Breakable Machine Models

• Breakable Machine is a class of models that incorporate threshold-governed, state-dependent failure directly into their governing equations.
• These models span diverse domains—from continuum fracture and discrete lattice damage to AI classifier vulnerabilities—demonstrating both localized and distributed failure mechanisms.
• The framework emphasizes irreversible transitions and mechanistic coupling, offering insights into damage localization, reconfiguration safety, and the interplay of local deficits with global behavior.

to=arxiv_search.search 출장안마json {"^^^^^^^^3query3 Machine\"^^^^^^^^3 OR ti:\3^^^^^^^^"Breakable Machine\"", "max_results": ^^^^3all:\3query3^^^^, "sort_by": "submittedDate", "sort_order": "descending"}

to=arxiv_search.search ԥхынjson {"^^^^^^^^3query3 OR id:(&&&^^^^3all:\3^^^^&&&) OR id:(&&&^^^^3 OR ti:\3^^^^&&&) OR id:(Pasquero, 15 Jan 2026) OR id:(Salman et al., 2020) OR id:(Piranda et al., 2020) OR id:(Zhang et al., 2024) OR id:(Uneyama, 2024)", "max_results": ^^^^3 OR ti:\3query3^^^^, "sort_by": "relevance", "sort_order": "descending"}

Breakable machine denotes a class of systems in which failure, rupture, brittleness, or threshold-dependent loss of function is treated as an explicit part of the model rather than as an external exception. In the literature, the term spans several distinct but technically related usages: physically based fracture simulation of three-dimensional solids; lattices of breakable load-bearing elements; geometric impact models with breakable kinetic constraints; modular robots that predict whether reconfiguration will break connections or lose stability; reduced thermodynamic models of breakable amyloid filaments; production–inventory systems for fragile items with stock-dependent breakability; coarse-grained polymer models with irreversibly removable brittle bonds; and, in a pedagogical reapplication, an XAI-based classroom game in which students learn AI literacy by “breaking” an image classifier (&&&^^^^3all:\3^^^^&&&). This range suggests that “breakable machine” functions less as a single formalism than as a family of models organized around state-dependent failure, irreversible transitions, and mechanistic coupling between local damage and global behavior.

^^^^3all:\3^^^^. Conceptual scope and defining features

Across these usages, a breakable machine is not merely a system that can fail. The defining feature is that breakage is internal to the governing description: the state variables, constitutive rules, or interaction topology explicitly determine when failure occurs and how the system continues afterward. In continuum fracture, the object deforms under non-linear finite element analysis and fails according to elastic fracture mechanics; in discrete-element models, a site breaks when the experienced force exceeds its threshold and neighboring thresholds are reduced; in geometric impulsive mechanics, a kinetic constraint changes branch when a dry-friction threshold is exceeded; in modular robotics, a planned reconfiguration step is declared unsafe if equilibrium predicts overload or instability; and in amyloid kinetics, fragmentation is a reversible process within the reaction network rather than an external perturbation (&&&^^^^3 OR ti:\3^^^^&&&).

A second recurrent feature is irreversibility or branch switching. In the Brazilian-test lattice model, an element “breaks irreversibly” and “never heals.” In the coarse-grained crystalline-polymer model, a B-bond is removed once its stored energy exceeds PRESERVED_PLACEHOLDER_^^^^3query3^^^^, and “no rebinding is allowed.” In frictional impact, the breakable instantaneous kinetic constraint remains stick-enforcing only while

PRESERVED_PLACEHOLDER_^^^^3all:\3^^^^

and switches to a slipping branch when that inequality is violated. In the classroom game, the sense of breakability is epistemic rather than material: the “failure” of the classifier is the intended pedagogical event (Pasquero, 15 Jan 2026).

A third commonality is localization versus redistribution. Some models seek realistic localization of cracks, while others are designed precisely to prevent localization. The fracture-animation system computes crack initiation at high-stress regions and propagates discontinuities through a tetrahedral mesh, whereas the metamaterial design for de-localizing brittle fracture introduces a zero-stiffness pantograph-like reinforcement so that “every breakable sub-element fails independently” and damage becomes distributed rather than concentrated into a single crack tip (Salman et al., 2020).

