communication into higher frequency bands, such as millimeter wave, terahertz, infrared, visible light, and ultraviolet. The wavelength of visible light ranges from 380 to 760 nm, corresponding to an unlicensed frequency band of 400 to 800 THz, which is free from electromagnetic interference and has considerable potential for high-speed transmission [1,2]. Furthermore, the visible light spectrum coincides with the transmission window underwater, offering higher data rates and longer transmission ranges compared to traditional underwater acoustic communication, thus becoming a new research hotspot in recent years [3–5].
Visible light communication (VLC) transmitters can be categorized into incoherent light sources, such as light-emitting diodes (LEDs), and coherent light sources, such as laser diodes (LDs). In comparison to LEDs, LDs possess several advantages including higher coherence, smaller divergence angles, greater power, more focused beams, and larger modulation bandwidths. Therefore, underwater visible light laser communication (UVLLC) is expected to play a significant role in the construction of integrated three-dimensional communication networks encompassing air, land, and sea [6,7].
Figure 1 summarizes the recent achievements in UVLLC systems [8–21]. High-speed and long-distance transmission remains the competitive focus of current UVLLC research. To achieve this goal, efforts must be directed towards the development of devices, system design, and optimization algorithms.
Sign up for Photonics Research TOC. Get the latest issue of Photonics Research delivered right to you!Sign up now
Figure 1.Recent achievable transmission distance and data rate of UVLLC systems.
At the device level, spontaneous-emission-based LEDs are primarily designed for illumination and are limited by their size and carrier lifetime, achieving modulation bandwidths of only about 1 GHz [22]. By reducing the size of LEDs, parasitic capacitance can be decreased, thus lowering the resistance capacitance (RC) time constant. This can also enhance the tolerable injection current density and consequently reduce the carrier lifetime. As a result, micro-LEDs can improve modulation bandwidth [23–28]. In contrast, the bandwidth of stimulated-emission-based LDs is a result of the interplay between carrier concentration changes and photon density changes. As the photon lifetime is much shorter than the carrier lifetime, the modulation bandwidth of LDs is not limited by the carrier lifetime [29]. Therefore, LDs can achieve modulation bandwidths significantly higher than those of LEDs. In the past decade, researchers have made significant advancements in underwater visible light communication based on LDs as Fig. 1 shows. Besides, in UVLLC systems, the attenuation of blue and green light is minimal [30,31], making the research on GaN-based blue-green laser diodes crucial for UVLLC applications.
Figure 1 also indicates that simple point-to-point communication based on a single wavelength struggles to attain breakthroughs in transmission capacity. The maximum achievable data rates for current single-wavelength point-to-point communication using red, green, and blue lasers are only 25 Gbps [13], 15.004 Gbps [18], and 19.02 Gbps [20], respectively. Merely focusing on device-level design and optimization is insufficient for a substantial increase in transmission capacity; therefore, at the system level, exploring multi-dimensional multiplexing of light beams, including wavelength, polarization, and mode, as well as the integration of multiple single-component devices, is becoming an essential trend. In previous studies, breakthroughs in underwater communication have been made through red, green, and blue (RGB) wavelength division multiplexing (WDM) and polarization division multiplexing (PDM), successfully achieving underwater visible light transmission speeds of 102.2 Gbps [21].
However, visible light communication encounters challenges such as device noise, nonlinear effects, and inter-symbol interference. Therefore, signal quality should also be enhanced at the algorithmic level. Traditional digital signal processing (DSP) algorithms, such as Volterra filters [32] and least mean squares (LMS) algorithms [33], have limited capabilities in compensating for nonlinearities. With the rise of artificial intelligence (AI), many researchers have introduced various AI algorithms for equalization in VLC, including those based on deep neural networks (DNNs) [34–36], bi-directional recurrent neural networks [37–39], and reservoir computing algorithms [40]. Given the presence of common-mode interference components in the interference of UVLLC systems, this type of interference can be eliminated by transmitting both the original and reverse signals and then performing differential processing of their corresponding receiving signals [41]. Unfortunately, this approach sacrifices half of the transmission rate; hence, we attempt to overcome this limitation. For example, we can introduce a neural network (NN) at the receiver that generates a receiving signal of the reverse signal based on the received signal of the transmitted signal, utilizing the NN-generated signal for differential processing with the received signal to achieve common-mode interference cancellation.
