Low power wide spectrum optical transmitter using avalanche mode LEDs in SOI CMOS technology

: This paper presents a low power monolithically integrated optical transmitter with avalanche mode light emitting diodes in a 140 nm silicon-on-insulator CMOS technology. Avalanche mode LEDs in silicon exhibit wide-spectrum electroluminescence (400 nm < λ < 850 nm), which has a signiﬁcant overlap with the responsivity of silicon photodiodes. This enables monolithic CMOS integration of optocouplers, for e.g. smart power applications requiring high data rate communication with a large galvanic isolation. To ensure a certain minimum number of photons per data pulse (or per bit), light emitting diode drivers must be robust against process, operating conditions and temperature variations of the light emitting diode. Combined with the avalanche mode light emitting diode’s steep current-voltage curve at relatively high breakdown voltages, this conventionally results in high power consumption and signiﬁcant heating. The presented transmitter circuit is intrinsically robust against these issues, thereby enabling low power operation. (2005).


Introduction
Many smart power applications require data communication with galvanic isolation. Currently this is achieved using inductive isolators (transformers), capacitive isolators or discrete optocouplers [1]. Integrated transformers are big and significantly add to cost and size while they are also prone to external electro-magnetic interference (EMI) [1]. Capacitive isolators can be integrated in the backend, but are relatively big when isolating between voltage domains that have a large voltage difference. Optocouplers are immune to EMI effects and monolithic integration of optocouplers is attractive for smart power and on chip communication applications [2].
Wide spectrum electroluminescence (EL) from silicon (Si) p-n junctions operating in avalanche mode has been reported earlier [3][4][5]. Avalanche mode light emitting diodes (AMLEDs) are fast with reported small signal modulation speed in the range of tens of GHz [6]. An AMLED as a light source in a CMOS integrated optocoupler has also been proposed [5,[7][8][9][10]. The coupling efficiency between the AMLED and an Si PD has been reported to be higher as compared to the same LED in forward mode of operation [10,11]. This is because of the stronger overlap between the emission spectrum of Si AMLEDs and the spectral responsivity of Si PDs [7].
For optocoupling applications with a sufficiently low bit error rate (BER), the AMLED driver must ensure a certain minimum number of photons at the receiver side for data communication. However, optoelectronic properties of AMLEDs are sensitive to process, voltage and temperature (PVT) variations [12]. Together with their steep current-voltage (IV ) curve at relatively high voltages, this easily results in high power consumption, and significant heating [10,11] which are bottlenecks to implement power efficient On-Off Keying (OOK) LED driver circuits in optocoupling applications. In this work, we introduce an AMLED driver circuit to solve these issues, enabling low power Si integrated optical transmitters. Section 2 of this paper describes several physics related properties of the AMLED that are relevant for this work. We present an estimate of the transmission efficiency of our designed optical link in section 3 which is essential to characterize the AMLED in terms of its photon flux output. The principle of the driver circuit is to drive the AMLED (per data bit) with a minimum quantity of avalanche charge required to get certain amount of detectable photons at the PD, independent of PVT variations. The circuit to implement these features is introduced in section 4. The robustness, low power consumption and emission properties of the AMLED integrated with the driver circuit are demonstrated using the measurement results in section 5. Section 6 discusses the potential of using the AMLED with the introduced driver circuit in a monolithically integrated CMOS optocoupler. We conclude our work in section 7. Figure 1(a) shows a schematic cross section of an AMLED (not to scale) in a 140 nm SOI CMOS technology [13]. The Medium Trench Isolation (MTI) regions and Buried Oxide (BOX) layer isolate the high voltage at the AMLED from the CMOS circuitry and provide galvanic isolation from the receiver. Figure 1(b) shows the TCAD simulated 2-D electric field profile (in the y-z plane), for the regions in the dashed box in Fig. 1(a). The field was simulated above breakdown: at a reverse bias (V BIAS ) of 18 V having a breakdown voltage (V BR ) of the AMLED of ∼ 17 V. Avalanche breakdown and hence avalanche mode light emission is initiated in the region with the highest electric field [14], indicated in Fig. 1(b). Further, light is emitted mainly from the n + periphery that is closest to the p + contact (along x-axis) in Fig. 1

