betaFeedback

WorksComputer Science

AI agents can redesign engineering parts with more than a thousand CAD parameters and nonlinear simulations

Abstract

Simulation-driven design of real engineering components has to meet three conditions at once. It runs on few solves, since a nonlinear finite-element simulation of a 3D part takes hours; its design space has many values, the hundreds to thousands of values of the part's CAD model; and a better design often needs a redesign, coordinated changes of the part's proportions rather than local polishing of its surface. Automated optimisers meet at most two of these conditions. Classical optimisers, Bayesian or derivative-free, need a person to cut the design down to a few parameters, and on all the values they need far more solves; industrial shape optimisers move every surface node of the mesh in one solve per cycle, but only in small steps that polish rather than redesign. We ask whether a general-purpose large language model (LLM) can meet all three by joining what it knows with what the simulation shows. GPT-6 Astra edits a magnesium stent through a CAD geometry parametrisation of about 1,650 values, with a budget of 15 simulations; every proposal is built in CAD, meshed and simulated with large-deformation plasticity and contact, and the LLM reads the full results of every solve. Within that budget it lowered the objective, a high percentile of the strains in the stent's bends, by 32%; on the same budget, the industrial shape optimiser Tosca improved the original design by a tenth and then stopped improving. An earlier GPT-6 Astra design, 25% lower in the same objective, was checked in an independent, higher-fidelity model of the full deployment with crimping and damage: it was 14% less strained than the original after the full chain, and its most damaged spot began to fail later and carried 63% less damage at full expansion. LLMs can thus join their learned knowledge with simulation data to adjust more than a thousand design values and reach valid designs with real improvements.

Introduction

Engineers improve a component by simulating it, reading the results and changing the design. For a nonlinear three-dimensional part, three conditions shape that work (Figure 1). Small simulation budget: each finite-element simulation takes hours, so a design study runs on tens of them. High parameter count: the CAD model has hundreds to thousands of values that could change, from the proportions of the part to the shape of every local feature. Freedom to redesign: a better design often needs coordinated changes of the proportions, not only local polishing of the surface. Automated optimisers meet at most two of the three. Classical optimisers, derivative-free or Bayesian [26, 35], redesign within tens of solves once a person has cut the design down to a few parameters, the usual parameter study. On all the values they would need more solves than the whole budget just to start: a simplex method needs one solve per value plus one, and high-dimensional Bayesian optimisers are designed and tested for budgets of hundreds to thousands of evaluations [15, 16, 22]. Industrial shape optimisers such as Tosca move every surface node of the finite-element mesh in one solve per cycle, guided by a stress controller or by sensitivities the solver computes [3, 13]. They take small steps on a mesh rather than a CAD model, so they polish a design but do not redesign it. None of them reasons about what an engineer reads off a simulation: where a design is strained, why, and which change would help.

Three conditions of simulation-driven design

Figure 1. Three conditions of simulation-driven design, in colour at the corners. Each family of automated optimisers meets the two conditions its side joins and gives up the third, marked in that condition’s colour; the agent of this paper meets all three.

Large language models (LLMs) bring physics knowledge, read rich results such as tables, fields and plots, write code to analyse them and act through tools. They have been placed in design loops with finite-element or flow feedback [18, 23, 52], and on engineering simulators their advantage shows at small budgets [8, 47]. Takehi et al. [43] showed that a frontier LLM can read the weights of a neural network and edit them much as a programmer edits code, without gradients. We ask the same of a design: working through the CAD model of a real part with a budget of fifteen simulations, can a general-purpose LLM reach a real improvement by using both the simulation results and its own knowledge?

We give GPT-6 Astra a balloon-expandable magnesium stent, 39 mm long and 16 mm across as cut. Expanding it is a large-deformation elasto-plastic problem with contact, and the strain concentrates in the crowns, the bends of its zigzag struts. A redesign must lower that strain within tight limits: the expanded stent must stay as wide inside (its lumen), shorten no more and use no more metal. The agent writes design files that set up to about 1650 values of a geometry parameterisation, from the stent’s shared proportions down to smooth bumps on every crown edge. Every proposal is built in CAD, meshed afresh and simulated, and the full fields of every solve are available in a sandbox where the agent analyses them with its own code. The agent ran on the stent twice: first in pilot sessions that helped to develop the loop, over which the parameterisation grew in stages at its request, and then in a clean run that tested the finished loop, with the complete parameterisation from the first call and nothing from the pilot sessions. In the clean run, within fifteen simulations of about two hours each on one CPU, it lowered the objective, a high percentile of the crown strains, by 32% (Figure 5b), for $19 of LLM calls; the gain is far larger than the change in score when the same design is built and meshed again, and it holds on every other measure we read. The agent uses the results, not only its prior knowledge: when the per-crown results it had just received were shuffled between crowns, it expected the highest strains where the shuffled numbers put them, and it no longer aimed its changes at the crowns that the simulation had actually found most strained (§ 5.3). The finite-element model we optimise on is a reduced one, the expansion alone. One of the improved designs was checked in an independent, higher-fidelity model of the full deployment with crimping and damage, which the agent never saw and which takes 17 to 39 hours per run; there, too, it was less strained than the original, and its most damaged spot began to fail later and carried less damage (§ 5.2).

Figure 5. The agent’s two runs on the stent: the score QQ of every design, the 90th percentile of its 60 crown strains (§ 4.3). The line is the best score among the valid designs so far. (a) The pilot sessions, which developed the loop; shaded and named, the stages over which the geometry parametrisation grew at the agent’s request. (b) The clean run, with the complete parametrisation from the first call; shaded and labelled R1–R6, its six rounds (§ 3.3), with the designs of each round in the order submitted. Dotted: the best score of the pilot sessions, which the clean run passes with its tenth design.

Among automated methods, only an industrial shape optimiser can compete at this level of detail and budget, so we ran Tosca from the same design with the same fifteen simulations. It polishes rather than redesigns: from the original design it gained a tenth and then stopped improving, but started from a design the agent had already redesigned, it polished that design well (§ 5.4). Two smaller problems add evidence (§ 5.5). On Tosca’s own bundled example, a plate with a hole described by up to 26 values, four LLMs beat Tosca as shipped in 9 of 12 runs. On a single crown with four values, fewer than the solves, a good Bayesian optimiser is a real competitor, and the LLMs’ lead depends on the run.

Contributions.