^^^^3 OR ti:\3^^^^. Continuum fracture and the mechanics of solid breakage

In solid mechanics, a breakable machine is often a three-dimensional body represented as a tetrahedral finite-element mesh whose motion and failure are both computed from mechanics. The fracture-animation formulation uses a continuum mechanics model discretized with tetrahedral finite elements, with four nodes per tetrahedron storing positions in material coordinates and positions and velocities in world coordinates. Deformation is described piecewise linearly over the mesh, from which the method computes strain, strain rate, stress, internal forces, and accelerations. The strain measure is the Green strain tensor, chosen because it is nonlinear and invariant to rigid motions, and stress is parameterized, for isotropic linear elasticity, by the Lamé constants PRESERVED_PLACEHOLDER_^^^^3 OR ti:\3^^^^ and $\lambda$ (&&&^^^^3all:\3^^^^&&&).

Its simulation pipeline couples deformation, collision, time integration, and fracture testing. Collision handling uses penalty forces based on the volume of overlap between tetrahedra, computed by clipping tetrahedral faces and applying a force proportional to intersecting volume at the intersection’s center of mass. Fracture is then triggered automatically when internal forces exceed a mechanical limit. After each time step, nodal internal forces are resolved into tensile and compressive components and assembled into a separation tensor; if one of the eigenvalues becomes sufficiently large, local failure is declared. A toughness parameter controls resistance to fracture, so crack initiation is material-dependent rather than hand-authored (&&&^^^^3all:\3^^^^&&&).

Crack orientation is derived from the mechanics rather than prescribed geometrically. Once a node fails, a fracture plane is computed from the separation tensor, following the principal directions implied by the tensor’s eigenstructure. The crack propagates by splitting the node into two copies and dividing all attached tetrahedra along the plane. Elements on opposite sides of the plane are attached to different node copies, creating a true discontinuity in the mesh. Dynamic mesh restructuring is central here: tetrahedra intersected by the fracture plane are split and adjacent elements are modified so that fractures can propagate in arbitrary directions, not merely along original element boundaries. The reported effect is the formation of irregular shards and edges rather than blocky fracture patterns (&&&^^^^3all:\3^^^^&&&).

The same domain also contains an opposing design logic: rather than simulating crack localization faithfully, one can engineer structures to suppress it. In the metamaterial model of de-localized brittle fracture, a chain of brittle springs is reinforced by a zero-stiffness pantograph made of inextensible but flexible beams connected by ideal pivots. The discrete bending energy

$$E_B({\bf u}) = a\sum_{i=1}^{N-1} \frac{\lambda_1^2}{2} \left( \frac{u_{i+1}+u_{i-1}-2u_i}{a^2}\right)^2$$

adds a nonlocal penalty on spatial variation, and the continuum limit

$$E(u)=\int_0^1 \left( f(u')+\frac{\lambda_1^2}{2}(u'')^2 \right)\,dx$$

is dominated by gradient or bending elasticity rather than classical stretching elasticity. This produces a brittle-to-effectively-ductile crossover in which diffuse damage replaces a single dominant crack (Salman et al., 2020).

These two strands clarify an important distinction. A breakable machine in continuum mechanics may be designed either to form physically plausible cracks or to avoid catastrophic localization. The common denominator is explicit constitutive control of how mechanical failure nucleates, propagates, or is redistributed.

3. Discrete breakable elements, thresholds, and mesoscale networks

A second major usage treats the machine as an assembly of discrete load-bearing units with local failure rules. In the Brazilian-test model, the system is a square lattice of size $L$ with one Hookean element per lattice site, random breaking thresholds $b_{(x,y)}$, a linearly decreasing force profile $F(x)=mx+F_0$ with cutoff $L_c$, and local threshold reduction after failure. All elements in a column experience the same load, PRESERVED_PLACEHOLDER_^^^^3all:\3query3^^^^, and an element breaks irreversibly when

PRESERVED_PLACEHOLDER_^^^^3all:\3all:\3^^^^

Because loading is quasi-static, the system advances by identifying the weakest surviving element through

PRESERVED_PLACEHOLDER_^^^^3all:\3 OR ti:\3^^^^

raising the load until that element reaches threshold, weakening its nearest surviving neighbors by multiplication with PRESERVED_PLACEHOLDER_^^^^3all:\33^^^^, and then allowing any newly unstable elements to fail (&&&^^^^3 OR ti:\3^^^^&&&).

