Avalanche Scenario: Mechanisms and Criticality

• Avalanche scenarios are a framework where a localized event triggers a chain reaction of subsequent activations across systems.
• They incorporate mechanisms like network branching, elastic redistribution, and crack propagation, with observables such as power-law size and duration distributions.
• Modeling avalanches informs predictions of criticality, crossover behavior, and hazard assessment in materials, interfaces, and natural events.

An avalanche scenario denotes a class of dynamical descriptions in which one event causes one or more subsequent events, which in turn may cause further events in a chain reaction. Across disordered elastic systems, network cascades, sandpiles, creep failure, coarsening materials, supercooled liquids, and snow-slab release, the term refers not to a single model but to a family of mechanisms that combine local instability, propagation, and intermittent collective response. Standard observables are avalanche size, duration, inter-event time, and temporal or spatial shape; standard theoretical questions concern criticality, universality, crossover behavior, and predictability (Gleeson et al., 2016, Dobrinevski et al., 2014, Burridge, 2013, Oyama et al., 4 Apr 2026).

1. Defining mechanisms and elementary ingredients

Avalanche dynamics are studied in many disciplines, but the basic construction is recurrent. A localized event changes the state of neighboring degrees of freedom or of the global driving field, thereby enabling a burst of subsequent events. In network language, an avalanche or cascade is a chain reaction of activations on a network; when the network is locally tree-like and the dynamics are unidirectional, the process can be approximated by a Markov branching process with offspring distribution $q_k$ (Gleeson et al., 2016). In driven elastic systems, the same logic appears as successive jumps of an interface pinned by substrate disorder, with bursts localized in both time and space (Dobrinevski et al., 2014, Zhu et al., 2017).

Several models make the mechanism explicit. In the lattice model of stabilizing failures, each site carries an instability number $p_{ij}$ that controls the local probability that failure spreads through that site; avalanches originate spontaneously at sites at a rate proportional to their instability and propagate to neighboring east, west, and south sites. When an avalanche passes through a site, its $p_{ij}$ is locally reduced by a fixed amount $\epsilon$, so failure is itself stabilizing (Burridge, 2013). In the Kadanoff Sand Pile Model, the avalanche is the sequence of column firings that follows a single grain addition; for parameter $D$, $D-1$ grains can fall from column $i$ onto the $D-1$ adjacent columns to the right if the height difference between columns $i$ and $i+1$ is greater or equal to $p_{ij}$0 (Perrot et al., 2011). In dry-snow slab avalanches, release starts with an initial failure in a weak layer that may propagate across the slope until the slab fractures and slides (Meloche et al., 2024).

These formulations suggest a common decomposition into four elements: a local instability criterion, a propagation kernel, a driving protocol, and a termination mechanism. The propagation kernel may be branching on a network, elastic redistribution in a continuum, crack propagation in a slab, or stress transfer in a yielding solid. The termination mechanism may be local stabilization after failure, geometric arrest, depletion of susceptible sites, or recovery of a subcritical state (Burridge, 2013, Castellanos et al., 2018, Meloche et al., 2024).

2. Statistical signatures, criticality, and diagnostics

The statistical signature most often associated with an avalanche scenario is a power-law distribution. In driven elastic systems, avalanches have power-law distributions of size, duration, and velocity; beyond mean field, these exponents can be written in terms of the roughness exponent $p_{ij}$1 and the dynamical exponent $p_{ij}$2. For short-range elasticity,

$p_{ij}$3

and the mean avalanche size at fixed duration obeys

$p_{ij}$4

These relations make avalanche statistics an explicit consequence of depinning exponents rather than an independent phenomenology (Dobrinevski et al., 2014).

Comparable exponents arise in other nonequilibrium media. During coarsening dynamics of a two-dimensional biphasic system, the avalanche size distribution follows $p_{ij}$5, the duration distribution follows $p_{ij}$6, and the inter-avalanche time distribution follows $p_{ij}$7 over a range, despite the fact that the system is internally rather than externally driven (Pelusi et al., 2019). In a mesoscale elastoplastic model of creep failure, the avalanche size distribution obeys $p_{ij}$8 with a decreasing $p_{ij}$9 that approaches $p_{ij}$0 at failure, the average avalanche rate follows an inverse Omori law as a function of time-to-failure, and the waiting-time statistics are consistent with the ETAS model of earthquake statistics (Castellanos et al., 2018).

