PaperlayerView exact revision on arXiv

Source-linked AI summary

Learning from fracture: Teaching language models to design through physical experience

Bo Ni

arXiv:2610.17804v1cond-mat.mtrl-scics.AIcs.LG

TL;DR

Physical design requires learning which architectural choices work in unfamiliar settings, where interacting failure mechanisms complicate inference. FracLearn organizes simulated fracture experience into conditional mechanics guidance and tests it in new designs. Its best composite score exceeds the experience-stage record by 11.0%, while the study’s results remain within a fixed two-dimensional computational setting.

  • Problem

    How to organize language-model learning so physical outcomes inform later design decisions and lessons remain useful beyond encountered examples is an open challenge.

    Source: §1 · p. 1, §3 · p. 11

  • Method

    FracLearn cycles through experience, abstraction, and generalization, retaining conditional mechanics knowledge and revising designs while keeping model weights fixed.

    Source: §1 · p. 2, §3 · p. 11

  • Results

    11.0% above the experience-stage record: the best composite score followed lesson combination, architecture refinement, and hypothesis revision.

    Source: §4 · p. 11

  • Takeaways & Limitations

    FracLearn links improved designs with conditional mechanics principles and an explicit record of how physical evidence informs design decisions.

    Source: §4 · p. 11

    The reported results concern a fixed two-dimensional computational setting; transfer under changed material contrast or loading remains a proposed test.

    Source: §3 · p. 11

Abstract

from arXiv · show

How can language models learn to optimize physical designs through experience? We introduce FracLearn, an experience-abstraction-generalization framework evaluated through a deliberately unfamiliar, Tetris-inspired fracture game. The model first proposes architectures and examines simulated fracture responses to accumulate experience. Abstraction organizes these observations into conditional principles linking reinforcement placement, crack evolution and load transfer, with supporting evidence and failure conditions. Generalization challenges the model to combine these principles in new architectures and surpass the earlier design record. Campaigns demonstrate progressive refinement and recovery through hypothesis revision, yielding a best composite score 11.0% above the experience-stage record. Fracture analyses connect improved performance to crack deflection and temporary trapping as toughening mechanisms, alongside greater residual load capacity. Learning occurs through external knowledge and decisions, without model-weight updates. This approach offers a route to scientific agents that turn computational experiments into reusable physical knowledge for materials discovery and engineering design.

1 Introduction

Designing damage-tolerant materials increasingly involves architectures beyond familiar motifs, where interacting failure mechanisms make physical insight difficult to obtain. FracLearn organizes language-model design experience into conditional principles and tests their use in new architectures, linking design improvement with mechanistic understanding.

1 Introduction

  • Architectural freedom expands design beyond familiar motifs, while established rules offer little guidance and trial-and-error becomes less tractable or affordable.This motivates seeking physical insight into unfamiliar arrangements.Source: §1 · p. 1
  • Load carrying, crack resistance, and residual capacity depend on interacting architectural features whose effects can reinforce or compete with one another.Energy-dissipating features may weaken connections needed to sustain load, while tested designs can provide evidence about failure resistance and guide subsequent designs.Source: §1 · p. 1
  • Language models can propose and explain physical hypotheses, but pretraining or coherent explanations do not establish how new interventions will behave under load.The challenge is to use physical outcomes to inform later design assumptions and retain lessons useful beyond encountered examples.Source: §1 · p. 2
  • Learning is constrained by the experience available and by what can be inferred from it, while experiments and detailed fracture simulations demand time, specimens, or computational resources.Improved responses can reflect interacting effects, making it uncertain which lesson will transfer to a new architecture.Source: §1 · p. 2
  • FracLearn organizes language-model learning into experience, abstraction, and generalization, using a fracture game to develop designs and assess conditional mechanics principles.It couples proposals with numerical search and fracture simulation, then challenges the model to combine lessons in new architectures; model weights remain fixed.Source: §1 · p. 2, Figure 1 · p. 3
  • FracLearn’s best composite score is 11.0% above the experience-stage record after combining lessons, refining architectures, and revising unsuccessful hypotheses.Fracture analyses connect the gains to reinforcement placement, crack deflection, temporary trapping, and retained load-bearing capacity.Source: §1 · p. 2

2 Results

FracLearn uses a fracture-design game to connect language-model design proposals with simulated physical feedback, then abstracts experience into conditional principles for new architectures. Across four generalization campaigns, the best score exceeded the experience-stage record by 11.0%.

