Seeing the Crack: Does Showing Predicted Crack Paths Reduce Fixation in Nanofiller Layout Design?

When engineers add graphene platelets to a brittle ceramic or a stiff epoxy, the toughness gain comes from a hidden, spatial mechanism. The advancing crack is forced to tilt and detour around the platelets, and that longer, rougher path costs more energy. Yet the tools engineers use to explore platelet arrangements usually return a single number per candidate, such as strength or work of fracture, shown in a list. Research on design fixation (the tendency of designers to lock onto early or familiar solutions and stop exploring) shows that list-style presentation pushes people into hill-climbing on obvious knobs. Showing information inside the problem space instead helps them build a model of how the problem works. We propose to test whether the same holds when the "information" is physics. A fast neural surrogate would draw the predicted crack path directly over the platelet layout an engineer is editing. The question is whether this moves engineers away from textbook arrangements toward layouts that use less obvious deflection and arrest mechanisms.

The bet in one sentence: the "obscure feature" of a nanocomposite is where the crack actually goes, and making that visible in context should reduce fixation more than giving engineers better scores or more options. We test this against the strong alternative that a well-tuned autonomous optimizer, given the same simulation budget, makes the human unnecessary. The study is designed so that a negative result on that comparison is still informative.

Core Claim

In a controlled 2D platelet-layout design task, engineers who see surrogate-predicted damage fields overlaid on their layout will find tougher and more mechanistically diverse designs than engineers who see the same surrogate's scalar toughness predictions in a ranked list. A placebo overlay that is visually rich but physically uninformative should not produce this effect. The claim is about information placement, not information quantity. Both human arms get identical predictions from the same model. Only the presentation differs.

Grounding in Prior Work

The motivating evidence comes from fixation research. Formulating or Fixating showed in the WildCat Wells grid task that examples in a scrollable list led to hill-climbing and lower best scores than examples overlaid on the problem space. Their task was a 2D resource search with no physics. We transfer the in-context versus list contrast to a task governed by fracture mechanics. The Obscure Features Hypothesis work argues that people filter out the physical properties needed for unusual solutions. Our hypothesis is that a crack-path overlay surfaces exactly such a property. Understanding Design Fixation in Generative AI found that GPT-4o and Midjourney homogenize outputs (novelty ratio 67.4% versus 78.2% for Red Dot human designs; Midjourney significantly more clustered, $p<0.01$) and that their innovation is surface-level. This motivates including a prompted language model as a baseline rather than as the design engine.

The constraining evidence is the "fidelity trap" result in Lode Encoder: designers who invested effort in polished starting points treated algorithmic suggestions as errors. We therefore start every participant from a cheap, low-commitment layout and never ask them to hand-craft a high-fidelity seed.

The methods come from computational mechanics. Learning the Stress-Strain Fields in Digital Composites using Fourier Neural Operator showed that a Fourier neural operator (a network that learns a map from an input field to an output field in one forward pass) can predict full stress tensors of 2D two-phase composites and transfer across resolutions. It covered only static mode-I tension without fracture, and it reports no pixel-wise error figure in the material available to us. We extend the same architecture to predict phase-field damage fields, which is the main technical risk. AutoMS named the "Partial Satisfaction Trap," an algorithmic form of fixation in which optimizers satisfy easy objectives and stall on hard ones. Its adaptive-weight evolutionary search is our strongest autonomous competitor. AutoMS works on single-material lattices built on MIND, which is limited to linear elasticity and explicitly excludes fracture. Neither touches platelet-reinforced brittle matrices or human interaction.

The physical motivation comes from experiments. Effect of Graphene on Microstructure and Mechanical Properties of Si3N4/SiC Ceramics attributes a rise in flexural strength from 329.11 to 443.19 MPa at only 0.3 wt% graphene to tortuous crack deflection. The same line of work notes that higher loadings agglomerate and become flaws. That gives our design task its central constraint: a small, fixed platelet budget, so performance must come from arrangement rather than quantity. Microstructure-Aware Bayesian Materials Design shows that optimizers improve when they model intermediate microstructure rather than treating the property map as a black box. That supports giving the autonomous baselines access to the surrogate's damage fields, not just scalar scores.

Research Program

Task. We first build the design task. A 2D brittle matrix strip with a pre-notch is loaded in mode I. The designer places a fixed number of rigid, elongated platelets, each with position, orientation, and length, inside a design window. A minimum-spacing rule stands in for the agglomeration limit, and a single interface-strength parameter is set per task. Ground truth comes from phase-field fracture simulation, which tracks a smeared crack as a damage field. From it we read the work of fracture (energy dissipated until failure) and the crack path. We define six task variants that differ in interface strength and notch position. In some variants the textbook layout (platelets aligned perpendicular to the expected crack) is near-optimal. In others, staggered or oblique arrangements that cause repeated deflection or arrest should win. These counter-textbook variants are where fixation should cost the most.

Surrogate. Second, we train a Fourier neural operator on roughly ten to twenty thousand phase-field simulations of random and perturbed layouts. It predicts the final damage field and the load-displacement work, and runs in well under a second. We report its damage-field error and toughness ranking accuracy on held-out layouts. We commit in advance to a usability threshold: rank correlation with simulated work of fracture of at least 0.8 on the counter-textbook variants. If the surrogate falls short, the human study does not proceed until it is fixed, because a misleading overlay would contaminate the test.

Interface. Third, we build a single editing interface in three versions that differ only in what is displayed after each edit. In the list arm, candidates and their predicted toughness appear in a sortable list. In the in-context arm, the predicted damage field and crack path are drawn over the layout being edited, together with the same scalar. In the placebo arm, an overlay is drawn in the same visual style but computed from the predicted elastic stress of the uncracked part, which does not show where the crack goes. Every participant gets the same number of surrogate queries and the same number of "verify" calls to the full phase-field simulator. The verify budget is the currency used to compare against the algorithms.

