What the tool computes from the model, the fits and the environment you choose, the assumptions it makes, how it was checked, and where each part of it comes from
A programmed NAND flash cell stores electrons on a floating gate. An ion that crosses the gate generally removes some of them at once (a prompt loss through the transient conductive path it leaves in the tunnel oxide [4]), an ion that passes nearby generally charges the surrounding oxide and compensates part of the stored charge [5], and ionizing dose does the same to every cell a little at a time [6]. Once a cell has lost more than its margin to the read reference it is read as erased: a bit flip without any upset of the sensing circuits. L. D. Edmonds put these three contributions into one statistical model, JPL Publication 16-9 [1]; Edmonds, Irom and Allen fitted it to beam and dose data for the Samsung 8 Gb SLC NAND flash memory, K9F8G08U0M [2]; Irom, Allen and Sheldon measured the proton cross sections of the same part [3]. The estimator takes that model and those fits, carries them over the site's CREME96 spectra, the dose at the die and the proton spectra, and returns the expected number of deprogrammed cells between a program (or refresh) and a read, then the expected number of codewords beyond the error-correcting code. The mechanism page shows the three routes to a flip in three dimensions.
The device: a parameter set (the published conservative set, one of the two single-part fits, the worked example of the publication, or five parameters of your own), the number of cells per die and the number of dies, and the fraction of the cells that are programmed (only charged cells can lose charge; the tests programmed every cell). The mission: a heavy-ion environment from the site's CREME96 tables or a pasted integral LET spectrum, the storage interval in days (the time from a program or refresh to the read, over which fluence and dose accumulate), the dose at the die over that interval in krad(Si), the number of worst-week and worst-day solar events in the interval, an angular factor and the LET conversion factor. The protons: a table, a pasted spectrum or none, with the two Bendel parameters prefilled from the measured points. The code: bits per codeword and bits corrected. The ground-test card takes an LET, a fluence and a prior dose of its own.
Losses are measured in units of the mean critical charge, the charge a typical cell must lose to read as erased. A strong interaction (a direct hit on the gate, cross section σS,sat per gate) removes L/b₁, where L is the ion's LET in SiO₂ and b₁ is a fitted constant; a weak interaction (a hit near the gate, σW,sat) removes L/b₂. The critical charge varies from gate to gate as a Weibull distribution with shape k and unit scale, so a gate that has lost x survives with probability exp(−xk). Dose D in rad(Si) is converted to LET times fluence by μ = 6.2415 × 10⁴ MeV per mg per rad (one rad is 100 erg per gram; the publication's rounded 6.25 × 10⁴ gives dose fractions 0.7 percent higher at k = 5) and multiplied by a dose factor DEF, so that it removes xD = μ DEF D (σS/b₁ + σW/b₂) from every gate. With Poisson numbers of strong and weak hits after a fluence H of one LET, the probability that a programmed gate reads as erased is
p = 1 − Σm,n P(m; HσS) P(n; HσW) exp[−(m L/b₁ + n L/b₂ + xD)k]
which is Eq. (50) of the publication with its dose terms, Eqs. (52) and (54), as reconstructed in the October 2026 review of the site's memory tools; the publication prints the single-hit limit of it as Eq. (62). The LET is in SiO₂; H is the fluence since the last program or refresh; losses are defined at normal incidence; p is per programmed gate. The strong-hit cross section is constant: every ion that crosses the gate is a strong hit, and its LET decides the loss, so the SEU-like cross section at small fluence is σS[1 − exp(−(L/b₁)k)] plus the same for the weak term, which rises as LET to the power k below b₁ and saturates at σS above it.