A recurring caution is that power laws are not sufficient by themselves to establish criticality. In neural systems, the existence of power-law distributions is described as only a first requirement; long-range spatio-temporal correlations are fundamental. Power spectra with $p_{ij}$1 decay, power-law scaling in Detrended Fluctuation Analysis, and structured inter-event time statistics are used as additional diagnostics of critical organization (Lombardi et al., 2018). A related misconception appears in self-organized criticality: the common belief that power-law avalanches are inherently unpredictable was challenged experimentally in a quasi two-dimensional sandpile, where large avalanches were uncorrelated in time yet were preceded by continuous, detectable variations in internal structure (Ramos et al., 2008).

3. Temporal and spatial morphology

The avalanche scenario is not exhausted by size and duration distributions. Average avalanche shapes provide a finer probe of universality classes and of deviations from mean-field behavior. For cascade dynamics on networks, Markov branching process theory yields the average avalanche shape for avalanches of duration $p_{ij}$2 as

$p_{ij}$3

where $p_{ij}$4 is the generating function of the offspring distribution and $p_{ij}$5 satisfies

$p_{ij}$6

At criticality, rescaled shapes collapse onto a universal curve. If $p_{ij}$7 has finite second moment, the critical shape is symmetric and parabolic; if $p_{ij}$8 with $p_{ij}$9, the shape is asymmetric and skewed left, with peak at $\epsilon$0 (Gleeson et al., 2016).

For elastic depinning beyond mean-field theory, the average velocity profile at fixed duration $\epsilon$1 is well approximated by

$\epsilon$2

with asymmetry

$\epsilon$3

Near the upper critical dimension, the negative sign skews the avalanche towards its end; this is a fluctuation effect absent in mean field (Dobrinevski et al., 2014).

The spatial analogue is the mean local advance at fixed spatial extension $\epsilon$4,

$\epsilon$5

In the one-dimensional Brownian force model, the behavior near the avalanche boundary is cubic:

$\epsilon$6

The second cumulant of the shape is also universal, with

$\epsilon$7

away from the boundaries (Zhu et al., 2017).

In fracture models, temporal shape asymmetry can itself be a precursor. In the fiber bundle model under quasi-static loading, avalanche shapes are asymmetric away from the critical point and become symmetric as the critical point is approached. The asymmetry measure follows

$\epsilon$8

independently of the disorder distribution. Under discrete loading, by contrast, avalanche shapes remain asymmetric even near breakdown (Bodaballa et al., 2023).

4. Driving protocols, shocks, and crossover regimes

The driving protocol is often the decisive ingredient in determining whether an avalanche scenario exhibits a single scaling regime or a crossover. In the lattice model of stabilizing failures, instability is driven globally by a process $\epsilon$9 that increases all sites’ $D$0 simultaneously and contains both slow drift and rapid random shocks. The system is typically subcritical, but shocks occasionally lift it into a near or super critical state from which it rapidly retreats due to large avalanches (Burridge, 2013).

At fixed average instability $D$1, the avalanche size probability mass function is

$D$2

with $D$3, and the cutoff diverges near criticality as

$D$4

Averaging over the dynamically varying $D$5 produces a distinct crossover,

$D$6

with $D$7 set by the most probable value of $D$8 in steady state. For the Poisson driving case, $D$9; for Gamma driving, $D-1$0 (Burridge, 2013). This distinguishes shock-driven systems from standard avalanche models with a single exponential cutoff.

A finite-temperature analogue appears in the creep of a driven elastic interface well below depinning. There, two distinct length scales govern the dynamics. The optimal activation scale $D-1$1 controls the bottleneck for escape from metastable states and remains essentially temperature-independent for fixed drive, while the avalanche scale $D-1$2 characterizes thermally activated avalanches and grows as temperature decreases:

$D-1$3

The relaxation time is Arrhenius and controlled by barriers associated with $D-1$4, whereas the four-point dynamical susceptibility and the large-scale crossover in the structure factor are governed by $D-1$5 (Russo et al., 19 Apr 2026). This is a separation between temporal control by activation and spatial control by depinning criticality.

5. Materials, interfaces, and glassy dynamics

In supercooled liquids, the avalanche scenario has been formulated as a zero-temperature criticality picture for dynamical heterogeneity. In the three-dimensional Kob-Andersen binary Lennard-Jones model, with system sizes $D-1$6 to 1500 and temperatures $D-1$7, the dynamical susceptibility

$D-1$8

has a peak $D-1$9 whose temperature and system-size dependence are described by

$i$0

with independently determined exponents $i$1, $i$2, and $i$3. Below $i$4, the scaling collapse is successful; above it, the scaling fails (Oyama et al., 4 Apr 2026).