  1. A demonstration that a general-purpose LLM can redesign a nonlinear three-dimensional component through about 1 650 CAD values within fifteen simulations, lowering its objective by a third for US$19 of LLM calls (§ 5.1); to our knowledge, the first stent designed by an LLM agent. The gain achieved by the LLM design holds in an independent, higher-fidelity model of the full deployment (§ 5.2).

  1. Evidence on how the agent approaches the problem, and why it works (§ 5.3): it picks its own structure, giving whole families of crowns the same shape and correcting single crowns only at the end, and the scale of each move, concepts first and polishing last. Its first idea comes from prior knowledge; after that it uses the results: its predictions sharpen as it reads them, and when its per-crown results are shuffled between crowns, it no longer aims its changes at the crowns that the simulation actually found most strained.

  1. To our knowledge, the first comparison of an LLM design agent with an industrial shape optimiser: Tosca on the stent, from the original design and from the agent’s first fix (§ 5.4), and Tosca’s own example run as shipped (Section C).

  1. A ladder of three problems at thirteen to fifteen simulations (§ 5.5), in which the values set the part’s geometry in increasing detail: four for a single crown (Section D); up to 26 for Tosca’s plate with a hole (Section C); and about 1 650 for the stent (§ 4). Where the values outnumber the solves, the agent leads; with fewer values than solves, a good classical optimiser catches up.

  1. Limits (§ 6): the simulation set-up is still written by a person, because weaker LLMs accept wrong set-ups and report success. Moreover, the geometry of the CAD model has to be parametrized before the LLM can change it.

Related work

LLMs as optimisers. LLMs can act as black-box optimisers: they propose new candidates from a prompt that lists earlier candidates and their scores [49], serve as evolutionary operators, including on low-dimensional engineering problems [6], evolve programs against automated evaluators [32, 39], or strengthen and steer Bayesian optimisation [7, 29]. A frontier LLM can also read the weights of neural networks and edit them like a program, without gradients or training [43]; this work started from that observation. In hyperparameter optimisation, classical optimisers beat pure LLM search, while a hybrid that shares CMA-ES’s internal state with an LLM did best [17]. On engineering simulators, physics-labelled prompts gave an LLM a strong early lead in calibration, but the lead was gone by about fifteen evaluations and varied widely between runs; a warm start handed to a classical optimiser worked best [47]. In these studies the LLM acts on a few parameters or on code, guided by scores or short diagnostics. Our agent edits a CAD model of about 1 650 values and reads the full field results of every solve.

LLM agents for simulation. LLM agents now write and run finite-element and CFD analyses from natural-language specifications [21, 31, 51], including Abaqus analyses [40] and complete geometry–mesh–simulation pipelines [36]. They automate set-up and execution; improving a design against a baseline is not their aim.

LLMs in the design loop. Jadhav and Barati Farimani [23] let an LLM refine 2D truss designs from finite-element feedback, with faster convergence and fewer evaluations than NSGA-II. LLM-PSO optimises parametric airfoil and axisymmetric body shapes in laminar and Stokes flow [52]. An MDO agent turns natural-language intent into parametric 3D CAD and optimises it through finite-element tool calls, demonstrated on a gas-turbine blade, a machine-tool column and a heat sink [18]. Agents also design electric motors with a surrogate and finite-element hybrid, beating finite-element-only search on the same budget [19]; design building energy systems, coming within 1% of the best design in about 20 simulations of a full year where the best classical baseline needed about 190 [8]; and shape cars with deep-learning surrogates [14]. In topology optimisation, LLMs set the continuation parameters of SIMP [50] or orchestrate a sensitivity-based solver [28]. Where these works state the size of the design that the LLM writes, it has 2 to 25 values [8, 52]; in most of the others the LLM writes the set-up, tunes a solver or steers a classical optimiser that does the search. Our agent’s geometry parametrisation has about 1 650 values, some sixty times the largest of these, and from its third design on the agent set about 800 to 1 200 of them (§ 5.3). To our knowledge, none of these compares an LLM agent with an industrial shape optimiser, works in the large-deformation elasto-plastic regime with the CAD in the loop, or checks its result in an independent, higher-fidelity model.

High-dimensional optimisation with few evaluations. Bayesian optimisation is the standard tool for expensive black-box functions. Its high-dimensional variants fit local surrogate models [16], assume that few of the variables matter [15], or scale the prior with the dimension [22]. They are designed for budgets of a few hundred to several thousand evaluations: even their smallest tests use 50 to 500, three to thirty-three times our fifteen, and their initial designs alone take 10 to 200 points. SAASBO needed about 20 to 30 evaluations just to find the two variables that mattered among 100 in a noise-free test. With fifteen evaluations in about 1 650 dimensions, any surrogate model of the objective sees two orders of magnitude fewer points than dimensions. Grey-box Bayesian optimisation models intermediate outputs, such as a strain per crown, rather than the score alone [2]. It is the closest classical counterpart of reading the results, but its surrogate models still grow with the number of design values. Gradient-based shape optimisation handles many values with few solves when an adjoint sensitivity analysis supplies the gradient, the derivatives with respect to all design values for about the cost of one more simulation [24, 44]; but such sensitivities are not available off the shelf for an objective like ours, through the whole chain from the CAD values to a new mesh, plasticity and contact.

Stent design optimisation. Wu et al. [48] optimised the shape of biodegradable magnesium stents with finite elements, and a design from this line was tested in vitro and in vivo [9]. Other work combines surrogates with multi-objective search on a few geometric parameters [10, 33, 46], including Bayesian, expected-improvement methods [37]. All of it relies on classical optimisers or surrogates; we found no stent designed with an LLM.

Industrial non-parametric shape optimisation. Industrial optimisers such as Tosca build on optimality-criteria and controller-based methods [3] and are established in industrial practice [20, 34]. In shape optimisation, Tosca moves the surface nodes of the finite-element mesh directly, guided either by a stress target or by sensitivities of the objective [11], and returns a reshaped mesh rather than CAD.

Overview of the LLM-driven optimisation loop

Two questions decide whether an LLM can design through finite-element simulations at all: how the results of a simulation reach it, and how it can specify a new geometry. Figure 2 answers both. The LLM reads results as text, organised by the design’s own named features, and writes designs as short files that set the values of a geometry parametrisation; building, meshing, solving and scoring run without it. The loop runs in rounds, in each of which the LLM proposes up to three candidate designs at once (§ 3.3). From here on we call the LLM working in this loop, with its tools, the agent. We write LLM for the language model itself, for instance when comparing LLMs; model on its own always means a simulation model, such as the finite-element model of a design. We describe the loop on the stent of § 4.

