Radiation hardness and lifetime studies of LEDs and VCSELs for the optical readout of the ATLAS SCT

Abstract We study the radiation hardness and the lifetime of Light Emitting Diodes (LEDs) and Vertical Cavity Surface Emitting Laser diodes (VCSELs) in the context of the development of the optical readout for the ATLAS SemiConductor Tracker (SCT) at LHC. About 170 LEDs from two different manufacturers and about 130 VCSELs were irradiated with neutron and proton fluences equivalent to (and in some cases more than twice as high as) the combined neutral and charged particle fluence of about 5×10 14 n (1 MeV eq. in GaAs)/cm 2 expected in the ATLAS inner detector. We report on the radiation damage and the conditions required for its partial annealing under forward bias, we calculate radiation damage constants, and we present post-irradiation failure rates for LEDs and VCSELs. The lifetime after irradiation was investigated by operating the diodes at an elevated temperature of 50°C for several months, resulting in operating times corresponding to up to 70 years of operation in the ATLAS SCT. From our results we estimate the signal-to-noise ratio and the failure rate of optical links using LEDs developed specifically for application at LHC.


Introduction
Two dierent technologies have recently been evaluated for the readout of the ATLAS SemiConductor Tracker (SCT): optical links and miniature shielded twisted pair cables (STP) as proposed for the readout of the Transition Radiation Tracker (TRT) [1]. In the case of optical links, one of the crucial issues is the radiation hardness and the life time of the optical emitters 1 , which w ould be mounted close to the SCT detector modules and thus be subject to uences on the order of 10 14 charged hadrons and 10 14 neutrons per cm 2 , and to an ionizing dose of about 10 5 Gy during 10 years of operation at LHC [1].
The radiation hardness and the long term reliability of dierent optical emitters suitable for a bre optic link for the SCT, namely two types of Light Emitting Diodes (LEDs) from dierent manufacturers and one kind of Vertical Cavity Surface Emitting Laser diodes (VCSELs), have been studied in detail.
The scope of these studies has been twofold: The short term behavior of LEDs and VCSELs after irradiation was studied by irradiating them under dierent operating conditions with neutron and proton uences comparable to (and in some cases exceeding) those expected at LHC. After the irradiation, the diodes were subjected to an annealing treatment which in most cases reduced the radiation damage considerably.
The long term reliability after irradiation was studied in an accelerated life time test, where a large number of devices were operated at an elevated temperature of 50 C for several months. In order to reach the required statistics, a dedicated measurement system (the scanning machine) was built [3,4], which allows for the long term operation and measurement of several hundred devices.
While it is well known that the light output of LEDs may be decreased after irradiation, little information seems to be available about the inuence of radiation on the aging properties and thus on the life time of LEDs. For the LEDs typically used for bre optic applications, i.e. high radiance devices operated at high current densities, the dominant degradation mechanism is the inhomogeneous development of crystal defects (dark line defects) acting as centers for non{radiative recombinations. Knowing that the development of dark line defects depends on initially present crystal defects and material impurities, the concern arises, that radiation{ induced displacement damage in LEDs might lead to an increased growth of dark line defects and thus t o a m uch shorter life time after irradiation.
Being a rather new type of device, the life time of VCSELs has been a major concern even without irradiation, and only recently VCSELs with a life time on the order of 10 7 hours could be produced. As far as radiation hardness is concerned, little experimental data is available for VCSELs, although from theoretical considerations VCSELs can beexpected to be very radiation hard devices.
In order to check if there is a large eect of irradiation on the life time of LEDs, we irradiated in an earlier life time test [3,5] a small number of GaAlAs LEDs (manufactured by ABB Hafo) with up to 1:4 10 14 n/cm 2 . After the irradiation, an almost complete annealing of the radiation damage was observed under forward bias. Five LEDs were subsequently operated in an accelerated life time test without seeing any degradation due to aging for a duration which was estimated to correspond to about 57 years of operation in the ATLAS SCT.
Although these results were very encouraging, the conditions required for the annealing had to be claried further, the tests had to be extended to include charged particle irradiation, and higher statistics were needed in order to make a prediction concerning the reliability of LEDs in the SCT. In addition, the use of VCSELs was suggested more recently as a cost eective and presumably more radiation hard replacement of the LEDs. Therefore we have continued our tests, studying both LEDs and VCSELs, and using an automatic measurement system in order to reach higher statistics. Complementary radiation hardness studies of LEDs have been performed by the University of Birmingham [6]. The radiation hardness of LEDs and of dierent types of laser diodes has also been studied by the ATLAS group developing the liquid argon calorimeter readout [7] and within the CMS collaboration [8].
While some rst results were reported previously [9], we present in this paper the nal results of our studies. The outline of this document is as follows: After a brief review of the mechanisms leading to the degradation of LEDs and VCSELs, we give some details about the devices studied in section 3 and we describe the experimental procedure and the scanning machine in section 4. The results obtained are presented in detail in section 5.

Introduction
In the following sections we attempt to give a brief review of the mechanisms leading to the degradation of LEDs and VCSELs as far as relevant for the present work, including both normal aging eects and the inuence of irradiation. The aging processes leading to degradation and eventually to the failure of LEDs and VCSELs have been discussed in a large numberof papers, application notes and textbooks (see e.g. [10]{ [17] and references therein). In the following, only a short overview of the principal degradation mechanisms can bepresented.
Typical operating life times of commercial unirradiated LEDs under normal operating conditions are on the order of 10 5 10 7 hours, and it has been found experimentally that aging takes place only under forward bias conditions and not when the LED is turned o [10]. The data transmission protocol for the readout of the SCT was dened such that it would take advantage of this fact by switching the LEDs o when no data is transmitted [18].
2 For VCSEL, only a few life time studies of unirradiated devices are available. For the more common proton implanted VCSELs mean time to failure as high as 3 10 7 hours have been reported [19]. For oxide conned VCSELs the mean time to failure was estimated to be more than 3 10 5 hours [20].

Dark Line Defects
For the LEDs typically used for bre optic applications, i.e. high radiance devices operated at high current densities, the dominant degradation process is the inhomogeneous development of crystal defects acting as centers for non{radiative recombinations [10,11,15,21]. These defects, which occur also in semiconductor lasers, can beseen under high magnication as dark lines and are therefore often called Dark Line Defects (DLD).
The development of DLDs is due to the growth of dislocation networks by a climb mechanism under absorption or emission of point defects, apparently using the energy released under forward bias by non{radiative recombinations [22] { [28]. The growth and propagation of DLDs starts at initially present material impurities or crystal defects and, by increasing the non{radiative current, decreases the light output of the LED at a xed forward current. The rate of growth increases with current density and temperature, but seems to be also enhanced by mechanical stress, e.g. due to diode assembly or dicing{induced strain [29,30].

