Impact of Low-Energy Proton Induced Upsets on Test Methods and Rate Predictions

IEEE TRANSACTIONS ON NUCLEAR SCIENCE, VOL. 56, NO. 6, DECEMBER 2009 3085

Impact of Low-Energy Proton Induced Upsets on TestMethods and Rate Predictions

Brian D. Sierawski, Member, IEEE, Jonathan A. Pellish, Member, IEEE, Robert A. Reed, Senior Member, IEEE,Ronald D. Schrimpf, Fellow, IEEE, Kevin M. Warren, Member, IEEE, Robert A. Weller, Senior Member, IEEE,

Marcus H. Mendenhall, Member, IEEE, Jeffrey D. Black, Member, IEEE, Alan D. Tipton, Member, IEEE,Michael A. Xapsos, Member, IEEE, Robert C. Baumann, Member, IEEE, Xiaowei Deng, Member, IEEE,

Michael J. Campola, Member, IEEE, Mark R. Friendlich, Hak S. Kim, Anthony M. Phan, and Christina M. Seidleck

Abstract—Direct ionization from low energy protons is shownto cause upsets in a 65-nm bulk CMOS SRAM, consistent withresults reported for other deep submicron technologies. The ex-perimental data are used to calibrate a Monte Carlo rate predic-tion model, which is used to evaluate the importance of this upsetmechanism in typical space environments. For the ISS orbit and ageosynchronous (worst day) orbit, direct ionization from protonsis a major contributor to the total error rate, but for a geosyn-chronous (solar min) orbit, the proton flux is too low to cause asignificant number of events. The implications of these results forhardness assurance are discussed.

Index Terms—Direct ionization, hardness-assurance, MonteCarlo, proton, single event upset (SEU), soft error rate.

I. INTRODUCTION

A DVANCED complementary metal-oxide-semiconductor(CMOS) technologies are increasingly being used in

spacecraft systems because of the performance gains andreductions in size, weight, and power. The accompanyingdecrease in total gate capacitance as the device size decreasescorrelates to increased single-event susceptibility in staticrandom access memory (SRAM) cells. Upsets have been ob-served in these technologies for particles with low linear energytransfer (LET) (less than ). Characterizingthe low-LET response of advanced technologies is important,as the flux of particles exceeding a part’s LET threshold rises

Manuscript received July 17, 2009; revised August 30, 2009. Current ver-sion published December 09, 2009. This work was supported in part by theNASA Electronic Parts and Packaging Program and the Defense Threat Reduc-tion Agency under Grants HDTRA1-08-1-003 and HDTRA1-08-1-0034.

B. D. Sierawski and K. M. Warren are with the Institute for Space and DefenseElectronics, Vanderbilt University, Nashville, TN 37203 USA (e-mail: [email protected])

J. A. Pellish and M. A. Xapsos are with NASA Goddard Space Flight Center,Greenbelt, MD 20771 USA.

R. A. Reed, R. D. Schrimpf, R. A. Weller, M. H. Mendenhall, and J. D. Blackare with the Electrical Engineering and Computer Science Department, Vander-bilt University, Nashville, TN 37235 USA.

A. D. Tipton is with the Johns Hopkins University Applied Physics Labora-tory, Laurel, MD 20723 USA.

R. C. Baumann and X. Deng are with Texas Instruments, Dallas, TX 75243USA.

M. J. Campola, M. R. Friendlich, H. S. Kim, A. M. Phan, and C. M. Seidleckare with MEI Technologies (NASA/GSFC), Greenbelt, MD 20771 USA.

Color versions of one or more of the figures in this paper are available onlineat http://ieeexplore.ieee.org.

Digital Object Identifier 10.1109/TNS.2009.2032545

rapidly for devices with lower upset thresholds in much of thespace environment.

For highly sensitive devices, protons are able to generateenough charge through electronic stopping to cause single eventupsets (SEU). Evidence of direct ionization from protons hasbeen reported for 65 nm CMOS technologies [1], [2]. Severalchallenges prevent the use of low-energy proton test data forsoft error rate predictions in most cases. The limited range ofprotons, straggling, and uncertainty in stopping power makeupset cross sections difficult to obtain. In [3] the authors presenta methodology to characterize a chain of latches and an SRAMbuilt in an IBM CMOS Silicon-On-Insulator (SOI) process. Theauthors suggest a critical angle analysis whereby the criticalcharge of a device can be estimated through interpretation ofthe single event upset response as a function of angle. In [1],low-energy protons (1.0 – 1.5 MeV) were used to characterizethe angular response of the same part and in [2], data wereobtained using a wide range of accelerator energies.

