The mid-infrared spectral range is attracting more and more attention because of the potential it holds for various fields like spectroscopy [1] and high-speed free-space transmission [2]. These wavelengths benefit from very low attenuation in the atmosphere and are intrinsically resistant to degraded atmospheric conditions [3]. Various optical sources emitting at mid-infrared wavelengths, in the thermal atmospheric windows 3–5 μm and 8–14 μm, have been developed including optical parametric oscillators [4], supercontinuum generation in fibers and bulk crystals [5–8], and cascaded technology [9]. Historically, many efforts were dedicated to the development of quantum cascade lasers (QCLs), which rely on intersubband transitions in band-engineered quantum wells (QWs). The main advantage of QCLs is the availability of high output power over a wide range of wavelengths and, in addition, a theoretical potential for very high-speed direct modulation due to their ultra-fast transitions [10], which is yet to be confirmed experimentally [11,12]. In contrast, interband cascade technology is based on interband recombination in a broken gap configuration, employing the same cascading principle as QCLs [13]. Technological advantages of interband cascade lasers (ICLs), compared to their QCL counterparts, include the possibility to circumvent the fast phonon scattering losses and the polarization selection rule constraints of the QCL [14] while still achieving sufficiently high output powers of tens of milliwatts [15]. Due to its low lasing threshold, this technology is extremely attractive for many applications requiring low-power consumption. Several aspects of the working principles of ICLs were recently investigated, uncovering limiting factors for their performance like valence intersubband absorption in the active region [16] or relaxation oscillations, which can limit their maximum modulation speed [17]. A key step toward a wider adoption of ICL technology could come from the development of interband cascade (IC) photonic integrated circuits (PICs), which has seen a growing interest [18–20] since the first demonstrations of ICLs on silicon in 2018 [21,22]. Also, the potential and feasibility of integrating III–V optoelectronics on Si have also been demonstrated for QCLs [23]. Recent works showed the possibility to mitigate a very high dislocation factor, while still keeping the same characteristics as regular ICLs grown on GaSb [22], paving the way for spectroscopic or high-speed chip-based systems at mid-infrared wavelength. The ICL is already a mid-infrared source of choice for molecular spectroscopy [24,25], and recent experimental efforts are poised to extend the capabilities of ICLs to cell ablation in a medical context [26] and to private optical links [27] allowed by mid-infrared hyperchaos [28]. Moreover, simulations based on a semi-classical rate equation model for ICLs are promising for the generation of squeezed states and for quantum-oriented mid-infrared applications [29]. ICLs are also of paramount interest for regular free-space communications. State-of-the-art transmission rates in the 3–5 μm window using cascaded technologies have been demonstrated at 6 Gbit/s with a QCL emitting at 4.65 μm [30]. However, while wavelength conversion can potentially achieve very high data rates, it currently lacks sufficient output power (milliwatts) and requires high power consumption [31]. Conversely, high-speed communication based on ICLs is lagging behind, with a record transmission rate of 300 Mbit/s so far [32]. In the latter, the bandwidth limit is mostly due to the parasitic capacitance of the ICL structure lacking a dedicated high-frequency launcher. Most of the recent articles about mid-infrared communication are also limited by the bandwidth of their mercury cadmium telluride (MCT) detector. This technology is renowned for its high sensitivity, but so far commercial devices are slower than the 2.4 GHz bandwidth one can be expected from our radio-frequency (RF)-mounted ICLs [17]. Interestingly, novel mid-infrared photodetectors based on a uni-traveling carrier process showed impressive performances, such as 3 dB bandwidth of tens of GHz and high responsivity up to 600 mA/W [33,34] at 300 K. We decided to investigate the potential of a more mature technology, named interband cascade infrared photodetectors (ICIPs) for free-space communication in a full interband system (ICL and ICIP). Since the first demonstration of this cascaded type of detector employing a type-II superlattice as an absorber in 2010 [35], this promising technology has raised attention from both industry and researchers. Especially, the high-speed behavior has lately moved into the research spotlight with demonstrations of 3 dB cutoff frequencies in the GHz range [36–39]. Moreover, in comparison to a single stage absorber, the cascading principle greatly benefits the detectivity of the ICIP [40]. Compared to the widely used MCT detectors, ICIPs also offer numerous advantages like suppressed Auger recombination, reduced tunneling currents, and high uniformity in their growth [41]. In contrast to their intersubband counterparts, named quantum cascade detectors (QCDs), where 3 dB bandwidths of more than 20 GHz were recently demonstrated [42,43], ICIPs exhibit a narrower electrical bandwidth but a significantly broader spectral response. This allows for a higher degree of flexibility when designing a free-space optical (FSO) communication setup, enabling the use of transmitters over a broader range of wavelengths. In general, ICIPs combine fast detection with high responsivity and a high saturation level when biased [39]. Yet, moderate voltage is required, and the power consumption of the overall IC system is still as low as 1 W at maximum conditions of operation (900 mW for the ICL [17] and 10 mW for the ICIP). In this article, we report on a full interband free-space communication system, allowing for data rates up to 14 Gbit/s when combining a directly modulated ICL emitting around 4.18 μm wavelength and an ICIP. We investigate three different modulation formats: two-level on–off keying (OOK), 4-level pulse amplitude modulation (PAM-4), and 8-level pulse amplitude modulation (PAM-8) [44].

