Magnetic retention: why 60 kBT is not an expiry date

First specify the barrier and the clock

Write the dimensionless stability factor as Delta = E_b / (k_B T). Here E_b is an energy barrier in joules and T is absolute temperature in kelvin. For a simplified uniaxial grain at zero effective field,E_b = K_u V: anisotropy energy density times grain volume. This is not a universal barrier formula for interacting recording media. [3]

For this worksheet, choose a constant transition rate in each direction:

lambda = f_0 exp(-Delta)

f_0 is an attempt frequency in s^-1. The mean dwell time before leaving one state is 1 / lambda. Naming that time matters: it is not automatically the decay time of an ensemble's magnetization.

Rippard and colleagues measured two-state dwell times in a magnetic tunnel junction, fitting exponential dwell-time histograms and a modified Neel-Brown model. Their equation 1 reduces to this rate convention at zero effective field and negligible bias. The manuscript discusses device-dependent attempt times and suitability checks, including symmetry. Its MTJ experiment is not an HDD retention test. [1, pp. 4-7]

Three questions, three answers

Our derivation starts every hypothetical grain in state A. Transitions A to B and B to A have the same constant rate. Define the dimensionless exposure x = lambda t. Independent exponential waiting times then make the number of transitions a Poisson variable with mean x.

• Has it ever reversed? One or more transitions: P(first reversal by t) = 1 - exp(-x).
• Is it opposite at time t? An odd number of transitions: P(B at t) = (1 - exp(-2x)) / 2.
• How much magnetization remains? For equally weighted, aligned two-state moments: M(t) / M(0) = exp(-2x).

The second equation also follows from dp_B/dt = lambda(1 - 2p_B)with p_B(0) = 0. Consequently the magnetization relaxation time is 1 / (2 lambda), half the mean dwell time under our convention. Other authors may absorb that factor into a prefactor; compare definitions, not just the symbol used for time.

At x = 1, first-reversal probability is 63.21%, while opposite-state probability is 43.23%. At long times the latter approaches 50%, not 100%: repeated switching erases information about the initial state, but does not leave every grain opposite. For rare transitions both probabilities are approximately x.

What does a barrier of 60 buy in this model?

Choose f_0 = 10^9 s^-1 and ten years, with 365.25 days per year. These are illustrative inputs, not fitted media parameters. Each probability below concerns one hypothetical grain initially in A, with constant temperature and barrier throughout. Expected transitions count all flips, including returns.

For Delta = 60, first-reversal probability is 2.763e-9. Requiring at most one-in-a-billion first-reversal risk over the same interval instead needs a barrier of about 61.02in this model. The inverse calculation is:

Delta = ln(f_0 t) - ln(-ln(1 - p))

This is the equality boundary for a selected first-reversal target 0 < p < 1; increasing Delta lowers the risk. Raising the exposure time or attempt frequency tenfold requires an additional ln(10) = 2.303 at fixed target probability.

The familiar value near 60 has context. Weller and Moser's 1999 abstract ties its projected ten-year requirement to grain-size dispersion and a particular recording-density scenario. It does not establish a material-independent pass/fail boundary. We use the IBM-hosted abstract only for that qualified historical statement, not as a full-method reproduction.[2]

A mean barrier hides the weak population

Keep the same frequency and duration. Compare identical Delta = 60 grains with an invented population containing half Delta = 50 grains and half Delta = 70 grains. Both populations have mean barrier 60, but averaging the two first-reversal probabilities gives 3.043e-5, about 11,013 timesthe uniform population's probability.

This follows from averaging probabilities after applying the exponential, rather than inserting an average barrier first. It is an exact result for this two-point example, not a measured grain distribution. For expected transition counts the ratio is exactly cosh(10); for first-reversal probabilities it is slightly smaller because probabilities saturate. Neither ratio is a universal law for real media.

What this cannot predict

A recording bit need not be one grain. Readback, coding, grain interactions, barrier distributions and applied or stray fields connect microscopic switching to a detected bit error. Multiplying a per-grain probability by a disk's advertised bit count skips that model. These results are not HDD bit-error rates, AFR estimates, or archival-life guarantees.

For fixed E_b, raising absolute temperature lowers Delta. But real anisotropy and prefactors can themselves change with temperature; extrapolating a fixed-barrier model through a HAMR write pulse is unjustified. HAMR addresses writing high-anisotropy media through localized heating. Retention after cooling and thermal writability during writing are different questions. [3]

No statistical confidence interval is reported: no grains or drives were sampled. Uncertain material parameters would require a separate sensitivity or uncertainty model. The illustrative two-state assumptions are not a claim that all physical grains follow exponential waiting times.

Reproduce and inspect

Download the dependency-free JavaScript model,full-precision JSON results and provenance and numerical tests. The tests independently sum odd Poisson probabilities and check inverse targets, limits and invalid inputs. The implementation uses expm1 and log1p so tiny risks are not lost to subtraction rounding.

node --input-type=module -e "import('./magnetic-retention.mjs').then(m => console.log(JSON.stringify(m.examples(), null, 2)))"
node --test magnetic-retention.test.mjs

Definitions: thermal stability factor,Neel-Arrhenius model, and superparamagnetic limit. Continue with the HAMR technology card.

Sources and evidence scope

1. Rippard et al. Thermal relaxation rates of magnetic nanoparticles in the presence of magnetic fields and spin-transfer effects. Physical Review B 84, 064439 (2011), doi:10.1103/PhysRevB.84.064439; arXiv:1107.5007v1, pp. 4-7, equation 1 and Figure 1. Accessed 2026-10-17.
2. Weller and Moser. Thermal effect limits in ultrahigh-density magnetic recording. IEEE Transactions on Magnetics 35(6), 4423-4439 (1999), doi:10.1109/20.809134. IBM-hosted abstract only. Accessed 2026-10-17.
3. Zhou et al. Plasmonic near-field transducer for heat-assisted magnetic recording. Nanophotonics 3(3), 141-155 (2014), doi:10.1515/nanoph-2014-0001; introduction and equation 1. Accessed 2026-10-17.

Original derivation and editorial synthesis by Hesela. No source figure or experimental dataset is reproduced. The numerical worksheet is not a replication of any cited experiment.