Radiation effects on floating-gate memory cells

1 downloads 0 Views 152KB Size Report
particular, high LET ions, such as iodine used in this paper, can produce a bit flip. Since the ... been the subject of several studies in the past years, regarding total ionizing dose ... tunnel oxide, thus reducing the FG MOSFET . 3) Electrons ..... Sperandei, M. Ricci, and P. G. Picozza, “TID on 16 MbFlash memo- ries in radiation ...
2222

IEEE TRANSACTIONS ON NUCLEAR SCIENCE, VOL. 48, NO. 6, DECEMBER 2001

Radiation Effects on Floating-Gate Memory Cells Giorgio Cellere, Student Member, IEEE, Paolo Pellati, Andrea Chimenton, Jeff Wyss, Alberto Modelli, Luca Larcher, Student Member, IEEE, and Alessandro Paccagnella, Member, IEEE

Abstract—We have addressed the problem of threshold voltage ( TH ) variation in flash memory cells after heavy-ion irradiation by using specially designed array structures and test instruments. After irradiation, low TH tails appear in TH distributions, growing with ion linear energy transfer (LET) and fluence. In particular, high LET ions, such as iodine used in this paper, can produce a bit flip. Since the existing models cannot account for large charge losses from the floating gate, we propose a new mechanism, based on the excess of positive charge produced by a single ion, temporarily lowering the tunnel oxide barrier (positive charge assisted leakage current) and enhancing the tunneling current. This mechanism fully explains the experimental data we present. Index Terms—EPROM, gate leakage, single-event transients (SETs), transient effects.

I. INTRODUCTION

F

LOATING-GATE (FG) semiconductor memories, such as EEPROM and flash [1], are of interest for space applications thanks to their high information density [2]. They have been the subject of several studies in the past years, regarding total ionizing dose effects [3]–[6], single-event upsets (SEUs) [6], [7], or wearout of a limited number of cells [8]. The most radiation-sensitive part of commercial flash memories is the complex circuitry external to the FG cell array. Different functionality failures have been detected in commercial devices depending on the operation mode during heavy-ion irradiation [3], [4], [6], [9]. On the other hand, these works have shown that bit flips due to heavy ions are seldom observed in devices submitted to heavy-ion tests, indicating that the FG transistor is quite robust against radiation damage. This robustness is related to the readout protocol: no bit flip is measured if the variation of the ) is below that needed to FG MOSFET threshold voltage ( exit the 0/1 programming limits. Nevertheless, ionizing radiation definitely produces some ) of the FG MOSFETs. variation of the threshold voltage ( This problem was addressed by Snyder et al. [10], who identified three mechanisms responsible for large charge loss from FG. 1) Hole–electron pairs are generated by radiation inside the tunnel oxide or the oxide–nitride–oxide interpoly oxide Manuscript received July 17, 2001. This work was supported in part by MURST-Italy. G. Cellere and A. Paccagnella are with DEI, Università di Padova, 35100 Padova, Italy (e-mail: [email protected]; [email protected]). P. Pellati and A. Chimenton are with Università di Ferrara, 44100 Ferrara, Italy (e-mail: [email protected]; [email protected]). J. Wyss is with Facoltà di Ingegneria, Università di Cassino, 03043 Cassino (FR), Italy (e-mail: [email protected]). A. Modelli is with ST Microelectronics, Agrate Brianza (MI), Italy (e-mail: [email protected]). L. Larcher is with DSI, Università di Modena e Reggio Emilia, 41100 Modena, Italy (e-mail: [email protected]). Publisher Item Identifier S 0018-9499(01)10715-X.