Overview of the LLM-driven optimisation loop

Figure 2. Overview of the LLM-driven optimisation loop, which runs in rounds of up to three designs. Top: the agent’s input, the LLM with its knowledge, and the agent’s output, the design files it writes. Bottom: the pipeline that turns each file into a scored simulation without the LLM; orange: the parts each problem supplies (Table 1). Example entries from the stent.

Simulation results as agent input

The agent works in a multi-turn session and keeps its messages, tool calls, results and reasoning across turns. Its opening message (Section A) states the problem in words: the part and its load case, the geometry parametrisation with its ranges and rules, the objective, the limits and the budget, and the results of the start design. After each round it receives, for every design it submitted, a digest: the score, every limit with its margin, and tables of the results by the design’s own named features. On the stent these are the strains of every crown, link and ring, named by ring and angle as in the design files. Because the results and the design files use the same names, a strain peak in the results points to the feature that can change it. The agent also gets the full results: the fields of every node and element in every exported frame are available in a Python sandbox with a small library to read them, and the agent writes its own analysis code there and looks at the plots it draws. It never operates the solver or its graphical interface: everything reaches it as text, tables and the images it draws itself. Table 1 lists the input and output for the stent and for the two smaller problems of § 5.5.

Stent (Section 4)Plate (Section C)Crown (Section D)
Agent writesShared settings, every crown and link, bumps on the crown edges; about 1 650 values (Table 2); outlines wrapped onto a tubeThe hole’s two semi-axes and up to eight bumps on its edge; up to 26 values; outline extrudedThree widths and a shift of the apex; four values, written into the solver input
Free checkRules, CAD build, measuresRules, CAD build, measuresRules on the widths
SimulationBalloon expansion and free recoil: large deformation, plasticity, contact; about two hours per solveTosca’s example: linear elastic, two load cases; under a secondOne crown opened and released: plasticity; one to two minutes
It receivesStrain per crown, link and ring; lumen, shortening, volumeStress along the edge of the hole, both load casesPlastic strain along the crown, edge by edge
Full resultsEvery node and element of every frame, in a Python sandboxThe sameThe solver’s output files
Objective and limits90th percentile of the crown strains; lumen, shortening, volume, length, build rulesPeak von Mises stress at the hole; material areaMean strain of the most strained 2% of the metal; opening, shortening, volume
Budget15 simulations, the start’s included15, the start’s included13, the start’s included
LLMsGPT-6 AstraGPT-6 Astra, Claude Fable 5.2, Kimi K3.1 Pro, Mistral Large 4GPT-6 Astra, Mistral Large 4, Kimi K3.1 Pro, DeepSeek V4.1 Flash, Qwen 3.5 397B, Qwen 3.8

Table 1. Input and output of the agent for each problem. The plate runs through the same tools and geometry code as the stent; the crown ran in a lighter implementation of the same protocol, without sandbox or CAD (Section D).

Design files as agent output

The agent writes design files: short JSON files that set the values of the problem’s geometry parametrisation. The parametrisation names the features of the part and what may change about them, with a range for every value and rules that keep every design buildable. A file lists only what differs from shared defaults, and anything left out takes the shared value. A parametrisation can be as small as a few shared parameters, the form a classical optimiser needs, or reach down to single features of the part; the stent’s offers both, about 1 650 values in all. The problem’s design code, which the agent can read, turns a file into a CAD solid (§ 4.4). Before it spends a simulation, the agent can check a file for free: the check applies the rules, builds the geometry and reports its measures, such as the metal volume, and what changed.

Rounds of candidate designs

A session runs in rounds. In each round the agent submits one to three candidate designs at once, each with its reasons and its predictions of the main results. The pipeline simulates them, side by side as far as the solver licences allow, and the agent receives the results of all of them together before it proposes the next round. A candidate must pass the free check and must not repeat a design already simulated; a round with a rejected candidate is returned as a whole and costs nothing. Every candidate uses one simulation of the budget, which also counts the simulation of the start design. The agent decides how many candidates to submit, up to three. Several candidates at once let it compare alternatives on the same footing: different concepts in an early round, variants of one idea later, or a single candidate for a last correction. They can also shorten a study when they run side by side; in the clean run they mostly ran one after another, as the licences we share with colleagues rarely left room for two (Section B). Its predictions show whether it expects what the simulations then give (§ 5.3). The session ends when the budget is spent or the agent finishes, returning its best design, a summary and the changes it wanted but could not express.

Pipeline and protocol

A pipeline turns each submitted design into a scored simulation: it builds the geometry, meshes it, writes the solver deck with the problem’s set-up, solves it, exports the fields and computes the digest; on the stent it meshes and solves with Abaqus [12]. A scoring script, which we call the judge, then rates the design against the objective and the limits. A problem enters the loop through four parts: the geometry parametrisation with its code, the simulation set-up, the digest, and the objective with its limits. They are the orange boxes of Figure 2, where the digest shares a box with the judge that applies the objective and the limits. The agent’s tools and the rules below are the same for the stent and the plate; the crown ran in a lighter implementation with a smaller budget (Table 1). On the stent we use GPT-6 Astra through OpenRouter, at its highest reasoning effort in every session with tools (Table 5). It ran there twice: in pilot sessions that helped to develop the loop, over which the geometry parametrisation grew at the agent’s request, and in a clean run that tested the finished loop, with the complete parametrisation from the first call and nothing from the pilot sessions (§ 5.1).

Table 5: The agent’s runs and the settings of their protocols; changes during a run are given in the text. Reasoning effort as sent to OpenRouter; the crown’s LLMs ran at temperature 0.2 with their default reasoning. Budget: simulations of the agent’s designs, beyond the start design’s own; checks: free design checks. Caps and costs in US$ of LLM calls, per session unless stated; a turn is one LLM call. Costs as billed, all sessions of a row together; the single call’s include its three variants, and the crown’s its set-up sessions, which were billed together.

