Detachment-Class Failures

• Detachment-class failures are thresholded transitions where systems lose attachment or operational connectivity, leading to abrupt changes in observable behavior.
• Studies across domains—from divertor plasmas and GPUs to vehicle platooning and biofilms—demonstrate that onset indicators differ from the final detachment event.
• Key analyses emphasize that geometry, transport properties, and advanced diagnostics are crucial for deciphering these failure modes beyond gradual parameter drifts.

Detachment-class failures are failure modes whose decisive transition is organized around loss of attachment, accessibility, connectivity, or sustainable detached-state operation. In the cited literature, the term spans several technically distinct settings: divertor plasmas that either fail to enter or fail to deepen detachment; GPUs that become unavailable at the driver or interconnect level with little numeric precursor; vehicle platoons whose inter-vehicle spacing exceeds communication range; and physical systems in which bubbles, grafts, cracks, particles, or bonded structures separate from a substrate or stable branch. Taken together, these works suggest a cross-domain class of thresholded failures in which the critical event is not gradual parameter drift alone, but a state transition to a qualitatively different regime, often with altered observability, altered geometry, or loss of a steady solution (Zhao et al., 19 Dec 2025, Bidollahkhani et al., 17 Mar 2026, Somarakis et al., 2018).

1. Scope and operational definitions

Across domains, “detachment” does not denote a single mechanism. In magnetic-fusion edge physics, it denotes a divertor state in which target heat and particle fluxes are strongly reduced, and failure may mean inability to access that state, inability to deepen it, or loss of its controllability. In GPU operations, detachment-class failure denotes abrupt driver- or interconnect-level unavailability, operationally described as “GPU fell off the bus,” where structural telemetry collapse dominates the observable signature. In platooning, detachment is loss of connectivity when two consecutive vehicles separate beyond communication range. In adhesion, biofilms, rock fines, electrolytic bubbles, and viscous filaments, detachment denotes material or interfacial separation under stress, flow, or capillarity (Zhao et al., 19 Dec 2025, Bidollahkhani et al., 17 Mar 2026, Somarakis et al., 2018, Zhang et al., 2023).

Domain	Attached state	Failure signature
Tokamak divertors	Attached outer target or sustainable detached branch	delayed onset, shallow detachment, rollover stall, or thermal collapse
GPU infrastructure	Observable device at driver/interconnect level	metric disappearance, payload collapse, scrape degradation
Vehicle platooning	Neighbor connectivity within range	spacing exceeds communication limit
Electrolytic nanobubbles	Pinned finite bubble	unbounded growth until buoyant detachment
Corneal grafts after DMEK	Apposed graft	partial graft detachment on AS-OCT
Adhesive, biofilm, porous-media systems	Bonded or attached material	crack growth, erosion, sloughing, breakage, particle release

A recurring distinction is between onset and deep or terminal failure. In DIII-D negative-triangularity divertor modeling, detachment onset is operationally identified by rollover of outer-target ion saturation current, whereas deep detachment corresponds to $T_e \le 1\ \mathrm{eV}$ with strong recombination (Zhao et al., 19 Dec 2025). In electrolytic nanobubbles, the onset criterion is loss of any steady pinned state; after threshold crossing, growth becomes irreversible until final detachment (Zhang et al., 2023). In GPU failures, onset is effectively the beginning of structural telemetry collapse rather than a large excursion in temperature or power metrics (Bidollahkhani et al., 17 Mar 2026). This suggests that many detachment-class failures are bifurcation-like: crossing the first threshold does not necessarily imply progression to the desired or final detached regime, and the observable marking onset may differ from the physics that governs late evolution.

2. Divertor detachment access, shallow detachment, and loss of steady branches

In divertor physics, detachment-class failure has a particularly specific meaning: the plasma either remains attached when detachment is desired, enters only a shallow detached state, or cannot sustain a smooth path to a deeply detached branch. The DIII-D negative-triangularity study reports that approximately $40\%$ higher density is required to reach detachment onset for forward $B_T$ compared to reverse $B_T$, and that deep detachment is not achieved for reverse $B_T$. In the experimental NBI cases, the forward-$B_T$ configuration showed clear rollover of outer-target $J_{\mathrm{sat}}$ around $$n_{e,\mathrm{la}} \sim 8\times 10^{19}\,\mathrm{m}^{-3}$$, corresponding to about $1.3\,n_G$, whereas the reverse-$B_T$ case reached onset around $40\%$0 but stalled in a “shallow detachment” state. The modeling attributes the increased difficulty of detachment in negative triangularity to a shorter midplane-to-target connection length, a reduced outer divertor leg length, and lower cross-field transport relative to positive triangularity (Zhao et al., 19 Dec 2025).

