Pulse-by-Pulse Hazard of Runaway Amorphization in Phase-Change Memory

Phase-change memory cells store data by switching a small piece of material, such as germanium telluride (GeTe), between electronic states with electrical pulses. A recent in situ microscopy study found that, before the crystal melts and turns amorphous, single-crystal GeTe nanowires pass through intermediate states: metal, then a "dirty metal", then an insulator, all while the crystal stays intact. These intermediate states could supply extra, low-energy memory levels. The obstacle is control. A few pulses too many and the device runs away into melt-quench amorphization, which destroys the intermediate state (Direct Observation of Metal–Insulator Transition in Single-Crystalline Germanium Telluride Nanowire Memory Devices Prior to Amorphization). This note proposes to treat each pulse as a trial in a survival problem. We estimate the per-pulse probability that the next pulse triggers runaway, given that the device has survived so far, and let that probability depend on which electronic stage the device is in right now. The stage is read cheaply between pulses from resistance and the sign of its temperature coefficient.

The claim is specific. The risk of runaway amorphization is not a smooth, increasing function of accumulated pulse energy, which is what standard reliability fits assume. It jumps when the device changes stage. A model that is told the current stage, and has separate coefficients for each stage, should forecast runaway better than dose-only models. A stopping rule built on it should land more devices in the crystalline insulating state before they amorphize. The survival-analysis side supplies the tools: likelihoods that handle censoring, and discrete-time hazards. The physics supplies the covariates that make the hazard predictable.

Core Claim and Model

Survival analysis models the hazard, the chance of an event in the next interval given survival so far. It uses censored records, meaning devices whose test stopped before they failed, rather than discarding them. The Cox model assumes that covariates scale a shared baseline hazard by a constant factor over time. We expect that assumption to fail here by construction. The same pulse that is harmless in the metallic stage may be dangerous once dislocations and antiphase boundaries have pushed the device into the insulating stage. The core model is a discrete-time, multi-state hazard. For device $i$ at pulse $k$, let $s_{ik}$ be the current stage (metal, dirty metal, or crystalline insulator), let $u_{ik}$ be the pulse amplitude and width, and let $r_{ik}$ be the read features: log-resistance, its change since the last pulse, and a two-temperature TCR proxy. Then

$$\mathrm{logit}\, h_{ik} = \alpha_{s_{ik}}(k - \tau_{ik}) + \beta_{s_{ik}}^{\top} [u_{ik}, r_{ik}],$$

where $h_{ik}$ is the probability that pulse $k$ causes amorphization given survival to $k$, $\tau_{ik}$ is the pulse at which the current stage was entered, $\alpha_s(\cdot)$ is a flexible baseline hazard within stage $s$ (a spline in pulses-since-stage-entry), and $\beta_s$ are stage-specific coefficients. Stage-to-stage transitions are modeled the same way, as additional competing hazards. So are non-physical failures, such as electrode open circuits, which are treated as a competing risk rather than silently censored. Survival over pulses follows the product rule $S_k=\prod_{j\le k}(1-h_j)$, which is the same construction Deep Recurrent Survival Analysis uses for its discrete hazards. We reuse that likelihood. Its recurrent network is a comparator, not the core.

The low-capacity form is deliberate. Device campaigns will produce hundreds of devices, not the thousands of patients behind the clinical deep survival results. The proposal's bet is that physically meaningful stage covariates carry the signal, not model capacity. The two-temperature reads stay below about 70 °C because the crystalline insulating state is reported to be unstable above roughly that temperature (Direct Observation…).

Relation to Prior Work

The GeTe nanowire study motivates the work. It establishes the metal, dirty-metal, and insulator sequence, ties it to Ge-vacancy condensation into planar defects, and reports that over-pulsing triggers runaway. It analyzes switching deterministically, from hand-tuned pulse sequences and TEM snapshots, and gives no probability of runaway as a function of pulse or stage. We add exactly that. Two related works treat phase-change dynamics without time-to-event statistics. Unraveling the crystallization kinetics of the Ge2Sb2Te5 phase change compound with a machine-learned interatomic potential simulates crystallization with a neural force field and fits growth with Wilson–Frenkel laws, which needed empirical Jackson corrections to match high-temperature rates. Ultrafast Charge Carrier Dynamics in Vanadium Dioxide shows that electronic transitions driven by defect states can decouple from the lattice transition. Both warn that single-mechanism kinetic laws miss stage-dependent behavior, and in that sense they constrain the design. Accurate deep neural network inference using computational phase-change memory treats conductance drift and programming noise as nuisances to compensate for in neural-network inference. It does not model when devices fail or switch.

On the statistics side, two papers supply methods. DeepSurv replaces the Cox linear risk with a neural network, but it keeps proportional hazards, which is the assumption we expect to break. DRSA drops proportional hazards through recurrent discrete hazards. Both were evaluated only on clinical and web data. Our contribution is not "deep survival on a new domain". It is the claim that stage-conditioned non-proportionality, with stages identified by electronic signatures, is what makes runaway predictable.

