What the tool computes from the counts you enter, the assumptions it makes, how it was checked, and where each part of it comes from
A stuck bit is a DRAM cell whose retention time has fallen below the refresh period, generally after a single proton or heavy ion left a cluster of displacement damage in or near the cell's storage node [1 to 5]. The cell then fails until the damage anneals, which can take hours or years depending on the temperature. The estimator takes the counts of distinct failing cells read after each beam exposure, with how they changed over time and storage temperature, and returns bit errors against mission duration at the flight temperature, the expected number of codewords with more than one stuck bit, a plain-language measurement plan for the test that would cover the mission, and a statement of which parts of the result are measured, assumed or missing and which way each gap pushes the number. What one particle damages to make a cell stick, and why the position of the damage decides whether it sticks, is shown in three dimensions on the mechanism page.
One row per read: the part, the beam (ion and energy; protons as H with the energy in MeV), the LET if the ion is not one the engine knows, the cumulative fluence the part had received, the hours since the end of the exposure, the temperature the part was stored at since its previous read (blank for the test room), and the number of distinct failing cells counted at that read. Rows of one part after one exposure form its time series. The device is described by the number of bits scanned per device, the flight temperature of the memory, the room temperature of the reads and the attempt frequency of the annealing kinetics; the mission by its length, place, shielding, solar level, trapped-proton source and solar-event model; the memory architecture by the codeword layout and correction.
For each beam the first count after exposure divided by the bits scanned and the fluence gives the cross section per bit, σ, with the 1/√N counting uncertainty. Exposures of one beam at several fluences are pooled as one linear estimate, which is what published proton data support below roughly 10 krad(Si) [3, 7]; the page reports the proton dose at the die over the mission and flags it beyond that. Heavy ions are carried in LET: between measured beams the cross section follows a straight line in the logarithms, below the lowest beam it falls in proportion to the LET (a single-recoil tail) and above the highest it is flat. Protons are carried in energy, flat below the lowest measured energy down to a 15 MeV floor, below which protons count nothing, and flat above the highest. The creation rate is the omnidirectional flux at the die integrated against σ, which is exact when creation is proportional to path length in a thin sensitive layer, as it is for recoil production; no angular factor is applied.
Galactic cosmic rays come from the ISO 15390 model as modified at DLR [13], at a chosen solar level, through the LET engine shared with the radiation environment page: one-dimensional transport through carbon dioxide and aluminum with nuclear losses and fragments. In Earth orbit the vertical cutoff rigidity is sampled along the orbit with the SAMPEX-measured cutoff law in the IGRF eccentric dipole [23, 24], and the Earth's shadow is the open-sky fraction at the orbit radius; the result was checked against SPENVIS within 2 percent for the station orbit and GEO and 3 to 7 percent for an 800 km polar orbit. Trapped protons are AP-8 solar-minimum orbit averages from SPENVIS [21, 22] for the preset orbits, or a pasted spectrum. Solar particle events come from the ESP model (cumulative fluence and largest event, with the ECSS parameters [14 to 17]) or the JPL model with its updated parameters [18, 19], at a confidence level that is the accepted risk [20], transported through the same shielding and, in orbit, through SPENVIS transmission tables for event protons. Electrons, neutrons and event heavy ions are not included; the page says where that matters.
Recovery of a stuck cell is the annealing of its damage, a process with a spread of activation energies. For first-order release with one attempt frequency ν, a hold of t seconds at temperature T releases every cell whose activation energy lies below the demarcation energy E* = kT ln(νt) [9]; a sequence of holds at different temperatures reaches the energy that solves Σj Δtj ν exp(−E*/kTj) = 1. The tool converts every read to its E*, so parts stored at different temperatures fall on one curve, and converts the mission age at the flight temperature the same way. A hot hold therefore stands in for years at the flight temperature: 1000 h at 65 °C reaches the energy that three years at 40 °C reaches. The persistence curve is pooled per particle class (protons; heavy ions), chain-linked across the arms on the union of the energies reached and weighted by first counts; a class without a time series borrows the other's curve. Beyond the highest energy measured the curve is held at its last value, which is an upper bound on the count when recovery is the only process and the flight is not colder than the reads, and not a bound otherwise.