2.1 A fracture-design game for learning from physical experience

  • The design game fixes constituent properties and loading while tetromino reinforcement placement varies, testing how architecture affects fracture and residual load capacity.Thirteen tetrominoes occupy 52 cells of a 20 × 9 grid; a precracked specimen is evaluated with finite-element simulations under a traveling opening field.Source: §2.1 · pp. 3–4, Figure 2 · p. 4

2.2 Learning through experience, abstraction, and generalization

  • Six motif-focused campaigns supply fracture experience; cross-campaign abstraction records conditional principles, evidence, scope, predictions, and proposed tests for generalization.In Stage II, the LLM combines those lessons in new architectures and uses physical feedback to refine, repair, or reconsider them.Source: §2.2 · pp. 5–6, Figure 3 · p. 5

2.3 Physical experience guides focused architectural refinement

  • The Stage I mean cycle-best score rose 16.2%, from 0.4785 to 0.5560, across 270 evaluations, although two campaigns peaked before their final cycle.The highest score, 0.5916, established the experience-stage record; focused refinement did not guarantee improvement in every later cycle.Source: §2.3 · p. 6, Figure 4 · p. 7

2.4 Abstraction distills fracture experience into conditional design principles

  • Crack arrest and separated daughter cracks accompanied a higher peak reaction, while relocating six cells among otherwise similar layouts redirected fracture and changed peak reaction.These examples support conditional principles about crack engagement and precise reinforcement placement, rather than rules based only on overall motif.Source: §2.4 · pp. 6–7, Figure 5 · p. 8
  • Additional fracture traces coincided with similar peak reactions but final reactions of 0.087 and 0.865 nN, so their significance depends on location and remaining load paths.The textbook also advises interpreting fracture paths alongside the objective and component measures; crack deflection or added fracture surface alone need not imply higher effective toughness.Source: §2.4 · pp. 7, 9

2.5 Generalization advances fracture design beyond the experience record

  • Campaign 4 combined a staggered core, flank guidance, and terminal relief; after Campaign 3’s terminal-bypass setback, revision restored crack interception and raised its score 9.4% over cycle 6.Cycle 8 retained 34.4% of peak reaction, compared with approximately one-fifth as much final reaction in cycle 6.Source: §2.5 · p. 9, Figure 6 · p. 10
  • 11.0%: the best Stage II score rose from 0.5916 to 0.6566, and three of four campaigns surpassed the Stage I record.The campaigns combined lessons at the outset, refined designs cumulatively, or recovered through hypothesis revision.Source: §2.5 · p. 9, Figure 6 · p. 10

3 Discussion

FracLearn links design improvement to reusable mechanics knowledge: retained principles guide new architectures, while cycles test and revise physical hypotheses against simulated outcomes.

3 Discussion

  • Campaigns 1 and 4 surpassed the experience-stage record in their first cycle, with retained knowledge and design decisions changing while model weights stayed fixed.The framework links improvements in individual architectures to knowledge that can guide subsequent designs.Source: §3 · p. 11
  • Relocating six of 52 reinforcement cells within the same two-cluster motif changed the fracture route and raised peak reaction from 2.65 to 4.11 nN.The comparisons connect placement to crack engagement and secondary-fracture consequences to the load-bearing connections that remain.Source: §3 · p. 11
  • In Campaign 3, renewed crack interception and temporary trapping accompanied a final reaction increase from 0.283 to 1.406 nN relative to the recalled design.The cycle process moved from terminal bypass to renewed crack engagement.Source: §3 · p. 11
  • Across three cycles and 45 FEM evaluations, the no-learned-prior control fell from 0.5265 to 0.5005 and 0.4911, while three of four Stage II campaigns exceeded its best.Stage II best scores were 0.6046, 0.5065, 0.5659, and 0.6001, respectively.Source: §3 · p. 11
  • FracLearn makes physical rationales explicit and revisable, but generated explanations need not faithfully describe internal model reasoning; their value rests on evaluated predictions.The LLM can reconsider the mechanism and design subspace pursued in the next cycle.Source: §3 · p. 11
  • Results concern a fixed two-dimensional computational setting; testing transfer across changed material contrast or loading requires models or experiments that resolve relevant failure processes.The discussion identifies those changed conditions as useful next tests of learned relationships.Source: §3 · p. 11