The same literature establishes that negative triangularity is not merely “hard to start.” In reverse $40\%$1, the outer-target $40\%$2 falls only to about $40\%$3–$40\%$4 and then stops decreasing meaningfully as upstream density is raised further. Continued fueling then drives particle accumulation and radiation growth elsewhere, culminating in a broader thermal collapse rather than a robust low-$40\%$5, recombining, deeply detached state. The DIII-D comparison with positive triangularity further shows that negative triangularity requires line-average densities at or above the Greenwald limit to access detachment, whereas a typical positive-triangularity DIII-D H-mode begins rollover around $40\%$6 (Zhao et al., 19 Dec 2025).

The TCV comparison of Ohmic positive- and negative-triangularity plasmas independently reaches a similar conclusion. For $40\%$7, the outer divertor leg could not be cooled below $40\%$8 through core density ramps alone. Upstream variables such as $40\%$9 and $B_T$0 were broadly similar between NT and PT over the relevant density range and therefore did not explain the disparity. Geometry contributed, because NT generally had a shorter outer divertor leg and smaller effective outer-to-inner connection-length ratio, but matched-shape cases still detached more readily in PT. Langmuir-probe and IR analysis found $B_T$1 roughly constant within $B_T$2 across an upper-triangularity scan, whereas the spreading factor $B_T$3 was lower in NT by up to $B_T$4 from LPs and $B_T$5 from IR, implying a smaller $B_T$6 and a narrower, more sheath-limited-like exhaust channel (Février et al., 2023).

A further clarification comes from TCV and SD1D studies of rollout physics. The experimentally observed target-ion-current rollover is not primarily a recombination effect at onset. Rather, $B_T$7 follows the divertor ionization source $B_T$8, and rollover begins when the divertor becomes power-limited: $B_T$9, equivalently $B_T$0, with the threshold corresponding to $B_T$1. Recombination remains small or negligible at rollover and becomes important only when $B_T$2 (Verhaegh et al., 2018). The SD1D formulation sharpened this into a simultaneous power-balance and momentum-balance problem: $B_T$3 A steady detached state exists only where these two relations intersect. Depending on the temperature dependence of momentum and power losses, there are regions of $B_T$4 where no solution exists and the plasma “jumps” from high- to low-$B_T$5 states. In this framing, detachment-class failure may mean either no rollover at all or the loss of a smooth steady detached branch (Dudson et al., 2018).

The TCV X-Point Target divertor adds a control-oriented perspective. System-identification experiments with D$B_T$6 fueling, N$B_T$7 seeding, and ECRH modulation found an inherent disturbance-rejection capacity near the secondary X-point. Upstream of that point, the dynamic response of the detached state in XPT and single null was similar, but near the secondary X-point the XPT front response was reduced by about a factor of 3 in the strongest case, and in one ECRH-modulation case the response did not exceed the noise floor. This geometry-driven buffering is beneficial for passively absorbing disturbances, but it also makes the detached state harder to monitor near the secondary X-point (Winkel et al., 22 Jun 2026).

3. Structural detachment in observability and connectivity

In GPU infrastructure, detachment-class failure denotes abrupt loss of device accessibility rather than thermal or efficiency drift. The defining operational symptom is that the GPU becomes unavailable at the driver or interconnect level—repeatedly described as “GPU fell off the bus”—while numeric telemetry may remain nominal until the event. The primary observable is structural collapse: device metrics disappear, scrape latency increases, sample loss appears, scrape payloads shrink or truncate, and time-series gaps emerge. The paper states that in all processed detachment cases, GPU telemetry disappeared at or immediately after the alignment time $B_T$8, and that the dominant signal was structural rather than numeric (Bidollahkhani et al., 17 Mar 2026).