The parameter sets on the page, as printed in [2]:
| Set | b₁ | b₂ | σS,sat, cm² | σW,sat, cm² | k | DEF |
|---|---|---|---|---|---|---|
| Conservative, Table IV | 23.2 | 1155 | 1.80 × 10⁻¹⁰ | 1.00 × 10⁻⁸ | 5.057 | 1.65 |
| Device A (Ar, LET 9.0), Table III | 23.2 | 1155 | 6.00 × 10⁻¹¹ | 1.00 × 10⁻⁸ | 5.057 | not fitted |
| Device B (Ne, LET 4.55), Table III | 22.8 | 2271 | 7.74 × 10⁻¹¹ | 2.00 × 10⁻⁹ | 7.419 | not fitted |
| Pub 16-9 worked example | 49.0 | 3811 | 1.10 × 10⁻¹⁰ | 7.21 × 10⁻⁹ | 7.643 | 1.0 |
b₁ and b₂ are in MeV cm²/mg (LET in SiO₂). Each Table III row is a fit to one part exposed to one ion at fluences from 10⁶ to 10⁹ per cm²; the two parts, from the same date code, differ by a factor of two in b₂, five in σW,sat and by 7.4 against 5.1 in k, and no parameter uncertainty is published. The conservative set starts from Device A, raises σS,sat to the saturation cross section measured after 150 krad(Si), and sets DEF from one Ar point (one part, 5 × 10⁶ per cm²) after 150 krad(Si) of cobalt-60; the authors state it as conservative for electron doses up to 150 krad(Si), not beyond, and only when programming precedes the dose, since the DEF carries an empirical correction for dose effects in the sensing circuits. The page carries the Table IV DEF with the Device A and B sets, since none was fitted for them; the dose term reuses the ion parameters, so with the conservative set it removes 1.46 times more charge per rad than with Device A.
Each LET bin of the spectrum contributes strong hits at rate σSΦj with loss Lj/b₁ and weak hits at rate σWΦj with loss Lj/b₂, Φj being the fluence in the bin over the interval. The total loss is a compound Poisson variable, and the page evaluates its distribution exactly on a grid of loss (step 0.0005 of the mean critical charge, up to 3, beyond which the survival is below exp(−3k)) by the Panjer recursion [11], splitting each hit's loss between the two grid points around it so that the mean loss is exact. The failure probability is then the sum over the grid of the loss distribution times the Weibull failure at that loss plus xD. At one LET the recursion reproduces the double Poisson sum above to better than 0.01 percent at fluences from 10⁶ to 10⁹ per cm². Where hits per gate are rare, as they are in every built-in space environment (a few thousandths of a direct hit per gate over three years at GEO), the result reduces to fluence times the single-hit cross section plus the dose-only fraction, and the page prints both so that the agreement is visible; the interval chart uses the single-hit form.
The heavy-ion tables are those of the site's SEE rate tool: CREME96 [9] integral LET spectra of ions Z = 2 to 92 behind 100 mil of aluminum, solar quiet (1977) galactic cosmic rays, quiet magnetosphere unless stated, run on 11 and 12 September 2026 for the orbits listed; the lunar and Mars tables carry the half-sky factor. The worst-week, worst-day and peak-five-minute solar event tables are at GEO, with the GCR background included, and are applied without geomagnetic attenuation, which is a bound in low orbits; when events are added to an interval, the background of the event days is counted twice, a small excess. Each table interval is split into four log-spaced parts with log-log interpolation of the integral flux, so that the LET assigned to a bin is within 1.5 percent of every ion in it.
The tables are in LET in silicon; the model wants LET in SiO₂, and the page multiplies by one factor, 1.08 by default, which is what [2] quotes for Ar (9.0 against 8.3) and Ne (4.55 against 4.2); for Xe it quotes 54 against 51.5, a factor 1.05. Because the cross section rises as LET to the power k below b₁, this factor alone moves the ion term by about 1.08k, a factor 1.5 for the Samsung sets, so it is on the page and in the Conclusions card.
The fluence over the interval is 4π·I per cm² (I the CREME96 integral flux per m² s sr, times 4π × 10⁻⁴), times the angular factor, times the seconds in the interval, and it multiplies the normal-incidence cross section. The factor 1 is the convention of the publication's space routine as far as the review could reconstruct it; it counts each crossing of a thin plate twice, so 0.5 is the geometric count. What neither captures is the longer path of an oblique crossing, which raises the charge removed and so the sensitivity to the lighter, more numerous ions; the model defines losses at normal incidence and the gate geometry needed to do better is not published. Grürmann et al. [10] measured the omnidirectional cross section of a Samsung 4 × 8 Gbit SLC stack (K9WBG08U1M, 2008 date code) at about 2.1 times the normal-incidence value at LET 10.1 MeV cm²/mg, with the cross section rising up to seven times at grazing incidence along the long axis of the die and most of those errors in multi-bit clusters; a Micron 8 Gbit part gave 0.8 times. Whether the stack shares the die of the part fitted here is not confirmed. The factor is editable so that such a measurement can be applied.