Sign up for Photonics Research TOC. Get the latest issue of Photonics Research delivered right to you!Sign up now
In this paper, we integrated a high-speed and compact
2. DEVICE FABRICATION AND CHARACTERIZATION
A. Design and Fabrication of the
Figure 2(a) illustrates the structure of the
Figure 2.(a) Structure of the high-speed
B. Polarization Characteristics of the
Next, we analyze the polarization characteristics of the
Figure 3.(a) Degree of polarization and polarization extinction ratio of the
C. Photoelectric Characteristics of the
To investigate the optoelectronic characteristics of the
Figure 4.Electroluminescence spectra with the variation of bias current of (a) 685 nm, (b) 638 nm, (c) 520 nm, (d) 450 nm, and (e) 405 nm lasers.
Figure 5 displays the laser output optical power versus bias current and the I-V curves for the five wavelengths. The output optical power gradually increases with rising bias current and approaches linearity. The slopes in the linear region for the 685, 638, 520, 450, and 405 nm lasers are 0.887, 1.107, 0.524, 0.434, and 1.425 W/A, respectively. At the same bias current, the output optical power for red and violet lasers is relatively high because, in underwater environments, red and violet lights are typically subject to more attenuation compared to blue and green lights [30,31], thereby necessitating higher output power for red and violet emissions. In Fig. 5(b), the I-V curves exhibit a monotonically increasing trend; the voltage increases rapidly at low bias currents, while at high bias currents, especially upon reaching the lasing state, the equivalent resistance decreases, resulting in a slower rate of voltage increase.
Figure 5.(a) Output optical power versus bias current and (b) the
3. PRINCIPLE AND EXPERIMENTAL SETUP
A. Experimental Setup of the
With the
Figure 6.(a) A photograph of the
In the DSP algorithm at the transmitter, we first generate a random bit stream. Every seven bits are mapped to corresponding constellation points using 128-QAM mapping. Since the signals transmitted in an IM/DD system must be real-valued, we apply carrier-less amplitude and phase (CAP) modulation [48] to the constellation points. The real and imaginary components of the constellation points are convolved with a set of orthogonal shaping filters and combined to form the transmitted signal. The detailed information on the specific parameters used for modulation and transmission is shown in Table 1. Due to significant attenuation of high-frequency components in visible light communication systems, mainly originating from the bandwidth limitations of the laser and the photodetector, we introduce a pre-emphasis operation before transmission to enhance the high-frequency components of the signal. This results in the final transmitted signal, referred to as
Modulation and Transmission Configuration
Parameter | Value | ||||
---|---|---|---|---|---|
Wavelength (nm) | 685 | 638 | 520 | 450 | 405 |
Modulation bandwidth (GHz) | 2.3 | 2.5 | 2.3 | 2.4 | 2.6 |
Center frequency (GHz) | 1.40875 | 1.53125 | 1.40875 | 1.47000 | 1.59250 |
Roll-off factor | 0.105 | ||||
Filter length | 64 | ||||
Oversampling rate (samples per symbol) | 4 | ||||
Modulation format | 128-QAM | ||||
Communication burst length (symbols) | 131072 |
Due to the limited number of devices including AWG, EAs, and optical lenses, we conducted separate tests for each wavelength. In future study, we can use dichroic mirrors or diffraction gratings to combine five wavelengths and design appropriate control circuits for the lasers to simultaneously modulate all wavelengths. Figures 6(c)–6(e) show a photograph of the experimental setup for one of the wavelengths. The output light beam is initially split into two using a beam splitter (BS). Each beam then passes through a half-wave plate (HWP), adjusting their polarization directions to horizontal and vertical. Finally, a polarization beam combiner (PBC) combines the two beams into one. Due to a significant optical path difference between the two beams, they can be considered uncorrelated. After transmitting through a 1.2 m water tank, the beam is re-split by the polarization beam splitter (PBS) into two beams with orthogonal polarization directions. These beams are focused by two lenses and converted into electrical signals by two self-developed avalanche photodetectors (APDs,
The electrical signals, after being amplified by the EA, are captured by an oscilloscope (OSC, MSO9404A, 20 GSa/s, Agilent) and sent to a computer for reception DSP processing. We first apply the proposed ResDualNet for post-processing, followed by CAP demodulation and QAM de-mapping to obtain the output bit stream.