The neighbor-weakening rule is the core source of spatial correlation. When a site breaks, the model does not redistribute its load in a fiber-bundle sense; instead, it reduces nearby thresholds according to PRESERVED_PLACEHOLDER_^^^^3all:\3^^^^4, with PRESERVED_PLACEHOLDER_^^^^3all:\3^^^^5. This creates damage localization and clustering: breakage progressively lowers the resistance of adjacent intact elements, so failure tends to organize into a vertical splitting zone consistent with the Brazilian-test morphology. The stress measure is

PRESERVED_PLACEHOLDER_^^^^3all:\3^^^^6

and the critical strength for a realization is defined by the maximum of the stress–strain curve, PRESERVED_PLACEHOLDER_^^^^3all:\3^^^^7. Finite-size scaling is assumed in the form

PRESERVED_PLACEHOLDER_^^^^3all:\3^^^^8

For thresholds uniformly distributed between PRESERVED_PLACEHOLDER_^^^^3all:\3^^^^9 and PRESERVED_PLACEHOLDER_^^^^3 OR ti:\3query3^^^^, the asymptotic strength is reported as PRESERVED_PLACEHOLDER_^^^^3 OR ti:\3all:\3^^^^ with PRESERVED_PLACEHOLDER_^^^^3 OR ti:\3 OR ti:\3^^^^; for the Weibull distribution PRESERVED_PLACEHOLDER_^^^^3 OR ti:\33^^^^, the asymptotic strength is PRESERVED_PLACEHOLDER_^^^^3 OR ti:\3^^^^4, again with PRESERVED_PLACEHOLDER_^^^^3 OR ti:\3^^^^5. Strength fluctuations decay as PRESERVED_PLACEHOLDER_^^^^3 OR ti:\3^^^^6 for uniform thresholds and PRESERVED_PLACEHOLDER_^^^^3 OR ti:\3^^^^7 for Weibull thresholds (&&&^^^^3 OR ti:\3^^^^&&&).

A related but more structurally detailed mesoscale construction appears in the coarse-grained model for crystalline polymer solids. Here the machine is a network of large particles, each representing a mesoscopic chunk comparable to the crystalline layer thickness. Particles carry positions PRESERVED_PLACEHOLDER_^^^^3 OR ti:\3^^^^8 and directors PRESERVED_PLACEHOLDER_^^^^3 OR ti:\3^^^^9, and they are connected by two bond classes: D-bonds, which are soft, ductile, harmonic, and non-breakable,

$\lambda$^^^^3query3^^^^

and B-bonds, which are hard, brittle, and composed of stretching, tilt, and bending terms. The brittle bond energy for bond $\lambda$^^^^3all:\3^^^^ is

$\lambda$^^^^3 OR ti:\3^^^^

and the break rule is irreversible: if $\lambda$3, the bond is removed from the network. Under uniaxial elongation, B-bond breakage begins around $\lambda$4, while the macroscopic yield point is around $\lambda$5, linking yield to the onset of brittle-bond rupture (Uneyama, 2024).

These discrete and coarse-grained models share a strong “machine” interpretation. The global response is not prescribed directly; it emerges from local thresholds, local weakening, and evolving topology. A plausible implication is that breakable-machine models are especially useful when the scientifically relevant object is not a single crack surface but a damage process distributed across many interacting sub-elements.

