Space-RHA explainer
How a single particle flips a bit in NAND flash
A floating-gate string and a 3D charge-trap string. Charge leaves through the dielectrics around it: in one heavy-ion track, in the added tracks of many light ions, or a little at a time under dose. What leaves lowers the threshold voltage, and a cell whose threshold crosses a read reference reads wrong; an erased cell, net positive, can only move toward neutral and keeps its state. The numbers come from the JPL charge-loss model fitted to the Samsung 8 Gb SLC part.
Cell
Planar floating gate
3D charge trap
Levels
SLC
MLC
TLC
Route to the flip
One heavy ion
Many light ions
Dose
Particle
Heavy ion
Proton
Light ions through this cell 10 ions
Forced onto one cell to show that the losses add; the caption gives how rare that is.
Electrons, protons and photons: each too lightly ionizing to matter alone, together a dose. The exposure shows the gate losing charge as the dose grows.
LET in SiO2 9.0 MeV cm²/mg
Where the track passes through the floating gate
Through the floating gate (strong), beside it (weak), far away (miss).
Reach of the conductive path 6 nm
How far from the track the tunnel oxide conducts: it sets the share of a charge-trap ring that one hit empties.
Tilted track (crosses a few plates instead of all)
Environment since the last program
Fluence H 1.0×10^7 /cm²
Added dose 0 krad(Si)
Model parameters
Samsung 8 Gb SLC, conservative (TNS 2017 Table IV)
Samsung 8 Gb SLC, Device A fit (Table III)
Samsung 8 Gb SLC, Device B fit (Table III)
Example device of JPL Pub 16-9 (Table II)
Run the fluence Erase and program
Readout
0 V threshold shift of this hit
0 cells of this level flipped by one such hit
0 cm² one-hit cross section at this LET
0 / 0 strong / weak hits per cell at H
0 pCLE per programmed cell at H and dose
0 errors expected in an 8 Gb die
0 loss the dose adds to every cell
0 cm² apparent cross section p/H
programmed
The model behind the numbers
Geometry: a generic floating-gate NAND string of the 50 nm class and a generic charge-trap string with eight plates, drawn to explain, not to the scale of any product. The numbers are the charge-loss model of JPL Publication 16-9 (Edmonds 2016) with the parameters fitted to the Samsung 8 Gb SLC NAND flash (Edmonds, Irom and Allen, RADECS 2016 and IEEE TNS 2017):
Two kinds of interaction: strong, a track through the floating gate, cross section σS per cell and loss LET/b1 ; weak, a track nearby, σW and LET/b2 . Losses are in units of the Weibull scale of the critical loss and add.
The critical loss a cell can take before it reads wrong is Weibull distributed over the cells with shape k; the fraction failing after a total loss x is 1 − exp(−xk ).
Over a fluence H the hits are Poisson with means HσS and HσW ; pCLE = 1 − Σm,n P(m) P(n) exp[−(m LET/b1 + n LET/b2 + xD )k ] (its Eq. 50 with dose).
Dose enters as a deterministic loss for every cell, xD = μ DEF D (σS /b1 + σW /b2 ) with μ = 6.25×104 MeV/(mg rad) and the empirical factor DEF of the paper.
Parameter sets: conservative (b1 23.2, b2 1155 MeV cm²/mg, σS 1.80×10-10 , σW 1.00×10-8 cm², k 5.057, DEF 1.65); Device A (σS 6.00×10-11 ); Device B (22.8, 2271, 7.74×10-11 , 2.00×10-9 , 7.419); the example device of the publication (49.0, 3811, 1.10×10-10 , 7.21×10-9 , 7.643).
Three routes to the flip: one heavy ion (its LET sets the loss of one strong hit); many light ions, forced onto one cell to show the losses adding, with the caption giving the Poisson rarity of that on any one cell; and dose, the deterministic loss xD that shifts every cell, with the dose-only error fraction 1 − exp(−(xD /m)k ) for a level of margin m.
Drawn, not fitted: the scale is shown as 2.5 V, the SLC distance from the programmed level to the read reference. MLC and TLC levels share the same window, so each level’s margin is a fraction of the scale; that fraction rescales the model for the level shown. The fits are for an SLC part. The neutral threshold, the cell with no net charge on its gate, lies between the states in NAND, near the read level: cells are erased to about −3 V by design, past neutrality, since over-erase is no concern in a NAND string (Aritome 2016; Brewer and Gill 2008). The view places it 0.2 V below the lowest read reference of each scheme (an assumption; the JPL papers do not give the window of the part), so that every programmed level can fail by losing its electrons and the erased level, which carries a positive charge and loses it toward neutral, cannot reach a reference. The model has no term for erased cells; the view applies its losses to the positive charge, capped at neutral, as an illustration. The electron count is drawn at 550 per volt, the published figure for a 90 nm cell, about twice what the drawn 60 nm gate would hold (Marinella 2021; Aritome 2016). An erased cell loses nothing in the model; the published tests loaded every cell with 0, the charged state.
Protons: the measured cross sections of the same part, 5.1×10-20 cm² per bit at 105 MeV and 1.2×10-19 at 480 MeV (Irom, Allen and Sheldon 2016); the recoil shown has an LET of 10 MeV cm²/mg for illustration.
Charge-trap cells: the share of a ring within reach of the track is geometry, not a fit; no published parameters of this model exist for them.
The published record qualifies the model: measured losses scale with the charge still stored, the loss distributions are broad, errors partly anneal, the response depends on angle, and dose effects depend on bias and history; there are generally exceptions in radiation effects. The page below the frame gives the sources.
Made with Claude, Anthropic's AI model, and reviewed by Greg Allen. Three.js r147 (MIT license) is inlined in this file; nothing is loaded from another host.
This view needs WebGL. The mechanism is described in the text below the frame.