Barak [4] provides a simple calculation of the high-energyproton single event upset cross section by combining theprobability of secondary particle production with experimentalheavy ion cross section data. The approach provides predictionof high-energy proton effects in the absence of proton data.

Inguimbert et al. [5] built upon the work of Warren et al. [6]in constructing a single sensitive volume model for subthresholdheavy ion data and predicting the high-energy proton responsewith GEANT4. Their analysis showed good agreement with thehigh-energy proton SEU cross section and threshold.

Further, Edmonds [7] considered the contribution of protonsto SEU from direct ionization by incorporating them into ana-lytical path length calculations. He suggests that a constant-LETapproach is a sufficient approximation to the energy depositedby a proton as it moves through a target with dimensons on theorder of one micrometer.

For space applications, one needs to assess the impact oflow-energy protons on error rates. With 100 mils of aluminumshielding, the flux of low-energy protons that reach the part canbe significant for proton-rich environments. The two parameterBendel equation [8], which is only appropriate for empiricallyfitting nuclear reaction driven upsets, inadequately describes thecircuit response.

For devices with low critical charge, such that the direct ion-ization from protons is able to induce SEU, experimental mea-surements contain a mix of SEU mechanisms that are not easilyidentifiable from the proton cross-section curves. Energy loss

0018-9499/$26.00 © 2009 IEEE

3086 IEEE TRANSACTIONS ON NUCLEAR SCIENCE, VOL. 56, NO. 6, DECEMBER 2009

from electronic stopping, Coulomb scattering, and nuclear in-elastic collisions contribute to the measured cross section. Inthis case, the necessity of a detailed physical analysis to quan-tify the contribution of each mechanism makes a Monte Carloapproach appropriate.

In this work, we identify the mechanisms of low-energy,proton induced SEU in a 65 nm bulk CMOS SRAM usingphysics-based simulation models. The physics-based simu-lation used in this analysis was calibrated with experimentaldata acquired from well-controlled and characterized radiationsources to avoid measurement errors typically associated withlow-energy protons. The implications for SEU rate prediction inthe trapped proton belts and inter-planetary solar environmentsare determined. The rate prediction demonstrates that, for afixed sensitive volume geometry, as critical charge decreases,low-energy proton effects will dominate the error rate. Finally,we present the implications for hardness assurance.

II. EXPERIMENTAL RESULTS

The 4 Mbit SRAM used in this experiment was a test arrayfabricated in a Texas Instruments 65 nm bulk CMOS process.The SRAM consisted of 6-transistor, high-density cells. Theoverlayers were approximately 5 thick and the device wasbonded as a chip-on-board to allow for front side testing. Thetest board was biased at 1.2 V and operated by a “low cost” FieldProgrammable Gate Array (FPGA) tester designed by NASAGoddard Space Flight Center.

Proton single event upset data, shown in Fig. 1, were collectedover a wide energy range at four facilities. The cyclotron at In-diana University was used to obtain cross sections at 198 and 98MeV. A 63.3 MeV proton beam at UC Davis was used (with de-graders) to obtain data as low as 19.8 MeV and a 14.6 MeV pri-mary beam was degraded to obtain data down to 2.6 MeV. Thelow-energy ( ) data were acquired at NASA GSFC’sVan de Graaff facility. The LBNL cyclotron was used to confirmthe measured proton cross section at 32.5 MeV and a 6 MeV

beam was broken up and degraded through air to obtain aproton spectrum with a peak at 1.2 MeV. While this beam had adistribution of lower energy protons, the 400 X increase in crosssection confirms the upset mechanism is not due to facility ordosimetry differences.

The proton cross-section curve demonstrates the characteris-tics expected when direct ionization from protons is one of theupset mechanisms. The three orders of magnitude increase incross section going from high energy ( ) to low en-ergy ( ) indicates a transition from upsets due to nu-clear events to direct ionization. At higher energies, the upsetcross section is consistent with the probability of inelastic col-lisions. At low energies, the proton cross section is on the orderof the dimensions of the physical features of the cell, consistentwith direct ionization. We will show that, for certain missions,this energy range may represent the largest contribution to thesoft error rate of the device.

