Optimized Fragility: Tuning and Control

• Optimized fragility is a concept that defines and tunes system sensitivity by treating fragility as a controllable design variable across diverse fields.
• It employs methods such as molecular dynamics, surrogate modeling, and statistical analysis to quantify failure probabilities, cooperative dynamics, and risk under various conditions.
• Research shows that optimization can inadvertently induce brittleness, necessitating deliberate design strategies to balance performance and robustness.

Optimized fragility is a cross-disciplinary term used to describe situations in which fragility is not merely measured but explicitly tuned, constrained, or exploited as a design variable. In contemporary research, the phrase spans at least three recurring meanings: the control of dynamical fragility in glass-forming matter, the optimization of fragility models for failure prediction under uncertainty, and the management of brittleness induced by optimization itself in computational systems. The unifying theme is that fragility is treated as a structured response property—of relaxation, failure probability, timing margin, decision risk, or user behavior—whose dependence on temperature, loading, data shift, or guidance can be formalized and systematically modified (Starr et al., 2011, Saeednejad et al., 17 Jun 2026, Darvishi, 9 Jan 2026, Fan et al., 2021, Wannaz, 30 Sep 2025).

1. Conceptual scope and recurrent definitions

Across the literature, fragility denotes sensitivity to perturbation, but the measurable object varies by domain. In supercooled liquids and polymers it is the steepness of the growth of relaxation time or viscosity near the glass transition, commonly quantified by Angell’s steepness index

$$m=\left.\frac{d\log_{10}\tau}{d(T_g/T)}\right|_{T=T_g},$$

or by VFT parameters. In structural engineering it is the exceedance probability of a damage state conditioned on an intensity measure, such as

$P[\mathrm{Damage}\ge d\mid IM].$

In reliability-aware computing it becomes a violation probability, for example

$F_i=P(s_i-\Delta d_i<0),$

while in risk-aware machine learning it becomes a tail-sensitive index on the distribution of confident errors, and in optimization theory it can appear as heavy-tailed regret or sensitivity to model misspecification (Tarjus et al., 2014, Saeednejad et al., 17 Jun 2026, Darvishi, 9 Jan 2026, Yang et al., 18 Feb 2025, Fan et al., 2021).

Domain	Fragility object	Principal optimization lever
Glass-forming matter	$\tau(T)$, $\eta(T)$, yielding threshold	density, nanoparticles, annealing, interatomic repulsion
Structural/seismic assessment	$P[\mathrm{Damage}\ge d\mid IM]$	surrogate modeling, transfer learning, active learning, bootstrap
Digital systems	timing-violation probability	selective hardening, localized redundancy
Statistical/ML systems	tail risk, typical-case reversibility, confident-error risk	robust satisficing, exploration schedule, case-selection rules
LLM/HCI/strategic systems	reasoning brittleness, perceived fragility, tipping-point tension	prompting, motion/material design, interaction structure

A persistent conceptual distinction runs through these usages. Some works treat optimized fragility as a desirable tuning target: for example, making a melt stronger, broadening a processing window, or stabilizing a fragility model under domain shift. Others treat it as an unwanted side effect of optimization: a bandit algorithm can attain Lai–Robbins optimality yet acquire truncated-Cauchy regret tails, and an in-context guide can improve factual retrieval while degrading flexible reasoning (Fan et al., 2021, Wannaz, 30 Sep 2025). This suggests that “optimization” and “robustness” are not interchangeable: many systems become fragile precisely because they are optimized for a narrower criterion.

2. Glass-forming materials and the tuning of dynamical fragility

In the glass literature, fragility is the stronger-than-Arrhenius temperature dependence of relaxation and transport. A central methodological point is that fragility is not a uniquely invariant scalar: its value depends on the observable, the reference temperature, and whether it is measured isobarically or isochorically. The assessment by Tarjus and Alba-Simionesco emphasizes that isochoric fragility is often the more intrinsic quantity, whereas isobaric fragility inherits density effects through thermal expansivity and density scaling (Tarjus et al., 2014).