(ONO) separating the FG from the Si substrate and control gate (CG), respectively (a schematic view of the FG MOSFET is given in the Appendix). Some of the holes surviving the prompt recombination phase can drift into the FG, where they recombine part of stored electrons, thus reducing the FG negative charge. The FG negative charge itself produces the oxide fields driving the holes toward the FG, while radiation-generated electrons are driven toward bulk Si or CG. 2) Other radiation-generated holes can be trapped in the . tunnel oxide, thus reducing the FG MOSFET 3) Electrons stored in the FG can gain energy from the ionizing radiation and be emitted over the oxide barrier height toward the CG or the Si substrate. Noticeably, these results have been observed on EEPROM cells proper of the 1989 technology, featuring thick oxides (40 and 47.5 nm) around the FG. variation of FG cells, a single heavy ion Concerning the may lead to large effects. However, no data are available in literature to evaluate the cell threshold distribution after exposure to heavy ions. Besides, long-term reliability problems may arise on the irradiated cells, which are not immediately detectable by measurements after heavy ion tests. The first problem comes from the reduced electron charge present on the FG after irradiation, which may endanger the 10-y retention limit of the memory under normal operation. Also, transitory and dc leakage current paths may develop along the ion-induced damage across the tunnel oxide, further reducing the stored FG charge. These problems may increase with shrinking the cell size and the FG stored charge. Moreover, program/erase (P/E) cycles are well known to stress the tunnel oxide by producing local oxide defects [11], which can be enhanced by the ion-induced damage. Further, one of the peculiar reliability concerns of the FG memories resides in the threshold voltage distribution, which may show undesired overerased and overprogrammed tails external to the characteristic steep distribution in a Weibull plot [12]. At present, it is not clear if and how these reliability problems could be enhanced by heavy-ion irradiation. The aim of this paper is to investigate some of these issues. II. DEVICES AND EXPERIMENTAL DETAILS We have studied specially designed 4-Mbit flash arrays manufactured by ST Microelectronics, Agrate Brianza, Italy, whose characteristics are listed in Table I. Each device is organized in eight memory sectors, each featuring 256 rows and 2048 columns, for a global count of 2048 rows and 2048 columns. The array is organized as a NOR matrix, where programming is done at cell level using channel hot electrons (CHEs) at drain,

0018–9499/01$10.00 © 2001 IEEE

CELLERE et al.: RADIATION EFFECTS ON FLOATING-GATE MEMORY CELLS

TABLE I USED DEVICES CHARACTERISTICS

while erasing is done at block level using Fowler–Nordheim tunneling at source. The on-chip control circuitry is limited to sense amplifiers (one for each data bit) and row/column decoders. The of the memory cell in the programmed threshold voltage V and in the erased “0” state “1” state should be V. was read and programmed at difIn each memory cell, ferent nominal conditions by using a specially designed research instrument for flash evaluation (RIFLE) [13]. In particular, we at a fixed drain current before and after irradimeasured ation by using the on-chip sense amplifier with a read error of about 40 mV. RIFLE allows to perform direct measurements of the cell characteristics rather than rely on the circuit output [14], [15]. Devices were irradiated by using the Tandem Van De Graaff accelerator at Laboratori Nazionali di Legnaro, INFN, and Università di Padova, Italy. We used 276.7-MeV Iodine ions with an effective linear energy transfer (LET) of 62 MeV/cm mg and 134.28-MeV silicon ions with 9.27 MeV/cm mg effective LET. Ion flux was about 40 000 ions/cm for all devices, fluences, and ions, and was measured by using an array of diodes surrounding the device under test. Fluences were determined by integrating the measured ion flux. All device terminals were grounded during irradiation, corresponding to a possible bias condition of the memory in space applications when no operation is performed on the device. We also used a flash cell model recently developed (see the Appendix) to accurately evaluate the oxide field and the trapped . Through this model, we evaluated charge as a function of that the charge needed to produce 100-mV threshold shift is about 350 electrons. This number should be considered in the following, when radiation-induced threshold voltage shifts will be extensively treated. III. RADIATION EFFECTS ON THRESHOLD VOLTAGE A chip with all FGs programmed at “1” was irradiated with variation over a small 2 10 I/cm s. Fig. 1 shows the part of the irradiated array. Each square represents a single cell shift: the darker the featuring a gray level related to the square, the larger the threshold variation, which is always negdecreases). The struck cells appear randomly disative ( tributed across the chip surface. Considering the area occupied by FGs, it is easy to evaluate that about 3% of the cells should be hit after 2 10 I/cm . The corresponding total dose was 20 krad (SiO ), possibly dangerous for the control circuitry [9]. For this reason, we shielded