B. Principle of the Proposed ResDualNet
Before showing the principles of ResDualNet, we first introduce the acquisition of the pre-equalized signal. We initially transmit the signal without pre-equalization, represented as
Figure 6(f) illustrates the principle of the proposed ResDualNet. This algorithm consists of two subnets: the first subnet generates the received signal of the reverse transmitted signal, while the second subnet performs equalization on the received signal that has undergone differential processing. During the training phase, we transmit both original and reverse signals:
The corresponding received signals are
By applying a negation operation to
The label for the second subnet corresponds to the transmitted signal Tx after a sliding window and its output signal is
During the experimental stage, only the
4. EXPERIMENTAL PERFORMANCE AND DISCUSSION
A. Communication System Working Points
Before evaluating the performance of the testing system, we adjusted the peak-to-peak voltage (Vpp) and the magnitude of the bias current to change the operating state of the laser in order to find the optimal working point. Figure 7 presents heatmaps of the bit error rate (BER) as a function of the bias current and Vpp for lasers at 685 nm [Fig. 7(a)], 638 nm [Fig. 7(b)], 520 nm [Fig. 7(c)], 450 nm [Fig. 7(d)], and 405 nm [Fig. 7(e)]. For these five wavelengths, the heatmaps demonstrate a consistent trend. The BER initially decreases and then increases with the increase in either Vpp or bias current. This phenomenon occurs because a small Vpp or bias current results in low output optical power from the laser, limiting the system by noise. Conversely, a large Vpp or bias current drives the laser into a nonlinear operating region, causing waveform distortion and subsequently affecting the transmission quality of the system. Therefore, there exists an optimal working point for the system. The optimal Vpp values for the wavelengths of 685, 638, 520, 450, and 405 nm are 600, 650, 650, 450, and 700 mV, respectively; the corresponding optimal bias currents are 60, 110, 165, 85, and 90 mA.
Figure 7.The variation of BER with bias current and Vpp of the (a) 685 nm, (b) 638 nm, (c) 520 nm, (d) 450 nm, and (e) 405 nm lasers.
B. Performance of the Proposed ResDualNet in the PDM UVLLC System
Since ResDualNet is an algorithm that operates only at the receiver, the pre-equalization processing at the transmitter is also crucial for improving the overall system performance. Therefore, we first compared the spectral envelopes of the received signals and the transmitted signals under optimal operating conditions for four scenarios: without pre-equalization operations and without ResDualNet, without pre-equalization operations and with ResDualNet, with pre-equalization operations and without ResDualNet, and with pre-equalization operations and ResDualNet, as shown in Figs. 8(a)–8(e). Without ResDualNet, pre-equalization can result in a flatter spectral profile for the received signals. When ResDualNet is present at the receiver, pre-equalization improves the flatness of the in-band received signal while also significantly reducing the out-of-band noise. Figure 8(f) shows the BER for the received signals under four different scenarios. It is evident that the system without pre-equalization and without ResDualNet performs the worst, followed by the system without pre-equalization but with ResDualNet. The system with pre-equalization shows improved performance at the receiver without ResDualNet, while the system utilizing both pre-equalization and ResDualNet algorithms demonstrates the best performance.
Figure 8.Spectral envelopes of the transmitted signals and received signals under four scenarios (without pre-equalization operations and without ResDualNet, without pre-equalization operations and with ResDualNet, with pre-equalization operations and without ResDualNet, with pre-equalization operations and ResDualNet) for (a) 685 nm, (b) 638 nm, (c) 520 nm, (d) 450 nm, and (e) 405 nm lasers. (f) BER of the five wavelengths under four scenarios, with the dashed line indicating a threshold of 3.8 × 10−3.
Next, take the 685 nm laser as an example to compare the performance of the traditional equalization algorithm with our proposed ResDualNet, as illustrated in Fig. 9. The bandwidth of the transmitted signal is fixed at 2.3 GHz, with the bias current set to the optimal value of 60 mA.
Figure 9.Communication performance testing of the 685 nm laser. (a) The variation of BER with Vpp for both horizontal and vertical polarization directions when using ResDualNet or the traditional equalization algorithm. (b) Distribution of constellation points at the working points i (400 mV), ii (600 mV), and iii (700 mV). (c) Distribution of the constellation points in the first quadrant at the optimal Vpp, with the lower graphs showing the probability density curves of the two constellation points with the minimum and maximum amplitudes, (1,1) and (11,7). (d) Comparison of the received time-domain waveform and frequency spectrum with the transmitted waveform at the optimal Vpp. (e) Probability density curves of noise distribution and comparison of noise spectrum.