Temperature and Current Dependence of Life Time
The dependence of the mean life time 2 t of LEDs on operating temperature and current is usually given by an Arrhenius equation with a power law dependence on the forward current [10,11,15] t = C I n f exp E a kT j (1) where I f is the forward current through the LED, T j denotes the junction temperature, and the constants C, n and the thermal activation energy E a depend on the composition and the fabrication process of the device and must bedetermined experimentally. For LEDs, values found in the literature for n and E a are typically in the range from 1 to 2, and from 0.4 eV to 1.0 eV, respectively.
The temperature dependence of t of VCSELs can bemodelled by an Arrhenius equation, too, [31] and for proton implanted VCSELs, E a is typically in the range from 0.7 eV to 1.1 eV [16].
The current and temperature dependence of t may beused in a life time measurement in order to accelerate the aging and thus shorten the duration of the test.
For some values of E a , the acceleration factor that can be gained according to eq. (1) by raising the temperature with respect to aging at T j = 1 0 C is shown in gure 1.

Displacement Damage and Non{Ionizing Energy Loss
The degradation of LEDs and VCSELs by irradiation is primarily due to displacement damage (bulk damage). If a high energy particle interacts with an atom (called the primary knock{on atom, PKA) in the semiconductor lattice, enough kinetic energy may betransferred to dislodge the atom from its site. The PKA will rapidly loose its energy in the vicinity of the primary interaction site due to both ionization and the displacement of further atoms, eventually producing a cascade of collision processes. Energetic charged particles or fragments produced in inelastic nuclear collisions may contribute to the resulting displacement damage, too, and very low energy neutrons may cause displacement damage through nuclear reactions like e.g.
(n; ) even if their kinetic energy is below the elastic threshold of about 200 eV in GaAs. Even though in the end only a small fraction 3 of the total energy deposition goes into the displacement of atoms (non{ionizing energy loss, NIEL), depending on the energy of the PKA, a large numberof crystal defects may beproduced in a tree{like structure [34,35,36]. The initial displacement damage is produced on the fast collision time scale of about 10 14 s to 10 12 s [35] and is complete before any thermally activated atomic motions take place. At a slower time scale, the generated vacancies and interstitials will move around and most of them ( 95%) will recombine due to diusion processes (short term annealing). Those that remain will eventually form a variety of stable defect complexes, including di{vacancies, vacancy{impurity complexes, and larger clusters. Depending on the temperature, further annealing [37], but also reverse annealing may take place. The annealing processes may be enhanced under forward bias (injection annealing) [22], because electron{hole recombinations at a defect can increase the defect's mobility.
The presence of defect complexes changes the eective doping concentration and, by i n troducing additional states within the forbidden bandgap of the semiconductor, they act as recombination centers, possibly causing a decrease of the carrier density, the minority carrier life time, and the carrier mobility.
The dierent defect complexes might be expected to have dierent probabilities for their formation, depending on the type and energy of the radiation, and to possess dierent eciencies for producing macroscopic device degradation. However, it has been observed that the degradation due to displacement damage of a given semiconductor device at specic operating conditions is primarily a function of the non{ionizing energy deposition (NIEL hypothesis) and not of the particular type or energy of the incident particles. Although some deviations of measured device degradation from the one expected according to the NIEL hypothesis have been reported (see e.g. [38,39,40]), the NIEL hypothesis has been veried experimentally over a wide range of energies, for dierent incident particles, for silicon and to a somewhat less degree for GaAs [38] { [43].
For a particle passing through a given material, the NIEL 4 is given by where d=dE r is the dierential cross section for producing a recoil atom or fragment with energy E r , atomic weight A and atomic numberZ,and the sum extends over all types of recoil atoms and fragments. L(E r ) is the Lindhard partition function [32], which gives the fraction of the recoil energy E r that contributes to displacement damage. N a and A 0 are Avogadro's numberand the atomic weight of the material, respectively. NIEL calculations are available for dierent particles both for Si [41,44,45,46] and for GaAs [41,45,47], although for charged pions, which dominate the radiation environment in the SCT near the interaction point, and, to a smaller extent for protons, these calculations presently have rather large uncertainties [48]. Table 1 summarizes some values of the NIEL in GaAs. By comparing the NIEL with the one of particles chosen as a reference | like e.g. 1 MeV neutrons or 24 GeV protons | damage functions can be dened which correlate the displacement damage produced by dierent particles at dierent energies in a given semiconductor material. Such damage functions have been calculated by several authors for dierent incident particles and semiconductor materials [44, 4 5 , 4 6 ]. Using these calculations, one can e.g. calculate for a given radiation environment a n equivalent 1 MeV neutron uence, which is expected to produce the same amount of displacement damage, or the results of radiation hardness tests can be scaled for a particular application. For example, the NIEL values for 24 GeV protons and 300 MeV pions are equivalent within the large uncertainties. Therefore, 24 GeV proton uences can beconsidered as goodapproximation to the dominating 300 MeV Particle NIEL Ratio to Ratio to and energy ( k eV cm 2 = g) 1 MeV n 24 GeV p The uncertainty of the NIEL given for 1 MeV neutrons is 10%, while for 24 GeV protons and pions an uncertainty on the order of a factor of 2 is not unlikely.
pions in the ATLAS SCT environment. However, it must be emphasized that by doing so one relies on the validity of the NIEL hypothesis which should therefore bechecked thoroughly. Further it should benoted that eects from non{local energy deposition outside a device's sensitive v olume may be essential for high energy protons, and any additional long term ionization eects must be taken into account in order to obtain goodagreement of the damage functions with the experimental data.
While the displacement damage produced by dierent particles at dierent energies relative to e.g. 1 MeV neutrons is primarily a function of the NIEL, the degradation of a particular semiconductor device by a given amount of irradiation, measured as a change in a specic performance parameter, depends, of course, on many additional factors. These include e.g. the sensitivity of the device to a reduction of minority carrier life time, carrier density and mobility, the doping and impurity concentrations, and the inuence on the annealing processes of temperature, radiation ux and biasing conditions during and after the irradiation.