This failure mode motivated an observability-aware early-warning framework that jointly models GPU telemetry, monitoring-pipeline indicators, and node-level OS telemetry through the window-level feature vector

$B_T$9

The evaluation used production telemetry from GWDG and a fixed $B_T$0 alert budget. Average lead times in windows were reported as $B_T$1 for GPU-only Isolation Forest and $B_T$2 for Joint Isolation Forest; the corresponding maxima were $B_T$3 and $B_T$4 windows. Median lead was still often zero, indicating that many incidents remained hard to predict well in advance even with joint modeling. The key point is conceptual rather than algorithmic: conventional value-based anomaly detection fits drift-dominated faults, whereas detachment failures are primarily observable through missingness and monitoring degradation (Bidollahkhani et al., 17 Mar 2026).

Vehicle platooning presents an analytically cleaner but structurally analogous case. Here a detachment event occurs when the relative spacing between consecutive vehicles exceeds the communication range. In the delayed, noisy second-order platoon studied in (Somarakis et al., 2018), the steady-state relative distance between vehicles $B_T$5 and $B_T$6 is Gaussian with variance $B_T$7 determined by the Laplacian eigen-spectrum and eigenvectors. The detachment value-at-risk $B_T$8 is defined from the probability of entering a danger zone near the communication limit, and becomes infinite when

$B_T$9

The same work proves a lower bound

$B_T$0

which implies an irreducible delay/noise floor in detachment uncertainty. It also shows that stronger connectivity is not always safer when time delay is present: weakening connectivity may reduce risk, and increasing connectivity may increase it (Somarakis et al., 2018).

Taken together, these two literatures separate detachment-class failures from both gradual drift and explicit hard faults. In GPUs, the failure is “quiet” because structural observability collapses before or instead of numeric alarms. In vehicle networks, the failure is spectral and stochastic rather than localized to a single spacing error. A plausible implication is that detachment-class failures often demand state estimation on the loss-of-observability manifold itself, not just extrapolation of healthy-state trends.

4. Mechanical, capillary, and geomechanical separation

In electrolytic nanobubbles, the decisive question is whether a pinned bubble admits a finite equilibrium. The generalized stability theory augments pinned-bubble mass balance with an electrolytic gas influx $B_T$1 at the contact line: $B_T$2 If $B_T$3 intersects the diffusive outflux $B_T$4, a stable finite $B_T$5 exists; if not, no steady state exists and the bubble grows without bound until buoyant detachment. The threshold influx is

$B_T$6

and the corresponding minimum current is

$B_T$7

In the detachment regime, the growth law crosses over from $B_T$8 to $B_T$9, with the $J_{\mathrm{sat}}$0 regime arising from real-gas pressure dependence rather than diffusion-limited transport (Zhang et al., 2023).

Capillary thinning in particle-laden viscous liquids shows a different thresholded geometry effect. Even a single trapped particle strongly modifies filament thinning by creating a large droplet of fluid around itself, effectively splitting the thread into shorter sub-filaments that thin independently. The resulting detachment occurs earlier than in pure fluid. In dilute suspensions with $J_{\mathrm{sat}}$1, the effective viscosity remains close to that of the interstitial fluid, so early-time thinning is nearly unchanged; the late-stage acceleration is therefore a discrete particle effect rather than a bulk-rheology effect. The paper reports that the time shift caused by a single particle is significant and amounts to about $J_{\mathrm{sat}}$2 of the time shift observed for a $J_{\mathrm{sat}}$3 suspension (Deen et al., 2013).

Porous-media fines detachment introduces a bonded-failure analogue. For authigenic fines, release is governed not by DLVO detachment but by mechanical breakage of the particle–rock bond under Darcy flow. The model combines 3D Timoshenko beam theory, CFD-derived drag and moment factors, and tensile/shear failure criteria: $J_{\mathrm{sat}}$4 The resulting critical breakage Darcy velocity is

$J_{\mathrm{sat}}$5

Upscaling leads to a maximum retention function for breakage, and a two-population colloidal transport fit to Castlegate sandstone corefloods yields $J_{\mathrm{sat}}$6 for breakthrough concentration and $J_{\mathrm{sat}}$7 for impedance, supporting the interpretation that detrital and authigenic fines detach by distinct mechanisms (Hashemi et al., 2023).