2223

the control circuitry by using Al foils. The Al shield left exposed to irradiation two zones containing a different number of cells at the opposite sides of the chip, the first one on top (Sector 1) and the second one toward the bottom (Sector 6). After irradiation, in Sector 1 we observed only one cell with V out of the 187 200 cells exposed to V and radiation. Two cells in Sector 6 featured , respectively, out of 249 600 cells exposed to the ion beam. These three cells are read as “erased” by the sense amplifier, corresponding to the irreversible corruption of stored information. In the next section, we will see that many more cells could actually fail under test. both on isolated cells and on cells We measured large whose neighbors were hit. The largest shifts can derive from multiple hits on a single FG. An evaluation of the multiple hit probabilities is possible from classical probability theory [16] but not feasible due to the huge numbers involved in this case. Hence, we have evaluated the probability of double and triple ion hits on a single cell by computer simulations. We designed a two-dimensional pattern reproducing the FG array: each ion was of its identified by the randomly generated coordinates strike position on the array surface. We simulated different ion fluences and repeated 10 times the simulation for each fluence. Results are reported in Fig. 2. At the highest ion fluence used in this work (i.e., 2 10 ions/cm ), the number of triple hits is almost negligible, but the number of double hits can be as large as 1.5% of the total number of hit cells. In Sector 1 (Sector 6), we got 7907 (11 812) hit cells with mV (i.e., the minimum detectable variation) over a total of 187 200 (249 600) cells exposed to radiation. Based on our previous evaluation, about 120 (180) cells were hit twice. On the average, about 4.5% of cells have been hit by one or more ions, in contrast to the 3% expected. This means that ions hitting “close enough” (but not exactly over) the FG can still produce threshold shift. In this case, part of the radiation-induced charge can be collected by the fringing field around the FG. Extending by just 50 nm the FG linear dimension will increase the hit cell percentage to 4%, not far from our experimental results. IV. THRESHOLD VOLTAGE DISTRIBUTION Even though just three cells were observed to flip from “1” to “0” in the previous section after irradiation, many more suffered the threshold decrease illustrated in Fig. 3, showing the distribution in a Weibull plot before and after irradiation. is distributed between 6.4 and 7.4 V, due to The prerad cell-to-cell differences and to the programming protocol, which V but allows for some stops the CHE injection when distribution shape is typical for overprogramming. This the memories we studied. Results from Sector 1 and Sector 6 nicely overlap even after irradiation, indicating good uniformity of different parts of the memory chip. After irradiation, a long tail appears at low voltages, while the upper part of the curve is not appreciably modified for V (Fig. 3). This tail represents cells that suffered a large shift after irradiation, but not necessarily all hit cells, since some hit cells may have preserved a high . The tail can be

2224

IEEE TRANSACTIONS ON NUCLEAR SCIENCE, VOL. 48, NO. 6, DECEMBER 2001

2

Fig. 1. Spatial distribution of threshold voltage for a small subset of cells from Sector 6 of a chip after 2 10 Iodine ions/cm . The gray scale (right) indicates the amount of V ; the two cells indicated by arrows have V V. Each square represents an FG.

1

1

>2

Fig. 2. Number of (simulated) double and triple hits as a function of fluence.

divided in three sections: the first, almost flat, is between 6.7 and 6 V; the second has the steepest slope and spans between 5.6 and 6 V; in the third, below 5.6 V, the slope decreases again. For clarity, an alternative representation of the same data is supplied in Fig. 4, where we show the probability density on the -axis. V and Noticeably, in about 1.8% of all cells V the threshold lies between “0” and “1,” i.e., between and . How many of these cells will actually be detected as flipped bits in commercial devices will depend on the circuit, including the sense amplifier and the reference current source, temperature, noise, etc. distribution shows Quite surprisingly, in irradiated cells, a secondary peak around 6 V (see Fig. 4) at the border between the first and second region identified in Fig. 3. This effect cannot be related to double hits on the same FG, which affect only 300 cells (as deduced from Fig. 2), that is, much less than the number of cells forming the peak at 6 V. For converse, we may assume distribution for simplicity that the lowermost part of the comes entirely from double-hit cells. In fact, 180 double-hit cells over 249 600 correspond to 0.07% of all the cells exposed V in Fig. 3, to the ion beam, and 0.07% identifies that is, the upper limit of the third part of the tail. The secondary

Fig. 3. Cumulative distributions of threshold voltages for exposed areas of Sector 1 and Sector 6 of the device represented in Fig. 1 (after 2 10 iodine ions/cm ).

2

Fig. 4. Probability density of threshold voltages for Sector 6 before and after irradiation with 2 10 iodine ion/cm .