Radiation Damage in LEDs and VCSELs
The most important eect of radiation on LEDs is the introduction of stable defect complexes which act as non{radiative recombination centers (see e.g. [50,51]), so that, at a xed operating current, the fraction of non{radiative recombinations is increased and therefore the minority carrier life time decreases. Thus, the minority carrier life time is sensitive to the non{ionizing energy loss (NIEL) inside the active semiconductor area and can be used to verify experimentally the NIEL hypothesis.
The total pre{irradiation minority carrier life time 0 may bewritten as where r and nr denote the initial life times associated with radiative and non{ radiative processes, respectively. The radiation{induced non{radiative recombination centers decrease the minority carrier life time , where v th is the average thermal velocity of the minority carriers, n i and i denote the density and capture cross section, respectively, of the defect complexes of a given type, and the sum extends over the dierent t ypes of radiation{induced defects. As long as there is no signicant overlap of the regions where the individual incident particles produce crystal defects 5 , the resulting initial 6 defect densities n i will be proportional to the uence (or, equivalently, t o t h e dose), n i = c i (5) where the coecients c i denote the density of defects of a given type produced per unit uence.
Dening a damage constant K, eq. (4) can be rewritten as If it is assumed that the total current density in the LED junction is dominated by diusion currents, the ratio 0 = can be directly related to the relative light output RLO after irradiation [50]: This relation shows that a degradation of the minority carrier life time , and thus of the light output, is expected if the product 0 K becomes signicant compared to 1. The damage constant K is determined by the NIEL of the radiation, by the composition of the semiconductor material, and by the amount of annealing that has taken place. Therefore K cannot be changed easily in order to obtain more radiation hard LEDs. However, the pre{irradiation minority carrier life time 0 may be minimized by increasing the radiative recombinations through heavily doping the optical emitting region and by operating the LED at high current densities [51]. This has not only the advantage of increasing the radiation hardness, but improves also the light yield and the speed of the LED.
Since for VCSELs the minority carrier life time in the lasing regime is dominated by stimulated emission, it is several orders of magnitude smaller than for LEDs and much higher uences are required to produce a substantial change of 0 =. Below the lasing threshold, however, the minority carrier life time is determined by s p o n taneous emission as is the case for LEDs. Therefore the presence of radiation induced non{ radiative recombination centers decreases the light yield and higher currents are required to enter the lasing regime. Thus the lasing threshold current o f V CSELs is expected to increase after irradiation.

Annealing of Radiation Damage
Signicant annealing of radiation damage of LEDs under forward bias (injection annealing) has been reported for dierent types of radiation and LEDs (and also for laser diodes) by several authors [5,50,53,54,55,56], but in some special cases, mostly concerning SiC and Si{doped GaAs devices, no annealing could be obtained [50,57]. In GaAs p + n junctions defect annealing under forward bias was studied after 1 MeV electron irradiation [22] with Deep Level Transient Spectroscopy [58], and it was found that the annealing process is directly related to recombinations at the defect.
Thermal annealing of LEDs was reported, too, but seems to require temperatures above 200 C for the annealing of defects which are stable at room temperature [37,55].

Overview
In our tests we studied two t ypes of Light Emitting Diodes (LEDs) from ABB Hafo (Sweden) and from GEC{Marconi (U.K.), respectively, and Vertical Cavity Surface Emitting Laser Diodes (VCSELs) from Sandia National Laboratories, Albuquerque (NM, USA). An overview of the devices is given in table 2. Except for the dierent packaging, the ABB LEDs correspond to those used in the previous life time test [3,5], where they showed an excellent radiation hardness under neutron irradiation. The GEC LEDs are custom devices, which w ere developed together with a dedicated radiation hard package for possible use at LHC. The kind of VCSELs tested is still under development and we examined devices from three dierent production lots.
All devices were purchased without optical bres and without any packaging, and were mounted on ceramic boards serving as carriers for the operation in the scanning machine. A picture of the dierent t ypes of devices mounted on the ceramic boards used in the scanning machine are shown in gure 9 (section 4, p. 18).
Before the irradiation, the devices were installed in the scanning machine (see section 4.2) and an initial characterization of the devices was performed. The default measurement conditions and the electrical characteristics of the devices are summarized in table 3 and will be described in the following sections in more detail.    n/a n/a n/a Table 3: Electrical characteristics of operational devices at the default operating conditions chosen for our tests. The light output is given in terms of optical power that could be coupled into a passively aligned multimode bre. The quoted errors are RMS errors and do not include systematic errors. The fall and rise times are corrected for the intrinsic rise time of 2.1 ns of the measurement system.

Required Light Output
The LEDs have to provide an initial optical power of at least 10 W i n to a multimode bre in order to make sure that the optical links remain operational during 10 years at LHC in spite of the anticipated degradation [1]. Since not all of the LEDs purchased were tested by the manufacturer, not all LEDs satised this requirement at the default operating current as can be seen from gure 2. For this test only devices with a light output of less than 5 W w ere rejected for not loosing too much in statistics.

ABB Hafo LEDs
The ABB LEDs were chosen initially for possible use in the ATLAS SCT because of their radiation hardness, their outstanding light yield, and their fast response. However, they are not available commercially in a package which is suitable for the application in the SCT. Recently, ABB Hafo was sold to Mitel Corporation and the production of this LED type was stopped.
The distribution of the pre{irradiation light output of all 144 ABB LEDs at 10 mA and +10 C is shown in gure 2a. Due to the fact, that the ABB LEDs were tested by ABB Hafo, all devices were working. However, 9 out of 144 LEDs yielded less than 5 W of optical power and were screened out.

GEC{Marconi LEDs
The second type of LED tested was developed by GEC{Marconi (U.K.) on behalf of the Oxford group. These LEDs were specially designed for use in the SCT, and a suitable radiation hard package is available. However, in contrast to the ABB LEDs, they show a marked threshold eect and produce little light below a forward current of about 10 mA. Therefore, a 20 mA forward current was chosen as the default operation current (see table 3).
The GEC LEDs were not tested by the manufacturer, and 17 out of the 176 tested devices gave no light at all. This is mainly due to mechanical defects like broken or not existing bonding wires. The distribution of the light output is shown in gure 2b. As can be seen another six devices were below the 5 W threshold and had to be screened out.

VCSELs from Sandia National Laboratories
Vertical Cavity Surface Emitting Lasers (VCSELs) are an attractive alternative to LEDs as emitters for an optical link. In contrast to the older edge emitting lasers, the mirror is grown as paired layers of semiconductor materials into the laser structure itself. Therefore, no mirrors have to be cleaved out of the crystalline structure, making the production and the testing of laser diodes much easier, since testing is possible already on the wafer. 11 Figure 3: Sketch o f a n o xide{conned VCSEL with distributed Bragg reectors (DBR) grown above and below the optical cavity. The lateral oxidized layers conne the current to a small region of the cavity.
Three types of VCSELs from three dierent production runs were purchased from Sandia National Laboratories, Albuquerque (NM, USA), in collaboration with the ATLAS group developing the liquid argon calorimeter readout [59]. The three types are labeled by roman numbers in the following. At +10 C the lasing threshold current is about 2 mA (or a bit less) for VCSEL types Ia and Ib, and a few devices have a threshold current below 1 m A (VCSEL type II). Due to the dierent lasing threshold currents, each t ype of VCSEL was assigned its own default forward current as listed in table 3. As can beseen in the table, the spread of the light output of the individual devices is large for all three types of VCSELs. The fall and rise time within the lasing regime is much faster than the one of the electronics of the scanning machine ( 2 ns). The thermal behavior of VCSELs depends strongly on the design of the device and will be discussed in section 3.3.2.
In contrast to the more commonly available proton implanted VCSELs, the VCSELs used in our test were made with a selective oxidation technique [60]. Figure 3 shows a sketch of an oxide conned VCSEL. The oxidized layers conne the current to a small region of the optical cavity, increasing the current density and thus decreasing the lasing threshold. Above and below the optical cavity, distributed Bragg reector (DBR) mirrors are built from interleaved layers of GaAs and Al x Ga 1 x As. The variable x parameterizes the refractive index and ranges from 0.96 in the innermost to 0.16 in the outermost layers.
All VCSELs tested were manufactured in the form of twentyfold arrays and mounted onto ceramic boards for use in the scanning machine. However, the geometrical arrangement of the driver electronics in the scanning machine is not adapted to these arrays and only 8 electronics channels are available for each ceramic board.
Thus only 8 out to the 20 VCSELs can be operated simultaneously, although by shifting the ceramic board relative to its connector another 8 devices become accessible. Therefore, 16 out of the 20 devices on each array could be studied.