We used a 3 mm long and 10 μm wide ridge Fabry–Perot ICL emitting around 4.18 μm as shown in Fig. 1(c). Details about this laser were already presented in a previous article [17], but additional manufacturing information can be found in Appendix A. One can find in Fig. 1(a) the calculated band structure of the W-shape structure of the ICL. This laser was operated at a constant temperature of 20°C. The bias current for the ICL was supplied by a homemade battery source to minimize the current noise in the laser. For a temperature of 20°C and a drive current of 200 mA, a maximum output power of 8 mW can be reached with an operating voltage of 4.5 V. The consumption of the ICL is as low as 0.9 W at maximum output power.

The ICIP was grown by molecular beam epitaxy (MBE) on a compensation-doped GaSb substrate (n≤5×1016cm−3) to facilitate reduced free-carrier absorption when illuminating the detector from the substrate side. A 500 nm thick GaSb bottom contact was grown with a doping of n=1×1018cm−3 since the compensation doping lowers the conductivity of the substrate. The bottom contact is followed by three active stages, which are each made of the electron barrier, the active region based on a type-II superlattice, and the hole barrier, in this order, as shown in Fig. 1(b). The absorbers consist of 54 superlattice periods of 1.5 nm GaSb/0.8 nm AlSb/1.5 nm GaSb/2.7 nm InAs, following a design previously reported by Lei et al. [41]. The entire active region was left undoped. In order to balance the strain in the structure, 0.15 nm thin InSb layers were grown at the InAs interfaces. This modification leads to a redshift of the cutoff wavelength of about 500 nm compared to the original design. The electron and hole barriers consist of four GaSb/AlSb pairs and eleven InAs/AlSb pairs, respectively. The highly doped (n=2×1018cm−3) top contact was realized as a lattice matched InAs0.91Sb0.09 layer of 50 nm thickness. Transition layers were employed toward the bottom and top contacts to avoid potential barriers at the interface and allow for a smooth transition of the carriers. Circular detector mesas with a diameter of 75 μm were fabricated using reactive ion etching. A Si3N4 passivation layer is deposited and selectively removed on top, where the Ti/Au top contact is realized afterward. Forming an Ohmic contact to the n-GaSb bottom contact, Ge/Au/Ni/Au is deposited and thermally annealed. Subsequently, the mesas are mounted on a copper substrate with a slit to facilitate backside illumination and connected to a coplanar waveguide and an end-launch connector. The responsivity is displayed in Fig. 1(c), exhibiting a rather flat spectral response between 2 and 4.5 μm. At the lasing wavelength of 4.18 μm, the responsivity is evaluated to be around 130 mA/W. As shown in Fig. 1(c), the emission wavelength of the ICL is located near the peak detection of the ICIP. In addition, the intensity–voltage (IV) response of the ICIP as a function of the size of the detector is shown in Fig. 1(d). The 75 μm ICIP exhibits a dark current around 1.6 mA at 4.5 V.

