Detectors shoot for terahertz processing speeds

With the ever-continuing drive to boost computer processing speeds, the data communications links between processors and memory or other processors are producing serious bottlenecks. Electrical interconnects, the conventional solution, face bandwidth limitations, so more manufacturers are exploring the use of optical interconnects. Reasons include their capability to accommodate far higher bandwidths within smaller cross-sections and the fact that they do not suffer from electromagnetic interference. With the growing demand for large-bandwidth optical interconnects comes a need for detectors at the receiver end that can handle tens of gigabits or even terabits per second.

The difficulty in meeting these goals revolves around the need to integrate intrinsically high-speed detectors and the required electronics on a single chip. On one hand, typical high-speed detectors use materials that are not very compatible with silicon complementary metal-oxide semiconductor (CMOS) circuits. On the other, CMOS detectors tend to be slow because charge carriers generated by photons deep within the material take a while to migrate to the collection points. To resolve these issues, manufacturers either must modify fast detectors to improve their compatibility or increase the detection speed of CMOS sensors. Researchers are actively pursuing both approaches.

Basics of high speed

The workhorse of high-speed photodetection is the avalanche photodiode (APD), which combines detection and initial amplification of the signal. In APDs, as in any photodiode, a p-n junction of two semiconductor types allows current to flow in only one direction. The photodiodes include a layer of silicon doped with atoms carrying one valence electron less than neutral (a p-type semiconductor) on top of a layer doped with atoms carrying extra valence electrons (n-type semiconductor). Charge migration creates a depletion region with an electric field directed toward the p region, allowing current to flow in only one direction.

With a reverse bias potential applied across the diode, in the absence of light, only a dark current due to thermal generation of electrons is present. Exposure to light produces electron-hole pairs, a process that generates a current. With a conventional photodiode, the process ends there. With an APD, the bias potential is higher, and there is sufficient acceleration of electrons to ionize the semiconductor atoms and create new free electrons. The electrons again are accelerated, producing more electron-ion pairs in an avalanche process. The current produced by many such avalanches combines into a steady stream measurable by external circuits.

Because fiberoptic communications require fast response time, there usually is a need for a reach-through APD with a very narrow junction where the multiplication occurs at the back end of the device, far from the light source. Most of the thickness of the device consists of low-doped silicon or germanium, in which the field strength is very low. This region absorbs the photons, which then drift in a few picoseconds towards the acceleration junction. Since almost no photons are absorbed in the multiplication region, the device has almost no variations in gain. On reaching the p-n junction, all electrons undergo exactly the same acceleration and multiplication. The resulting device provides both high-speed response and low noise.

Depending on their primary semiconductor material, APDs span a wide spectral range. Silicon APDs work well between 400 and 1100 nm, germanium between 800 and 1550 nm, and indium gallium arsenide (InGaAs) between 900 and 1700 nm. For areas of the spectrum in which the different materials overlap, InGaAs provides the lowest noise and best frequency response, but the highest price. Germanium is the intermediate option in both respects. Fiberoptic applications generally use the InGaAs material.

Integrating detectors and circuits

The basic problem to be overcome when integrating optoelectronics and electronic devices is the incompatibility of the materials involved. For the infrared frequencies used in fiberoptic communications, silicon is transparent, so most detectors are based on GaAs, InGaAs, indium phosphide (InP), or related materials. Electronics, however, are generally silicon-based because of the low cost and the enormous industrial base for silicon processing.

Several incompatibilities exist between the two types of materials. For one thing, the lattices are mismatched so that it is difficult to grow one layer on top of another without producing massive numbers of defects. Photodetectors also require very pure layers of absorption material, while typical transistors, such as field-effect types (FETs), require narrow layers doped at levels hundreds of times higher than is allowed in photodetector absorption layers. In addition, photodetectors normally are grown on negatively doped substrates and have metal contacts on both top and bottom, while FETs are grown on semi-insulating substrates with electrodes only on the top.

One of the easiest and oldest ways to overcome material incompatibilities is flip-chip bonding, in which two devices are bonded back-to-back on the same chip with indium or solder bonds. The assembly is rugged, and little modification of either device is necessary. The process also can work for higher-speed applications.

Researchers at Arizona State University (Tempe, AZ), for example, are using flip-chip bonding to build massively parallel optical interconnects with multiteraherz data transmission capacity.¹ One outcome is a 32 × 32 array of InGaAs/InP photodiodes on an iron:InP substrate that is bonded flip-chip fashion to a silicon preamplifier circuit.

Creating this design required that the team calculate the optimum thickness for the active absorption region. The thicker the region, the smaller its electrical capacitance, which increases both the electrical bandwidth and the speed of the device. For the optical bandwidth, though, the thicker the region, the longer the charge-carrier transit time, and the slower the response time. The scientists found what they considered an optimum compromise using a thickness of 400 nm. The result is that each diode has a bandwidth of 10 GHz, and the entire array, the size of a dime, can carry 10-THz signals.

Faster CMOS

One alternative to linking incompatible detector materials with silicon circuits is to improve the response time of silicon-based CMOS detectors themselves. A conventional CMOS detector generates carriers over a depth of 15 µm, the absorption length of 860-nm light. The charge carriers produced in shallow levels are detected immediately, but those generated in the substrate at greater depths diffuse slowly to reach the p-n junction some microseconds later. This blurs the pulses because current continues to rise long after a pulse begins and continues to decline after it ends, creating low-frequency gain.

The problem is not unsolvable, though. A research team at the University of Brussels (Brussels, Belgium) reportedly has developed a spatially modulated light detector that cancels out the low-frequency gain, while maintaining high-frequency gain.^{2, 3} The detector consists of a row of rectangular p-n junctions, every other one of which is covered by light-blocking material. The masked detectors combined form the deferred detector. The unmasked ones form the immediate detectors.

When a light pulse arrives, the unmasked sensors immediately detect shallow-generated carriers, while those diffusing from greater depths are detected almost equally by masked and unmasked detectors. Subtracting the response of the deferred detectors from that of the immediate detectors thus eliminates the effect of the slow, deep carriers. Using this technique, the Brussels team has fabricated a 100-channel array on a 1-mm chip that has a total capacity of 20 Gbit/s.

Researchers still must deal with the basic problem that the infrared radiation used for fiberoptics penetrates too far into the silicon, producing charge carriers that take too long to migrate back to the collector. One option under consideration to resolve this issue involves thinning the detector layers and sacrificing sensitivity for speed—an approach that would collect all carriers promptly. So far, however, efforts to develop this technology at the University of Texas at Austin (Austin, TX) have not been successful.⁴

REFERENCES