Optoelectronic properties of the AMLED
show the TCAD simulated electron and the hole current density for the dashed region in Fig. 1(a).
A schematic top view of the AMLED and the integrated PD, including their dimensions is shown in Fig. 2(a). Two identical diodes have been used, one acting as an AMLED and the other acting as a PD. Hence the schematic cross section of the PD is same as shown in Fig. 1(a). Figure 2(b) shows the micrograph of the AMLED and PD. This PD is used only to measure photon flux of the AMLED as discussed in section 5.2. Figure 2(c) shows an EL-micrograph of the AMLED demonstrating emission at one side (along the x-axis indicated in Fig. 2(a)) as explained by our TCAD simulation results in Fig. 1 Figure 2(d) shows the emission spectrum of the AMLED as measured vertically. Fabry-perot interference in the back-end causes the ripples in this vertical emission spectrum.
Measured reverse IV characteristics of the AMLED are shown in Fig. 3(a); for circuit simulations, these characteristics were used in a table-based (interpolated) IV model combined with a junction capacitance model. V BR is defined as the voltage at which the I AMLED starts to sharply increase. Figure 3(b) shows a linear relation between the AMLED avalanche charge (Q AMLED ) and the vertically emitted number of photons as calculated by Eq. (2) (section 3). For Fig. 3(b), the AMLED was biased at several DC currents (I AMLED ) using an Agilent B2901A source and measurement unit (SMU) (with measurement integration time = 1 s) and the vertical emission spectrum was measured using an Avantes ADC-1000-USB spectrometer with measurement integration time (T integration ) of 30 s at each setting. Q AMLED is calculated as   the AMLED per bit (N electrons,AMLED ). For estimating N photons,PD in response to N electrons,AMLED , the total coupling quantum efficiency of the link (η system ) should be determined; we define η system as η system = . This η system can be written as the product of two efficiencies. The first one is the internal quantum efficiency (IQE) of the AMLED which relates the number of photons emitted per bit from the AMLED (N photons,AMLED ) to N electrons,AMLED [15]. The second efficiency component is the transmission efficiency of the optical link between the AMLED and the PD, denoted as η TE .
The IQE is estimated from DC measurements in section 5.2. In this section, we estimate the η TE . Figure 4 represents the schematic structure of our design. The following steps describe our η TE estimation procedure: 1. The total number of photons emitted by the AMLED (N photons,AMLED ) is calculated. As shown in Fig. 2(c), light emission occurs predominantly at the lower edge of the AMLED.  Fig. 4. Schematics of the AMLED-PD to estimate the optical transmission efficiency (η TE ). For the estimation, each light emitting point is assumed to emit isotropically. For clarity, important dimensions are enlarged. x LED is the width of the AMLED (=45 µm), x 2 (= 7.2 µm) is the separation between the AMLED and the PD and z PD = 1.5 µm is the thickness of the SOI layer.
We model this light emitting region as a line of length x LED , of which each point is emitting uniformly and isotropically [16,17]. Hence, some photons are emitted towards the top and bottom of the chip considering the isotropic nature of the emission. Photons emitted towards the top of the chip enable e.g. the micrograph of the light emission in Fig. 2(c) and Φ(λ) in Fig. 2(d). For simplicity reasons, any waveguiding effect through the BOX layer has been neglected. The total number of photons emitted by the AMLED (N photons,AMLED ) is given by: where Φ(λ) is the photon spectral flux density, shown in Fig. 2(d).
2. The total number of photons received at the PD (N photons,PD ) is calculated; it is assumed that N photons,PD is limited only by [18]: • absorption losses in the AMLED. The absorption coefficient (α) is λ dependent and for this layout of the AMLED and the PD (as shown in Fig. 2(a)), most of the photons emitted at short λ are absorbed in the AMLED itself before reaching the PD.
• the finite solid angle of the PD over the AMLED. The solid angle is limited either by the critical angle at the Si-SiO 2 interface or by the (apparent) size of the PD (Fig.  4). Due to refraction at the Si-SiO 2 interface, the apparent height of the PD seen at each point along the x-axis of the AMLED is different: the apparent height of the PD (y PD ) is reduced to y PD (x). Similarly, the width of the PD (z PD ) is reduced to z PD (x).
Along the x-direction, the lateral dimensions of the PD are much larger than the absorption length of the photons in the spectral region of interest. Therefore it is assumed that all the photons that reach the PD are detected.
Along the z-direction, for the DC measurements (section 5.2), any photon absorbed within a diffusion length from the depletion edge of the PD will contribute to the photocurrent [18].
Under these assumptions and using Fig. 4, the received photon spectral flux density per unit length at the PD is: where: x LED is the emitted photon spectral density per unit length. • r 1 (x, y (x)) = x 2 + (y (x)) 2 is the distance traveled by photons emitted at (x, 0) within the AMLED, while propagating towards (x 2 , y).
• T X (0, y (x)) is the transmittance of the AMLED-SiO 2 interface and T X (x 2 , y) is the transmittance of the SiO 2 -PD interface. • is the solid angle per unit length subtended by the PD over the dx section of the AMLED.
The total number of photons received at the PD because of emission from the entire light emitting area of the AMLED (N photons,PD ) is given by: 3. Using numerical integration in Eq. (4), the η TE = ( It is emphasized that the η TE of this link is low due to mainly the sub-optimum (side-by-side) layout of the AMLED and the PD, as shown in Fig. 2(a) and 2(b). A possible direction to improve η TE will be discussed in sec. 6.