The attempt frequency can be fitted from parts stored at two or more temperatures: the arms are compared where they overlap in E*, and the value that overlays them best is reported with the range that overlays them within twice the best misfit. The fit is applied only when the arms determine it; otherwise the page reports it and keeps the entered value. A decade of error in ν changes the flight time a read stands for by a factor 10 to the power (Tread/Tflight − 1), about 1.1 to 1.4 for the usual arms, which the page states for the case entered. Reads are taken at the room temperature with the refresh period extended to the flight equivalent, 64 ms × 2 to the power ((Tflight − Troom)/10) [11, 12], so that the count is the flight count; the same refresh-period analog of temperature is used in published stuck-bit work [7].
With creation rates λc per device-day and persistence V(a) at the age a, the failing population at time t is Σc λc ∫0t V(a) da; the event fluence is added either spread uniformly over the mission or as arriving late with no time to recover. Stuck bits are placed at random over the codewords of the memory, so codewords with more than the correctable number follow the binomial tail and grow as the square of the stuck population; the stuck-plus-upset rate uses an entered upset rate, and the largest single event adds its cells at once late in the mission as the worst moment. Clusters from one ion, row and column faults and the physical-to-logical mapping are not represented.
The page writes a plan for the device, mission and temperature entered: which beams, which parts go into the ovens (one proton and one heavy-ion beam carry the temperature arms, since one curve is pooled per particle class), fluence per part for a first count of about 400 cells with the dose that gives, one read procedure kept identical for the whole test, the first read within an hour of exposure, storage arms at the room temperature, the flight temperature, a hot temperature chosen so that 1000 h reaches the end-of-mission energy, and a hotter one for margin, the read schedule, and a table of what each arm reaches. ν is fixed by how well neighboring arms overlay where they overlap in E*; the plan gives the overlaps the schedule produces and says that keeping the cooler arm of a pair running widens them.
Every number on the page comes from the measurements entered, the environment chosen and the temperatures given; no device parameter is built in. The choices that stand between those inputs and the number, and what each one costs, are these.
The environment engine was compared with SPENVIS 4.6.14 in October 2026: cosmic-ray spectra in free space and behind the cutoff of the station, sun-synchronous and geostationary orbits, the ESP cumulative fluences against the ECSS tables (within 0.2 percent), the ESP worst event, the event transmission in orbit, and the AP-8 orbit averages that the presets carry. The creation and persistence machinery was run on a three-temperature proton anneal data set of an SDRAM (internal source: three proton energies, parts stored at three temperatures from the room to about 90 °C, reads over several weeks at one refresh period). The cross sections agreed with that source's own fit within 10 percent, the creation rate at 1 AU behind 100 mils of aluminum came out 10 to 25 percent below the source's, attributable to the cosmic-ray model and to using per-part first counts rather than a fitted function, and the long-term persistence fell inside the part-to-part spread of the source's own number. The arms of that set did not overlay with one attempt frequency over their whole range, and the page reported so; two of its features were not recovery at all (reads of the room-temperature parts at an ambient that varied, and an interruption of the storage), which is what the plan's controls on read temperature and thermal history are for. No measured heavy-ion stuck-bit set was available, so the heavy-ion path has been checked only for arithmetic. The example loaded with the page is synthetic proton data patterned on that set, not a measurement.
The tool predicts distinct failing cells at the refresh period and temperature of the reads, mapped to the flight condition by the 10 °C retention rule; it does not predict variable retention time, single-event upsets, functional interrupts or total-dose failure of the periphery. It does not know the device, the board, the enclosure or the mission beyond what is entered. A result is a preliminary analysis aid in the sense of the site's disclaimer: use it to size a test, to compare designs or to check someone else's number, and verify it before it enters a design decision.