2

peak around 6 V in threshold distributions will be further discussed in Sections V and VI. We also irradiated a different device containing sectors programmed at “0” and at “1” levels with 2 10 Si/cm . Results on “1” programmed sectors will be shown in the next sec-

CELLERE et al.: RADIATION EFFECTS ON FLOATING-GATE MEMORY CELLS

Fig. 5. Cumulative distributions of threshold voltages for irradiated sectors distributions, after growing silicon fluences. Inset: with identical initial V Percentage of cells in the tail as a function of fluence.

tion. In the erased sectors (“0”), practically no difference can be observed between the fresh and the irradiated devices. In V, less than 5000 electrons these cells, featuring are trapped in the FG, and no FG net charge is present for V. Results are different from those previously obincreased in erased tained on EEPROM cells [10], where cells after gamma irradiation. Differences must be attributed to shift active in the the different mechanisms leading to the two experiments, as will be clarified in Section VI. In addition, in that case the “0” condition corresponded to a FG net positive charge. V. ION FLUENCE EFFECTS We have investigated the effect of ion fluence on the distribution, only in the case of programmed cells, as shown in Fig. 5. Three devices were subjected to irradiation with 2 10 , distribu10 , and 2 10 Si/cm , respectively. Pre-rad tions overlap each other. As expected, after irradiation, tails at grow with the Si fluence, qualitatively reproducing the low three-slope behavior already observed in Fig. 3. For converse, a secondary peak such as that shown in Fig. 4 is not detectable when using low LET ions such as Si (Fig. 5), due to the smaller . These results are in agreement with the Si ion-induced smaller charge released by Si in comparison with I. The number (that is, the probability at which the tail of cells in the tail joins the pre-rad distribution) is also shown in inset of Fig. 5. has a linear dependence on the fluence. Moreover, the number of cells forming the radiation induced tail is very similar for Si and I ions for the same ion fluence. Hence, the LET affects the threshold voltage variation but not the number of cells modified by irradiation, for the same programming with conditions of the FG. VI. THRESHOLD VOLTAGE EFFECTS ON THE CHARGE LOSS plays a fundamental role in determining the radiation response of irradiated FGs. As previously observed [10], [17], [18], the amount of radiation-induced electron and holes collected by the FG depends on the field intensity in the dielectric

2225

Fig. 6. Cumulative distributions of threshold voltages for irradiated sectors after 10 silicon ions/cm .

layers (tunnel and ONO) sandwiching the FG. Different pro) can be selected for the memory gramming levels (i.e., array, corresponding to different negative charges in the FG and, consequently, different fields in the tunnel oxide and ONO around the FG. For this purpose, we irradiated a device with secvalues. The voltage wavetors programmed at different forms during programming were kept the same for all sectors, and only the minimum threshold voltage for the “1” state was changed. This way, a second or eventually a third program pulse was applied to cells with threshold voltage still too low. This procedure is useful to simulate real threshold distributions in commercial parts, where part-to-part variation, as well as the history of the device, leads to large threshold voltage dispersion [12]. distributions before and after 10 Fig. 6 shows the Si/cm irradiation of two sectors programmed at different values (measured at 63% cumulative probability), namely, V and V. The 63% cumulative probability value identifies the average value of in a Weibull distribution [19], but it is a useful marker data not following a Weibull distribution [20]. even for the distribution tail is longer for V, The as expected, and also involves many more cells (1.8% against V. This does not mean that 0.2%) than , fewer cells are struck by ions for low programmed shift which is physically meaningless, but only that the decreases (and fewer cells show large shift) when the electric field in the tunnel and ONO is reduced. As just observed in Fig. 6, the distribution probabilities are useful for reliability studies plotted as a function of variation of the single cell hit by but say nothing on the ion(s). To clarify this point, we have investigated first of all the statistical behavior of the threshold voltage shift evaluated on each cell after irradiation, as illustrated in Fig. 7. In this case, data are reported for two sectors of a device after 2 10 Si/cm . The cumulative probability range of the -axis is compressed at high values, as most of the cells show variation after irradiation. Further, on the average, no cells with lower threshold before irradiation [Sector 1, V] undergo a smaller when compared V). with cells from Sector 3 (

2226

IEEE TRANSACTIONS ON NUCLEAR SCIENCE, VOL. 48, NO. 6, DECEMBER 2001

the most probable charge loss peaks at 1200 lost electrons for curve , 850 for , and 500 for . For the other curves, the peak is undetectable and compressed on the -axis. Hence, the largest charge losses occur from those FGs trapping the largest electron charge. Moreover, on the average, group loses electrons more than group , and group loses electrons more than group . The falling edge of curve is again 350 electrons on the left of the falling edge of curve . electrons is exactly the pre-rad threshold difference between two adjacent groups. VII. A NEW CHARGE LOSS MECHANISM

Fig. 7. Cumulative distributions of threshold voltage shifts for irradiated sectors, after 2 10 silicon ions/cm .