III. TCAD AND SPICE ANALYSIS

A three-dimensional Technology Computer Aided Design(TCAD) model of the device was used to investigate the SEUmechanisms. The dimensions and device parameters were ap-

Fig. 1. Proton single event upset cross-section curve for a 65 nm bulk CMOSSRAM. Proton datasets with dramatic increases in cross section are notamenable to traditional rate prediction models; two parameter Bendel fit shownfor comparison.

Fig. 2. TCAD model of a 65 nm CMOS SRAM. Oxides have been omitted forillustration. Sensitive drain areas are on the order of 1�� . Ionizationcore from ion strike to reverse-biased drain diffusion indicated.

proximated with values representative of the technology node.An ion track with a stopping power ofwas introduced into the device and simulated in the SynopsysSentaurus Device TCAD solver. Fig. 2 contains the TCADstructure with all materials other than silicon omitted and theion strike to one of the N-type Metal-Oxide-Semiconductor(NMOS) transistors. The results of the transient simulationsconfirmed the inherent sensitivity of scaled technology nodes.This investigation of the physical processes of charge collectionestablishes that the electronic stopping of protons, which havea peak LET near , may induce upsets if thepeak occurs near the sensitive device regions.

Further, the critical charge of the SRAM was estimated byintroducing a double exponential current pulse into the SPICEnetlist. In the case of a strike to the off-state NMOSFET, thedrive current of the PMOSFET was used to specify the plateaucurrent, as described in [9]. The rise and fall time constants werefixed at 1 ps, and the fall time was varied. Integrating the current,the method found the critical charges on the sensitive NMOSand PMOS transistors to be 1.4 and 1.8 fC, respectively.

IV. RADIATION TRANSPORT MODELING

The simulation tool used for this investigation is MRED(Monte Carlo Radiative Energy Deposition, developed at Van-

SIERAWSKI et al.: IMPACT OF LOW-ENERGY PROTON INDUCED UPSETS ON TEST METHODS AND RATE PREDICTIONS 3087

TABLE IEXPERIMENTAL ION SPECIES AND ENERGIES

derbilt University). MRED is based on GEANT4 [10], whichcomprises reliable and well-calibrated computational physicsmodels for the transport of radiation through matter. GEANT4is a library of C++ routines assembled by an internationalcollaboration for describing radiation interaction with matter.MRED includes a model for screened Coulomb scattering ofions [11], tetrahedral geometric objects [12], a cross-sectionbiasing and track weighting technique for variance reduction,and a number of additional features relevant to semiconductordevice applications. MRED is structured so that all physicsrelevant for radiation effects applications are available andselectable at run time.

The MRED modeling methodology begins with calibration ofa weighted sensitive volume model for Monte Carlo simulationusing a subset of the heavy ion datasets described in Table I. Thecalibration is based on the device response to low-LET heavyions and high-energy protons; the same calibrated model is usedto predict the upset mechanisms and cross sections for protonsat low energies where experimental data are difficult to obtain.

Fig. 3 shows the measured single event upset cross sectionsfor the ions used at TAMU and LBNL. Each ion has sufficientenergy and range to deposit a known quantity of charge withinthe device region. We make the assumption that the single bitupset cross section corresponds to a physical device area that issensitive to energy deposition events. Therefore only ions withstopping power between 0.6 and were usedfor calibration, as the cross section must be physically justifi-able and we avoid using data where multiple bit upsets are sig-nificant.

Sensitive volumes represent regions of sensitivity within thesemiconductor materials. A weighted sensitive volume model[13] can be used to describe intracell variation in charge collec-tion. The following equation relates energy deposited in thevolumes used to describe a single device to the charge collectedon the corresponding circuit node

(1)

The model quantifies the charge collection at a node by anindividual particle event as a linear combination of the energy

deposited in each volume, , scaled by the respective co-efficient , which is related to the collection efficiency. Whenthe model contains nested volumes — that is, one or more vol-umes encapsulated within another — we make the distinctionbetween the coefficient of an individual volume’s energy de-position and the collection efficiency of the region . Ourmethod for calibration to normally-incident heavy ion crosssections proceeds as follows.

1) Choose the dimensions of a sensitive volume placed inthe active region and consistent with , device features,and isolation of the region. Assign to indicate100% charge collection in this volume.