AMLED driver circuit for an optocoupler
The main idea of the circuit is to drive the AMLED with a minimum amount of avalanche charge per data bit (Q b ), required to get a certain amount of photons at the PD, independent of PVT variations, with a relatively small area and relatively low demands on the driver circuit (including timing demands). Figure 5 shows the principle of the self-quenched AMLED driver circuit (using idealized time domain waveforms) that limits the Q b by dynamically quenching the avalanche process. We denote the voltage across the AMLED as V AMLED (= V BIAS − V CAP ). The excess bias voltage across the AMLED (V EX ) is defined as the extra voltage above V BR [19], V EX = V AMLED − V BR . The current through the AMLED is denoted as I AMLED and the series resistance of the AMLED for V AMLED > V BR as R AMLED . R AMLED (∼ 1.45 kΩ) is estimated from the measured IV characteristics of Fig. 3(a) and is assumed constant for V AMLED > V BR for simplicity [19]. The junction capacitance of the AMLED in and near avalanche is modeled by capacitance C AMLED . The resistance of the driver circuit is negligible in comparison to R AMLED and hence is ignored. A simplified model for the AMLED is also shown in Fig. 5 including the capacitance of the AMLED (C AMLED ) [19]. For this section, we assume C AMLED << C Q for simplicity reasons.
We now describe the operating principle using Fig. 5. Initially, the voltage across the C Q (V CAP -V DRV ) is 0. First reset switch M1 is opened (using control signal RST) after which the input (IN) is set high at t = 0. As IN is set high, the V DRV becomes low and instantly V CAP also becomes low, the initial V AMLED = V BIAS − V CAP ≈ V BIAS and V EX = V BIAS − V BR (assuming C AMLED << C Q ). The initial V EX yields an initial I AMLED = V EX / R AMLED .  [19]. Important voltage and current waveforms are also shown on the right.
After the AMLED goes into the avalanche, I AMLED (t) charges the quenching capacitor C Q which results in approximately an exponentially increasing V CAP , hence in approximately an exponentially decreasing V AMLED and V EX . As V AMLED approaches V BR , the avalanche is quenched.
After the on time (T ON ), first IN is set low turning the driver off and then M1 is closed to reset the V CAP to V DD . Since the avalanche quenches itself by reducing V AMLED to V BR , we denote this circuit a self-quenched driver circuit.
V AMLED (t), V EX (t), I AMLED (t) and Q b (t) can be estimated: assuming the total resistance to be R AMLED and the total capacitance to be C Q , Equations (4) describe the mentioned physical quantities for 0 ≤ t ≤ T ON . The initial V EX has been denoted as V EX,0 .
The complete schematics of the implemented circuit are shown in Fig. 6(a). The driver circuit comprises of a chain of inverters (with enable functionality to switch C Q ) and a reset transistor. Selecting a C Q value is achieved using a number of parallel identical drivers (D1 to D7) connected to the AMLED anode which can be controlled using their respective enable (EN) signals. Using D8 (identical to D1 to D7), the AMLED can be operated without any quenching (Fig. 10(c)-(d)).
To enable the measurement of fast AMLED current transients at the onset of the avalanche, a differential structure is adopted and measurement circuit was implemented using an open drain PMOS differential amplifier [20]. The measurement setup enables us to measure at frequencies The chip was packaged without any top cover to measure vertical light emission from the AMLED. Also, no tiling was used over the AMLED area to avoid light blocking by the backend. Total area of the chip is 0.38 mm 2 , the area of the main driver is 0.024 mm 2 and the area of the AMLED is 0.0033 mm 2 . upto 2.8 GHz. A micrograph of the implemented chip in a 140 nm CMOS SOI technology is shown in Fig. 6(b) [13]. We will show that this driver circuit is robust to many physics issues related to the power dissipation and PVT variations of an AMLED.