RunLLMBudget and toolsCapsCost
*Stent: pilot sessions (§ 5.1)
Single callGPT-6 Astra, highOne reply, no tools, whose design is simulated; also three one-call variants (§ 5.3)6 per call, 15 in all2.86
Four shared settingsGPT-6 Astra, highThree proposals, one call each, each after the results of the one before6 per call, 15 in all1.69
Crown by crownGPT-6 Astra, maxThree simulations, one design at a time; 40 checks; tools as in Table 6150; 12 per call; 150 turns7.31
Edge profilesGPT-6 Astra, maxNine simulations in rounds of up to three; 60 checks150; 12 per call; 150 turns17.47
Stent: the finished loop
Clean runGPT-6 Astra, max14 simulations in rounds of up to three; 60 checks150; 12 per call; 150 turns19.44
Shuffled-results controlGPT-6 Astra, maxEight continuations of the clean run, each ending at its next round, whose designs are built but not simulated25 per continuation, 60 per arm; 20 turns14.99
*Plate (Section C)
On the judge’s fine modelGPT-6 Astra, Claude Fable 5.2, Kimi K3.1 Pro, Mistral Large 4, max; three runs each14 simulations in rounds of up to three; 60 checks200; 25 per call; 300 turns243.60
On Tosca’s coarse modelThe same, but Claude Fable 5.2 one runThe sameThe same48.36
*Crown (Section D)
Crown sessionsSix LLMs, 20 sessions12 simulations in rounds of up to three, after a critic’s review; no sandbox, no CAD0.25 to 30 by LLM and round; 200 or 1 000 steps; 150 or 240 minutes98.61
Set-up sessionsSix LLMs, 20 sessionsThe LLM writes the study, then 12 simulations in rounds of up to three0.50 to 30 by LLM and round; 200 or 1 000 steps; 150 minutes

Table 5. The agent’s runs and the settings of their protocols; changes during a run are given in the text. Reasoning effort as sent to OpenRouter; the crown’s LLMs ran at temperature 0.2 with their default reasoning. Budget: simulations of the agent’s designs, beyond the start design’s own; checks: free design checks. Caps and costs in US$ of LLM calls, per session unless stated; a turn is one LLM call. Costs as billed, all sessions of a row together; the single call’s include its three variants, and the crown’s its set-up sessions, which were billed together.

Before the first LLM call of each run we fix its protocol (Section A): the LLM, the opening message, the tools, the budget, the cost caps and the questions the run is to answer. Every message, tool call, result and cost is recorded; a run can be paused and resumed without repeating an LLM call or a simulation; and no human intervenes during a session. On budgets of a few simulations, small choices can decide a comparison, so the protocol and the judge close four loopholes. (i) The objective, the limits and their tolerances are fixed before the first call; a reading chosen after results exist, such as a looser tolerance, is reported beside the fixed one and marked as such. (ii) A design whose build, mesh or solve fails uses up its simulation and is never feasible, so no method gains by proposing designs that break the pipeline. (iii) Every limit is an absolute value or the start design’s own value, never a percentage, which a design can meet by changing the denominator: a longer stent, for instance, shortens less in percent without any real improvement (§ 5.3). (iv) Two designs count as different only if their scores differ by more than the noise of a rebuild, the change in score when the same design is built and meshed again (§ 5.1); likewise, what would count as a result across repeated runs, such as a lead in at least two of three, is fixed with the protocol.

Stent design problem

This section explains in detail the headline non-linear problem the agent worked on. We first describe what a stent does, why simulating its deployment is demanding, and which reduced model we optimise on (§ 4.1), and outline that finite-element model (§ 4.2), with its details in Section B. We then state what the redesign should achieve, as an objective and limits (§ 4.3). Last comes the geometry parametrisation, through which the agent changes the stent’s CAD model (§ 4.4).

Stents and their simulation

A stent is a small tubular scaffold that holds a narrowed vessel open. A balloon-expandable stent is cut from a thin-walled tube as rings of zigzag struts joined by links. It is crimped onto the balloon of a catheter, carried to the narrowing and expanded by inflating the balloon, which bends the struts open at their crowns until the stent presses against the vessel wall. When the balloon is deflated, the stent springs back a little (its recoil) and must still hold the vessel open (Figure 3). The stent of this paper is cut from a tube 16 mm across, crimped to about 10 mm and expanded to about 23 mm, all as outer diameters. It is made of magnesium, which dissolves in the body over time, but magnesium alloys are much less ductile than the alloys of permanent stents, so the strain that deployment puts into the crowns largely decides whether the stent is damaged [41, 48].

Figure 3. Deployment of the original design in the full-cycle model of Steglich and Besson [41], with the parameters of the magnesium alloy ZX10: (a) as cut, (b) crimped onto the balloon, (c) expanded by the balloon and (d) after the balloon’s release. Side views at one scale, axis upright, front half only, each with the stent’s outer diameter; the crimper and the balloon are not drawn. Colour: cumulative plastic strain, each element’s largest integration-point value. Bottom: the steps each model runs, with its run time on one or two CPUs. The reduced model used in the optimisation (§ 4.2) covers only (c) and (d): it leaves out crimping, and a rigid cylinder expands the as-cut stent until the cylinder is 21.6 mm across and is withdrawn.

Simulating a deployment is demanding. The struts deform far beyond the elastic range, and the plastic strain concentrates in the bends of the crowns; the stent touches the crimper, the balloon and itself; it springs back after each step; and a magnesium alloy behaves differently in tension and compression and accumulates damage, which matters because crimping and expansion bend the crowns in opposite directions. A full simulation therefore follows the whole deployment, crimping, release, expansion and release, with a material model calibrated on cyclic tests that captures load reversals and damage. Steglich and Besson [41] built such a model for this stent; a run took 17 hours for the original design and 39 hours for one of the agent’s designs, on one or two CPUs (Section B).

For the optimisation we use a reduced model that solves in about two hours on one CPU. It leaves out crimping: it expands the stent from its as-cut state and lets it recoil, with plasticity but without load reversals or damage (Figure 3, bottom). It keeps what makes the problem hard: large deformation, plasticity concentrated in the crowns, contact and recoil, all strongly nonlinear in the design. Whether the reduced model is a reliable surrogate for the full deployment is beyond the scope of this paper; we check one of the agent’s designs in the full model (§ 5.2).