4 Conclusion

FracLearn uses physical feedback to develop design capacity and conditional mechanics principles. In the fracture prototype, combining lessons and revising hypotheses yielded a best composite score 11.0% above the experience-stage record.

4 Conclusion

  • The fracture prototype’s best composite score was 11.0% above the experience-stage record, with gains connected to reinforcement placement, crack deflection, temporary trapping, and retained load-bearing capacity.The learning process also produced conditional mechanics principles and an explicit record of how physical evidence informs design decisions.Source: §4 · p. 11
  • Learned relationships could support damage-tolerant materials that balance crack resistance and load-bearing capacity in architected composites and additively manufactured metamaterials.The proposed engineering value includes knowing which interactions to preserve as loading, material properties, or fabrication constraints change.Source: §4 · pp. 11–12
  • The authors envision scientific agents using accumulated experience to select informative experiments, distinguish mechanistic explanations, and develop design principles with researchers.Earlier physical outcomes would inform both proposed designs and the questions asked next.Source: §4 · p. 12

5 Materials and Methods

The methods define a fixed two-phase fracture-design specimen and a damage-and-loading procedure for comparing reinforcement architectures.

5.1 Material properties and design space

  • The specimen was a 100 × 50 nm membrane with a 25 nm edge crack and a 20 × 9 design grid containing 52 reinforcement cells arranged as thirteen tetrominoes.Protected matrix bands transmitted boundary displacements, and placement checks enforced geometry, material budget, bounds, and protected regions.Source: §5.1 · p. 12
  • The biologically inspired model composite used fixed constituent properties and small-strain isotropic plane-stress elasticity, with perfectly bonded phase boundaries and no separate debonding or plasticity law.The constituent properties were not calibrated to a specific tissue; properties, mesh, and specimen geometry were fixed across designs.Source: §5.1 · p. 12, Table 1 · p. 12

5.2 Fracture law and loading

  • Element damage evolved irreversibly from maximum principal stress relative to phase-specific critical tensile stress; elements were deleted at damage ≥1, with no independent Griffith surface-energy parameter.The vertical displacement profile used a ramped one-sided amplitude of 0.70 nm, then advanced its imposed loading position to 95 nm over 60 nominal quasi-static increments.Source: §5.2 · p. 12

5.3 Response measures and composite score

Six normalized response measures capture fracture performance and manufacturing simplicity, and their weighted benefits and deductions form a composite score.

5.3 Response measures and composite score

  • Peak and residual loads are the maximum and final sums of reaction magnitudes at the loading boundaries; residual capacity uses a diminishing-return map.Zero peak force gives zero load retention.Source: §5.3 · pp. 13–14
  • Total crack length is estimated from deleted-element area using a 1.5 nm reference width, while connected crack extension is measured from the initial crack tip.The reference width is used for scoring.Source: §5.3 · p. 13
  • Energy absorption uses boundary work and stored strain-energy changes between ramp-end and final states, normalized with a fixed 75 nm ligament length.Boundary work is accumulated by integrating nodal reactions against imposed displacements.Source: §5.3 · p. 13
  • Manufacturing simplicity uses phase-boundary fraction and reinforcement-cluster count, while the edge and secondary-crack deductions track near-boundary deletions and crack increments.The edge fractions use a 4.0 nm boundary zone and exclude protected nondesignable elements.Source: §5.3 · p. 13
  • The composite score weights energy absorption, crack growth, stable growth, peak load, residual load, and manufacturing simplicity by 0.25, 0.20, 0.15, 0.15, 0.15, and 0.10, respectively.Instability also contributes to stable-growth scoring and to a direct deduction.Source: §5.3 · pp. 13–14
  • Nonfinite or nonpositive energy and force inputs contribute zero, and polygon area is excluded from the score.The instability measure contributes both to stable-growth scoring and its direct deduction.Source: §5.3 · p. 14

5.4 Homogeneous fracture benchmarks

Single-phase specimens provide energy benchmarks under the same mesh and loading protocol; the comparisons assess consistency, not mesh-independent toughness.