Temperature Dependence of the Lasing Threshold Current
The lasing threshold current of VCSELs is determined by the optical gain at the lasing wavelength. Since both the Fabry{Perot cavity wavelength and the gain spectrum depend on the operating temperature, the lasing threshold current is not necessarily a monotonic function of temperature 7 , in contrast to most edge{emitting lasers. However, for each t ype of VCSEL there is an optimal operating temperature where an ideal alignment between the gain spectrum and the lasing wavelength takes place, which, in most cases results in a minimum threshold current. Therefore, VCSELs may betuned for a given operating temperature.
The VCSELs used in our test were optimized for room temperature operation, and they show a very uneven behavior at 10 C. However, similar VCSELs could be produced for operation at 10 C, and recent developments lead to laterally oxidized VCSELs with a sub{milliamp threshold current and goodlight output in a temperature range from 77 K to 370 K [62].

Modal Noise and Transverse Modes
VCSEL cavities are tuned to lase in a single longitudinal mode. However, in the presence of reective external surfaces like e.g. an uncoated optical bre, the effective length of the laser cavity may bealtered, possibly giving rise to additional longitudinal modes. This eect was observed for feedback strengths as low as 1% [63], and the modal noise induced by reections was increased in some cases by more than 40 dB above the noise level observed without any feedback [64].
When measuring VCSELs in the scanning machine, a large amount of noise was observed initially, presumably due to backreection from the lens built into the case of the PIN diode used for measuring the light output. A lter with an attenuation of a factor 8 at a wavelength of about 850 nm placed on top of the PIN diode improved the signal quality considerably, as can be seen in gure 4. Therefore, all VCSEL measurements were subsequently made with lters with attenuation factors between about 8 and 22.
Another well known problem of VCSELs concerns the development of higher order transverse modes. Due to the lateral diameter of the active optical cavity of typically >5 m, which is large compared to the lasing wavelength of 0:3 m in GaAs, higher order transverse electromagnetic modes (TEM) may develop. Since the reectivity of the distributed Bragg mirrors becomes lower for higher order TEM modes, the ground mode is strongly favored [65]. Nevertheless, at elevated forward currents (i.e. for higher pump levels), higher order TEM modes may appear due to dierent eects like e.g. spatial hole burning of the lateral gain prole due to thermal lensing [66]{ [70]. shows the transverse light output prole of a single VCSEL of type Ib as a function of the forward current as measured in the scanning machine with the PIN diode placed at a distance of a few millimeters in front of the VCSEL. The development of higher order transverse modes above 2 mA is clearly visible. In the corresponding light output vs. current curve (L{I curve), the development of higher order modes manifests itself as characteristic changes in the slope whenever a new transverse mode starts to develop (gure 6a). If not all of the emitted light is collected, e.g. because the PIN diode is placed too far away from the VCSEL or because of a misaligned optical bre, the development of higher order transverse modes may lead to a non{monotonic dependence of the light output from the forward current as shown in gure 6b.
Both the modal noise and the development of higher order transverse modes represent a potential problem for the coupling of VCSELs to bres. One solution is to use a special optical interface like e.g. the GUIDECAST used in Motorola's OPTOBUS system [71]. The drawback of such a solution is the increased cost and the possible lack of radiation hardness of the interface. However, a passively self{aligned plastic package with a MT connector for a 16 channel two{dimensional VCSEL array has recently been developed [72, 7 3 ] and a bit error rate of 10 11 was achieved for a 1 G b = s optical link. The development of radiation hard VCSEL{to{ bre connectors suitable for the application in ATLAS is currently pursued by the group developing an optical readout for the liquid argon calorimeter [74].  In order to make the necessary long term measurements of a large number of irradiated LEDs feasible, an automatic measurement system, called the scanning machine, was built with space and driver circuits for 448 LEDs or VCSELs. Before the irradiation, the devices were temporarily installed in the scanning machine for measuring the light output as a function of the forward current and the operating temperature. Most of the ABB LEDs were subjected to a burn{in treatment at 50 mA and 50 C before the irradiation. Unfortunately, due to time constraints, no burn{in period could be allocated for the GEC LEDs and for the VCSELs. After the irradiation, the devices were reinstalled in the scanning machine. Following a rst measurement at the default conditions (see table 3), the devices were operated at dierent forward currents and at either +10 C o r 10 C for several days or weeks in order to investigate the annealing behavior. During this annealing period, frequent measurements were made. In order to investigate the long{term reliability of the irradiated devices, they were operated (and measured regularly) for several months at an elevated temperature of 50 C in order to accelerate the aging.

The Scanning Machine
A picture of the scanning machine with its electronics and with the data acquisition system is shown in gure 7.
Since in the rst test considerable inaccuracy and handling problems were introduced by the use of optical bres and connectors, and since the radiation hardness of optical bres was suciently demonstrated [56], the light output of the individual LEDs and VCSELs is measured in the scanning machine directly with a moving PIN photodiode at a distance of a few millimeters in front of the diodes, thus avoiding the need of optical bres.
A longitudinal sectional view of the scanning machine is shown in gure 8. Two shifters are used to move an optical receiver board with two PIN photodiodes (one for measuring the LEDs and the other equipped with a lter for the much brighter VCSELs) in the longitudinal (X) and transverse (Y) directions in steps of 0:1 m m and 0:04 mm, respectively. The LEDs or VCSELs are mounted on temperature controlled supports called blocks (gure 10), emitting the light downwards to the PIN diodes. A precise mechanical mounting ensures a well dened vertical distance (which can beadjusted for each block as needed) between the devices and the PIN diodes. By moving the PIN diode to the position yielding the maximum signal below a given device, the measurement of the light output of the diodes is reproducible to about 2%.
The scanning machine is equipped with 7 blocks for which the temperature can beset individually between 10 C and 50 C using Peltier elements for cooling or heating. In order to keep the humidity low and to protect the diodes from ice at low temperatures, the scanning machine provides a tight volume with a nitrogen atmosphere. 16   Each block has space and driver electronics for 64 devices. The LEDs are assembled as surface mounted devices in groups of 8 on ceramic boards (called modules), which ensure a goodthermal contact with the temperature controlled blocks. The VCSELs are produced as 20{fold arrays which are mounted on the same kind of ceramic boards but providing 16 connector pads at half the pitch for an alternate operation of eight devices. Electrical contacts are made by wire{bonding from the diodes to the modules, and by special connectors from the modules to signal terminating boards, and from there via at ribbon cables to the driver boards located outside of the scanning machine. Thus for the irradiation the modules could be easily removed from and reinserted into the machine.
The operating current of the LEDs or VCSELs can be set independently for each module between 0 m A and 100 mA, and each device can be individually switched on, o, or set to pulsed mode. For each diode the current and the forward voltage, the light output (including the spatial prole of the emitted light cone), and the fall and rise time can bemeasured automatically under computer control, using a digital oscilloscope (LeCroy 9450) with a bandwidth of 350 MHz and a VME based 16{bit ADC card. Additional measurements were made of the temperature of each block, of the humidity inside the scanning machine, and of the temperature of the PIN diodes.
The operation of the scanning machine is controlled by a data acquisition system running on a dedicated OS/9 computer, allowing completely automated long term measurements. The data acquisition software was written in C++ based on an object{oriented design, where much emphasis was put on reliable long{term operation and comprehensive logging of all measurements and of the machine state. 19