Finite element model

The original design is the stent of Steglich and Besson [41], derived from the Open Stent Design [5]: five rings of twelve crowns each, joined by eight links that sit on sixteen of the 60 crowns, cut from a tube 39 mm long with a wall 0.6 mm thick (Figure 4). In the reduced model a rigid cylinder stands in for the balloon: it opens the stent from its as-cut state until the cylinder is 21.6 mm across, which takes the stent to an outer diameter of about 22.9 mm, and is withdrawn, so that the stent recoils freely. The magnesium alloy, Mg-4Y-3Gd, follows a plasticity model fitted to one tensile test [42], without damage and with the same response in tension and compression; the full-cycle runs of Figure 3 and § 5.2 use the parameters of another magnesium alloy, ZX10. The two differ for a practical reason: ZX10’s model is a custom material model, calibrated on cyclic tests, so the optimisation uses a standard plasticity model of an alloy whose test data are public. Three points on the middle ring, the supports, keep the stent from sliding or turning as a whole while leaving it free to expand; they load the six crowns next to them artificially (§ 4.3). Each design is meshed afresh with about 35 000 brick elements and solved in

Figure 4. The stent’s geometry parametrisation, from a file to the solid. (a) An example design file as the agent writes it (its schema tag omitted): shared settings, where crown_offset_mm is the tip narrowing dd, an entry for one crown with two bumps (red) and an entry for one link (blue). (b) The cut pattern it describes, unrolled onto the plane at equal scale. Crowns are named (ring, θ\theta) and links (gap, θ\theta), where gap gg lies between rings gg and g+1g + 1. (c) The solid that FreeCAD builds from the file, with the same crown and link coloured. (d) The shared settings at one crown: strut width ww, tip width w−dw - d and leg length LL. (e) The two bumps of crown (2, 150°): material added on the outer edge and removed from the inner edge. (f) The link in gap 1 at 120°, with a narrower and longer waist, shifted by 0.20 mm around the tube. Dashed in (e) and (f): the same file without its crown and link entries.

Abaqus/Standard on one CPU core of Strand, the HPC cluster at Hereon, in about two hours (1.9 to 2.3 hours in the clean run; Section B). Not modelled are crimping, load reversals, damage, the vessel, balloon folding and manufacturing tolerances. The digest the agent receives lists, per crown, link and ring, the strains and motions, together with the lumen, the shortening and the volume, which § 4.3 defines. Section B gives the material data, the supports, the contact and the mesh.

Objective and limits

The redesign should lower the strain in the crowns, where the stent deforms most and where damage begins, without giving up what the stent must do: hold the vessel open as wide as the original design, shorten no more as it expands, and use no more metal. In agreement with established practice [25, 27, 38, 53], we set the objective as a high percentile of the strain rather than its peak: QQ is the 90th percentile of the strains of the 60 crowns after recoil; lower is better. A crown’s strain is the mean of the five most strained elements that belong to it, each element read at its most strained integration point (Section B). A single peak depends strongly on the mesh and on the supports (Section B), whereas QQ is exceeded by only six crowns and can be lowered only by bringing all the others below the new level; it follows the most strained crowns as a group. Since QQ does not see how far those six lie above it, we also read the most strained crown and the most strained element of each best design (Table 3). And since the supports load the six crowns next to them artificially, QexQ_{\mathrm{ex}}, the objective without those six crowns, checks for that artefact. For the original design Qex=0.1604Q_{\mathrm{ex}} = 0.1604.

Best valid design ofQQMost strained elementMost strained crownMedian crownBeyond the data
Agent, clean run−32%−32%−32%−31%0
Tosca, from the original design−10%+5%−3%−11%0
Tosca, from the agent’s one-shot fix−20%−17%−21%−20%0

Table 3. Each method’s best valid design on the stent on five measures: the change against the original design, solved in the method’s own pipeline (lower is better), and the number of elements strained beyond the end of the material data. For Tosca’s run from the agent’s one-shot fix, the change includes the fix’s own gain.

A design is feasible if (i) its lumen, the mean inner diameter after recoil, is at least the original’s minus 0.01 mm (21.249 mm); (ii) its shortening, the as-cut length minus the length after recoil, is at most the original’s plus 0.03 mm (1.3605 mm); (iii) its metal volume, computed exactly from the CAD solid, is at most the original’s (229.19 mm3^{3}); (iv) its as-cut length is at most 41 mm; and (v) it obeys the rules of the geometry parametrisation (Table 2). A design that cannot be built, meshed or solved counts as infeasible, with Q=0.25Q = 0.25. The limits are engineering choices rather than clinical criteria; shortening, or foreshortening, for instance, is reported in regulatory testing without a pass–fail value [45].

LevelControlRange
SharedStrut width ww0.76–1.10 mm
Tip narrowing dd0–0.40 mm
Leg length LL0.40–1.40 mm
Ring pitch, once or per gap7.60–8.75 mm
Link waist width0.57–1.20 mm
Link flat0–2.0 mm
Per crown (60), named (ring, θ\theta)Width0.76–1.20 mm
Tip narrowing−0.30 to 0.63 mm
Extension along the axis−0.30 to 0.40 mm
Edge profile: up to six bumps, each on the inner or outer edge, with centre φc\varphi_c, half width Δ\Delta and amount aa∣φc∣≤80∘\lvert\varphi_c\rvert\le80^\circ, 5∘≤Δ≤60∘5^\circ\le\Delta\le60^\circ, ∣a∣≤0.30\lvert a\rvert\le0.30 mm
Per link (8), named (gap, θ\theta)Waist width0.57–1.20 mm
Flat0–3.0 mm
Shift around the tube−0.50 to 0.50 mm
FixedThe tube (15.99 mm outer diameter, 0.593 mm wall), five rings of twelve crowns, and the sixteen crowns that carry the eight links
RulesTip and local crown width at least 0.57 mm; straight part of every leg at least 0.2 mm; link flank radii at least 0.3 mm; facing unlinked tips at least 0.5 mm apart; link attachments at least 3∘3^\circ inside the crown arcs; bumps within ±85∘\pm85^\circ of the apex, ending 2∘2^\circ before the legs and 5∘5^\circ clear of a link attachment; as-cut length at most 41 mm; metal volume at most the original’s

Table 2. The stent’s geometry parametrisation: the values a design file can set, and the rules every design must pass before it is simulated. Angles θ\theta are measured around the tube, angles φ\varphi from a crown’s apex.

The shortening limit has three readings, and every design is judged with each of them. The strict reading is the original’s own value, 1.3305 mm. The main reading is (ii), with its tolerance of 0.03 mm. That tolerance, like the 0.01 mm on the lumen in (i), covers the noise of the set-up: moving only the supports changed the original’s lumen by 0.0009 mm and its shortening by 0.028 mm. A rebuild, with a new CAD solid and a new mesh, scatters the shortening by more than its tolerance (Section B). The comparison between methods therefore uses a third, the head-to-head reading: the judge counts a design as breaking the shortening limit only if it exceeds the main limit by more than the noise of a rebuild, 0.17 mm, that is, beyond 1.5305 mm. The judge scores Tosca’s designs against the same limits as far as they apply. Tosca’s designs are reshaped meshes, not files of the geometry parametrisation, so their metal volume is read from the mesh, scaled to the start design’s exact volume, and of the rules in (v) only the two that can be measured on a mesh are checked, the width of the narrowest crown tip and the flank radii of the links. Tosca itself sees none of these limits except the volume, which it holds at its start’s value (§ 5.4).