Research Program

The first component is data. We will fabricate GeTe nanowire and confined-cell devices across several geometries. Each device is pulsed with randomized amplitude and width drawn from a design that spans the regimes where the metal-to-insulator sequence was observed. After every pulse the device receives a low-bias resistance read at two stage temperatures below 70 °C. Randomizing the pulse parameters matters: it lets the fitted pulse coefficients be read as causal effects within a stage rather than as artifacts of an adaptive protocol. A subset of devices goes to post-mortem TEM or electron energy loss spectroscopy. These label the true stage, so the stage classifier built from reads can be checked against structure. Runs end by design at a fixed pulse budget, which gives administrative censoring, or at electrode failure, which is the competing risk.

The second component is modeling. We fit the stage-conditioned hazard, and we check whether the proportional-hazards assumption fails by comparing stage-specific $\beta_s$ against a pooled $\beta$.

The third component is control. We turn the fitted hazard into a stopping and pulse-selection rule. At each step the rule pulses only if the predicted runaway probability for the next pulse is below a threshold, and it chooses the pulse that maximizes the modeled chance of entering the insulating stage. A synthetic generator, a hidden Markov stage process with known stage-dependent hazards in the spirit of DeepSurv's simulated benchmarks, is used to validate estimators before any hardware data exist.

Baselines

The naive reference point is the rule practitioners would use today: a fixed pulse schedule stopped when resistance crosses a preset ratio. It sets the floor for both prediction and control.

The strongest competitor for prediction is a parametric accelerated-life model: a Weibull or discrete log-log hazard in cumulative pulse energy plus current log-resistance as a time-varying covariate. This is the standard reliability-engineering approach, and it already uses the resistance read. It will be fit on identical training devices, with the same censoring and competing-risk handling, and tuned by the same cross-validation over devices. We also run linear time-varying Cox and a DRSA-style recurrent hazard given the same read sequence and the same hyperparameter-search budget. DeepSurv is included for completeness as a proportional-hazards neural model.

The key ablation removes stage information by pooling $\beta_s$ and $\alpha_s$ into single functions. Everything else stays fixed, including the raw resistance features. If the stage-specific form does not beat this ablation, the central claim is wrong.

No reported result was found on Lacuna for hazard prediction of phase-change switching or amorphization, so there is no published number to beat on this data. As context for how large model-class gaps have been elsewhere, DeepSurv reports a C-index of 0.867 against 0.816 for linear Cox on METABRIC, and DRSA reports 0.911 against 0.858 for DeepHit on BIDDING. These are clinical and auction data, not comparable targets. Beating a baseline means one of two things, each on held-out devices, with a confidence interval from device-level bootstrap that excludes zero. For prediction, better time-dependent concordance and lower out-of-sample negative log-likelihood of the runaway pulse. For control, a higher fraction of devices reaching the crystalline insulating state without amorphizing at an equal or lower pulse count.

Why not just…?

Why not just fit a Weibull accelerated-life model, as reliability engineers do? This is a real threat. If runaway risk really grows smoothly with cumulative dose, Weibull will match our model with fewer parameters. The physics suggests otherwise: the GeTe study describes discrete changes in conduction mechanism, not gradual wear. But that is a hypothesis, not evidence. The pooled-versus-stage-specific likelihood-ratio test, and the head-to-head prediction comparison, decide it. If Weibull wins, we report that stages carry no extra hazard information, which is itself a useful negative result for device design.

Why not just stop pulsing when the temperature coefficient of resistance turns negative? For control, a TCR-sign threshold rule is the most likely competitor to match us. It uses the same physical signal with no statistics. It cannot, however, rank pulse amplitudes, and it says nothing about how close the device is to runaway within the insulating stage, which is where over-pulsing does its damage. The control experiment runs this rule on the same devices and budget. If the rule reaches the insulating state as often as the hazard-guided policy, the model's value reduces to forecasting and interpretation, and we will say so.

Why not just use a recurrent deep survival network on raw read traces? DRSA avoids the proportional-hazards assumption and learns stage structure implicitly. With hundreds of devices, though, it risks overfitting, and it produces no per-stage pulse effects that a device engineer could use. It is run as a baseline with an equal tuning budget. If it wins decisively, we adopt it and use the post-mortem TEM labels to test whether its hidden states align with physical stages.

Decisive Checks

Three results would confirm the idea. First, stage-specific pulse coefficients differ significantly across stages, which shows that proportional hazards fails for a physical reason. Second, the stage-conditioned model beats the Weibull and pooled-ablation baselines on held-out devices. Third, the hazard-guided policy lands more devices in the crystalline insulating state than the TCR threshold rule.

Two results would refute it: a pooled model that matches the stage model, or a threshold rule that matches the policy. Either outcome is informative about whether this intermediate state is controllable at all.