A major line of work treats fragility as a tunable consequence of cooperative dynamics. In nanoparticle-filled polymer melts, molecular dynamics on bead–spring chains of length $M=20$ showed that attractive nanoparticle–polymer interactions increase $\tau$, raise $T_g$, increase fragility, and enhance string-like cooperative motion $L(T)$, whereas non-attractive interactions produce the opposite trend. With the Adam–Gibbs identification $P[\mathrm{Damage}\ge d\mid IM].$0, the relaxation obeys $P[\mathrm{Damage}\ge d\mid IM].$1, and the fragility relation becomes

$P[\mathrm{Damage}\ge d\mid IM].$2

Because $P[\mathrm{Damage}\ge d\mid IM].$3 is approximately constant in the simulated nanocomposites, the dominant control is the derivative term, not the absolute size of $P[\mathrm{Damage}\ge d\mid IM].$4. The paper therefore interprets fragility primarily as the temperature sensitivity of cooperativity rather than the static magnitude of cooperative motion (Starr et al., 2011).

A structurally analogous conclusion emerges from machine-learning work on glassy liquids. There, a single linear SVM-defined softness variable,

$P[\mathrm{Damage}\ge d\mid IM].$5

transfers across densities in a 3D binary harmonic mixture that spans strong to extremely fragile behavior. Rearrangement probabilities follow

$P[\mathrm{Damage}\ge d\mid IM].$6

and the onset temperature $P[\mathrm{Damage}\ge d\mid IM].$7 marks where dynamics become structure-sensitive. As density increases, fragility increases together with the temperature dependence of the mean softness $P[\mathrm{Damage}\ge d\mid IM].$8 and the steepness of both $P[\mathrm{Damage}\ge d\mid IM].$9 and $F_i=P(s_i-\Delta d_i<0),$0 with respect to softness. The resulting free-energy barrier,

$F_i=P(s_i-\Delta d_i<0),$1

rises more rapidly on cooling in the fragile regime (Tah et al., 2022).

In supercooled metallic melts, fragility is connected directly to the steepness of short-range interatomic repulsion. Using nonaffine lattice dynamics and a repulsive-flank approximation $F_i=P(s_i-\Delta d_i<0),$2, the high-frequency shear modulus and viscosity were written in closed form, leading to

$F_i=P(s_i-\Delta d_i<0),$3

Larger $F_i=P(s_i-\Delta d_i<0),$4 implies steeper repulsion and larger $F_i=P(s_i-\Delta d_i<0),$5, while softer repulsion lowers fragility. The fitted Born–Mayer overlap scale $F_i=P(s_i-\Delta d_i<0),$6 shows a linear relation with fragility across the studied alloys, providing a composition-level design route for tuning $F_i=P(s_i-\Delta d_i<0),$7 (Krausser et al., 2015).

Vitrimeric star-polymer networks extend this tunability across an unusually broad range. Decreasing bulk density drives a crossover from fragile to strong and then to superstrong behavior. For $F_i=P(s_i-\Delta d_i<0),$8, the reported fragility is approximately $F_i=P(s_i-\Delta d_i<0),$9, while $\tau(T)$0 shifts from about $\tau(T)$1 at high density to about $\tau(T)$2 at the lowest densities considered. Microscopic MCT reproduces this trend and attributes it to the weak temperature sensitivity of the static-structure-factor peak $\tau(T)$3 at low density, where the dominant length scale shifts from excluded-volume packing to intrachain attraction (Ciarella et al., 2019).

Fragility also governs nonequilibrium yielding. In oscillatory shear of harmonic-sphere glasses, higher density produces larger kinetic fragility and stronger annealing dependence of the yielding threshold. Strong glasses at $\tau(T)$4 show only an approximately $\tau(T)$5 increase in $\tau(T)$6 relative to $\tau(T)$7 with annealing, whereas fragile glasses at $\tau(T)$8 show approximately $\tau(T)$9–$\eta(T)$0. The proposed elastoplastic model rationalizes this through barrier growth, with

$\eta(T)$1

This links optimized fragility to the deliberate control of barrier statistics and reversible loading windows (Chatterjee et al., 2024).