2

X

Fig. 8. Probability densities of the number of lost electrons (lower -axis) and (upper -axis) for cells with different initial stored charge. Cells are taken from Sector 4 of the device depicted in Fig. 6.

1V

X

Even though the impact on the threshold shift is qualitatively illustrated in Fig. 7, still the comparison is quite rough, being made among all cells featuring a large dispersion over a range 1 V wide. Having confirmed that the , we decided to perform threshold shift is dependent on comparisons within groups of homogenous cells with similar threshold values. For this reason, we partitioned the memory into several groups of cells, depending on their prerad value. For instance, the values of a device were divided ), each one 100 in five adjacent ranges, named ( mV wide: spans between 7.0 and 7.1 V, between 7.2 and 7.3 V, and so on. The five cell groups have been selected for statistical purposes, as they contain a large number of cells, . By which is not the case for much smaller or higher using our FG cell model (see the Appendix), we have evaluated value: the minimum the FG stored charge for each V corresponds to 15 500 electrons in the FG, V to 17 250. while the maximum We have drawn in Fig. 8 the normalized probability density for the five cell groups after of electron charge loss and irradiation with 10 Si/cm . All curves increase when the charge loss approaches zero, indicating that small charge losses from the FG are reasonably probable due to the ion hit. In addition,

The impact of on the radiation-induced charge collected at the FG is strongly dependent on the fractional yield (i.e., the number of holes surviving the prompt recombination), which increases with the oxide field [17], [21], controlled in turn by the charge stored on the FG. However, the results of Fig. 8 cannot be naively explained by claiming that more holes reach the FG when the electron charge on the FG is larger. In fact, when considering, for instance, a single Si ion hitting the FG, it generates approximately 1200 and 2100 electron–hole pairs in the tunnel and ONO oxides, respectively (Si generates 1.17 10 pairs/cm) [21]. Moreover, we can evaluate the field in the oxides using the model described in the Appendix. For a FG V, the oxide field in the tunnel oxide is with MV/cm, while in the SiO layer of the ONO MV/cm. Under these conditions, less than 5% of the holes generated in the tunnel oxide and less than 1% of those generated in ONO will survive prompt recombination [17], driving into the FG less than 100 holes from both the tunnel oxide and ONO sides. These holes may well account for the behavior of the close to zero), curves in Fig. 8 for small lost charges (i.e., where high probability values are found. However, the few radiation-induced holes cannot produce the large FG charge loss (up to 1500 electrons) observed in Figs. 3, 4, and 8, just by recombination with FG electrons. According to [10], the large FG charge loss in principle could be attributed to another physical mechanism, i.e., electrons emitted from the FG over the Si/oxide barrier. Owing to the high poly doping, the effective FG n –poly/oxide barrier height is not modified by the FG negative charge. Therefore, almost the same electron charge should be emitted from the FG ), independently of the FG charge, in all cases of Fig. 8 ( and then swept away by the negative oxide fields. Therefore, ) should appear, in contrast no difference between curves ( with experimental results (Figs. 6–8). For this reason, we believe that a fourth mechanism, in addition to those listed in the Introduction, must be claimed to account for the FG charge loss, based on an enhanced leakage from the FG. Some of the holes, generated close enough to the FG, are soon collected at the FG during or after the thermalization phase. Instead, other holes can be trapped and re-emitted several times from defects in the oxide layer before reaching the FG. These residual holes are accumulated along the track of the ion, whose radius is a few nanometers, and can produce an enhanced leakage from the FG along the ion track in two possible ways.