2) For each heavy ion cross section calibration value ,where and , ( or selected pointsalong a Weibull curve ), create a sensitive volume inthe substrate with a surface area equal to . Assume thedepth, , of is the same as the lateral dimension, .

3) Assign each sensitive volume , where , a chargecollection efficiency where

is in units of and in . This relation-ship essentially uses to modify the quantity of chargegenerated in required to exceed the circuit (i.e.,a strike further from the device requires a particle with alarger to upset the cell than a strike near the device).

4) Create a larger outer volume with low efficiency for high-energy proton secondary particles.

5) Assign , where , to obtain thecoefficients necessary for (1).

6) Simulate the experimental conditions and adjust parame-ters until good agreement is achieved with both heavy ionand high-energy proton datasets.

Following this method, we created the model illustratedin Fig. 4. Volume was assumed to encompass a transistorbounded in width and depth by shallow trench isolation andassigned . Volumes and were assigned individualcoefficients 0.52 and 0.25 respectively. The outer volume wascalibrated to with upset cross sections from protonswith energy greater than 32.5 MeV, corresponding to nuclearreaction-driven upsets.

Within the simulator, we create an array of device models ina silicon block with representative overlayer materials, and irra-diate the target with normally incident particles. The underlyingGEANT4 code models the energy deposition in a device pro-duced by direct ionization, recoils, and nuclear reactions. Eachevent is tallied in a charge collection histogram for the partic-ular ion and energy simulated. The single event upset cross sec-tion versus collected charge is obtained by dividing the reverseintegral of the histogram (i.e., the definite integral fromto ) by the fluence in the simulation. Using the experimentalSEU cross sections for heavy ions and high-energy protons, weuse a least squares fit to validate the assumed . This is il-lustrated for a subset of the data in Fig. 5. As we scan acrosscritical charge values, we can extract the simulated SEU crosssection for each of the species. As the critical charge, or ver-tical line, decreases, note that the simulated SEU cross sectionincreases. As increases, the SEU cross section decreases.This method of fitting estimates to be 1.3 fC which is ingood agreement with the SPICE analysis. The sensitive volume


Fig. 3. Heavy ion single event upset cross sections measured at LBNL andTAMU compared to MRED simulated cross sections. Protons, for comparison,have a maximum LET of approximately �� .

Fig. 4. The weighted sensitive volume model used to model the response of a65 nm CMOS SRAM cell. The surface area of each volume is correlated withnormally-incident heavy ion cross sections.

model and associated critical charge are used through the re-mainder of the paper.

V. MRED SIMULATION RESULTS

The strength of first-principles Monte Carlo simulation is thatthe calibrated device model can be used to study mechanismsand predict the device response in other radiation environments.The model may be used to predict the device response where ex-perimental measurements may have systematic errors or cannotbe made.

A. Proton Response

The low-energy proton response was computed using MREDwith the calibrated model by extending the simulation of pri-mary energies below 32.5 MeV. With only a few key calibra-tion points to establish the charge collection volumes and crit-ical charge we are able to capture the variety of mechanismsthat result in the SEU cross sections in Fig. 6. The high-energyproton cross sections show good agreement with the measureddata and the simulation predicts the steep increase in cross sec-tion below 2 MeV. While we can estimate the 1.4 MeV protonswill be slowed and pass through the active silicon with a mostprobable energy of 430 keV, systematic errors in the energy ofthe proton as it passes through the device are not easily quan-tifiable. These errors are discussed further in Section VI.

Fig. 5. Simulated single event upset cross sections as a function of collectedcharge. Vertical dashed line indicates a critical charge value providing the leasterror to experimental cross sections.

Fig. 6. Simulated and experimental proton cross sections. Model provides goodagreement at high energy and captures the steep increase in cross section at lowenergy.

The nature of a physics-based simulation allows one to cap-ture events resulting in an upset. Fig. 7 illustrates examples ofthe different physical mechanisms leading to upset as predictedby MRED. The 1.4 MeV proton on the left passes through arepresentative back-end-of-line and exceeds the critical chargesolely by electronic stopping. The event in the middle illus-trates a short range silicon recoil that has been displaced by a4.6 MeV proton and traverses the device. At higher energies,such as 32.5 MeV, spallation reactions produce secondary frag-ments that easily upset the device as they pass through.