Normalization and Calibration of the Light Output Measurements
For practical reasons the light output of the LEDs was usually measured at the actual operating temperature and forward current during the annealing and the aging. However, normalization to the default operating conditions (see table 3) was possible by measuring for each device the light output as function of the forward current and the temperature. Since the dependence of the light output from the operating current of LEDs changes during irradiation and annealing, measurements have been made before the irradiation and during the annealing when necessary. The normalized light output thus obtained can be directly related to the pre{irradiation measurements made at the default operating conditions in order to calculate the relative light output (RLO). For both types of LEDs the normalization procedure works well, and the errors introduced by the normalization are on the order of a few percent. For VCSELs, however, a similar normalization procedure is not practical due to the strong and more complicated temperature dependence of the lasing threshold current, which would result in much larger errors of the normalization procedure. Both for VCSELs and for LEDs, only measurements made at the default operating conditions have been used for producing the histograms and the statistics reported thereafter in order to avoid normalization errors.
The measurement of the light output with the PIN diode inside the scanning machine results in a voltage signal which must becalibrated to obtain the optical power coupled into a multimode bre. For this purpose, a few devices of each type were operated on a test bench and manually coupled to a short piece of 50/125 multimode bre. The bres were aligned in three axis to nd the maximum light output, which was measured at the end of the bre with a calibrated optical head. The LEDs were equipped with a micro{lens. Therefore, the maximum light output was found at a short distance from the LED. For VCSELs Butt coupling, i.e. the bre is in direct contact with the emitter, was used, and a slight pressure was applied in order to avoid modal noise resulting from back reection from the cleaved end of the bre into the laser cavity. Not surprisingly, for VCSELs the light coupling into bres was found to bevery sensitive to the pressure applied, and therefore only an approximate calibration constant could be determined.
For the GEC LEDs, the optical power measured at the end of the actively aligned bre was compared to the optical power obtained from packaged GEC LEDs with passively aligned bres. The latter was in average 20% lower which was attributed to the dierent alignment technique. Since for the SCT the light output of the packaged devices is the relevant quantity, all calibration factors for the scanning machine were corrected for this 20% loss due to passive alignment.

Irradiation
Most of the LEDs and VCSELs were irradiated with either neutrons or protons, reaching uences as high as 510 14 Table 4 Figure 11: Neutron energy spectrum at the RAL ISIS facility, for three dierent distances from the target [76] at ISIS is shown in gure 11. The /n and the p/n uence ratios are about 10% and 10 3 , respectively [77]. The neutron ux is rather low, varying between about 1:5 10 12 and 4:3 10 13 n c m 2 day 1 depending on the operating conditions of the proton synchrotron and the position of the samples. In addition to the 1 MeV neutrons, there is a large background of thermal neutrons which, according to NIEL calculations [45], cause about two orders of magnitude less displacement damage in GaAs than 1 MeV neutrons. For most of the neutron irradiations done at ISIS, the contribution of the thermal neutrons to the total displacement damage is estimated to be less than 5%, and, in the case of the devices irradiated with the most unfavorable ratio of fast to thermal neutrons, less than 10%. Given an uncertainty o f the dosimetry of typically 15%, the contribution of the thermal neutron background was neglected.
In addition to the neutron irradiation at ISIS, for the reirradiation studies a few devices were irradiated at the Dynamitron in Birmingham [78], where neutrons are produced by deuterons hitting a 0:75 mm thick beryllium metal target. The energy of the incident deuterons can be adjusted in the range from 2.6 to 7.0 MeV. The resulting neutron energy spectrum is shown in gure 12. Very high uxes up to approximately 8 10 14 n c m 2 day 1 can beachieved [79].
All proton irradiations were done at CERN, where 24 GeV protons are available from the CERN PS (proton synchrotron) and uences of approximately 10 14 p c m 2 day 1 can be reached. Due to the extraction of the protons from the PS, they are delivered in bunches of 10 11 protons every 14 seconds. During the irradiation the modules with the LEDs or VCSELs were mounted one behind the other in slide holders aligned with the beam axis, usually behind devices irradiated by other groups. Because of the rather large size of about 2 cm of the modules in the transverse direction, the beam was not always uniform over the whole module, making the dosimetry with aluminum foils dicult and leading to higher uncertainties in the dosimetry than usual for irradiation at the PS. After the irradiation period of typically a few days for the proton irradiation and several weeks for the neutron irradiation, the irradiated devices were stored at room temperature until their activity had decayed to a level allowing the shipping of the devices back to the laboratory.

Measurement Procedure for Irradiated Devices
After shipping the irradiated devices back to the laboratory, they were reinstalled into the scanning machine and cooled down to either +10 C or 10 C for the annealing treatment. The rst measurement of all irradiated devices was done at the default forward current given in table 3. In the case of VCSELs, for some (but not all) devices on each array, light output vs. current and forward voltage vs. current curves were measured, too.
Once the rst measurement after irradiation was made, the forward current w as set to the desired annealing current and the devices were operated in DC mode for several days or weeks and frequently measured in order to monitor the progress of the annealing. When no further annealing was observed at a given forward current, the annealing treatment w as either stopped or continued at a higher current in order to investigate the dependence of the injection annealing on the forward current.
In order to study the long term reliability o f irradiated LEDs and VCSELs, the aging was accelerated by operating the diodes at an elevated temperature of 50 C for several months once the annealing was nished. However, for some of the GEC LEDs the annealing was extremely slow at the chosen forward current and would have taken many months to nish. Therefore, before the annealing was nished, these devices were set at 50 C where the annealing was continued at the elevated annealing current. For all other devices the long term studies at 50 C w ere performed at the default operating current. 24 5 Results

Overview
In the following sections we describe the results obtained in our studies. In section 5.2 we present the damage constants obtained for GEC LEDs from the rst measurements made after irradiation. The annealing behavior of LEDs and VCSELs is described in section 5.3 and 5.4, respectively, followed by the results of the reirradiation studies in section 5.5. The long term reliability is discussed in section 5.6 (life time) and 5.7 (estimated relative light output after 10 years of operation at LHC). The overall statistics are presented in section 5.8 and the behavior of failing devices is described in more detail in section 5.9.