Geometry parametrisation

The geometry parametrisation is how the agent changes the stent’s CAD model (§ 3.2). The agent writes values for the stent’s named features, so the parametrisation decides what it can change. It has three levels (Table 2). The shared settings are the strut width ww; the tip narrowing dd, which leaves a tip width w−dw - d; the length LL of the straight legs; the ring pitch, set once or gap by gap; and the waist width and flat of the links (Figure 4d). Each crown can take its own width and tip narrowing and move along the axis, which lengthens its two legs when it moves outwards. Each link can take its own waist width and flat, the length of its straight middle part, and move around the tube (Figure 4f). Still at the level of the crown, edge profiles add up to six bumps to its inner or outer edge, with the angle φ\varphi measured from the crown’s apex (Figure 4e). Crowns are named by ring and angle around the tube, links by the gap they bridge and their angle (Figure 4b), the names under which the agent also receives its results. The tube, the five rings of twelve crowns and the sixteen crowns that carry the eight links are fixed. In all, a design can set up to about 1 650 values, most of them in edge profiles. A parameter study would use four of the shared settings: the strut width, the tip narrowing, the leg length and the ring pitch.

The design code draws each crown as a U-bend between an inner and an outer circular arc. The legs are common tangents of neighbouring arcs, and each link is an hourglass whose flanks are arcs tangent to its crown and to its waist, so every outline is tangent-continuous. At the angle φ\varphi along an arc, a bump with centre φc\varphi_c, half width Δ\Delta and amount aa moves the edge along its normal by a(1−t2)3a(1 - t^2)^3, where t=(φ−φc)/Δt = (\varphi- \varphi_c)/\Delta and ∣t∣<1|t| < 1, so the edge returns to its arc with continuous curvature; a positive amount adds material and a negative amount removes it. The outlines lie on the unrolled outer surface, with the axial position and the arc length as coordinates (Figure 4b). FreeCAD wraps them onto the outer surface of the tube as B-spline curves, extends them radially to the inner surface as ruled walls, splits the tube with all the walls and keeps the piece that holds the struts (Figure 4c). The metal volume follows exactly from the outlines, and a build counts only if it gives a single valid solid whose finely tessellated volume matches the exact volume to 0.05% and which survives export to STEP and re-import. Tests confirm that a bump changes only its own edge and adds the volume of its integral to within 2%. The plate with a hole of § 5.5 uses the same code. The parametrisation grew over pilot sessions at the agent’s request (§ 5.1 and § 5.3); in the clean run it is complete from the first call.

Results

Stent optimisation by the agent

The agent worked on the stent in two runs with different roles (Figure 5). The pilot sessions were part of developing the loop. The agent worked on the stent while the loop was still being built, and between its sessions we extended the geometry parametrisation to cover the changes it asked for but could not yet express: after a single call and a round on the four shared settings came crown-by-crown edits with the sandbox, and then edge profiles (Figure 5a). Along the way its designs lowered the objective QQ, the 90th percentile of the crown strains, from the original design’s 0.1604 to 0.1181 with its fifteenth design, 26% less.

The clean run then tested the finished loop. The agent received it complete from the first call, with the full geometry parametrisation, the sandbox and the tools, and nothing from the pilot sessions. It lowered QQ from 0.1604 to 0.1085 with its thirteenth design, 32% less, valid under every reading of the limits. It spent its fourteen simulations in six rounds, of three, three, three, three, one and one designs. Three quarters of the gain came from its first round, three designs proposed together from the original design’s results alone, the best of which cut QQ by 24% (Figure 5b). The later rounds added eight percentage points in steps of up to 0.006, and the last two were pure polishing. Thirteen of its fourteen designs were valid under the head-to-head reading. The agent picked its best design itself. Its LLM calls cost US$19.44, against about two hours of cluster time for each of its fifteen simulations. The cost was low because the run needed only 30 calls and the LLM wrote little, building its design files with short Python programs: all it wrote cost US$1.81, and the rest paid for re-reading the conversation, which grew with every round. Only GPT-6 Astra ran on the stent; the two smaller problems carry the comparison between LLMs (§ 5.5).

Verification. Every design gets its own CAD model and a new mesh, so even two practically identical designs do not score exactly the same. To measure how much they differ, we rebuilt four designs with their struts a thousandth of a millimetre thinner, a change too small to matter, and simulated them again: their scores changed by at most 0.003. We take 0.004, 2.5% of the original design’s score, as the noise of a rebuild: two designs count as different only if their scores differ by more, which rebuilds of the same design rarely do (Section B). The clean run’s best design, rebuilt once more, scored 0.1088 instead of 0.1085 and still met every limit. The agent’s gains are an order of magnitude larger than this noise, 32% in the clean run and 26% in the pilot sessions, and even the clean run’s lead over the pilot sessions is more than twice it.

Validation in the high-fidelity full-cycle model

The full-cycle model of Steglich and Besson [41] follows the whole deployment of this stent (§ 4.1 and Figure 3), with a Cazacu–Plunkett–Barlat yield criterion for the tension–compression asymmetry of magnesium, kinematic hardening for load reversals and Gurson–Tvergaard–Needleman damage, calibrated on cyclic tests. One of the agent’s pilot designs, depicted in the Figure 6, was run through through crimping and expansion with the parameters of ZX10, the alloy that in their study accumulates damage fastest and fails before full deployment; both expansions come from the same model set-up (Figure 7).

Figure 6. Geometric differences between the agent’s design and the original, as cut. This design gives all crowns of a kind the same shape. Grey: metal in both designs; blue: only in the original; red: only in the pilot design. (a) The 44 unlinked crowns: thinner struts, a larger inner radius and, from two bumps on the outer edge, a wider tip; crosses: where each design’s most damaged element sits in the full-cycle model (Figure 8). (b) The 16 linked crowns: thinner struts, with two bumps on the inner edge that widen the shoulders; the link is unchanged.