Biofilm detachment under shear likewise cannot be reduced to a surface-only loading criterion. In the immersed-boundary formulation of (Sudarsan et al., 2015), the biofilm is a spring-network surrogate for a deformable solid, but detachment is triggered by a node-based averaged equivalent continuum stress tensor and a 2D von Mises yield stress,

$J_{\mathrm{sat}}$8

Detachment is initiated when $J_{\mathrm{sat}}$9 exceeds a location-dependent threshold, with $$n_{e,\mathrm{la}} \sim 8\times 10^{19}\,\mathrm{m}^{-3}$$0. The principal conclusion is that interfacial shear stress alone can lead to incorrect or inaccurate results, because maximum fluid shear and maximum internal von Mises stress need not coincide; sloughing and erosion are better understood as internal failure modes of a deformable colony (Sudarsan et al., 2015).

Axisymmetric bi-material adhesives add a fracture-mechanics counterpart. Two interfacial crack modes govern detachment: center crack propagation and edge crack propagation. For sufficiently thin soft tips, center-crack growth dominates over a finite crack-size interval and can be stable, making the interface flaw tolerant. The maximum theoretical adhesive strength is given empirically in the nearly incompressible limit by

$$n_{e,\mathrm{la}} \sim 8\times 10^{19}\,\mathrm{m}^{-3}$$1

and in the center-crack-dominated regime the detachment stress becomes independent of flaw size for $$n_{e,\mathrm{la}} \sim 8\times 10^{19}\,\mathrm{m}^{-3}$$2 (Benvidi et al., 2021). A common thread across these mechanical systems is that geometry creates or destroys stable branches: contact-line influx, trapped inclusions, bonded aspect ratio, layer thickness, and internal stress redistribution control whether separation is gradual, stable, accelerated, or catastrophic.

5. Biomedical, molecular, and information-transfer manifestations

In postoperative corneal transplantation, graft detachment after DMEK is a clinically important partial-failure mode that must be localized and quantified rather than merely declared present. The AS-OCT pipeline of (Heslinga et al., 2020) uses scleral-spur localization, ellipse refinement across the 16 radial scans, centered cropping, U-Net segmentation of detached graft regions, and extraction of two biomarkers: detachment length via skeletonization and horizontal projection of detached regions. On the test set, mean scleral spur localization error was $$n_{e,\mathrm{la}} \sim 8\times 10^{19}\,\mathrm{m}^{-3}$$3, detachment-length estimates were within $$n_{e,\mathrm{la}} \sim 8\times 10^{19}\,\mathrm{m}^{-3}$$4 pixels ($$n_{e,\mathrm{la}} \sim 8\times 10^{19}\,\mathrm{m}^{-3}$$5) of ground truth in $$n_{e,\mathrm{la}} \sim 8\times 10^{19}\,\mathrm{m}^{-3}$$6 of cases, and horizontal-projection Dice scores were $$n_{e,\mathrm{la}} \sim 8\times 10^{19}\,\mathrm{m}^{-3}$$7 for the model and $$n_{e,\mathrm{la}} \sim 8\times 10^{19}\,\mathrm{m}^{-3}$$8 for the second DMEK expert when empty masks were excluded (Heslinga et al., 2020). Here detachment-class failure is spatially distributed and requires reconstruction from multiple slices; the technical problem is therefore one of quantification and mapping rather than onset theory.

Templated polymer copying treats product detachment differently: separation from the template is essential for useful information transfer, yet it also changes which failure modes are available. In the heterogeneous-copying model with explicit detachment behind the leading edge, only a finite local neighborhood remains template-bound. As a result, the system does not exhibit the subdiffusive near-equilibrium behavior seen in heterogeneous templated self-assembly without detachment. Heterogeneity in correct monomer interactions tends to result in slower, less accurate copying, whereas heterogeneity in incorrect monomer interactions tends to result in faster, more accurate copying. The paper also distinguishes thermodynamic efficiency,

$$n_{e,\mathrm{la}} \sim 8\times 10^{19}\,\mathrm{m}^{-3}$$9

from information efficiency,

$1.3\,n_G$0

and shows that increased thermodynamic efficiency does not always translate into increased efficiency of information transfer (Guntoro et al., 2024).