3. Structural and seismic fragility as an optimization problem

In structural engineering, fragility is a conditional exceedance probability. Standard parameterizations include the lognormal form

$\eta(T)$2

and logistic/probit variants, but much recent work treats optimized fragility as the problem of obtaining calibrated, transferable, and uncertainty-aware estimates under sparse labels, domain shift, and state dependence (Saeednejad et al., 17 Jun 2026).

One major direction replaces expensive nonlinear simulation with surrogate models of the full conditional distribution. Stochastic polynomial chaos expansions model

$\eta(T)$3

by learning the conditional law of $\eta(T)$4 given SGMM parameters. In the reported three-story shear-frame and steel-frame examples, SPCE outperformed log–log linear cloud models, KCDE, and probit in estimating both conditional distributions and fragility functions, particularly for $\eta(T)$5 and for higher thresholds where tail fidelity matters (Zhu et al., 2022).

A second direction uses active learning on SVMs to minimize the number of expensive structural analyses. Earthquake excitation is embedded in a multivariate feature vector, SVM scores are calibrated to probabilities through Platt scaling, and uncertainty sampling targets points near the decision boundary. In the reported experiments, 100–1000 labeled runs were sufficient to obtain PGA- and $\eta(T)$6-based fragility errors $\eta(T)$7 of about $\eta(T)$8–$\eta(T)$9 at $P[\mathrm{Damage}\ge d\mid IM]$0 and about $P[\mathrm{Damage}\ge d\mid IM]$1–$P[\mathrm{Damage}\ge d\mid IM]$2 at $P[\mathrm{Damage}\ge d\mid IM]$3, while score-based fragilities were sharper but more calibration-sensitive (Sainct et al., 2018).

Random forests provide a related nonparametric route for bridge classes under stripe-based nonlinear time-history analysis. For multi-span concrete bridges, RF demand surrogates eliminated the need for a lognormal demand assumption, exposed the relative importance of uncertain variables such as $P[\mathrm{Damage}\ge d\mid IM]$4, $P[\mathrm{Damage}\ge d\mid IM]$5, $P[\mathrm{Damage}\ge d\mid IM]$6, $P[\mathrm{Damage}\ge d\mid IM]$7, and $P[\mathrm{Damage}\ge d\mid IM]$8, and enabled rapid recomputation of fragility curves for updated parameter sets without rerunning the full simulation campaign (Mangalathu et al., 2018).

Transfer learning generalizes this idea to low-data target domains. The methodology-centered framework for “optimized fragility modeling” combines instance-based importance weighting, parameter-based fine-tuning, hierarchical Bayesian partial pooling, and multi-source fusion. In the Katrina bridge case, selected-source adaptation improved AUC from $P[\mathrm{Damage}\ge d\mid IM]$9 to $M=20$0 and Brier score from $M=20$1 to $M=20$2; in the Hurricane Ian residential-building case, the extended fine-tuned model improved F1 from $M=20$3 to $M=20$4; in the Nisqually bridge case, adapted multi-source fusion raised macro F1 to about $M=20$5 (Saeednejad et al., 17 Jun 2026).

The state-dependent framework for industrial components adds another layer: fragility becomes a transition probability between initial and final damage states,

$M=20$6

estimated via PCA-reduced IMs, sparse PCE, and bootstrap percentile bands. For the vertical tank in the SPIF braced-frame mock-up, this produced state-conditioned DBE/SSE fragility functions at far lower computational cost than sequential NLTHA. The same framework introduces local and global IM-efficiency criteria,

$M=20$7

to select an optimal scalar IM; PGA emerged as the global choice in both the benchmark MDoF and the SPIF tank example (Nardin et al., 2024).

4. Generalized fragility metrics, tails, and state spaces

Outside structural engineering, optimized fragility frequently denotes the construction of a mathematically controlled fragility measure. In nonlinear dynamical systems, finite-amplitude fragility can be defined through the directional failure distance

$M=20$8

and the corresponding fragility curve

$M=20$9

The pre-failure predictor is the boundary-normalized gain

$\tau$0

which yields the leading-order relation

$\tau$1

The critical result is that breadth of the response spectrum matters beyond the worst direction: in the reported 12-dimensional non-normal network, two systems with matched $\tau$2 nevertheless differed in nonlinear fragility, with mean fragility-curve difference $\tau$3 and $\tau$4 (Limkumnerd, 30 May 2026).