CELLERE et al.: RADIATION EFFECTS ON FLOATING-GATE MEMORY CELLS

1) Being positively charged, they effectively reduce the oxide barrier between FG and Si, strongly enhancing the electron tunneling from FG into Si along the ion track. In this case, we may speak of positive charge assisted leakage current (PALC) (incidentally, how the conduction through the ONO layer may be enhanced by trapped holes is a problem we do not face here). Tunneling across the tunnel oxide (10 nm thick) will occur until the FG charge decreases and conduction is stopped by the corresponding decrease of the tunnel oxide field. 2) After capturing an electron, the trapped hole can be converted into a neutral trap [22]. If two or more neutral traps are spatially aligned, they give rise to a conductive path reproducing and multiplying that found in the case of radiation-induced leakage current (RILC) [23]. A single neutral trap typically produces a negligible RILC in oxides thicker than 8 nm. The first mechanism will control the FG discharge immediately after the formation of the conductive path, and it is much more effective than the second in reducing the oxide barrier and enhancing the leakage current. The second mechanism is activated only when the trapped holes are neutralized, and it may play only a marginal role in the FG discharge. Remarkably, our results qualitatively reproduce one of the most dangerous retention problems of Flash memories, well known as “erratic bit” or “anomalous SILC” phenomenon [24]. In that case, after program/erase cycles, an anomalous charge loss from a programmed FG is observed on cells of the distribution tail, leading to bit flip. This problem has been attributed to local coordination of positive charge [24] (that is, a cluster of positive charges, which act together in lowering the oxide barrier, is temporarily formed in a small oxide region), which enhances the leakage current across the tunnel oxide. This fact further supports the PALC as the FG discharge mechanism. Moreover, in both mechanisms 1) and 2), the tunneling conduction is self-limiting, since it stops when a tunneling turnoff is eventually reached on the FG. From the results voltage of Fig. 8, the charge losses measured at the curve peaks (500 electrons for , 850 for , and 1200 for ) are coherent with an around 7 V. This result justifies the different tails average (63%) values: only in observed in Fig. 6 depending on V) the tunnel oxide the case of Sector 2 ( field is high enough to support PALC (RILC), which drives large . The value is not a modification (up to 0.5 V) of universal constant for a cell but depends on the ion LET. The V for I irsecondary peak in Fig. 4 indicates that MV/cm. Lower radiation, corresponding to values correspond to more conductive leaky paths turned off at lower FG voltages, as expected in the case of ions with higher LET. The PALC (RILC) mechanism was not previously reported on EEPROMs [10] due to radiation used and memory structure. In that case, memories with tunnel and interpoly oxides 47.5 and 40 nm thick, respectively, were exposed to Co60 gamma rays. Holes produced by gamma rays are unable to form localized conductive paths across 40-nm-thick oxides. In fact, each Compton electron (responsible for the Co60 gamma ionizing damage) can generate at most only one cluster of three e–h pairs

2227

across those oxides [21]. The combined effect of thicker oxides (four times thicker than in our devices) and few holes (10 times less) inhibits the formation of leaky paths responsible for the PALC (RILC) related effect measured on our devices. To evaluate the long-term impact of ion hits on FG, we have measured an irradiated device after three months of room-temperature storage with all terminals grounded. After this anneal period, the number of bits in the tail did not change, but the number of bits with very low (below 4.5 V) slightly increased, indicating some leaky path still active in the FG with the . Moreover, hit cells may actually degrade faster largest if subjected W/E cycles, as in the case of “erratic erase,” but we have not yet performed these tests on a statistically significant basis. VIII. CONCLUSION We have addressed the problem of threshold voltage variation in flash memory FG cells after heavy-ion irradiation by using specially designed array structures and test instruments. decreases when the In the case of “1” programmed cells, FG is hit by an ion, corresponding to a reduction of the FG stored negative charge. On the contrary, in the case of “0” erased devices, no threshold variation is observed. Irradiated cells are uniformly distributed across the devices. After irradiation, low tails appear in distributions, growing with ion LET and fluence. In particular, high LET ions, such as iodine used in this work, can produce a bit flip. The effect of double ion hits on very large threshold shifts, as in the case of the observed bit flips, can be quantified in about 1.5% of the hit FGs after 2 10 ions/cm . Even when not producing a bit flip, we have observed many large variations of the FG stored charge, which cannot be simply attributed to hole injection into the FG, as the ion-induced hole number too small. We have proposed a new mechanism, based on the formation of a leakage path across the tunnel oxide along the ion track. This leakage current can be produced by the local excess of positive charge produced by a single ion, temporarily lowering the tunnel oxide barrier (PALC) and enhancing the tunneling current. The leakage current appears self-limiting, leading to the formation of a secondary peak in distribution tail. Excess leakage affects the postirradiation , and depends on cells with high FG stored charge, i.e., high cells the ion LET value. For this reason, controlling high (not just low cells, as usually done) is essential to guarantee flash memory reliability in a radiation environment. Ion effects may be a severe threat to flash memories in space applications distriwhen considering multilevel cells, where very tight variations can butions are needed and even relatively small lead to bit errors. Many aspects of the ion effects on flash memory cells remain to be addressed, such as the impact of FG size, readout protocol, multibit programming, voltage supply limits, tunnel oxide technology, program/erase cycles, and retention characteristics, just to name a few important issues related to scaling. Further, the impact of ion-induced transient leakage from isolated nodes, as PALC is for FG, is a potentially dangerous phenomenon in aggressive CMOS technologies. In these devices, temporary accumulation of positive charge inside the gate oxide (along an ion

2228

Fig. 9. cells.