Figure 7. Damage of the agent’s pilot design and the original in the full-cycle model of Steglich and Besson [41] with ZX10: both crimped, released and expanded, the two expansions from the same model set-up. The porosity of each design’s most damaged element at full expansion, followed through the expansion. Dots: where it reaches 0.3%, the onset of void coalescence; right: the values at full expansion.

After the full chain, the crowns of the agent’s design were less strained throughout: the 90th percentile of its crown strains was 14% lower and its most strained crown 24% lower. Its most damaged spot reached the porosity at which void coalescence begins later, at a stent diameter of 16.1 mm instead of 15.2 mm, and carried less damage at equal diameter: a third less at 20 mm, about half at 22 mm, and 63% less at full expansion, 1.2% against 3.4%. In both designs that spot sits on the inner edge of a crown’s bend, on the cut wall (Figure 8). There the two designs differ most (Figure 6): the agent’s design has thinner struts, 0.78 against 0.96 mm, but bumps on the outer edge widen the tips of its unlinked crowns from 0.785 to 0.935 mm.

Figure 8. Damage fields at full expansion, in the model of Figure 7: (a) the original, (b) the agent’s pilot design, each element coloured by its porosity read as there, on one scale; the darkest blue is below 0.3%, the onset of void coalescence. Left: the half of the stent around its most damaged element, seen from outside; circle: that element’s crown. Right: that crown, looking into its bend onto the cut wall where the element sits; arrow: the most damaged element and its porosity.

Our score targets the most strained crowns, not the amount of strained metal: as the design spread the strain from its peaks, more of its metal passed the coalescence porosity by full expansion, 8.6% against 7.0% of the stent. After crimping and release alone, the design was about 12% less strained than the original but needed about 16% more crimping force, and facing crown tips touched at full crimp.

The agent’s approach to the redesign

This subsection asks three questions of the clean run: how much of the design the agent changed, whether its changes redesign the stent or merely reshape it, and how they draw on prior knowledge and on the simulation results, from which the agent learned as the run went on.

Extent of the changes. The parametrisation lets a design set about 1 650 values, because each of the 60 crowns can be given its own shape. From its third design on, the agent set about 800 of them, and from its tenth about 1 200; over the whole run it set 1 263 of them at least once, nearly all in the edge profiles of the crowns (Figure 9). It did not set them one by one. Most designs gave the crowns only two or three shapes, each shared by a whole family of crowns: one for the unlinked crowns, one for the linked crowns and usually a third for the twelve crowns at the stent’s ends. The many values come from writing each shape into every crown of its family; written down one shape at a time, each as its differences from one already written, such a design takes only 27 to 40 values. Three designs also varied single crowns: one in the middle of the run varied the crowns’ extension along the stent one by one, and the last two corrected single crowns, the best design 20 of them, mostly with an extra bump of 0.01 to 0.02 mm. The design validated in the full-cycle model kept to two shapes, one for the unlinked and one for the linked crowns (Figure 6). Choosing such a structure is what a person does before handing a classical optimiser a handful of settings. The size of the changes followed the course of the run: the first round tried three different concepts, the middle rounds refined one of them, mostly by changing hundreds of values per design, and the last two rounds changed 95 and 201 values but moved no crown edge by more than 0.04 mm, too little to change QQ beyond the noise of a rebuild.

Figure 9. Number of design values the agent set in its clean run, out of the stent’s about 1 650, design by design. Light: the values in which a design differs from the original; dark, with their number: those in which it differs from the design submitted before it, within a round its sibling. Shaded: the six rounds. Dashed: all values of the geometry parametrisation; dotted: the values set at least once over the run, nearly all of them in edge profiles. Logarithmic scale.

Redesign rather than reshaping. We call a change a redesign when it alters a design’s proportions, such as the width of the crown tips against the struts or the length of the legs, and a reshaping, or polishing, when it moves the surface locally to follow the current strain peaks within the given proportions, as a shape optimiser does (§ 5.4). Neither changes the layout of rings, crowns and links, which the parametrisation fixes. The original design needs a redesign: its crown tips are almost a fifth narrower than its struts, so each crown bends mostly at its tip, and it sits exactly on its volume and shortening limits. The agent’s first round compared three redesigns: much longer legs, which lowered QQ by 15% but broke the lumen limit; far less narrowed tips with shorter legs, which made QQ worse; and new edge profiles on every crown with the eight links changed, which lowered QQ by 24%. The middle rounds kept the third concept and lengthened the legs, first by 13% and then by 21%. The best design therefore differs from the original in its proportions: it adds 0.11 mm of metal at the tip of every unlinked crown and 0.07 mm on the inner edges of the linked crowns, takes it from the shoulders, and has legs a fifth longer, while its metal volume stays within 0.3% of its limit and its lumen ends 0.003 mm inside its limit. Only the last two rounds reshaped. In the agent’s own summary: “The largest improvement came from spreading crown bending, not simply thinning struts. Longer legs reduced strain but consumed recoil-diameter margin; excessive shoulder relief created new strain peaks.”

Prior knowledge in the first move. In the pilot sessions, a single call to the LLM with the original design’s results proposed uniform 0.86 mm struts without the narrowed tip, the one-shot fix from which Tosca also started (§ 5.4); it lowered QQ by 7.5%, as the LLM had predicted. Three variants of that call, one call each, show where the idea came from. Without the table of per-crown results, the LLM proposed the same kind of change, about half as far. With an anonymised prompt, the same numbers without the physics, it moved a link away from a supported crown: it dodged an artefact of the supports instead of improving the design, with no gain and 66% more shortening. A repeat of the original call gave the same core change. The first move therefore came from prior knowledge, and the results read without the physics led the LLM astray. Two of the three answers that knew the part was a stent also lengthened it, which lowers its shortening as a percentage without any real improvement; this is why the shortening limit is absolute (§ 4.3).