B. Rate Predictions

The model was used to predict the single event error rate inisotropic space environments. Flux versus energy spectra wereobtained from CREME96 and transported through 100 milsof aluminum shielding. All species from to 92 weresimulated. The computation included both electronic and nu-clear energy loss and the contribution of each particle specieswas recorded. Energy filters required a minimum of 10 keV


Fig. 7. Events causing single event upsets for 1.4, 4.6, and 32.5 MeV incident protons. The material dimensions have been reduced for illustration purposes.

Fig. 8. Simulated error rate as a function of critical charge for CREME96International Space Station orbit, AP8MIN, magnetic quiet, solar minimum,with 100 mils of aluminum shielding. Direct ionization from protons contributeheavily to the error rate.

deposited in a device. The on-orbit error rate is predicted as thetotal errors from all species.

1) International Space Station Orbit: Fig. 8 shows the pre-dicted error rate in the ISS orbit with 100 mils of aluminumshielding. Similar to the cross sections, we plot the error rate asa function of generated charge. This allows us to see the sensi-tivity to the critical charge parameter. In this proton-rich envi-ronment, we focus on the proton contribution. Two curves areshown for the protons. The curve labeled “ ( )” repre-sents the contribution from proton direct ionization only, andthe second labeled “ (all processes)” includes recoils and nu-clear inelastic events. The vertical dashed line once again marksthe critical charge. The model predicts that direct ionizationfrom protons is a large contributor to the total error rate, whichfollows the “all processes” curve. The predicted error rate forthis environment is 4.1 upsets per bit per day. If devicescaling continues, the error rate will continue to rise from pro-tons.

Fig. 9. Simulated error rate as a function of critical charge for CREME96geosynchronous orbit, worst day, with 100 mils of aluminum shielding. Total� � � to 92 environment and major contributors to the error rate shown only.Direct ionization from protons and alphas dominate error rate.

2) Geosynchronous Orbit (Worst Day): In a geosynchronousorbit with the worst day environment and 100 mils of aluminumshielding, we obtain the proton contribution shown in Fig. 9.This environment is both proton and alpha rich and these parti-cles dominate the error rate. The critical charge is at the pointwhere proton direct ionization is significant and will continueto increase in the future. The large flux of protons requires thatwe assess the impact of increasing event rates on higher levelsof error mitigation. The predicted total error rate in this environ-ment is 3.0 upsets per bit per day.

3) Geosynchronous Orbit (Solar Min): In a geosynchronousorbit, during solar min, we find that the proton flux is too low tobe an issue in the total error rate and the contribution in Fig. 10is negligible. Instead, we can examine the contributions of ionspecies and see that the contributions of iron and oxygen con-tinue to dominate the error rate. Here we can see that while theprotons are not significant, the error rate will continue to climbfor decreasing critical charge values. The GCR spectra contain


Fig. 10. Simulated error rate as a function of critical charge for CREME96geosynchronous orbit, solar minimum, with 100 mils of aluminum shielding.Total � � � to 92 environment and major contributors to the error rate shownonly. Abundant low-LET ions dominate error rate.

Fig. 11. Simulated error rate due to direct ionization from protons as a functionof critical charge for CREME96 International Space Station orbit, AP8MIN,magnetic quiet, solar minimum, with 100, 400, and 1000 mils of aluminumshielding.

large numbers of low-LET, light ions, one of which is oxygen.The predicted total error rate in this environment is 1.2upsets per bit per day.

4) Effect of Shielding: Spacecraft shielding is an effectiveway to prevent particles from reaching internal components, butwhile the lowest energy particles are stopped, the shielding si-multaneously slows higher energy particles. Fig. 11 presents theeffect of increased shielding on the error rate due to direct ion-ization from protons in the ISS orbit. The predictions show thatthe increase in equivalent aluminum shielding from 100 to 1000mils reduces the rate by 5X, but does not eliminate upsets fromlow-energy protons. For other external environments, the levelof mitigation gained with shielding will be dependent on theproton energy spectrum. Therefore, this upset mechanism maystill have a significant contribution to the total error rate in aproton-rich environment, even behind heavy shielding.

Fig. 12. A comparison of experimental and calculated mass stopping power forhydrogen on silicon. The experimental data are taken from H. Paul’s database[14] and the calculated values are from SRIM-2008.04 [17]. Note that as protonenergy decreases, the uncertainty in the experimental mass stopping power in-creases.