In classification, the Fragility Index is defined within a robust satisficing framework as the smallest slack $\tau$5 such that expected ranking error remains below $\tau$6 for all distributions in an ambiguity set. Under KL divergence,

$\tau$7

and $\tau$8 is the unique root of $\tau$9. This produces explicit tail and VaR bounds,

$T_g$0

and

$T_g$1

making fragility a directly optimizable tail-risk quantity (Yang et al., 18 Feb 2025).

In statistical hypothesis testing, optimized fragility appears as a correction to the classic Fragility Index. The stochastic generalized fragility indices replace existential case selection by a probability threshold over randomly chosen subsets:

$T_g$2

This converts a rare-case reversal criterion into a typical-case one. In the electoral example, a deterministic generalized fragility of $T_g$3 Florida nonvoters becomes an $T_g$4 of about $T_g$5 once random selection is enforced. In the smoking-cessation example, the classic FI is $T_g$6, whereas $T_g$7 (Baer et al., 2022).

Game-theoretic fragility has also been formalized. In chess, the position-level fragility score

$T_g$8

combines directed-graph betweenness centrality with the attack status of pieces. Across $T_g$9 human games and $L(T)$0 engine games, maximum fragility typically peaks around ply $L(T)$1, pawns account for approximately $L(T)$2 of key attacked pieces and knights approximately $L(T)$3, and the average fragility curve exhibits a universal buildup and slow decay around the tipping point (Barthelemy, 2024).

5. Optimization-induced brittleness in computational systems

A separate lineage uses optimized fragility to describe vulnerabilities created or exposed by optimization. In FPGA soft processors, timing fragility is the probability that route-dependent delay perturbations eliminate available slack:

$L(T)$4

The practical proxies are the BER-versus-phase transition width $L(T)$5 and the transition-location variability $L(T)$6. In the reported XCZU7EV RISC-V implementation, EX and MEM were classified as High fragility, with $L(T)$7 and $L(T)$8, respectively. Selective hardening of EX+MEM achieved robustness close to full hardening: selective duplication incurred area $L(T)$9 and dynamic power $P[\mathrm{Damage}\ge d\mid IM].$00, while selective TMR incurred $P[\mathrm{Damage}\ge d\mid IM].$01 area and $P[\mathrm{Damage}\ge d\mid IM].$02 power, compared with a full-TMR reference at area about $P[\mathrm{Damage}\ge d\mid IM].$03 and normalized reliability gain about $P[\mathrm{Damage}\ge d\mid IM].$04 (Darvishi, 9 Jan 2026).

In bandit optimization, the classical objective of minimizing expected regret has a built-in fragility cost. For exponential-family problems, Lai–Robbins-optimal algorithms induce heavy regret tails; under discrimination equivalence, the second-best arm exhibits a truncated Cauchy tail with exponent $P[\mathrm{Damage}\ge d\mid IM].$05, and for every $P[\mathrm{Damage}\ge d\mid IM].$06 the $P[\mathrm{Damage}\ge d\mid IM].$07th moment of regret grows polynomially rather than polylogarithmically. The paper further shows that slight misspecification, such as Gaussian variance mismatch or AR(1) dependence, can destroy logarithmic expected-regret scaling. A robustified KL-UCB replaces $P[\mathrm{Damage}\ge d\mid IM].$08 by $P[\mathrm{Damage}\ge d\mid IM].$09, and the choice $P[\mathrm{Damage}\ge d\mid IM].$10 tightens the tail exponent to about $P[\mathrm{Damage}\ge d\mid IM].$11 at the cost of linearly larger expected regret (Fan et al., 2021).