The experimental setup is shown in Fig. 2(a). First, we evaluated the optimal ICL parameters for lowering the non-linearity and achieving the best modulation depth. It is a compromise between the bandwidth, the level of thermal parasitic effects (which both increase with the bias current of the laser), and a sufficient laser output power. Experimentally, the optimum configuration is to bias the ICL at 140 mA for a temperature of 20°C. In order to achieve high-speed electrical modulation of the ICL, we used a Socionext arbitrary waveform generator (AWG), which can provide 120 GS/s, with an analog 3 dB bandwidth of 30 GHz. We generate sequences corresponding to three different modulation schemes: OOK with a non-return to zero (NRZ) format, PAM-4, and PAM-8 [44]. Since 215-bit long pseudo-random-binary sequences (PRBSs) have been used and they guarantee the occurrence of more combinations of bit sequences than shorter PRBSs, we can study how the transmission system reacts to these different combinations. Moreover, we shaped our signal with a root-raised-cosine (RRC) filter, which decreases the spectral occupation of the transmitted signal with the goal of increasing the data rate of the transmission. The occupied bandwidth of our optical signal can be defined by B=(1+ρ)Rs, in which ρ is the roll-off factor of the filter and Rs is the symbol rate. The roll-off factor satisfies the condition 0≤ρ≤1, and an unfiltered signal corresponds to ρ=1, thus to the maximum bandwidth occupancy. The modulated electrical signal from the AWG has an amplitude of 1.5V_pp, which has been amplified upstream with a 10 dB gain D837C SHF differential amplifier. After approximately 2 m free-space propagation, a silicon lens focuses the optical signal on the ICIP. Interestingly, contrary to QCDs, the ICIP detector does not have a polarization requirement for detection. Toward the detection, the ICIP is connected to a Pasternack bias tee. The DC part is used to bias the detector with a Keithley source, and the AC contribution is amplified with a 25 dB gain 826H SHF amplifier before being recorded by a 20 GHz oscilloscope with a sampling rate of 50 GS/s. The resulting DC current of the ICIP when illuminated by a 5 mW laser signal can be found in Fig. 1(d) showing a responsivity in accordance with the announced values of 130 mA/W. After being detected by the oscilloscope, the signal is processed offline with a MATLAB code. The processing consists in RRC filtering followed by a feed-forward equalizer [46] that will be described below. Then, we compare the recovered bits to the transmitted ones to compute the bit error rate (BER). As shown in Fig. 2(c), one can visualize the superposition of the generated PRBS signal in an OOK configuration in green and the signal received by the ICIP detector in red. In this example, the generated signal is a 10 Gbit/s OOK PRBS pattern. At this data rate, there are strong inter-symbol interferences (ISIs) when you keep alternating a “0”-bit and a “1”-bit, and this shows the frequency limitation of our system, hence justifying the need for equalization algorithm.

The bandwidth of the full setup constituted by the amplifiers, laser, and detector was evaluated using a 40 GHz Rohde & Schwarz ZVK vector network analyzer (VNA), in the case where the ICIP detector was used. The VNA is made of an electrical source signal and an electrical receiver. It calculates the input–output relationship between ports in the system as a function of the source signal frequency. In our case, we plugged the source signal of the VNA on the AC input of the laser’s bias tee. We plugged the receiver of the VNA at the 25 dB amplified response signal of the ICIP. That gives the amplitude response of our system as a function of the modulating frequency. This response is plotted for four voltage conditions for the ICIP and is quite flat until reaching the cutoff frequency, as shown in Fig. 2(b). By increasing the bias voltage of the detector, one can increase its bandwidth by 1 order of magnitude, from 140 MHz at 0 V to 1.5 GHz at 5 V bias. For the sake of completeness, the response of the full system using a QCD with a 3 dB bandwidth of 20 GHz instead of an ICIP is presented in Appendix C. The 3 dB bandwidth of this mixed system is 2.5 GHz, which is the bandwidth of the ICL structure itself because the detector is no longer the limiting device. This 2.5 GHz bandwidth is in very good agreement with our prediction based on the study of the relaxation oscillation frequency [17]. With the ICIP, the system is then limited by the speed of the detector, even at high-voltage bias. As shown in Fig. 1(d), increasing the bias of the detector significantly increases its dark current, which can limit the detection capabilities at low optical power.

After characterizing the setup, we evaluate the capacity of our system to transfer information at different data rates and various modulation schemes. To evaluate the capacity of our system, we mapped a 215-bit long PRBS to the three aforementioned formats (OOK, PAM-4, and PAM-8). For each transmission, we used an RRC pulse-shaping filter with different roll-off factors in order to reduce the spectral occupation of the transmitted signal. We then chose the roll-off factor minimizing the BER. At the receiver side, we apply a matched filter, which is an RRC filter with the same roll-off factor. Then, the signal is equalized by using fractionally spaced feed-forward equalization (FFE) [46] at a rate of four samples per symbol. The first step of the algorithm learns the coefficients of the channel equalizer by using a least mean squares gradient descent algorithm with a convergence parameter μ, which controls the speed and the accuracy of the learning step. The number of filter coefficients is called ntap and should be large enough to cover all the channel memory. Once the algorithm has converged to achieve an error below a given threshold, the estimated equalizer is applied over the received signal to remove ISI. We evaluated the impact of those parameters on the quality of the transmission in Figs. 3(b) and 3(d). We also plotted the eye diagram in Figs. 3(a) and 3(c), which allows us to represent the capability of our system to distinguish between different bit levels. One can evaluate the presence of non-linearity or systematic errors in the transmitted signal by looking at the distortions of the eye. In addition, a good indicator can be the presence of compression between different bit levels (bit levels are not equally spaced) or a level with an ill-defined eye. The purpose is to optimize the parameters of our equalization in order to reach the lowest BER. The latter has been recovered using the whole traces constituted by 10 million data points sampled at 50 GS/s. In Figs. 3(b) and 3(d), one can see the importance of optimizing the coefficients of the equalization. For the two-level modulation (OOK), μ has an optimal value around 10−3 and ntap has to be above 11 to comply with a tabulated threshold BER. Only for particular well-chosen parameters, the equalization process leads to a great improvement in terms of transmission quality. By optimizing these coefficients, we achieve an OOK transmission at a bit rate of 12 Gbit/s with an acceptable BER under 0.38% and 14 Gbit/s with a BER below 4%. The result toward BER satisfying the <4% criterion can be found in Appendix B. Forward error correction (FEC) coding and decoding are used to correct bit errors by introducing redundancy. Hence, an FEC overhead should be taken into account in the dimensioning of the transmission system. For a maximum tolerated pre-FEC BER of 4%, an FEC code with 27% overhead and hard-decision (HD) decoding [47] is assumed to be used to achieve an error-free communication, whereas for a maximum pre-FEC BER of 0.38%, an FEC code with an overhead of only 7% is used. Toward the 4-level format (PAM-4), we achieve a transmission at 14 Gbit/s with a BER of 0.23% and 16 Gbit/s with a BER below 4% (also found in Appendix B). It is relevant to note that the optimal value of the convergence parameter μ has to be lower when considering multi-level modulation formats.