Damage Constants for GEC LEDs
By measuring the light output after irradiation, LEDs may be used to measure displacement damage. As shown in section 2.5, the relative light output after irradiation with a uence can be related to the damage constant K, or, equivalently, to k 0 K: In order to determine a physically meaningful value of k, it is important that either no injection annealing takes place, or that the amount of annealing can be quantied, since the defect concentration and thus the damage constant changes during the annealing 9 . Due to the fact, that the GEC LEDs exhibit no or only a very slow annealing at the default current o f 2 0 m A , (almost) no injection annealing takes place during the few ms of operation required for the light output measurement. Therefore, those GEC LEDs that were not operated during the irradiation can be used to determine the parameter k both for neutron (k n ) and for proton (k p ) irradiation. According to the NIEL hypothesis the ratio k p =k n is determined only by the ratio of the proton and neutron NIEL, and therefore a measurement of k p =k n provides a direct test of the NIEL hypothesis and of the theoretical NIEL calculations.
Damage constant ts for both neutron and proton irradiation of GEC LEDs are shown in gure 13. The resulting proton{to{neutron damage ratio k p =k n = 3 : 2 0 : 1 for 24 GeV protons and ISIS neutrons is in excellent agreement with the NIEL calculations for GaAs summarized in table 1 (in section 2.4) giving a ratio of 3.2. This result provides an important conrmation of the NIEL hypothesis, which is crucial for the extrapolation of radiation hardness studies to the LHC environment. 9 The damage constant K could be determined more directly by measuring the minority carrier life time in the frequency domain. This can be achieved by modulating a DC bias current with an AC signal and by measuring the AC frequency response [55]. Since such measurements can be made at low currents, the inuence of injection annealing can be minimized. However, the electronics of the scanning machine does not support this technique, and direct measurements of the fall time of light signals turned out to be nearly impossible given the very low light output of LEDs after irradiation.  If a similar t is done for GEC LEDs which were pulsed with a 20 mA forward current during the irradiation, a lower value ofk p = ( 6 : 8 0 : 2) 10 14 cm 2 is found for proton irradiation. As expected, the damage constant is signicantly lower due to the injection annealing already taking place during the irradiation.

Injection Annealing of ABB Hafo LEDs
A typical behavior of an ABB LED that was not biased during the irradiation is shown in gure 14a: After an irradiation with 1:610 14 p=cm 2 , the relative light output (see section 3.2.1) of this particular LED was decreased to a few percent. The operation of this LED at 10 mA and +10 C for about 12 days did not change the light output signicantly. However, when the forward current was increased to 50 mA, a fast annealing occured, increasing the relative light output to about 40% within a few hours. A dierent behavior is shown in gure 14b for another ABB LED irradiated with 3:210 14 n=cm 2 . In this case, annealing took already place at a current of 10 mA (and at +10 C), raising the relative light output from 10% to 60%. However, when the operating current was increased from 10 mA to 30 mA after 6 days, anti{ annealing was observed which decreased the relative light output to 40%.
Both annealing and anti{annealing have been observed both after proton and after neutron irradiation. Anti{annealing of ABB LEDs has been seen at dierent forward currents between 10 mA { 50 mA, although it was more pronounced at higher currents and not all devices showed the same amount o f a n ti{annealing. While the behavior described above can be understood qualitatively as being due to diusion processes involving dierent t ypes of defects in the LED junction, a quantitative modeling would require microscopic information about the defect types involved, their electrical properties and their concentrations in the LED junction. Such information could be obtained e.g. from Deep Level Transient Spectroscopy (DLTS) measurements [58,80].

Injection Annealing of GEC{Marconi LEDs
The annealing behavior of GEC LEDs is less favorable than the one of ABB LEDs. At currents below 40 mA there is almost no or only a very slow annealing (gure 15a). An annealing current of 40 mA { 80 mA is required to produce a signicant amount of annealing. While at 40 mA the annealing process is rather slow (gure 15b), there seems to belittle dierence in the annealing at currents between 50 mA { 80 mA. Figure 15c shows a GEC LED irradiated with 4:410 14 n=cm 2 , which annealed within a few days at 50 mA and 10 C. At currents higher than 80 mA, anti{annealing was observed even for unirradiated devices (gure 15d).

Temperature Dependence of the Annealing
All devices were at room temperature during irradiation, while being stored after the irradiation, and during the transport back to the laboratory. However, in complementary irradiation studies done at the University of Birmingham, a few GEC LEDs were irradiated at about 7 C without observing any signicant dierence in the resulting radiation damage [75]. Further, except for a yet unexplained drop 27 in light output for a single ABB LED (see section 5.9) no signicant annealing or anti{annealing was observed during room temperature storage periods in the laboratory. As mentioned in section 2.6, thermal annealing is not expected at room temperature for LEDs [37,55].
Most of the annealing studies were carried out at either +10 C o r 10 C. Comparing the annealing behavior of devices irradiated with the same uence and annealed at the same forward current but at a dierent temperature of either +10 C or 10 C, no signicant dierence was found in the nal relative light output at the end of the annealing treatment.
However, in a few cases the annealing of LEDs was accelerated when the operating temperature was increased from 10 C to 50 C in order to accelerate the aging. For most of the ABB LEDs, the annealing was completely nished when the high temperature operation was started. Therefore, only a slight or no additional annealing was observed at 50 C. The only exception is shown in gure 16a. This ABB LED showed a rather slow annealing after the irradiation with 5.310 14 n=cm 2 . When the temperature was raised, the annealing was accelerated and reached a relative light output of 40%, which is in goodagreement with the other LEDs on the same module that annealed completely before the operating temperature was increased. However, for GEC LEDs the annealing takes much more time and for most of the devices the annealing was not completely nished before the temperature was raised to 50 C. Therefore, the relative light output continued to increase during the high temperature operation. Obviously, the amount of annealing observed at 50 C depends on the forward current, too. If the forward current was reduced to the nominal 20 mA, the annealing was usually stopped (gure 16b). A few exceptions have been observed like e.g. the GEC LED shown in gure 16c which exhibits a very strong annealing at 50 C and 20 mA after 40 days of annealing at 80 mA. 29 While in most cases the aging of the devices was studied at the default current (see table 3), some GEC LEDs (for which considerable more annealing was expected) were continued to beoperated at an elevated forward current during the high temperature aging period. For these LEDs the annealing was accelerated at 50 C (gure 16d). However, the nal relative light output when the annealing was nished does not dier from those devices completely annealed at 10 C.