VI. HARDNESS ASSURANCE METHODS

Our results show, for some environments, that direct ioniza-tion from low-energy protons will be a significant factor af-fecting the overall error rate. If it is believed that a componentis susceptible to upset from the direct ionization of a proton,characterizing the low-energy proton SEU cross section will beimperative. Therefore we consider the implications for hardnessassurance and present the following sequence of steps to qualifythe component for space applications.

Step 1: Measure the upset cross section with long range,low-LET light ions. In general, soft error rate predictiontechniques derived from accelerated testing require that thestopping power of the incident particle is known as it traversesthrough the device. Low-energy proton testing is conductedusing proton beams with energies at or below 2 MeV, suchthat the particles are at or near end-of-range. This is done toachieve mass stopping powers – LETs – large enough to causesingle-event upset through direct ionization from the primarybeam.

The SRIM-calculated mass stopping power of protons inci-dent on silicon is compared to data from Helmut Paul’s database[14] in Fig. 12. It agrees well with the majority of data presented,though the large spread in experimental mass stopping powersbelow 1 MeV is apparent. It is generally true that experimentalerrors in measuring stopping powers increases with decreasingenergy [15]. For transmission measurements at low energy, thinfoils are needed. This makes the presence of pin holes, surfaceimpurities, and thickness variations detrimental to the measure-ment. Furthermore, the critical angle for channeling increasesat low energy along with the importance of multiple scattering.These facts translate to uncertainty in stopping power formula-tions that rely on these data, which includes SRIM and codesbased on ICRU Report 49 [16], such as GEANT4.

Alternatively, ions having greater mass, must be acceleratedto higher energy to achieve a similar stopping power. Thus theyhave increased range and reduced straggling, both issues thatplague low-energy proton testing. The energy loss of higher-


energy ions will be less sensitive to variations in materials asthey are transported through the back-end-of-line or a flip-chipsubstrate.

Step 2: Create an event model using low-LET data and tech-nology information. An event model, for Monte Carlo applica-tion or otherwise, should be derived from well-controlled andwell-characterized radiation sources to be a valid predictor offailure rates in other environments. The predictable transportand energy loss of low-LET heavy ions, and even high-energyproton data should be used to build and calibrate the model. Byapplying a physics-based code to the model, one can predictstochastic processes. It is easier to construct the convolution ofcircuit response with the variable interaction of stopping ratherthan deconstruct data.

The model should be constructed with technology informa-tion to reproduce the various material layers and thicknesses thatwill affect low energy particles.

Step 3: Validate the model by comparing it with measuredlow-energy proton response. Most proton and heavy ion ra-diation test facilities use cyclotron accelerators. Cyclotronscan achieve high particle energies, but have minimum energyrequirements below which resonant beams do not exist. Van deGraaff accelerators can produce low-energy beams with muchtighter energy resolution than cyclotrons – a few keV wideversus several hundred keV wide – however, the upper energylimits of Van de Graaff accelerators make them less useful formany accelerated SEE testing applications.

Since cyclotrons are often used to accelerate protons andthere is a lower limit to the energies they can produce, theproton beams must be degraded to achieve energies belowapproximately 7 MeV for in-air testing. Aluminum and Mylarare common degrader materials for protons, though higher-materials like tantalum are also employed. Additional materialsin the beam line, including the device-under-test, will alsodegrade the energy of the beam. However, degraded beamenergy comes at the cost of increased energy-loss straggling.A charged particle passing through matter loses most of itsenergy through random Coulomb collisions along the lengthof the track [16]. The total energy lost per particle along itstrack length is a stochastic quantity, so some particles willlose more energy and stop sooner than others, which is thebasic concept behind energy-loss straggling. Degrading a beamremoves its quasi-monoenergetic characteristic and therebymakes it difficult to tie proton energy to a specific single-eventupset cross section since the cross section was produced by aspectrum of proton energies, as shown in Fig. 13. It is moreadvantageous to tune the beam to the lowest possible energybefore degrading in order to minimize straggle.

Based on low-energy proton testing experience gained so far,the following suggestions should be considered.

1) Measure and record materials in the beam line upstreamfrom the device-under-test. Higher density and atomicnumber increase the importance of these materials forsubsequent transport calculations.