5.4 Homogeneous fracture benchmarks

  • The energy consumed per projected crack area was compared with a reference deletion-energy estimate using the same effective deletion-band width for both phases.The estimate used a 1 nm mesh spacing.Source: §5.4 · p. 14
  • Both matrix-only and reinforcement-only specimens developed essentially straight through-cracks with 74.83 nm connected extension.The specimens used the same mesh and surfing protocol.Source: §5.4 · p. 14
  • The measured-to-reference energy ratios were 1.049 for the matrix and 0.943 for reinforcement; the reinforcement-to-matrix ratio was 3.24 versus a reference prediction of 3.60.These benchmarks assess consistency at the selected mesh and loading protocol, not mesh-independent material toughness.Source: §5.4 · p. 14

5.5 Harness-mediated feedback and learning tasks

A harness mediates LLM proposals, deterministic design search, simulated evaluations, and retained feedback while keeping model weights fixed.

5.5 Harness-mediated feedback and learning tasks

  • The design agent used the OpenAI Responses API with model alias gpt-5.4, and its weights remained fixed.Proposals and reflections were stored as structured records; one schema-repair call was allowed for an invalid proposal.Source: §5.5 · p. 14
  • The LLM proposed reinforcement regions, piece budgets and patterns, and up to two design degrees of freedom; a deterministic compiler instantiated and checked layouts.The selector sampled a fixed pool in batches of five, using coverage initially and tested-design neighbors plus diversity afterward.Source: §5.5 · p. 14
  • Scores, component measures, fracture diagnostics, and LLM reflections were compressed into external memory for later decisions.Human audit prose and the researcher-facing high-score archive were excluded from agent context.Source: §5.5 · p. 14
  • During generalization, predefined rules requested refinement, repair, or reconsideration, while the LLM selected the mechanical hypothesis and architectural changes.Mechanical interpretations were assessed separately from workflow checks and rule-based summaries.Source: §5.5 · p. 14

5.6 Experience acquisition campaigns

Six experience-acquisition campaigns shared a design cycle and accumulated mechanism memory while refining distinct design motifs.

5.6 Experience acquisition campaigns

  • The six Stage I campaigns began with 12 qualitative FEM observations from the same setup, separate from the 270 reported Stage I evaluations.The initial observations used unrestricted random clusters with 54 reinforcement cells; reported evaluations used tetrominoes and 52 cells.Source: §5.6 · p. 14
  • Campaigns were not memory-isolated replicates: accumulated mechanism memory could be retained between campaigns even as a diversity directive requested a distinct motif.Instructions also required successful features to be preserved or their alteration justified as a test.Source: §5.6 · pp. 14–15
  • Each campaign comprised three cycles, with 20 admissible variants compiled and 15 evaluated per cycle.The LLM revised the design family between cycles while retaining its campaign motif.Source: §5.6 · p. 14

5.7 Performance and fracture comparisons

The section reports score comparisons across six Stage I campaigns and describes cross-campaign abstraction alongside selected fracture cases. It also specifies how the performance and fracture comparisons were interpreted.

5.7 Performance and fracture comparisons

  • 16.2% gain was calculated from the first- and third-cycle means; the six campaigns explored different motifs and were not replicate runs of one hypothesis.The experience-stage record was the maximum across all 270 evaluations.Source: §5.7 · p. 15
  • The retained fraction is the final-to-peak reaction ratio, and Campaign 1’s first-batch best differs from its first-cycle best because the latter includes two more batches.Crack and reaction histories share the traveling loading-field position as their coordinate; the LLM’s recorded interpretations were checked against numerical outputs.Source: §5.7 · p. 15