4. Breakable constraints, unsafe reconfiguration, and operational failure

Not all breakable machines are solids that fracture. In geometric impulsive mechanics, breakability can refer to the constitutive behavior of a constraint. The model of frictional impacts introduces a unilateral contact constraint $\lambda$6 and a suitable instantaneous kinetic constraint $\lambda$7 expressing vanishing tangential velocity of the point in contact. The state jumps from left velocity $\lambda$8 to right velocity $\lambda$9 through

$$E_B({\bf u}) = a\sum_{i=1}^{N-1} \frac{\lambda_1^2}{2} \left( \frac{u_{i+1}+u_{i-1}-2u_i}{a^2}\right)^2$$^^^^3query3^^^^

where the reactive impulse $$E_B({\bf u}) = a\sum_{i=1}^{N-1} \frac{\lambda_1^2}{2} \left( \frac{u_{i+1}+u_{i-1}-2u_i}{a^2}\right)^2$$^^^^3all:\3^^^^ is defined constitutively from a geometric decomposition relative to $$E_B({\bf u}) = a\sum_{i=1}^{N-1} \frac{\lambda_1^2}{2} \left( \frac{u_{i+1}+u_{i-1}-2u_i}{a^2}\right)^2$$^^^^3 OR ti:\3^^^^ and $$E_B({\bf u}) = a\sum_{i=1}^{N-1} \frac{\lambda_1^2}{2} \left( \frac{u_{i+1}+u_{i-1}-2u_i}{a^2}\right)^2$$3. The breakable law is piecewise: if

$$E_B({\bf u}) = a\sum_{i=1}^{N-1} \frac{\lambda_1^2}{2} \left( \frac{u_{i+1}+u_{i-1}-2u_i}{a^2}\right)^2$$4

the tangential demand is small enough to enforce sticking; if

$$E_B({\bf u}) = a\sum_{i=1}^{N-1} \frac{\lambda_1^2}{2} \left( \frac{u_{i+1}+u_{i-1}-2u_i}{a^2}\right)^2$$5

the constraint can no longer enforce full sticking and slip occurs. In this usage, “breakable” means that the instantaneous kinetic constraint changes operational branch when the dry-friction threshold is crossed, thereby yielding natural stick–slip impact behavior while restoring determinism and avoiding direct analysis of contact-point friction forces (Pasquero, 15 Jan 2026).

A different operational sense appears in modular robotic programmable matter. There, the machine becomes breakable because reconfiguration can mechanically overload inter-modular junctions or destabilize the whole structure under self-weight. Each module is represented by a node with six degrees of freedom,

$$E_B({\bf u}) = a\sum_{i=1}^{N-1} \frac{\lambda_1^2}{2} \left( \frac{u_{i+1}+u_{i-1}-2u_i}{a^2}\right)^2$$6

and each connection is approximated by a linear-elastic beam. Static equilibrium is assembled as

$$E_B({\bf u}) = a\sum_{i=1}^{N-1} \frac{\lambda_1^2}{2} \left( \frac{u_{i+1}+u_{i-1}-2u_i}{a^2}\right)^2$$7

with gravity load

$$E_B({\bf u}) = a\sum_{i=1}^{N-1} \frac{\lambda_1^2}{2} \left( \frac{u_{i+1}+u_{i-1}-2u_i}{a^2}\right)^2$$8

The predicted perturbed configuration is solved distributively with a weighted Jacobi scheme using $$E_B({\bf u}) = a\sum_{i=1}^{N-1} \frac{\lambda_1^2}{2} \left( \frac{u_{i+1}+u_{i-1}-2u_i}{a^2}\right)^2$$9, local neighbor-to-neighbor communication, and a regularization factor $$E(u)=\int_0^1 \left( f(u')+\frac{\lambda_1^2}{2}(u'')^2 \right)\,dx$$^^^^3query3^^^^ (Piranda et al., 2020).

Unsafe reconfiguration is then declared in two cases. First, stability fails if active supports cannot prevent rigid-body motion. Second, a connection is overloaded if internal force and moment exceed the connector criterion used for Blinky Blocks,

$$E(u)=\int_0^1 \left( f(u')+\frac{\lambda_1^2}{2}(u'')^2 \right)\,dx$$^^^^3all:\3^^^^

The experimental connector thresholds are $$E(u)=\int_0^1 \left( f(u')+\frac{\lambda_1^2}{2}(u'')^2 \right)\,dx$$^^^^3 OR ti:\3^^^^ for vertical connections and $$E(u)=\int_0^1 \left( f(u')+\frac{\lambda_1^2}{2}(u'')^2 \right)\,dx$$3 for lateral connections. The framework was verified in VisibleSim and on real Blinky Blocks hardware, using the same implementation, with an average time of about $$E(u)=\int_0^1 \left( f(u')+\frac{\lambda_1^2}{2}(u'')^2 \right)\,dx$$4 for one weighted Jacobi iteration on hardware (Piranda et al., 2020).