Measurement results
In this section, to demonstrate the functionality of the circuit, we show an example of the measured transient avalanche current (I AMLED ) waveform. Further, we show the measured energy-per-bit (E b ) of the presented transmitter and its photon flux output with respect to E b . We also demonstrate the robustness of E b against process, voltage, temperature, design and pulse width variations.

Functionality
With an off-state V AMLED ≈ V BR , in the on-state the AMLED is driven to V AMLED ≈ V BR + V EX,0 where V EX,0 is V DD divided between C Q and C AMLED at t = 0 (using the simplified model of the AMLED shown in Fig. 5).
Measured transient waveforms of I AMLED at T ON = 35 ns, pulse repetition rate ( f s ) = 10 Mbit/s are shown in Fig. 7 for three different C Q settings at V DD =1.8 V. C AMLED was measured to be about 650 fF at V AMLED = 15 V (close to V BR ). For a lower C Q , V EX,0 is lower (Eq. (9)) and hence the magnitude of the I AMLED (= V EX,0 /R AMLED ) decreases ( Fig. 3(a)). The charging rate of C Q is limited by the R AMLED (estimated as ∼ 1.45 kΩ from Fig. 3(a)). Simulated transient data (using Spectre [21]) were obtained using a lookup table based model of the DC AMLED IV characteristics ( Fig. 3(a)) in combination with a junction capacitance model for C AMLED (Fig. 5).