5.8 Cross-campaign abstraction and mechanical assessment

  • Cross-campaign abstraction organized records into mechanism summaries, synthesized conditional principles with predicted signatures and falsification tests, then compressed them into a four-entry textbook.The textbook retained evidence links, design practices, and diagnostic warnings, and was supplied as external guidance for Stage II.Source: §5.8 · p. 15
  • Figure 5A summarizes all 270 Stage I architectures, mapping reinforcement occupancy, element-deletion activity, and unsmoothed reaction histories.Crack activity is evaluated on the common undeformed mesh; reaction sums the magnitudes of upper and lower boundary forces.Source: §5.8 · p. 15
  • Representative Stage I cases were assessed with layouts, fracture maps, and reaction records; plateau events support an interpretation of main-crack arrest with daughter-crack formation, not a reconstructed nucleation sequence.Deleted elements adjoining retained shared nodes were treated as numerical connectivity that can transmit load, not physical crack bridging; proposed falsification tests were not completed validations.Source: §5.8 · p. 15

5.9 Generalization campaigns and response analysis

Stage II combined retained experience and the abstraction textbook with the shared design-and-evaluation procedure. The section describes stage comparisons, a guidance-omission control, and supplementary movie visualization.

5.9 Generalization campaigns and response analysis

  • Stage II supplied the abstraction textbook and retained Stage I experience as guidance, asking the LLM to combine principles across motifs and predict their mechanical roles and responses.The candidate pool grew from 20 to 24 architectures per cycle, while the 15-simulation evaluation budget remained fixed; four campaigns ran for 32 cycles and 480 evaluations.Source: §5.9 · p. 15
  • Stage comparisons included 270 Stage I and 480 Stage II evaluations; the larger Stage II total reflects longer campaigns at the same 15-evaluation cycle budget.Score distributions used common bins but were normalized separately by stage, with repeated evaluations retained as separate events.Source: §5.9 · p. 15
  • The control used the Stage II workflow and its own feedback and memory, but omitted prior FEM examples, Stage I evidence, the textbook, and motif and benchmark guidance.It was compared over the first three cycles against four Stage II campaigns, with 45 FEM evaluations per campaign.Source: §5.9 · p. 15

5.10 Supplementary movie visualization

  • Movies S1 and S2, rendered in OVITO 3.15.4 (60) from deformed FEM nodes attached to surviving elements without displacement magnification, synchronize configurations and reaction histories by traveling opening-field position without interpolating accepted simulation states.Marker size does not represent ligament width.Source: §5.10 · p. 16

Supplementary Materials

The supplementary materials document homogeneous fracture-test benchmarks, connect textbook entries to representative evidence, and provide synchronized movies of selected campaigns. They also delimit how the entries and mechanical interpretations should be read.

Supplementary Materials

  • Figure S1 benchmarks matrix-only and reinforcement-only specimens; energy consumed per projected crack area divided by its deletion-energy estimate is 1.049 for matrix and 0.943 for reinforcement.The figure also shows final crack maps and reaction force versus connected crack extension, normalized by the matrix specimen’s peak reaction force.Source: Figure S1 · p. 20
  • Table S1 lists initial design hypotheses for the six experience campaigns, while Table S2 presents LLM-generated textbook principles alongside their mechanical interpretation.Source: Table S1 · p. 20, Table S2 · p. 21
  • Textbook entries describe proposed interactions and anticipated responses, not confirmed mechanisms; channels, gates, and the local notch are matrix-filled regions excluded from reinforcement.These features are not additional precracks or voids, and campaign numbers match Fig. 4.Source: Supplementary Materials · p. 20
  • The mechanical interpretations apply only to the present material model, loading, and objective; matrix occupies all unreinforced regions.Source: Supplementary Materials · p. 21

Supplementary Movies

  • Movie S1 shows Stage I Campaign 4 cycles 2–3 during crack–reinforcement interaction, while Movie S2 shows Stage II Campaign 3 cycles 6–8 during design revision and restored crack engagement and load-bearing capacity.Both synchronize fracture evolution and total reaction to the traveling opening-field center.Source: Supplementary Movies · p. 21

Related papers

Loading 2610.17804v1…