These cases help correct a common misconception. A breakable machine need not literally fracture into pieces. In some literatures, the relevant event is loss of admissible constraint action, overload of a connection, or loss of support stability. The “break” is operational and constitutive, but it is still encoded as a precise threshold transition inside the model.

5. Kinetic, thermodynamic, and optimization models of breakability

Breakability also appears in systems whose fundamental state is chemical or economic rather than mechanical. In amyloid kinetics, breakable filaments are protein aggregates that undergo nucleation, elongation, and fragmentation through the reversible network

$$E(u)=\int_0^1 \left( f(u')+\frac{\lambda_1^2}{2}(u'')^2 \right)\,dx$$5

The full mass-action system is infinite-dimensional, with conserved total protein mass

$$E(u)=\int_0^1 \left( f(u')+\frac{\lambda_1^2}{2}(u'')^2 \right)\,dx$$6

The Maximum Entropy Principle is used to reduce this system to macroscopic variables $$E(u)=\int_0^1 \left( f(u')+\frac{\lambda_1^2}{2}(u'')^2 \right)\,dx$$7 and $$E(u)=\int_0^1 \left( f(u')+\frac{\lambda_1^2}{2}(u'')^2 \right)\,dx$$8. For the discrete model, entropy maximization under the constraints on $$E(u)=\int_0^1 \left( f(u')+\frac{\lambda_1^2}{2}(u'')^2 \right)\,dx$$9, $L$^^^^3query3^^^^, and $L$^^^^3all:\3^^^^ yields

$L$^^^^3 OR ti:\3^^^^

with effective average length

$L$3

The reduced equations are

$L$4

and the paper shows that the reduced discrete and continuous models retain nonnegative entropy production and free-energy dissipation, with equality between entropy production and free-energy dissipation under detailed balance (Zhang et al., 2024).

In production–inventory optimization, breakability is modeled as stock loss caused by accumulated stress in heaped inventory. The state variable is stock $L$5, the control variable is production rate $L$6, and the dynamics are

$L$7

with quadratic demand $L$8, breakability rate

$L$9

and fixed-endpoint conditions

$b_{(x,y)}$^^^^3query3^^^^

The objective is to maximize total profit,

$b_{(x,y)}$^^^^3all:\3^^^^

where $b_{(x,y)}$^^^^3 OR ti:\3^^^^, $b_{(x,y)}$3, and $b_{(x,y)}$4. For the linear-breakability case $b_{(x,y)}$5 with $b_{(x,y)}$6, the Euler–Lagrange condition reduces to

$b_{(x,y)}$7

and the numerical study reports profit $b_{(x,y)}$8 for $b_{(x,y)}$9 versus $F(x)=mx+F_0$^^^^3query3^^^^ when $F(x)=mx+F_0$^^^^3all:\3^^^^. Reported sensitivity analysis shows monotonic profit decrease as $F(x)=mx+F_0$^^^^3 OR ti:\3^^^^ increases from $F(x)=mx+F_0$3 to $F(x)=mx+F_0$4 (&&&^^^^3 OR ti:\33^^^^&&&).

These examples broaden the meaning of breakable machine. Breakability may refer to fragmentation reactions in a nonequilibrium thermodynamic system or to inventory loss in an optimal-control problem. What remains invariant is the formal treatment of breakage as a state-dependent process that alters evolution equations and observable macroscopic behavior.

6. Breakable Machine in AI literacy and model brittleness

In AI education, Breakable Machine is the name of a browser-based classroom game that teaches critical, transformative AI literacy by having learners break an image classifier rather than build one. The teacher chooses a target label such as “doctor,” “astronaut,” or “bear,” and students use webcams and mobile devices to manipulate their appearance or environment in order to trigger high-confidence misclassifications. The game is designed for grade 4–9 students, roughly ages ^^^^3all:\3query3^^^^–^^^^3all:\3 and is organized around adversarial play, a shared classroom leaderboard, a projector-based control interface, QR-code session joining, and switches between live play, heat-map view, and training-data view (&&&^^^^3query3^^^^&&&).