In this work, the 8-level modulation format (PAM-8) was also investigated. Unfortunately, due to noise coming from the amplifier and the detector in our system, we were not able to reach a BER below 0.38%. The parameters were adequately optimized, and the optimum μ factor was again lower compared to lower order multi-level modulation formats as shown in Fig. 4(b). In that case, it is mandatory to use a high ntap value in order to keep a low BER. The presented result in Fig. 4(a) is a 15 Gbit/s transmission associated with a BER of 1.9%. This can be corrected by employing an HD-FEC algorithm using a 27% overhead, which can accommodate a BER up to 4%. It should be noted that using a longer overhead may also introduce more complex decoding associated with higher latency.

We summarize the results of the transmission for the different modulation formats in Fig. 5. We found that good performance is achieved when we have a ρ value of 0.3 for the OOK and PAM-8 modulation formats and a ρ value of 0.4 for the PAM-4 format. This means that we decrease the bandwidth occupation of a modulated optical signal by 35% and 30%, respectively, with the following relationship: ΔBΔBunfiltered=1−1+ρ2. It allows us to achieve a BER below the 7% HD-FEC threshold for the OOK and PAM-4 configurations.

This paper shows the possibility of building a full IC system in one of the transparency windows of the atmosphere with the laser and detector adapted for 4.18 μm wavelength at room temperature. We shaped the signals to be transmitted with an RRC routine before injecting them into the laser in order to optimize the achievable data rate. In addition, the received signal detected by the ICIP was equalized by an FFE algorithm. We performed an optimization of the FFE parameters to minimize the BER. We successfully sent 12 Gbit/s and 14 Gbit/s with the 2-level modulation format (OOK) with a BER suitable for 7% and 27% HD-FEC, respectively. Furthermore, we investigated the possibility of using multi-level modulation formats and saw the effectiveness of equalization to compensate the ISI. The resulting data rates were at 14 Gbit/s and 16 Gbit/s with a 4-level modulation format (PAM-4) with a BER suitable for 7% and 27% HD-FEC, respectively. Hence, moving from a 2-level to a 4-level modulation format improved the achievable data rates by 2 Gbit/s while ensuring comparable performances. Finally, we investigated the possibility to use an 8-level modulation format (PAM-8). We successfully implemented a 15 Gbit/s transmission with a BER suitable for 27% HD-FEC correction. Those results are unprecedented with direct modulation of semiconductor lasers in the 3–5 μm window, while benefiting from 10-time lower power consumption of the IC devices compared to their quantum cascade counterparts. In addition, this study assessed that ICIPs are very suitable for FSO communication experiments because of the high electrical bandwidth, wide spectral response, and increased sensitivity. This work is a key step in order to build an energy-efficient system using cascaded interband devices for communication applications. Indeed, such a system can be very interesting in conditions where energy is scarce, like satellites sharing data information or autonomous rovers exploring distant planets. Also, it opens the way for a compact design of this communication apparatus, all integrated on a silicon platform. Still, improvements have to be made in order to increase the speed of both the laser and the detector. This is the subject of ongoing work and will be reported in due course.

Acknowledgment. The authors thank Prof. Jean-Christophe Cousin (Télécom Paris) and Prof. Wolfgang Elsäßer (Technische Universität Darmstadt) for lending the Vector Network Analyzer and the battery source, respectively.