Energy-per-bit and emitted photon flux per bit of the AMLED
From an electrical point of view E b is a key Figure of Merit (FoM) [2]; E b is defined as the energy required to transfer Q b through the AMLED: where V BIAS is the DC bias voltage at the AMLED cathode (Fig. 6). E b is limited by C Q (Eq. (8), Eq. (9)): Fig. 8(a) shows an example of the measured E b as a function of C Q . The simulated E b in Fig.  8(a) was obtained using Spectre [21], showing a good agreement with the measured E b . The slight difference could be explained by the variations in C AMLED with V AMLED , which has been assumed constant in our simplified model (Fig. 5). The loss in the driver circuit was estimated to be about 21 pJ/bit. For on-chip optical data communication, the lateral photon transmission to a nearby PD is relevant. To measure this, we integrated a calibration PD next to the AMLED (see Figs. 2(a)-2(b)). The AMLED is operated in OOK mode at f s and the average PD photocurrent (I PD ) is measured.
Using I PD and f s , the number of photons received at the PD per bit (N photons,PD ) can be estimated (Eq. (12)). We have assumed the quantum efficiency of the PD (η PD ) to be unity in the wavelength region of interest because the dimensions of the PD are much larger than the absorption length of the photons [18].
where q e (= 1.6 × 10 −19 C) is the elementary charge. Figure 8(b) shows an example of the measured N photons,PD (on left y-axis) at different E b settings at T ON = 35 ns and f s = 10 Mbit/s. It shows that N photons,PD depends linearly on E b . Using η TE = 3 × 10 −4 (as derived in section 3 for our sub-optimum AMLED-PD layout) and N photons,PD , the total number of photons emitted by the AMLED per bit (N photons,AMLED ) for different E b settings can be estimated (Eq. (13)); an example of the estimated N photons,AMLED is also shown in Fig. 8(b) (on right y-axis).
N photons,AMLED = N photons,PD η TE (13) Like N photons,PD , N photons,AMLED also depends linearly on E b . Using N photons,AMLED and the number of electrons through the AMLED per bit (N electrons,AMLED = Q b q e = E b V BIAS q e ), the internal quantum efficiency (IQE) of the AMLED (Eq. (1)) is estimated to be about 1.4 × 10 −5 , comparable to what was also reported for Si earlier [22]. Using IQE and n TE (section 3), the total coupling quantum efficiency of the link (η system ) is estimated as 4 × 10 −9 .
As an estimation for continuous mode operation, we measured an I PD = 8 pA at an I AMLED =2 mA for the same optical link including identical diodes confirming the η system as 4 × 10 −9 for DC conditions [11]. Figure 9(a) shows the measured Φ(λ) of the vertical emission of the AMLED (along the z-axis in Fig. 4), driven at T ON = 35 ns, f s = 10 Mbit/s, for several values of C Q , demonstrating that the Φ(λ) is almost proportional to C Q , with a minor effect on the spectrum. Only four C Q settings are shown for clarity. The increase in intensity with increasing C Q is because of increasing Q b . Figure 9(b) shows the vertical photon flux density per bit, Φ b = 1 f s λ Φ(λ)dλ, as a function of E b , demonstrating a linear relation between E b and Φ b .

Robustness of E b
We demonstrate the robustness of our circuit to process, voltage, temperature, design and pulse width (T ON ) variations using E b as the FoM. Note that using the results shown in Fig. 8(b), E b can be translated into N photons,PD .
Figures 10(a)-10(b) show measured E b as a function of C Q across PVT and AMLED design variations, demonstrating that E b is hardly affected by these variations but can properly be tuned by setting different C Q values. For process variations, three different samples (D1, D2 and D3 from the same processing batch) were used. To explicitly show robustness of the proposed driver circuit with design variations, a number of n + p diodes and a single p + n diode were measured, driven by a replica driver circuit. E b is non-linear with respect to C Q because the pulse width (T ON ) is not sufficient to charge large C Q completely (Eq. (8) and Fig. 5). For the p + n AMLED, V BR ≈14.4 V, R AMLED ≈ 560 Ω and C AMLED ∼ 585 fF. A slight difference between the E b of the n + p AMLED and the p + n AMLED is because of the different R AMLED , C AMLED and V BR . The higher linearity in E b with respect to C Q in our p + n AMLED is because of its low R AMLED . Figure 10(c) shows E b for conventional (non-quenched) OOK drivers that show a large sensitivity to PVT variations and (especially) design variations. The measured E b for the two AMLEDs are drastically different which is because of the different IV characteristics.
In data communication using OOK, the pulse width (T ON in Fig. 5) is of importance. Figure  10(d) shows the impact of the pulse width (T ON ) on E b : for the self-quenched driver the E b is mainly determined by the selected C Q , for pulse repetition rate ( f s ) to ∼ 10 Mbit/s. At lower T ON , the E b drops due to mainly R AMLED that determines the maximum Q b and hence the E b (Eq. (8), Eq. (11)).
For comparison, the non-quenched operation of the AMLED in Fig. 10(d) shows a strong T ON (hence timing) dependency. The simulated E b (using Spectre) shows a good agreement with the measured E b . For the non-quenched driver circuit, E b can be estimated as E b ≈ V BIAS I AMLED T ON = V BIAS (V EX /R AMLED )T ON . Using the self-quenched circuit, E b (thus power dissipation) is always lower than C Q V DD V BIAS independent of T ON (Eq. (11)).