Summary
The results of the short term annealing studies of LEDs are summarized in gure 17 by plotting the relative light output of individual LEDs after irradiation as a function of the uence. The uences for the neutron irradiation (with 1 MeV neutrons) at the RAL ISIS facility, and for the proton irradiation (with 24 GeV protons) at the CERN PS, are given by the top and the bottom axis, respectively. Using the NIEL calculations summarized in table 1, these axes were scaled such that the NIEL is the same for the corresponding neutron and proton uences. Thus if the radiation damage in LEDs after the annealing can be parameterized by the NIEL, the relative light outputs after neutron and proton irradiation should concur in gure 17. Indeed, within the fairly large spread of the data, which is partly due to the uncertainties in the dosimetry, the proton and neutron irradiation results match fairly well.
The importance of injection annealing is evident from gure 17 when comparing the measurements of the relative light output before (dots) and after (large symbols) the annealing. A comparison of the ABB LEDs which were annealed with a maximum current of 30 mA (full circles) with those annealed at up to 50 mA (hollow circles), shows that the latter generally reach a higher relative light output. Thus, the degree of the annealing depends on the maximum annealing current applied.
If the LEDs are operated during the irradiation as will be the case at LHC, depending on the radiation ux and the operating current, some or all of the annealing will take place during the irradiation. Consequently, the light output measured shortly after the irradiation is much higher than if the LEDs are o during the irradiation. This behavior was observed for about one third of the LEDs which were operated in a pulsed mode during proton irradiation. For GEC LEDs which were biased during irradiation, a signicantly lower damage constant w as found (see section 5.2).

Annealing of VCSELs
The degradation of VCSELs by irradiation and the subsequent annealing is best discussed using light output vs. current plots. Figure 18 shows such plots for two VCSELs irradiated with 1:810 14 p=cm 2 . One was o during the irradiation (gure 18a) while the other was operated in pulsed mode with a current of 4 m A and a duty cycle of 25% (gure 18b). The rst measurement after the irradiation (squares) reects this fact: the un{biased one gives no light up to a forward current of 5 m A while the pulsed device shows very little degradation. After an annealing period of 5.9 days at 4 mA and +10 C both VCSELs recovered almost completely. The outstanding radiation hardness of VCSELs is evident from gure 19. VCSELs irradiated with uences up to 3.710 14 p=cm 2 yield, after a short annealing period of a few hours to days, as much light as before the irradiation. For some devices, even a lower threshold current and a better light output than before the irradiation were observed. This behavior is not yet understood.

Reirradiation Studies
The irradiation of most devices used in this test was done in one step without intermediate annealing periods. This is an unrealistic scenario with respect to LHC where the total uence is applied over ten years with long pauses in between. Therefore, a few devices have been irradiated twice with an intermediate annealing period in order to check if there is a signicant dierent relative light output with respect to a single irradiation with the same total uence. Table 7 summarizes the irradiation and annealing conditions of the reirradiated devices. The total uence of both irradiations is calculated by scaling the individual uences with the NIEL in GaAs (see section 2.4). As shown in gure 20 the relative light output of the LEDs irradiated twice with a high total uence tend to be somewhat higher than the LEDs having the total uence applied at once. At l o w er uences, i.e. at an equivalent uence of about 210 14 p=cm 2 , no signicant dierence is found. In order to make a n y prediction on device life time in a reasonable amount of time, the aging of the devices studied has to be accelerated. It is generally believed that aging of LEDs takes place only under forward bias. Thus we can achieve an acceleration factor of 12 with respect to LHC by operating the devices with a DC current. This factor is based on the estimate, that a) LHC will run about 100 days per year, b) the average link occupancy is about 50%, c) the LEDs are o when no data is transmitted, and d) the 0 and 1 bits are balanced.
The acceleration of the aging of VCSELs depends on the operating mode of the optical link. If we assume that a constant bias current is needed to keep the VCSEL near or above the lasing threshold when no data is transmitted, only a factor of three could begained with respect to the SCT by continuously operating VCSELs in DC mode.
As shown in section 2.3 an additional acceleration can be achieved by increasing the operating temperature. Because the thermal activation energy E a is not known for neither the LEDs nor the VCSELs tested, the acceleration factor for the operation at 50 C cannot becalculated with precision. However, if a nominal operating temperature of 10 C is assumed (the operating temperature of LEDs or VCSELs in the ATLAS SCT is expected to bebetween about 10 C and +10 C), the typical values of E a found in the literature result in acceleration factors from 8 to 160 for LEDs. The thermal activation energy for VCSELs is about twice as high than the one for LEDs. Therefore, as a conservative estimate, an additional acceleration factor of 10 (35) is assumed for LEDs (VCSELs) for the aging at 50 C. The elevated currents applied during the annealing and in a few cases during the aging lead to an additional acceleration factor according to the Arrhenius equation (see section 2.3). However, the exponent is not known for the devices used in this test and therefore no additional acceleration factor is taken into account.

Long Term Stability
The light output of most of the devices operated at 50 C and at their default current remained essentially stable. However, some devices exhibited a slow anti{ annealing of a few percent permonth during the high temperature operation, and some LEDs showed instabilities of the light output. An example of an ABB LED showing both behaviors after the irradiation with 0.8510 14 n=cm 2 is presented in gure 21a. A rather strong anti-annealing is observed by one GEC LED irradiated with 4.410 14 n=cm 2 during the aging period (gure 21b). However, the decrease of the light output would not aect the performance of the optical link. Figures 21c and 21d show two GEC LEDs both irradiated with 0.7710 14 n=cm 2 exhibiting a sharp drop in the light output during the high temperature operation, but showing no other irregularities. Since after the drop, the light output is still sucient for the operation of an optical link, these LEDs were not considered as failing, too.

Estimated Relative Light Output after Ten Years of Operation at LHC
The distribution of the relative light output of GEC LEDs irradiated with 24 GeV proton uences between 1.510 14 p=cm 2 and 2.510 14 p=cm 2 and annealed at forward currents between 50 mA and 80 mA is shown in gure 22. The left histogram shows the relative light output distribution after the annealing, and the right histogram shows the corresponding distribution after operating the LEDs at 50 C for up to about 7 months, corresponding to up to about 70 years of operation in the ATLAS SCT. As can be seen, during the high temperature operation the relative light output remained essentially stable. Although the LEDs included into the gure were irradiated with uences up to about twice as high as those currently estimated for the ATLAS SCT during 10 years of operation, the relative light output distribution represents a reasonable approximation to the relative light output distribution, given the large uncertainties in the uence calculation, that can be expected for GEC LEDs after 10 years of operation at LHC.    Devices with a light output less than 10% after the annealing treatment and after the aging are considered as dead. This numberismuch l o w er than the 50% used as the standard denition in industry and has been derived from the S/N requirement of the optical link proposed for the SCT.