The dose at the die over the interval is entered, not computed: it comes from the mission's shielding analysis (SHIELDOSE, SPENVIS or the project environment document). It enters the model as the deterministic shift xD above, scaled with the interval length on the chart. The published fits found the dose term to dominate the deprogramming of this part above a few tens of krad(Si): with the conservative set the dose-only deprogrammed fraction is 2.8 × 10⁻⁷ at 30 krad(Si), 9.7 × 10⁻⁴ at 150 and 3.2 × 10⁻² at 300. The measured fractions after 150 krad(Si) of cobalt-60 quoted in [2] were 1.1 × 10⁻⁶ to 2.4 × 10⁻⁶ for parts unbiased during the exposure and 0.051 to 0.055 for parts biased, the latter from dose effects in the sensing circuits rather than in the gates, which is why the conservative set's DEF is a fitting parameter rather than the physical dose enhancement factor. The model has nothing for the circuits themselves: a part whose sensing fails under dose fails outside this estimate.
Protons are not in the heavy-ion tables (Z = 2 and above). Their nuclear-recoil hits enter as a separate single-hit term: the single-bit upset cross sections per bit measured for the part with an all-zero (all programmed) pattern and no prior dose, 5.1 × 10⁻²⁰ cm² at 105 MeV and 1.2 × 10⁻¹⁹ cm² at 480 MeV [3] (no double-bit upsets were seen; the functional-interrupt cross section of 1.1 × 10⁻²² cm² per device at 105 MeV is not carried), joined by the two-parameter Bendel shape [8], σ(E) = σlim[1 − exp(−0.18 Y1/2)]⁴ with Y = (18/A)1/2(E − A), fitted through the two points: A = 22.6 MeV and σlim = 1.34 × 10⁻¹⁹ cm². The expected upsets per bit are the sum over the integral proton spectrum (CREME96 tables of the SEE rate tool, trapped AP8MIN plus galactic protons, per cm² per day, all directions) of the cross section times the fluence in each energy bin, times the days. The term has no dose dependence and no accumulation, which the data do not support either way; a proton's ionizing dose belongs in the dose entered. Below A the cross section is zero, so direct ionization by low-energy protons, which a cell this large does not respond to in the published data, is not represented.
With p the probability per programmed cell and f the programmed fraction, each bit of a codeword is in error with probability pf independently of the others, and the probability that more than t of n bits are in error is the upper binomial tail, summed term by term from t + 1 so that small values are exact. The expected number of uncorrectable codewords is that probability times the number of codewords in the dies. Cells deprogrammed together by one ion within a codeword, which would make the tail heavier, are not represented; the EDAC tool treats codes and scrubbing in more detail. The default codeword, 4224 bits, is a 528-byte sector, the correction granularity of SLC NAND of this generation; set it to the controller's code.
The full double Poisson sum at one LET for the cells entered, all programmed, after the fluence and prior dose given. With the Device A set at LET 9.0 or the Device B set at LET 4.55 it reproduces the large-fluence curves of [2] Fig. 6 (the counts rise faster than the fluence above about 10⁸ per cm² as weak hits accumulate); the apparent cross section, counts over cells over fluence, is what a test at that fluence would report, and the single-hit cross section is what the model gives for a small-fluence test in the same dose state.
The engine runs the same checks in Node.js (sitebuild/nandflip/test_nandflip_core.js, 30 checks) and the page was exercised in a browser with Playwright. Against the published numbers:
Two comparisons did not close, and the page says so rather than adjusting anything:
The model and its fits are for SLC floating-gate cells of one part type and date code, programmed before the exposure, read in static mode after a power cycle; MLC, TLC, charge-trap and 3D cells have no published fit in this model, and the parameters of other floating-gate parts generally differ by more than the two parts here differ from each other. The estimate is for prompt charge loss of the gates only: functional interrupts, upsets and dose failure of the control and sensing circuits, retention loss after dose, read and program disturb, cycling wear and temperature are outside it. The published parameters are not stated to be conservative above 150 krad(Si). Solar quiet GCR is the bound over the cycle; the event tables are GEO values. The proton term rests on two measured energies and an assumed shape. The angular treatment is a stated convention, not a measurement. The dose is yours to enter. Nothing on the page has been reviewed by any organization; it is a planning aid, not a qualification result.