The explainability layer is based on Class Activation Maps. The classifier is a MobileNet V^^^^3 OR ti:\3^^^^ model fine-tuned for the task, and the CAM is generated from the final convolution layer, which provides a classification result for each of the $F(x)=mx+F_0$5 positions. The heat map highlights regions that contributed most to the current prediction, with examples including a white shirt, a shoe string, a pen in a pocket, or an astronaut visor-like cue in the training data. The paper’s central pedagogical point is that misclassification, brittleness, and spurious correlation are not incidental bugs in the activity; they are the instructional content. Students are meant to see that a model can be confidently wrong and that training data, design choices, and social context shape AI behavior (&&&^^^^3query3^^^^&&&).

A technically sharper notion of model breakability appears in the study of out-of-distribution attacks on neural-network encoders. There the core observation is the dimensionality reduction map

$F(x)=mx+F_0$6

which allows an attacker to optimize an OOD input so that its latent code matches that of an in-distribution sample. The attack minimizes

$F(x)=mx+F_0$7

subject to an $F(x)=mx+F_0$8 perturbation budget, using a projected-gradient-style update

$F(x)=mx+F_0$9

The paper argues that encoder-based OOD detection has no theoretical guarantee because the latent map is many-to-one; if $L_c$^^^^3query3^^^^, then the detector score also matches. Reported experiments show that AUROC for several attacked detectors collapses to around $L_c$^^^^3all:\3^^^^, and most MAPE values are below $L_c$^^^^3 OR ti:\3^^^^, indicating extremely close feature matching while the samples remain visually OOD (&&&^^^^3 OR ti:\3^^^^6&&&).

Taken together, these two AI-oriented uses identify a non-material sense of breakable machine. Here the machine is a classifier or detector, and breakability refers to exploitable brittleness, bias, or representational collapse. The classroom game turns that brittleness into a pedagogical resource, whereas the OOD-attack study treats it as a structural security limitation.

7. Interpretive synthesis and limits of the term

The literature does not support a single universal formal definition of breakable machine. Instead, the term covers several technically distinct constructions whose shared logic is threshold-governed loss of integrity, admissibility, or reliability. In one domain, the threshold is an eigenvalue of a nodal separation tensor; in another, it is a local breaking threshold $L_c$3; in another, an energy criterion $L_c$4; in another, a Coulomb-type inequality for stick versus slip; in another, a connector-strength bound under self-weight; and in another, stock-dependent fragility or latent-space vulnerability (&&&^^^^3all:\3^^^^&&&).

The term therefore risks ambiguity if used without context. It may denote literal material fracture, distributed damage in a metamaterial, fragmentation in a chemical reaction network, operational unsafety in reconfigurable robotics, or educational interrogation of AI systems. A plausible implication is that “breakable machine” is best understood as a cross-domain modeling motif rather than a settled disciplinary term. What unifies these usages is the decision to represent failure mechanistically: breakage is neither ignored nor post-processed, but embedded in the constitutive, algorithmic, or pedagogical core of the system.

Several limitations recur across the literature. Discretization influences crack initiation and speed in tetrahedral fracture simulation; coarse meshes can produce unintuitive behavior, including a “button popping” effect. In the Brazilian-test lattice, damage evolution depends on lattice structure and on the local threshold-reduction rule. In modular robotics, incorrect prediction of failure location can arise from model simplification or omission of twisting torques. In the shell-type soft jig for robotic disassembly, which is designed to mitigate component damage by soft fixation, some objects still fail because of locking tabs, insufficient holding force, or unfavorable geometry. In the classroom game, empirical learning outcomes are not yet reported; the paper is explicitly technical and non-empirical (&&&^^^^3 OR ti:\3^^^^8&&&).

Within these limits, breakable-machine models provide a coherent way to study systems whose salient behavior begins precisely when ordinary operation ceases. They are models of failure not as residual error, but as a primary mode of structure, inference, and design.