IEEE TRANSACTIONS ON NUCLEAR SCIENCE, VOL. 48, NO. 6, DECEMBER 2001

Schematic illustration of the model we developed for flash memory

track) may reduce the oxide barrier, thus enhancing the leakage current from the MOS gate and discharging the gate capacitance. APPENDIX A NEW MODEL FOR FLASH DEVICES All the models of FG devices proposed in the past are based on a lumped element description of the FG cell, where the FG node is connected to the source, drain, body, and control gate through a capacitive network. In these models, the FG voltage, which is fundamental for the correct operation of the cell, is calculated from capacitive coupling coefficients [1]. However, these coefficients are difficult to evaluate (the floating gate is electrically isolated and cannot be directly accessed) and depend on the applied terminal voltages [27], [28]. The new SPICE-like model of the FG cell we developed is formed by three elements (see Fig. 9): a MOS transistor whose gate G is the FG of the cell; a capacitor connected between FG , between and CG; and a voltage-controlled voltage source ground and FG, needed to overcome the problem of simulating a contains the C code implecapacitive net in dc conditions. menting the FG voltage calculation procedure, which is the core is equal to the of our model. The charge on the MOS gate charge of the capacitor between the FG and CG, plus the charge forced in/out of the floating gate during cell program/erase op, and has been evaluated by using the charge equaerations tions of the compact MOS transistor model Philips MM91 [29]. With the model proposed, the dc behavior of EEPROM and flash memory cells manufactured in the used technology can be excellently simulated [29] under every bias combination (above and below threshold, with and without substrate bias), without the need of any free parameter to adjust fitting quality. A better model description can be found in [29]. REFERENCES [1] P. Pavan, R. Bez, P. Olivo, and E. Zanoni, “Flash memory cells—An overview,” Proc. IEEE, vol. 85, pp. 1248–1271, Aug. 1997. [2] S. K. Lai, V. K. Dahm, and D. Guterman, “Comparison and trends in today’s dominant E technologies,” in IEDM Tech. Dig., 1986, pp. 580–582. [3] D. N. Nguyen, C. I. Lee, and A. H. Johnston, “Total ionizing dose effects on Flash memories,” in Proc. 1998 IEEE Radiation Effect Data Workshop, pp. 100–103.