2) Experimentally determine the mean beam energy and beamenergy-width at the device-under-test location. This shouldbe carried out for the primary, undegraded beam as well asfor all degraded beams. Accurate and precise knowledge

Fig. 13. Simulated kinetic energy spectrum of a 14.6 MeV tune degraded toachieve mean values of 2.6, 8.1, and 12.2 and a 63.3 MeV tune degraded to 10.7MeV.

of the beam energy is critical for subsequent transport cal-culations since differences in beam energy on the order of100 keV can result in single-event upset cross sections dif-ferent by more than an order of magnitude.

3) In reference to item 1), it is important to complete transportcalculations using accurate and properly ordered materialstacks. It is inadvisable to collapse identical materials ap-pearing in different upstream locations. As a proton slowsdown, its stopping power increases non-linearly, so trans-porting a proton through an aluminum-air-aluminum stackis not the same as an aluminum-air stack with equivalentthicknesses.

4) Different levels of systematic error in the form of energy-loss straggling can be introduced depending on the typeof device-under-test package. Topside wire bond schemesare preferable since the semiconductor back-end-of-lineprocess is thin. Controlled collapse chip connection (C4),or flip-chip, packages are more common for commercial,highly-integrated parts. Flip-chip parts require irradiationthrough the backside of the die – i.e., substrate – and shouldbe uniformly thinned if possible to reduce straggling andlower the energy of possible beam tunes. Thicknesses lessthan 100 are preferable, but are fraught with their ownset of challenges. All parts must be delidded.

5) In reference to item 4), if the die is thinned, variations inproton stopping power can occur in different regions of thedevice if the die thickness is not uniform. The single-eventupset cross section can be altered by variations of less than10 of material. Mitigating this problem requires twothings: knowledge of die thickness and a way to mon-itor the physical location of single-event upsets. Die thick-ness can be determined non-destructively via x-ray crosssections or Rutherford backscattering spectrometry or, al-ternately, through destructive physical analysis followingthe experiment. Knowledge of physical upset location isachieved more easily for SRAM arrays, but can be quitechallenging for more complex devices such as SDRAMsand FPGAs. This topic should be incorporated into the testdesign.


Step 4: Use the calibrated model to predict the on-orbit errorrate. The final step to determine the part’s suitability in a spaceenvironment is to use the model to predict a soft error rate. Arate prediction that is derived from tests with unavoidable er-rors will inherit those errors. We advocate that the event modelbe derived from well-characterized radiation sources and onlyvalidated with more stochastic tests. This process of steps willadd assurance to the determined reliability.

VII. CONCLUSION

As devices become more sensitive to ionizing radiation, itwill be important to characterize the SEU response fully, sinceprotons and alpha particles constitute a large portion of the spaceenvironment. The abundance of these particles could potentiallyoverwhelm error mitigation techniques unless they are properlypredicted.

Testing can be used to confirm the existence of a low-energyproton problem, but the use of the results for rate prediction canbe difficult because of the variability in dE/dx, the limited rangeof the particles, and the necessity to have a detailed descriptionof the back-end-of-line. As a result, it is recommended that theexperimenter take precaution to limit the introduction of theseerrors.

In this work, we show that a model that has been calibratedwith high-energy, low-LET heavy ions, which have much longerrange than low-energy protons, can reproduce the low-energyproton data. Such a model can then be validated with low-en-ergy protons, but more importantly it can also be used with con-fidence for general rate prediction in arbitrary space environ-ments, including those in which protons dominate the flux.

ACKNOWLEDGMENT

The authors would like to thank P. Norris for her help at theLawrence Berkeley National Laboratory. This work was con-ducted in part using the resources of the Advanced ComputingCenter for Research and Education at Vanderbilt University,Nashville, TN.

REFERENCES

[1] K. Rodbell, D. Heidel, H. Tang, M. Gordon, P. Oldiges, and C. Murray,“Low-energy proton-induced single-event-upsets in 65 nm node, sil-icon-on-insulator, latches and memory cells,” IEEE Trans. Nucl. Sci.,vol. 54, pp. 2474–2479, Dec. 2007.

[2] D. Heidel, P. Marshall, K. LaBel, J. Schwank, K. Rodbell, M. Hakey,M. Berg, P. Dodd, M. Friendlich, A. Phan, C. Seidleck, M. Shaneyfelt,and M. Xapsos, “Low energy proton single-event-upset test results on65 nm SOI SRAM,” IEEE Trans. Nucl. Sci., vol. 55, pp. 3394–3400,Dec. 2008.