Application in opto-couplers
In this section, first we discuss the potential of the proposed optical transmitter for application in opto-couplers. In a further section, we discuss the aspects of the transmission bit rate.

Optocouplers in CMOS technology
Although our work mainly focused on the transmitter in a fully Si-integrated optical link, the demonstrated optical transmitter aims at applications in a complete optical link in Si. For the minimum E b setting (∼ 53 pJ/bit), the n + p AMLED emits 220 photons per bit isotropically ( Fig.  8(b)). Using Fig. 8 and the model presented in section 3, it is possible to optimize the optical link geometry to receive e.g. 5 photons at the PD out of N photons,AMLED . To receive more than 5 photons out of 220, it is required that η TE > 2 × 10 −2 . For instance, when the AMLED and the PD are aligned as shown in Fig. 11, the η TE of the link is estimated to be 2 × 10 −2 using the method in section 3. Further suggestions to improve η TE such as patterning the link, improving the AMLED design have been extensively discussed in [11]. It is also beneficial to have optical transmission paths for such optical links for waveguiding the photons from the AMLED to the PD. This will improve the coupling quantum efficiency, as well as reduces the crosstalk among multiple channels in a multi-channel communication environment. If 5 photons are received at the PD per bit, this is typically sufficient to operate a well-designed Single Photon Avalanche Diode (SPAD) as a PD [19,[23][24][25][26][27][28]. SPADs are p-n junctions biased above V BR so that an incoming photon in the depletion region can generate free carriers thus triggering an avalanche [25][26][27]. The macroscopic avalanche current can then be easily measured by using digital simple read out circuitry. The design of SPADs and SPAD read out circuits are well-known in CMOS technologies [25][26][27]. Recently, a SPAD designed in the same technology has been reported [29].
Further, in the current link geometry a large jitter would occur if 5 photons per bit were statistically received at the PD. This is because many free carriers are generated by the photons outside the depletion region which subsequently diffuse towards the depletion region triggering the SPAD after some ill determined diffusion time. Therefore, it is recommended to design links with a higher η TE and to capture photons mainly in the depletion region. This could be achieved by using e.g., a lower doping in the PD which results in a wider depletion region [18]. This is a topic of future research.

Transmission bit rate
The transmission bit rate is ultimately limited by the speed of the detector (SPADs or the PDs). AMLEDs have been reported to have a very high modulation speed, in the range of tens of GHz [6].
For SPADs, the achievable bit rate is limited by their deadtime requirement. To reduce the unpredictable and hence undesired aferpulsing phenomena in SPADs, the deadtime of the SPADs after each photon counting event has to be increased, which limits the bit rate of the SPADs [26], [29]. For the SPADs that were reported in this technology [29], the bit rate would be limited to about 10 Mbit/s based on their reported deadtime (∼ 100 ns). The major challenge would be improving SPAD designs to have a lower deadtime requirement.
Using conventional photodiodes, higher bit rates can be achieved, however at the cost of increased power consumption in the AMLED. This is because of its required continuous operation mode to increase the signal-to-noise ratio (thereby to reduce the bit error rate) at the PD. A data rate of 3 Gbit/s has been reported with an integrated Si receiver in a standard 180 nm CMOS technology using off chip illumination [30].
Measurements of maximum achievable data rate using the proposed detectors is also a topic for future research.

Conclusion
In this work, a low power monolithically integrated optical transmitter using avalanche mode LEDs (AMLEDs) was designed in a standard 140 nm SOI CMOS technology. The novel selfquenched driver circuit resolves many physics issues related to power dissipation and PVT variations of an AMLED. This work successfully demonstrates a low power wide spectrum optical transmitter in CMOS technologies that can be integrated with standard Si detectors. It further reinforces the promise of enabling AMLEDs as light sources for Si CMOS technology for monolithic integration of optocouplers in CMOS [5,10,11,22].