Measurement. Fourth, we measure fixation directly rather than only outcomes. Fixation is the fraction of a participant's evaluated layouts that lie within a fixed distance of the textbook layout or of their own first layout. Hill-climbing is the fraction of edits that change a single platelet parameter by a small step. Mechanism diversity is the number of distinct crack-path morphologies among a participant's verified designs, clustered automatically from simulated damage fields: single deflection, repeated zig-zag, arrest at a platelet tip, and branching.

Baselines

The naive reference point is uniform random dispersion at the same platelet budget, which approximates what a well-mixed lab composite achieves. Beside it sits the textbook layout with all platelets perpendicular to the expected crack. Both are evaluated with the full simulator on every task variant, and any design method must beat both by a margin larger than the seed-to-seed spread of the simulator.

The strongest autonomous competitor is an evolutionary search in the style of AutoMS, with adaptive objective weights that penalize stagnant solutions, run on our surrogate and given the same full-simulator verify budget as humans. On its own 17-task microstructure suite, AutoMS reports an 83.8% success rate versus 43.7% for NSGA-II. That number comes from single-material lattice tasks, not fracture. There is no reported result on Lacuna for any method on platelet-layout crack-deflection design, so all baseline numbers in our task are produced here. We also run plain NSGA-II and a Bayesian optimizer that uses features from the surrogate's damage field, following the motivation of Microstructure-Aware Bayesian Materials Design. All algorithms get the same surrogate, the same number of surrogate calls as the most active human in the study, and the same verify budget. Their hyperparameters are tuned on two task variants that are excluded from the human study, so they are not tuned on the test tasks. We report results over at least twenty seeds per algorithm.

The generative AI reference is GPT-4o prompted with the task, the constraints, and the textbook layout, asked for twenty diverse layouts that are then verified. Its role is to check whether the homogenization documented in Understanding Design Fixation in Generative AI carries over to mechanical layout.

The key ablations are the list arm, which removes in-context placement while keeping identical predictions, and the placebo arm, which keeps the visual richness while removing the crack information. The in-context arm counts as beating the list arm if it reaches higher verified work of fracture on the counter-textbook variants, with lower fixation and higher mechanism diversity scores, and if the placebo arm does not also show these gains. Human-plus-surrogate counts as beating the autonomous optimizer only if a hybrid condition is better at matched verify budget. In that hybrid, the AutoMS-style search is seeded with the layouts in-context participants found most promising.

Why not just…?

Why not just run the optimizer and leave humans out? This is the most serious threat, and we treat it as a hypothesis to test, not dismiss. With a fast surrogate and a modest design space, an adaptive evolutionary search may well beat every human arm on raw toughness. The literature shows only that weak optimizers fixate: NSGA-II's 43.7% success rate in AutoMS shows this, and the adaptive search fixes much of it. We therefore make the human-versus-algorithm comparison secondary. The primary claim is the within-human comparison of in-context, list, and placebo arms, which stands regardless of the optimizer result. The hybrid condition then asks whether human-found mechanisms help the optimizer escape its own traps on the counter-textbook variants. If the optimizer alone matches the hybrid, we will report that humans add nothing to toughness in this regime. The interface finding would then matter mainly for settings where the objective cannot be fully specified ahead of time.

Why not just give engineers a better surrogate and a sorted list? Accuracy and presentation are separable, and our design separates them: list and in-context participants receive exactly the same predictions. Formulating or Fixating suggests the list format itself drives hill-climbing even when the content is good. If the list arm matches the in-context arm, this alternative wins and our claim fails cleanly.

Why not just show any rich visualization, like a stress map? The fidelity and engagement literature allows that any vivid overlay might hold attention and broaden search. The placebo arm, which uses the uncracked stress field, tests this directly. Our mechanism predicts the placebo helps much less, because peak stresses in the intact part do not show the detour a crack takes. If the placebo performs as well, the effect is about engagement rather than physics, and we would say so.

Why not just prompt a language model or retrieve analogies from patents, as MUSE does? Finding your MUSE raised the share of creative solutions from 49% to 75% using text-based functional analogies. However, it has no representation of where a crack goes in a specific layout, and the GPT-4o evidence points to surface-level variation. The prompted GPT-4o baseline measures how far text-based generation gets on this task.

Proposed Experiments

The surrogate check runs first. It has to reach the preset ranking threshold on held-out counter-textbook layouts. Its predicted crack paths must match simulated paths at a pre-registered overlap level, since the overlay is useless if it draws the wrong path.

The main study recruits about sixty materials and mechanical engineering graduate students, randomized across the three human arms. Each completes four task variants (two textbook-favoring, two counter-textbook) in counterbalanced order, with forty minutes and a fixed verify budget per variant. The pre-registered primary outcome is verified work of fracture on the counter-textbook variants. The fixation, hill-climbing, and mechanism-diversity measures are secondary outcomes that test the proposed mechanism. With twenty per arm, we are powered only for a large effect. If the effect is moderate, we will report it as a pilot effect size rather than overclaim.

The decisive pattern for our claim has three parts. On counter-textbook variants, in-context participants beat list and placebo participants on verified toughness. Their logs show less clustering around the textbook layout. And the advantage largely disappears on textbook-favoring variants, where fixation is harmless. The autonomous and hybrid comparison, run at matched verify budget across twenty seeds, then settles whether this human-side effect turns into designs an optimizer alone would not find. The top designs from every arm are re-simulated at higher resolution and under perturbed interface strength, to check that gains are not artifacts of surrogate error.