In LLMs, “optimized fragility” names a prompt-induced trade-off between efficiency and reasoning flexibility. Six GPT-OSS:20b variants showed that unrelated ICL guides improved general-knowledge accuracy from $P[\mathrm{Damage}\ge d\mid IM].$12 in the baseline to $P[\mathrm{Damage}\ge d\mid IM].$13–$P[\mathrm{Damage}\ge d\mid IM].$14 across guided variants, while riddle accuracy dropped to $P[\mathrm{Damage}\ge d\mid IM].$15–$P[\mathrm{Damage}\ge d\mid IM].$16 compared with the baseline’s $P[\mathrm{Damage}\ge d\mid IM].$17. Timing differences were significant for general questions and riddles,

$P[\mathrm{Damage}\ge d\mid IM].$18

but not for the single olympiad geometry problem,

$P[\mathrm{Damage}\ge d\mid IM].$19

The interpretation given is that guides install heuristic scripts that improve direct retrieval while narrowing the space of reasoning strategies (Wannaz, 30 Sep 2025).

Taken together, these results suggest a recurring computational pattern: optimization against nominal averages, critical paths, or benchmark tasks often compresses variance in desirable regions while enlarging the tail or reducing adaptability elsewhere. In this sense, optimized fragility is not simply fragility that has been reduced; it is often fragility that has been relocated.

6. Design principles, misconceptions, and open problems

A central misconception is that fragility is a single, context-free scalar. The literature instead shows dependence on observable, thermodynamic path, initial state, perturbation ensemble, and even user interpretation. In glass physics, rankings can change when one moves from isobaric to isochoric conditions or from one transport observable to another. In structural engineering, the relevant fragility can be component-level, system-level, or state-transition-based. In ML and decision analysis, fragility may refer to confident-error tails, typical-case reversibility, or misspecification sensitivity rather than average loss (Tarjus et al., 2014, Nardin et al., 2024, Yang et al., 18 Feb 2025).

Another misconception is that fragility is always undesirable. In some domains it is intentionally designed. Shape-changing interfaces provide a clear example: perceived fragility is a manipulable signal that changes how users touch, restrain, or avoid an object. Material cues were the strongest explicit driver in the second SCI study: paper and silicone infinity cubes were rated more fragile than plastic, fabric, metal, and wood, while metal was rated less fragile than all others. Autonomous motion altered behavior even when explicit ratings were less affected, with hesitation occurring in $P[\mathrm{Damage}\ge d\mid IM].$20 moving trials, compared with $P[\mathrm{Damage}\ge d\mid IM].$21 folding and $P[\mathrm{Damage}\ge d\mid IM].$22 static trials. The design problem is therefore not merely to maximize robustness but to calibrate the perceived cost of interaction to the intended use (Mackamul et al., 9 Mar 2026).

Across fields, a common design logic can nevertheless be extracted. One route tunes an underlying structural mediator: cooperative strings in polymer nanocomposites, softness distributions in glassy liquids, response-spectrum breadth in nonlinear networks, or inter-class weight differences in robust classification. A second route measures state dependence explicitly and optimizes under that representation: damage-state-conditioned fragilities, phase-swept timing observables, or bootstrap/PCE uncertainty bands. A third route accepts that optimization can create brittleness and adds deliberate slack, such as increased UCB exploration, broader IM sets before PCA reduction, or prompting strategies that preserve exploratory capacity. This suggests that optimized fragility is best understood not as the elimination of sensitivity, but as the controlled allocation of sensitivity to the variables and regimes that matter most.

Open questions remain domain-specific. In glassy systems, the direct microscopic separation of energetic and entropic barriers and the quantification of deviations from $P[\mathrm{Damage}\ge d\mid IM].$23 remain unresolved (Tah et al., 2022). In vitrimeric polymers, the broader transferability of the MCT-based fragile-to-superstrong mechanism beyond the studied star-polymer architecture is still open (Ciarella et al., 2019). In structural fragility modeling, multi-hazard state vectors, richer uncertainty decomposition, and reliable extrapolation under severe domain shift remain active issues (Saeednejad et al., 17 Jun 2026). In AI, the problem of designing ICL guides “that do not have a cost” to reasoning flexibility is still explicitly posed as an open research direction (Wannaz, 30 Sep 2025).

Under this broad but technically coherent reading, optimized fragility names a contemporary research program: measure sensitivity rigorously, identify the latent variable that carries it, and then either tune that variable for performance or constrain the optimization that would otherwise make the system brittle.