These biomedical and molecular cases broaden the concept of detachment-class failure. In DMEK, the failure is literal lift-off of a graft that must be measured over a geometry-constrained image ensemble. In templated copying, the failure is not detachment itself but the residual errors and efficiencies of a system whose physics is fundamentally altered by allowing the product to detach behind the growing tip. This suggests that detachment can function either as the adverse event or as the design constraint that reshapes error modes.

6. Cross-domain patterns, thresholds, and recurring misconceptions

Taken together, the cited works suggest several recurring structures.

First, detachment-class failures are commonly thresholded rather than smoothly drifted. Divertor onset is associated with $1.3\,n_G$1 and practical $1.3\,n_G$2 rollover criteria, while deep detachment requires $1.3\,n_G$3 and strong recombination (Zhao et al., 19 Dec 2025). Electrolytic nanobubbles admit an explicit threshold influx $1.3\,n_G$4 and current $1.3\,n_G$5 above which no equilibrium exists (Zhang et al., 2023). Vehicle platoons admit a variance threshold beyond which detachment risk becomes infinite (Somarakis et al., 2018). GPU incidents are aligned not by a thermal limit but by scrape payload collapse and device-metric disappearance (Bidollahkhani et al., 17 Mar 2026).

Second, onset and full development are often non-equivalent. In negative-triangularity divertors, onset can occur without progression to deep detachment, producing shallow-detachment behavior (Zhao et al., 19 Dec 2025). In TCV detachment studies, recombination is not the trigger of initial current rollover; it matters later, once the target is very cold (Verhaegh et al., 2018). In copying systems, explicit detachment removes one near-equilibrium trapping mode but does not eliminate local sequence-selection errors (Guntoro et al., 2024).

Third, local proxy variables are often insufficient. Upstream density and $1.3\,n_G$6 did not fully explain NT–PT detachment disparity in TCV (Février et al., 2023). Interfacial shear alone was inadequate for biofilm detachment prediction (Sudarsan et al., 2015). GPU value telemetry alone missed a class of failures whose primary manifestation was structural missingness (Bidollahkhani et al., 17 Mar 2026). Thermodynamic efficiency alone was insufficient to characterize copying performance because it did not guarantee improved information transfer (Guntoro et al., 2024).

Fourth, geometry and transport repeatedly act as the decisive control parameters. In fusion, connection length, divertor leg length, and near-SOL transport shift detachment access (Zhao et al., 19 Dec 2025). In the XPT divertor, secondary-X-point geometry suppresses front sensitivity (Winkel et al., 22 Jun 2026). In adhesives, soft-tip thickness shifts failure from edge crack to stable center crack (Benvidi et al., 2021). In capillary thinning, a single inclusion creates a new internal geometric length scale and accelerates pinch-off (Deen et al., 2013). In porous media, particle aspect ratio and bond geometry determine tensile versus shear breakage (Hashemi et al., 2023).

A common misconception, therefore, is that detachment-class failure is merely the terminal visible event. The cited literature indicates otherwise. In many systems, the decisive physics lies in the existence or non-existence of a steady branch, the redistribution of mass or momentum before visible separation, or the collapse of observability that accompanies the transition. Another misconception is that “more detachment” is uniformly favorable: fusion devices require detachment, but not uncontrolled front motion; product detachment is necessary in templated copying, but heterogeneity still modulates speed and fidelity; XPT disturbance rejection is beneficial, but it also reduces measurement sensitivity (Winkel et al., 22 Jun 2026, Guntoro et al., 2024).

In this broader technical sense, detachment-class failures are best understood as regime transitions in which an attached, coupled, or observable state is lost—or, in divertor contexts, a detached state is inaccessible, incomplete, or unstable. Their analysis therefore tends to require three elements: an operational criterion for the transition, a model of the branch structure or threshold condition, and diagnostics that remain informative when the system is close to or already crossing the detachment boundary.