The potential-energy-landscape formulation sharpens this picture. Avalanches are interpreted as correlated sequences of local rearrangements associated with unstable saddle modes; the number and localization of unstable modes at saddles, the low-frequency nonphononic modes of stable configurations, and the statistics of inherent-structure energies are all used to quantify the landscape. In that description, the saturation of $i$5 near the mode-coupling temperature and the localization of unstable modes are tied to a crossover in the available extended unstable directions of the landscape (Oyama et al., 4 Apr 2026). The paper also notes that the scenario is not universal across all glass models.

Other materials systems show analogous avalanche phenomenology with different driving sources. In a biphasic coarsening system, diffusion-driven domain growth and topological rearrangements generate intermittent bursts with size and duration exponents close to mean-field values, indicating that long-range elastic interactions and local instability are sufficient even in the absence of explicit external loading (Pelusi et al., 2019). In creep failure of disordered solids, thermal activation, disorder, elastic coupling, and local softening generate a run-up to catastrophic shear-band formation in which the avalanche exponent approaches the critical branching-process value $i$6 (Castellanos et al., 2018). The word “avalanche” also appears in runaway-electron multiplication: in simulations of the DTT full power scenario, the avalanche multiplication factor scales as

$i$7

and at $i$8 MA reaches $i$9, large enough to convert a 5.5 A seed current into macroscopic RE beams of $D-1$0 MA under high impurity injection (Emanuelli et al., 31 Dec 2025).

6. Snow avalanches, forecasting, and human exposure

In snow mechanics, the avalanche scenario refers primarily to release, arrest, and runout of dry-snow slab avalanches. Release starts with an initial failure in a weak layer that may propagate across the slope until the slab fractures and slides. A depth-averaged Material Point Method is used for efficient elasto-plastic modeling of these processes, with pure-elastic and elasto-plastic slab scenarios. Weak layer strength heterogeneity is identified as a major factor in arresting crack propagation, and slab tensile failure becomes an additional arrest mechanism in elasto-plastic slabs. Results are interpreted through a scaling law relating crack arrest distance to two dimensionless numbers related to weak layer strength variability and slab tensile fracture. Dynamic crack propagation also lowers slab tensile stress relative to quasi-static theory, according to

$D-1$1

which helps explain why supershear cracks can promote larger releases than static models suggest (Meloche et al., 2024).

Forecasting and flow-regime classification form a second branch of the snow-avalanche literature. Material Point Method simulations identify four distinct flow regimes—cold dense flow, warm shear flow, warm plug flow, and sliding slab flow—and Support Vector Machines are used for avalanche forecasting with meteorological and snowpack variables as inputs. In that work, the initial 30-dimensional vector is expanded to 44 features after recursive feature elimination, and the selected variables encode snowfall characteristics, snow accumulation, rain interaction, snowdrift patterns, cloud dynamics, snowpack mechanics, and temperature distribution within the snowpack (Sharma, 2023). Hazard-zone mapping has also been posed as a segmentation problem for a rotation-invariant convolutional neural network using terrain maps, three-day snowfall maps, and official hazard-zone maps. The architecture uses 16 identical sub-networks, each with 4 convolutional layers and dense layers, and achieves a peak balanced validation top-1 accuracy of about 50% and top-2 accuracy of about 95% (Rauter et al., 2018).

Human exposure introduces an additional avalanche scenario that is behavioral rather than mechanical. In a serious-game study with 278 experienced backcountry skiers, an avalanche accident video reduced willingness to engage in a risky slope relative to a no-avalanche video, while high familiarity with the terrain increased engagement probability. The main effects in ANOVA were $D-1$2 for video type and $D-1$3 for familiarity, with no significant interaction (Couret et al., 2020). At the scale of multiple parties, the rate of inter-party avalanche involvements is estimated to increase quadratically with party density,

$D-1$4

and the fraction of all involvements that are inter-party is

$D-1$5

When the product $D-1$6 approaches one, inter-party involvements become a substantial fraction of all avalanche involvements, and the relative rate of such involvements increases with avalanche size (Hagedorn, 2019).

Taken together, these works support a broad but technically specific use of the term. An avalanche scenario is not merely the observation of heavy-tailed event sizes; it is a mechanistic and statistical framework linking local triggering, propagation, and arrest to identifiable observables such as $D-1$7, $D-1$8, $D-1$9, shape functions, crossover scales, and hazard metrics. The main controversies concern diagnosis rather than existence: whether power laws suffice, whether asymmetry is controlled by topology or dynamics, how broadly a given universality class extends, and under what conditions internal-state monitoring yields prediction. The literature reviewed here indicates that these questions are answerable only by combining event statistics with the structure of the driving process, the geometry of propagation, and the state variables that encode proximity to instability (Lombardi et al., 2018, Gleeson et al., 2016, Ramos et al., 2008, Burridge, 2013).