Tools / Single events / NAND bit flips

How many bits will this NAND flash lose to ions and dose before the next refresh?

How many programmed cells of a NAND flash memory the radiation environment deprograms between one program (or refresh) and the next read: the floating-gate prompt charge-loss model of JPL Publication 16-9, with its published fits to the Samsung 8 Gb SLC part, carried over CREME96 ion spectra, the dose at the die and the measured proton cross sections, then through the error-correcting code.

v1.0 · Nov 2026JPL Publication 16-9 model, TNS 2017 fitsCREME96 spectra, dose at the die, measured proton cross sectionsruns in your browserMechanism in 3DHelp & Reference
Published fits to one part type: the Samsung 8 Gb SLC NAND (K9F8G08U0M, 2010 date code)
Device
b₁ and b₂ are the LET (in SiO₂) that removes the mean critical charge in one strong or one weak hit; σ are the per-gate cross sections of the two hit types; k sets the gate-to-gate spread of the critical charge (larger is tighter); DEF multiplies the dose before it enters the model (the published value carries an empirical correction for dose effects in the sensing circuits).
All dies see the same environment and dose.
Only charged cells can lose charge. The tests used an all-zero (all programmed) pattern; random data would be 0.5.
Mission
Time from a program or refresh to the read: the fluence and dose accumulate over it. A refresh resets the clock; the chart shows every interval length at once.
At the die, from the mission's shielding analysis (SHIELDOSE or SPENVIS). Not computed here. Zero leaves the dose term out.
Each adds seven days of the CREME96 worst-week ion and proton flux at GEO.
Each adds one day of the CREME96 worst-day flux at GEO.
Multiplies the omnidirectional fluence 4π·I. 1 is the convention of the publication's space routine as far as it could be reconstructed; 0.5 counts each crossing of a thin plate once. Path length through the gate at oblique incidence is not modeled.
The tables are in silicon; the model wants SiO₂. TNS 2017 quotes 9.0/8.3 (Ar), 4.55/4.2 (Ne) and 54/51.5 (Xe).
Protons
Error correction
Default: a 528-byte sector, the correction granularity of SLC NAND of this generation. Set it to the controller's code.
0 for detection only.
Bit errors are placed at random over the codewords (binomial tail). Several cells deprogrammed by one ion within one codeword are not represented; the EDAC tool treats the codes in more detail.

Deprogrammed cells against the storage interval

Expected bit errors per die at the read, for every interval length from one day to ten years, in the environment above. Dose is scaled with the interval from the value entered; events are added at the counts entered whatever the interval. The curves use the single-hit form of the model, which agrees with the full form to better than one percent wherever hits per gate are rare (the tiles use the full form).

Where the errors come from

Cross section against LET for this parameter set

The single-hit (SEU-like) cross section per programmed gate, at zero dose, after the dose entered, and after 150 krad(Si), the dose at which the conservative set's DEF was fitted. The two points are the post-150-krad measurements quoted in TNS 2017: the Ar point the DEF was fitted to, and the Ne point taken at a fluence too large for a cross section to be defined, which the fit excluded and which the model does not reproduce.

Rates by environment

Single-hit deprogramming rate per programmed gate per day in each built-in environment, with the parameter set and dose state above, and the strong hits (direct gate hits) per gate per day that drive it. Event rows are rates during the event.

Ground test: counts after a beam exposure

The full model at one LET: the expected number of deprogrammed cells in a device of the cells entered above, all programmed, after the fluence given, with the prior dose given. This is the calculation that reproduces the large-fluence curves of TNS 2017 Fig. 6 when the Device A or Device B set is chosen with its beam.

Conclusions: what the number rests on

"May be low" means the true number is likely above the prediction; "may be high" the reverse; "either way" the direction depends on the device or the environment.

The model and the trades it makes, in plain language

The model

A programmed floating gate holds charge; an ion that crosses the gate itself (a strong interaction, cross section σS,sat per gate) removes a share L/b₁ of the mean critical charge, L being its LET in SiO₂; an ion that passes near the gate (a weak interaction, σW,sat) removes L/b₂ by charging the surrounding oxide, which compensates part of the stored charge. Dose removes μ·DEF·D·(σS/b₁ + σW/b₂) deterministically, with μ = 6.2415 × 10⁴ MeV per mg per rad. The losses add; the gate reads as deprogrammed once they exceed its critical loss, which varies from gate to gate as a Weibull distribution of shape k. With Poisson hit counts, for one LET:

p = 1 − Σm,n P(m; HσS) P(n; HσW) exp[−(m L/b₁ + n L/b₂ + xD)k],   xD = μ DEF D (σS/b₁ + σW/b₂)

Over a spectrum the summed loss is a compound Poisson variable; the page evaluates its distribution exactly on a loss grid (Panjer recursion), so many light ions, a few heavy ones and the dose shift are all in one calculation. Where hits per gate are rare, this reduces to fluence times the single-hit cross section plus the dose-only fraction, and the page shows both.