[3] D. F. Heidel, K. P. Rodbell, P. Oldiges, M. S. Gordon, H. H. K. Tang,E. H. Cannon, and C. Plettner, “Single-event-upset critical chargemeasurements and modeling of 65 nm silicon-on-insulator latches andmemory cells,” IEEE Trans. Nucl. Sci., vol. 53, pp. 3512–3517, Dec.2006.

[4] J. Barak, “Simple calculations of proton SEU cross sections from heavyion cross sections,” IEEE Trans. Nucl. Sci., vol. 53, pp. 3336–3342,Dec. 2006.

[5] C. Inguimbert, S. Duzellier, T. Nuns, and F. Bezerra, “Using sub-threshold heavy ion upset cross section to calculate proton sensitivity,”IEEE Trans. Nucl. Sci., vol. 54, pp. 2394–2399, Dec. 2007.

[6] K. Warren, R. Weller, M. Mendenhall, R. Reed, D. Ball, C. Howe, B.Olson, M. Alles, L. Massengill, R. Schrimpf, N. Haddad, S. Doyle, D.McMorrow, J. Melinger, and W. Lotshaw, “The contribution of nuclearreactions to heavy ion single event upset cross-section measurementsin a high-density SEU hardened SRAM,” IEEE Trans. Nucl. Sci., vol.52, pp. 2125–2131, Dec. 2005.

[7] L. Edmonds and K. Edmonds, “A method for estimating SEU ratesfrom protons by direct ionization,” IEEE Trans. Nucl. Sci., vol. 55, pp.2666–2678, Oct. 2008.

[8] W. Stapor, J. Meyers, J. Langworthy, and E. Petersen, “Two parameterbendel model calculations for predicting proton induced upset,” IEEETrans. Nucl. Sci., vol. 37, pp. 1966–1973, Dec. 1990.

[9] S. DasGupta, A. Witulski, B. Bhuva, M. Alles, R. Reed, O. Amusan, J.Ahlbin, R. Schrimpf, and L. Massengill, “Effect of well and substratepotential modulation on single event pulse shape in deep submicronCMOS,” IEEE Trans. Nucl. Sci., vol. 54, pp. 2407–2412, Dec. 2007.

[10] S. Agostinelli et al., “G4–A simulation toolkit,” Nucl. Instrum.Methods Phys. Res. A, vol. 506, no. 3, pp. 250–303, 2003.

[11] M. H. Mendenhall and R. A. Weller, “An algorithm for computingscreened coulomb scattering in GEANT4,” Nucl. Instrum. MethodsPhys. Res. B, vol. 227, no. 3, pp. 420–430, 2005.

[12] A. Kobayashi, D. Ball, K. Warren, R. Reed, N. Haddad, M. Menden-hall, R. Schrimpf, and R. Weller, “The effect of metallization layerson single event susceptibility,” IEEE Trans. Nucl. Sci., vol. 52, pp.2189–2193, Dec. 2005.

[13] K. M. Warren, B. D. Sierawski, R. A. Weller, R. A. Reed, M. H.Mendenhall, J. A. Pellish, R. D. Schrimpf, L. W. Massengill, M. E.Porter, and J. D. Wilkinson, “Predicting thermal neutron-induced softerrors in static memories using tcad and physics-based monte carlosimulation tools,” IEEE Electron Device Lett., vol. 28, pp. 180–182,Feb. 2007.

[14] H. Paul, Stopping Power for Light Ions [Online]. Available: http://www.exphys.uni-linz.ac.at/Stopping/

[15] H. Paul and A. Schinner, “An empirical approach to the stopping powerof solids and gases for ions from �� to �� – Part II,” Nucl. Instr. andMeth. B, vol. 195, pp. 166–174, 2002.

[16] Stopping Powers and Ranges for Protons and Alpha Particles Interna-tional Commission on Radiation Units and Measurements, 1993, ICRURep. 49.

[17] J. F. Zeigler, M. D. Zeigler, and J. P. Biersack, Stopping and Range ofIons in Matter SRIM [Online]. Available: http://www.srim.org/

Documents

Impact of Low-Energy Proton Induced Upsets on Test Methods and Rate Predictions