Statistics of Irradiated Devices
As mentioned in section 3.2.1, LEDs with a light output lower than 5 W coupled into a multimode bre were not included into the statistics in order to avoid any bias from very weak LEDs. In order to use the same criteria, VCSELs yielding less than 10% of the initial light output are considered as dead, too. However, a VCSEL with 10% of the initial light output will still yield more light than a bright LED before the irradiation.
As shown in section 5.3.2, the annealing current for the GEC LEDs must be in the range of 40 { 80 mA in order to have a considerable annealing after the irradiation. Therefore, GEC LEDs with a maximum annealing current below o r a b o v e this range are excluded from the statistics.
3 out of 91 ABB LEDs and 2 out of 65 GEC LEDs with a maximum uence of 2.510 14 p=cm 2 or 5.510 14 n=cm 2 failed right after the irradiation or during the annealing. After the irradiation with even higher uences, the GEC LEDs yielded almost no light and showed only a very slow annealing at 50 mA. Only one out of 12 devices has reached a relative light output above 10% after an annealing period of up to 70 days. However, after another 80 { 120 days of operation at 50 C and at forward currents of 50 { 80 mA, another 3 devices yield more than 10% relative light output.
The VCSELs of type Ia and Ib seem to be very radiation hard. Only 1 out of 113 devices irradiated with uences up to 3.710 14 p=cm 2 failed. However, the VCSELs of type II have shown 6 failures in a total of 21 devices irradiated with 3.710 14 p=cm 2 or 4.410 14 n=cm 2 . Although the statistics is low, it seems that this type of VCSEL is less radiation hard than the devices with a higher threshold current.
No devices failed during the long term operation at 50 C, but some anti{ annealing and unexplained drops in the relative light output were observed for a few LEDs (see section 5.6.2).

Behavior of Failing Devices
In this section the failure characteristics of the LEDs and VCSELs are summarized. As in the previous section, only devices irradiated with at least 210 14 n=cm 2 or 0.610 14 p=cm 2 and GEC LEDs annealed at a forward current between 40 and 80 mA are considered. An overview of the failing devices irradiated with less than 610 14 n=cm 2 or 2.510 14 p=cm 2 is given in table 9.
The following categories of failures were found: The LED behaves no longer as a diode but as an ohmic resistor. This is the case for one ABB LED after 0.710 14 p=cm 2 and one GEC LED after 2.210 14 p=cm 2 (gure 23a). The latter one gave some light after the irradiation, but died within about 30 min of operation as shown in gure 23b.    The light output of the device is very low and no signicant annealing at a given forward current was observed, in contrast to the other devices on the same module. One ABB LED after 5.310 14 n=cm 2 under a forward current o f 10 mA (gure 23c) and one GEC LED after 2.210 14 p=cm 2 under a forward current of 50 mA fall into this category.
The LED anneals after the irradiation, but after a storage period of three months at room temperature, the light output was decreased to a few percent. This behavior has been observed for one ABB LED after 2.110 14 p=cm 2 (gure 23d).
The VCSEL does not yield any light after the irradiation, but still shows the normal diode characteristics. This is the case for one VCSEL Ib after the irradiation with 1.810 14 p=cm 2 and one VCSEL IIafter the irradiation with 4.410 14 n=cm 2 .
The light output after the irradiation is comparable to the other devices irradiated with the same uence, but the VCSEL died within seconds when operated. This behavior was seen by three VCSELs of type IIafter the irradiation with 4.410 14 n=cm 2 .
There is no light output after the irradiation and the forward voltage is very low. This is the case for two VCSELs II after the irradiation with 4.410 14 n=cm 2 .

Conclusions
The short term annealing behavior and the life time of a large number of LEDs and VCSELs were measured after irradiation with neutron and proton uences beyond those expected at the inner tracker of ATLAS. The VCSELs produced by Sandia National Laboratories were found to bemuch more radiation hard than both kind of LEDs tested in this study which were manufactured by GEC{Marconi and ABB Hafo, respectively. Moreover, at an operation current of only 4 mA the light output of these low threshold VCSELs is nearly two orders of magnitude larger as for LEDs with a 3 to 5 times higher operating current. The measured reduction of the light output of the GEC LEDs after irradiation was used to determine the damage constants both for 24 GeV protons and for the 1 MeV spallation neutrons at the RAL ISIS facility. The resulting proton to neutron damage ratio of 3:2 0:1 agrees perfectly with the ratio obtained from calculations of the non-ionizing energy loss (NIEL) in GaAs. This result supports strongly the NIEL hypothesis on the degradation of LEDs by irradiation, which is crucial for the extrapolation of these radiation hardness studies to the LHC environment. If scaling of the radiation damage in GaAs according to the NIEL hypothesis is assumed, the currently estimated total uence for the innermost barrel layer in the ATLAS SCT corresponds to a 24 GeV proton uence of about 1:5 10 14 p=cm 2 , with an uncertainty of roughly 50%.
The light yield of the GEC LEDs, which w ere irradiated with about 210 14 p=cm 2 , was typically reduced by a factor of 50 to 200, and about a factor of 30 for the ABB LEDs after the same proton uence. However, the radiation damage could be partially annealed by driving the LEDs at a forward current of 40 to 80 mA continuously for a few days to several weeks, resulting in a relative light output with respect to the one before the irradiation of 10% to 20% for the GEC LEDs and of 20% to 40% for the ABB LEDs. Unfortunately the ABB LEDs are no more available, because the company was sold.
The VCSELs, which were pulsed with the normal operating current during the irradiation, showed essentially no degradation up to uences of about 410 14 p=cm 2 , whereas for those which w ere o during the irradiation the lasing threshold current increased. Complete annealing of this radiation damage was achieved by driving the VCSELs for at most a few days at the normal operating current.
A few LEDs died during or soon after the irradiation with uences in the range of 210 14 n=cm 2 to 810 14 n=cm 2 or 0.610 14 p=cm 2 to 2.510 14 p=cm 2 , namely 2 out of 65 GEC LEDs and 3 out of 91 ABB Hafo LEDs. Also 1 of the 113 VCSELs with a threshold current of 2 m A , which were irradiated either with about 410 14 n=cm 2 or with about 210 14 p=cm 2 to 410 14 p=cm 2 , was found dead after the irradiation. The 21 VCSELs with a threshold current below 1 m A turned out to bemuch less robust: 6 of them died after irradiation with about 410 14 n=cm 2 . All the LEDs and VCSELs, which survived the irradiation and the annealing treatment showed stable performance during the subsequent long term test at 50 C, which is equivalent to up to 70 years of operation at LHC.
This study showed clearly, that the tested VCSELs with a threshold current of about 2 m A w ould be the best choice for the optical readout of the SCT (and eventually of the pixel layers). The problem of the packaging, however, is more dicult for VCSELs than for LEDs, because reections can cause huge noise. At the time of this writing it is not sure if radiation hard and aordable VCSELs{to{ bre couplings can bedeveloped.
For the GEC LEDs the packaging problem is already solved. According to the ATLAS Technical Design Report [1], the anticipated reduced light output after annealing of these devices still allows for safe operation of digital optical links at 40 Mb=s, but requires PIN{diode receivers with excellent noise performance. LEDs with an initial light output of less than about 15 W should be used only at the outermost layers of the SCT or even berejected. A special annealing procedure is needed during the shut down periods of the LHC, during which the LEDs must be driven at a current of about 50 mA for several days to several weeks.
The observed failure rates of 1% for the VCSELs with 2 mA threshold and 3% for the GEC LEDs are rather high, but are tolerable in view of the foreseen redundancy scheme for the SCT readout. 44