Figure 9(a) illustrates the variation of BER with Vpp for horizontal and vertical polarization states when using ResDualNet and traditional equalization algorithms. The gray dashed line indicates the 7% hard-decision forward error correction (HD-FEC) BER threshold of 3.8 × 10−3. It can be observed that the system using the traditional equalization algorithm fails to meet this BER threshold, whereas the proposed ResDualNet satisfies the threshold at the optimal Vpp. Moreover, for all Vpp values, the BER of the system utilizing ResDualNet is lower than that of the system using a traditional equalizer. The BER performances of the signals in horizontal and vertical polarization states are nearly identical; therefore, in the analysis presented in Figs. 9(b)–9(e), we only show the results for horizontal polarization. Different background colors in Fig. 9(a) indicate the various operating states of the system: orange signifies the SNR constrained region, pink indicates the linear operating region, and blue denotes the nonlinear region. A Vpp value is selected from each of these three regions: 400 mV, 600 mV, and 700 mV, denoted as i, ii, and iii, respectively.
Figure 9(b) presents the constellation point distributions at operating points i, ii, and iii. Larger constellation points imply a higher probability of misclassification between points. Overall, after employing ResDualNet, the spacing between constellation points is significantly larger. Figure 9(c) displays the distribution of the constellation diagram in the first quadrant at the optimal Vpp. The lower graph shows the probability density curves of the received constellation points with the minimum and maximum amplitudes, specifically the distributions for the constellation points (1,1) and (11,7). It is observed that the peak values of the received probability density curves for the constellation point (1,1) with and without ResDualNet are 0.092 and 0.079, respectively, with FWHM of 0.88 and 0.96. For the constellation point (11,7), the peak probabilities are 0.072 and 0.048, with FWHM values of 1.05 and 1.52, respectively. This suggests that ResDualNet can reduce more interference of noise and nonlinearity on the system to some extent.
Figure 9(d) compares the received time-domain waveform and frequency spectrum with the transmitted signal at the optimal Vpp. From the perspective of the time-domain waveform, the normalized mean squared errors (NMSEs) between the received signals using ResDualNet and those using a traditional equalizer, relative to the transmitted signal, are 2.35% and 24.04%, respectively, indicating a significantly lower difference between the output of ResDualNet and the transmitted signal. In terms of spectrum, ResDualNet effectively suppresses low-frequency and high-frequency noise, a conclusion that is also supported in Fig. 9(e). This figure presents the probability density curves of the time-domain noise distribution and the noise spectrum. The peak values of the time-domain noise for the received signals with and without ResDualNet are 0.13 and 0.04, respectively, with FWHM values of 0.36 and 1.20. The amplitudes of the noise spectrum when utilizing ResDualNet are markedly reduced compared to those without it. This indicates that using ResDualNet can decrease the variance of noise, thereby increasing the SNR of the received signal.
Figure 10 illustrates the variation of the BER with Vpp for the other four wavelengths, exhibiting a trend similar to that observed in Fig. 9(a). Moreover, we discovered that ResDualNet demonstrates significant gain in both the linear and nonlinear regions; the BER curves for the 685 nm laser in Fig. 9(a) exhibit identical characteristics. However, in the SNR constrained region, the gain of ResDualNet is slight, as a substantial portion of the noise in this scenario is attributed to completely random environmental noise.
Figure 10.The BER performance of the system utilizing ResDualNet or traditional post-equalization algorithms as a function of Vpp for wavelengths at (a) 638 nm, (b) 520 nm, (c) 450 nm, and (d) 405 nm.
C. Exploring the Maximal Achievable Data Rate
Finally, we explored the maximum achievable transmission rates of the system under optimal operating points. Figure 11 presents the achievable data rates (bar chart) and the BER for the transmission of 128-QAM signals (line graph) for 685 nm [Fig. 11(a)], 638 nm [Fig. 11(b)], 520 nm [Fig. 11(c)], 450 nm [Fig. 11(d)], and 405 nm [Fig. 11(e)] lasers as a function of transmission bandwidth.
Figure 11.Achievable data rates (bar chart) and BER (line graph) for the (a) 685 nm, (b) 638 nm, (c) 520 nm, (d) 450 nm, and (e) 405 nm lasers as a function of transmission bandwidth.