Environment

  • CREME96 tables of the site's SEE rate tool. Ions Z = 2 to 92 behind 100 mil of aluminum, solar quiet GCR, quiet magnetosphere unless stated; events at GEO without geomagnetic attenuation, a bound in low orbits. LET in silicon, converted to SiO₂ by one factor.
  • Omnidirectional fluence times the normal-incidence cross section. 4π·I per cm², times the angular factor. The model defines losses at normal incidence; the longer path of an oblique crossing, which raises the loss and so the sensitivity to lighter ions, is not modeled. A measurement on a Samsung 4 x 8 Gbit SLC stack found the omnidirectional cross section about 2.1 times the normal-incidence one at LET 10.1, with up to 80 percent of the errors at grazing incidence in multi-bit clusters [10]; whether that stack shares this part's die is not confirmed.
  • Dose entered, not computed. The die dose comes from the mission's own shielding analysis. The model's conservatism is stated for electron doses up to 150 krad(Si) and not beyond, and only when programming precedes the dose.

Protons

  • A separate single-hit term. The measured all-zero-pattern cross sections at 105 and 480 MeV [3] are joined by the two-parameter Bendel shape and multiplied by the integral proton spectrum. The term carries no dose dependence and no accumulation; a proton's ionizing dose is in the dose entered, if the mission analysis included it, and its nuclear-recoil hits are here.

What is not traded

No parameter is invented: the sets are the published tables, the cross sections the published values, the spectra the site's CREME96 runs. The Conclusions card lists, for this run, which of these is measured, assumed or missing, and which way each one pushes the number.

References

  1. L. D. Edmonds, "A method for estimating the probability of floating gate prompt charge loss in a radiation environment," JPL Publication 16-9, Jet Propulsion Laboratory, March 2016.
  2. L. D. Edmonds, F. Irom and G. R. Allen, "Total ionizing dose influence on the single event effect sensitivity in Samsung 8Gb NAND flash memories," IEEE Transactions on Nuclear Science, vol. 64, no. 8, pp. 2046 to 2053, 2017 (the journal form of "Fitting a model to floating gate prompt charge loss test data for the Samsung 8 Gb SLC NAND flash memory," RADECS 2016).
  3. F. Irom, G. R. Allen and D. J. Sheldon, "Heavy ion, proton and electron single-event effect measurements of a commercial Samsung NAND flash memory," IEEE Radiation Effects Data Workshop, 2016.
  4. G. Cellere, A. Paccagnella, A. Visconti, M. Bonanomi and A. Candelori, "Transient conductive path induced by a single ion in 10 nm SiO₂ layers," IEEE Transactions on Nuclear Science, vol. 51, no. 6, pp. 3304 to 3311, 2004.
  5. S. M. Guertin, D. N. Nguyen and J. D. Patterson, "Microdose induced data loss on floating gate memories," IEEE Transactions on Nuclear Science, vol. 53, no. 6, pp. 3518 to 3524, 2006.
  6. G. Cellere, A. Paccagnella, A. Visconti, M. Bonanomi, P. Caprara and S. Lora, "A model for TID effects on floating gate memory cells," IEEE Transactions on Nuclear Science, vol. 51, no. 6, pp. 3753 to 3758, 2004.
  7. F. Irom, D. N. Nguyen, R. Harboe-Sorensen and A. Virtanen, "Evaluation of mechanisms in TID degradation and SEE susceptibility of single- and multi-level high density NAND flash memories," IEEE Transactions on Nuclear Science, vol. 58, no. 5, pp. 2477 to 2482, 2011.
  8. W. L. Bendel and E. L. Petersen, "Proton upsets in orbit," IEEE Transactions on Nuclear Science, vol. 30, no. 6, pp. 4481 to 4485, 1983 (the one-parameter form); the two-parameter form follows W. J. Stapor, J. P. Meyers, J. B. Langworthy and E. L. Petersen, "Two parameter Bendel model calculations for predicting proton induced upset," IEEE Transactions on Nuclear Science, vol. 37, no. 6, pp. 1966 to 1973, 1990.
  9. A. J. Tylka et al., "CREME96: a revision of the Cosmic Ray Effects on Micro-Electronics code," IEEE Transactions on Nuclear Science, vol. 44, no. 6, pp. 2150 to 2160, 1997 (the environment tables, run in September 2026 for the site's SEE rate tool).
  10. K. Grürmann, D. Walter, M. Herrmann, F. Gliem, H. Kettunen and V. Ferlet-Cavrois, "SEU and MBU angular dependence of Samsung and Micron 8-Gbit SLC NAND-flash memories under heavy-ion irradiation," IEEE Radiation Effects Data Workshop, 2011 (omnidirectional against normal-incidence cross section of a Samsung 4 x 8 Gbit SLC stack, K9WBG08U1M).
  11. H. H. Panjer, "Recursive evaluation of a family of compound distributions," ASTIN Bulletin, vol. 12, no. 1, pp. 22 to 26, 1981 (the recursion used for the spectrum).

This page is a planning aid. Its numbers come from the published model and fits, the environment chosen and the dose entered, carry the assumptions listed under Conclusions, and are not a qualification result for any part or mission. Nothing on it has been reviewed by any organization.

What one particle does to a cell, in three dimensions

The mechanism behind the model above. A programmed cell holds electrons on a floating gate, or in a nitride layer, that raise its threshold voltage above a read reference; for the nanoseconds an ion track lives, the dielectrics around that charge conduct along the track and part of the charge leaves. If the threshold falls below a reference the cell reads wrong, and an erased cell, with nothing above the reference to lose, does not care. Drag to turn the string and step through the strike, the conductive path, the loss and the read, by the three routes the model adds: one heavy ion, many light ions, and dose; the full page carries the explanation, the levels and Gray code, the model and the references.