1See

http://www-us.semiconductors.philips.com/Philips\_Models/

[4] D. N. Nguyen, S. M. Guertin, G. M. Swift, and A. H. Johnston, “Radiation effects on advanced Flash memories,” IEEE Trans. Nucl. Sci., vol. 46, pp. 1744–1750, Dec. 1999. [5] S. Bertazzoni, G. C. Cardarilli, M. Salmeri, A. Salsano, C. Bacis, G. Golla, M. L. Longo, F. Desideri, D. Di Giovenale, G. C. Grande, S. Sperandei, M. Ricci, and P. G. Picozza, “TID on 16 MbFlash memories in radiation environment,” in Proc. RADECS2000, pp. 229–232. [6] D. R. Roth, J. D. Kinninson, B. G. Karlhuff, L. R. Lander, G. S. Bognaski, K. Chao, and G. M. Swift, “SEU and TID testing of the Samsung 128 Mbit and the Toshiba 256 Mbit Flash memory,” in Proc. 2000 IEEE Radiation Effect Data Workshop, pp. 96–99. [7] H. R. Schwartz, D. K. Nichols, and A. H. Johnston, “Single-event upset in Flash memories,” IEEE Trans. Nucl. Sci., vol. 44, pp. 2315–2324, Dec. 1997. [8] T. Huang, F. C. Jong, T. S. Chao, H. C. Lin, L. Y. Leu, K. Young, C. H. Lin, and K. Y. Chiu, “Improving radiation hardness of EEPROM/Flash cell by N2O annealing,” IEEE Electron Device Lett., vol. 19, pp. 256–258, July 1998. [9] D. Krawzsenek, P. Hsu, H. Anthony, and C. Land, “Single event effects and total ionizing dose results of a low voltage EEPROM,” in Proc. 2000 IEEE Radiation Effect Data Workshop, pp. 64–67. [10] E. S. Snyder, P. J. McWhirter, T. A. Dellin, and J. D. Sweetman, “Radiation response of floating gate EEPROM memory cells,” IEEE Trans. Nucl. Sci., vol. 36, pp. 2131–2139, Dec. 1989. [11] A. Brand, K. Wu, S. Pan, and D. Chin, “Novel read disturb failure mechanism induced by FLASH cycling,” in Proc. 31st Annu. Int. Reliability Physics Symp., 1993, pp. 127–132. [12] D. Esseni and B. Riccò, “On the origin of the dispersion of erased threshold voltages in the Flash EEPROM memory cells,” IEEE Trans. Electron Devices, vol. 47, pp. 1120–1123, May 2000. [13] P. Pellati and P. Olivo, “Automated test equipment for research on nonvolatile memories,” , submitted for publication. [14] L. Z. Scheick, P. J. McNulty, and D. R. Roth, “Dosimetry based on the erasure of floating gates in the natural radiation environments in space,” IEEE Trans. Nucl. Sci., vol. 45, pp. 2681–2688, Dec. 1998. [15] L. Z. Scheick, P. J. McNulty, D. R. Roth, M. G. Davis, and B. E. Mason, “Measurement of dose with individual FAMOS transistors,” IEEE Trans. Nucl. Sci., vol. 46, pp. 1751–1756, Dec. 1999. [16] W. Feller, An Introduction to Probability Theory and Its Applications. New York: Wiley, 1950. [17] T. R. Oldham, “Recombination along the tracks of heavy charged particles in SiO films,” J. Appl. Phys., vol. 57, p. 2695, 1985. [18] H. E. Boesch and J. M. McGarrity, “Charge yield and dose effects on MOS capacitors at 80K,” IEEE Trans. Nucl. Sci., vol. 23, p. 1520, Dec. 1976. [19] K. P. Cheung, Plasma Charging Damage. Berlin, Germany: SpringerVerlag, Nov. 2000. [20] P. J. McNulty, L. Z. Scheick, D. R. Roth, M. G. Davis, and M. R. S. Tortora, “First failure predictions for EPROM’s of the type flown the MPTB satellite,” IEEE Trans. Nucl. Sci., vol. 47, pp. 2237–2243, Dec. 2000. [21] T. P. Ma and P. V. Dressendorfer, Ionizing Radiation Effects in MOS Devices and Circuits. New York: Wiley, 1989. [22] M. Walters and A. Reisman, “Radiation-induced neutral electron trap generation in electrically biased insulated gate field effect transistor gate insulators,” J. Electronchem. Soc., vol. 138, pp. 2756–2762, Sept. 1991. [23] M. Ceschia, A. Paccagnella, M. Turrini, A. Candelori, G. Ghidini, and J. Wyss, “Heavy ions irradiation of thin gate oxides,” IEEE Trans. Nucl. Sci., vol. 47, pp. 2648–2655, Dec. 2000. [24] P. Cappelletti, P. Cappelletti, C. Golla, P. Olivo, and E. Zanoni, Flash Memories. Norwell, MA: Kluwer Academic, 1999. [25] B. Eitan, R. Kazerounian, and A. Roy, “Multilevel Flash cells and their trade-offs,” in Proc. IEDM Tech. Dig., 1996, pp. 169–172. [26] B. Riccò, G. Torelli, M. Lanzoni, A. Mastandretta, H. E. Maes, D. Montanari, and A. Modelli, “Nonvolatile multilevel memories for digital applications,” Proc. IEEE, vol. 86, pp. 2399–2421, Dec. 1998. [27] R. Duane, A. Concannon, P. O’Sullivan, and A. Mathewson, “Advanced numerical modeling of nonvolatile memory cells,” in Proc. ESSDERC, 1998, pp. 304–307. [28] L. Larcher, P. Pavan, L. Albani, and T. Ghilardi, “Bias and W/L dependence of capacitive coupling coefficients in floating gate memory cells,” IEEE Trans. Electron Devices, to be published. [29] L. Larcher, P. Pavan, F. Gattel, L. Albani, and A. Marmiroli, “A new compact model of a Floating Gate nonvolatile memory cell,” in Proc. 2001 IEEE ICMSMSSA, SC.