The devices employed in the system are bandwidth-limited, and the visible light communication channel exhibits high-frequency attenuation characteristics. Thus, as the modulation bandwidth increases, inter-symbol interference and other disturbances become more pronounced, resulting in an increase in the BER. When the BER is below the threshold of 3.8 × 10−3, the system supports the transmission of 128-QAM. Since the baud rate of a 128-QAM signal is 7 Baud, the achievable data rate is equivalent to seven times the modulation bandwidth. Conversely, when the BER exceeds the threshold of 3.8 × 10−3, it indicates that the system does not support 128-QAM transmission, necessitating a reduction in modulation order, thereby leading to a decrease in the achievable rate.
Experimental results indicate that after implementing ResDualNet, the data rates for the 685, 638, 520, 450, and 405 nm lasers can reach 32.2, 35.7, 32.2, 33.6, and 36.4 Gbps, respectively. Ultimately, a total transmission rate of 170.1 Gbps was achieved, which represents a 16.1 Gbps increase compared to the system without ResDualNet.
It is noteworthy that previous studies have indicated that narrow-ridge lasers not only exhibit higher degrees of polarization but also have increased relaxation resonance frequencies, enabling higher modulation bandwidths [45,46]. However, due to the limited bandwidth of the current system’s electrical components and photodetectors, the full modulation bandwidth of the lasers cannot yet be sufficiently utilized. Furthermore, as higher-order mode emission is more significantly constrained, the output power of lasers with narrower ridges tends to decrease slightly [49,50]. Therefore, careful consideration must be given to the design of the laser ridge width to balance modulation bandwidth, output power, and beam quality. This provides guidance for our future research. Additionally, the development of broadband devices such as photodetectors is crucial for enhancing the capacity and quality of visible light communication systems.
Additionally, we noticed that the ellipticities of the visible light lasers are approximately 0.1, indicating potential crosstalk within the PDM system. In our experiment, we have employed a PBC to combine horizontal and vertical polarizations to address this issue. This is because the propagation rule in the PBC allows the transmission of the horizontal polarization component while reflecting the vertical component. In future research, we will further tackle this problem from both algorithmic and device aspects, for instance, incorporating MIMO demultiplexing algorithms into the post-processing network at the receiving end to mitigate crosstalk. Moreover, optimization of the laser diode structure can be considered, such as appropriately reducing the ridge width of the ridge waveguide laser to decrease the ellipticity of the laser output beam, as discussed above.
5. CONCLUSION
In this paper, we designed and fabricated a compact
References
[2] Y. Zhou, X. Lin, Z. Xu. Beyond 600 Gbps optical interconnect utilizing wavelength division multiplexed visible light laser communication [Invited]. Chin. Opt. Lett., 23(2025).
[15] C. Lu, J. Wang, S. Li. 60 m/2.5 Gbps underwater optical wireless communication with NRZ-OOK modulation and digital nonlinear equalization. 2019 Conference on Lasers and Electro-Optics (CLEO), 1-2(2019).
[18] J. Hu, F. Hu, G. Li. A 15 Gbps 520-nm GaN laser diode based visible light communication system utilizing adaptive bit loading scheme. 2021 IEEE 6th Optoelectronics Global Conference (OGC), 31-34(2021).
[20] T. Zhang, J. Tian, Y. Wang. 19.02 Gbps/25 m underwater wireless optical communication adopting probabilistic constellation shaping QAM-DMT transmission. 2023 Asia Communications and Photonics Conference/2023 International Photonics and Optoelectronics Meetings (ACP/POEM), 1-4(2023).
[21] Z. Lu, Z. Xu, Y. Zhou. 102.2 Gbps underwater visible light laser communication utilizing a tri-color laser transmitter and a neural network-based reverse signal generator. 50th European Conference on Optical Communications (ECOC 2024), 463-466(2024).
[27] L. Wang, L. Yu, Z. Li. InGaN quantum dots for micro-LEDs. APL Photonics, 9, 100904(2024).
[33] R. Martinek, L. Danys, R. Jaros. VLC channel equalization simulator based on LMS algorithm and virtual instrumentation. 2019 International Symposium on Advanced Electrical and Communication Technologies (ISAECT), 1-6(2019).
[39] Y. Zhu, C. Gong, J. Luo. Indoor non-line of sight visible light communication with a Bi-LSTM neural network. 2020 IEEE International Conference on Communications Workshops (ICC Workshops), 1-6(2020).

Set citation alerts for the article
Please enter your email address