Use of the simulation results. Later in the run the agent corrected single crowns, guided by strain peaks it found in the element fields. Whether such changes come from the simulation results is what the shuffled-results control tests, following the misdirected-prompt ablation of [47]. We replayed the clean run up to the results of a round and let the agent continue twice from the real results and twice from the same results shuffled between crowns of the same kind, with each crown’s element fields moved along with its numbers: the results look plausible, but the most strained crowns appear in the wrong places. Each continuation ends with its next design, which is built but not simulated, and the measures were fixed before any design was seen. Started after the first round, the agent expected the strain where the shown results put it: of the 18 crowns it predicted to be the most strained, three for each of its six designs, 14 were among the five that the shuffled results showed as most strained and only one among the five that the simulation had actually found most strained, against 12 of 18 with the real results. Its design changes in that round answered what the shuffle leaves intact, such as the broken shortening limit, so, as stated in advance, they could not differ between the variants. Started before the best design, where the clean run corrected single crowns, the variants differ in the designs themselves: with the real results, both continuations aimed their changes at the crowns that the simulation had found most strained, one widening crowns in line with their strain as the clean run did, the other lengthening three of the five most strained; with the shuffled results, neither did. The prompt, the geometry and the earlier rounds are the same in both variants, so these crown-level changes come from the per-crown results, not from prior knowledge. Two smaller decisions also followed the shuffled numbers to their new places: where the agent took margin on the stent’s diameter and, in one continuation, which linked crowns it reinforced. Its main reinforcement, however, did not follow the shuffled results to the crowns shown as most strained; instead it changed shared settings, shorter legs and once wider struts, for reasons the control does not show. With two continuations per variant and one round each, the control shows a consistent pattern, not a statistic.

Learning within the run. The agent’s predictions of its own results sharpened as the run went on. Its predicted QQ missed by 0.010 to 0.020 in the first round, two to five times the noise of a rebuild, and by at most 0.0035 from the fourth round on, below that noise, though it stayed slightly optimistic: 12 of its 14 predictions were below the outcome. Its predicted lumen missed by 0.03 to 0.045 mm in the first round, several times the lumen’s tolerance, and by at most 0.005 mm from the fourth, which let it end 0.003 mm inside the lumen limit. Its shortening predictions came down from misses of 0.09 to 0.13 mm to about the scatter of a rebuild.

Comparison with the industrial optimiser Tosca

Choice of comparison. Classical optimisers cannot work on all of the stent’s values within the budget (§ 1 and § 2). That leaves industrial shape optimisers, of which we ran Tosca [13]. It moves all the surface nodes of the mesh, about 17 600 on the stent, ten times the agent’s number of values, but it optimises what the solver computes on its node and element groups rather than the judge’s score and build rules, and its moves accumulate in one mesh rather than a CAD model, so it cannot change proportions such as the leg length. Of its modules, shape optimisation is the one for the task; topology optimisation would only remove metal inside the given struts or search a new layout from scratch.

Set-up. We ran Tosca with its standard controller settings on the original design’s mesh, with fifteen simulations and the same judge as the agent, from two starts: the original design, and the agent’s one-shot fix, a design with uniform crowns that the LLM proposed in a single call from the original design’s results (§ 5.3). Tosca moves the nodes on the rings’ cut faces within the tube surface; the links and patches of 0.5 mm around the links and supports stay fixed, and before each solve every node is returned to its radius, so the wall keeps its thickness. Its target is the peak von Mises stress on the movable faces at full expansion, with the metal volume held equal to the start’s. Tosca does not see the lumen and shortening limits; its designs are judged like the agent’s, and those that break a limit do not count. These choices follow Tosca’s installed examples and are ours, not limits of the tool, and no run was tuned after its results. Two set-ups closer to our score, with a plastic-strain target, the sensitivity-based method and the lumen, shortening and volume limits, did not beat the original design in the simulations they ran, so we report the out-of-the-box set-up, whose designs did.

Start from the original design. Tosca lowered QQ, the 90th percentile of the crown strains, by 10% within two updates, at its third simulation (0.1444, valid only with the shortening tolerance), and then drifted (Figure 10). Its own target kept falling, to 39% below the start, while the strain piled up in the fixed patches next to the reshaped regions, at the links and supports: the most strained element reached an equivalent plastic strain of 0.47, about 2.7 times the end of the material data. Beyond that end the material model holds the stress constant, so a stress target gets no signal, and the fixed patches lie outside Tosca’s target. QQ, set by the most strained tenth of the crowns, stayed between 0.146 and 0.155 for the rest of the run.

Figure 10. Optimisation progress on the stent: the score QQ of every design, the 90th percentile of its 60 crown strains (lower is better), against simulations, all judged by the head-to-head reading of the limits (§ 4.3). Each marker is one simulated design, filled if it meets every limit and hollow if it breaks one; the line is the best score among the valid designs so far, as in Figure 5. From the original design, fifteen simulations each: the agent’s clean run, with its rounds R1–R6, and Tosca with its out-of-the-box set-up (§ 5.4). The agent’s first simulation is the original design. Green: Tosca with the same set-up from the agent’s one-shot fix, a design with uniform crowns, whose own simulation is its first; one run of fifteen simulations, capped after five to free the cluster and resumed by Tosca’s own restart.

Start from a design reshaped by the agent. Started from the agent’s one-shot fix, whose crowns are uniform, Tosca lowered QQ by 14% within four simulations, steadily, and then held that level for its other ten without drifting: its best design, at its fifteenth simulation, scored 0.1276 against 0.1279 at its fifth, a fourteenth of the noise. Every design was valid under every reading of the limits (Figure 10, green). From the original design’s 0.1604, the agent’s fix and Tosca’s polishing together reach 20% lower, the agent’s clean run alone 32%. Tosca added metal where the agent’s edge profiles had added it in the pilot sessions, from a similar start with uniform crowns; the two patterns of crown width correlate at 0.96. A third of its shape change was on the legs: it thinned every leg, by 0.24 mm on average from 0.86 mm, and moved that metal onto the crown arcs. The agent’s geometry parametrisation cannot make that move, since a leg takes its crown’s width; independent leg widths were the first change the agent asked for at the end of its clean run (§ 6). The run was capped at five simulations to free the cluster, a decision taken before any of its designs was scored, and resumed to fifteen by Tosca’s own restart; we report it as one run. We report Tosca’s designs as its own meshes give them, but they are reshaped meshes, not rebuilt designs, so we checked its best design on a fresh mesh: rebuilt as CAD from its own outline, each edge a radial wall as a laser cut makes it, and meshed afresh, it scores 0.1295 against 0.1276, under half the noise apart, while its start, rebuilt the same way, moves only from 0.1484 to 0.1476. Measured on fresh meshes at both ends, the polish is 12% rather than 14%.

Other measures. The objective is ours, so we also read each method’s best valid design on four other measures, each against the original design (Table 3). The three strain measures leave out the six crowns next to the supports, as QexQ_{\mathrm{ex}} does: the original’s most strained element and crown sit there. The agent’s design leads on all of them. On the most strained element it is 32% below the original, while Tosca’s best design from the original is 5% above it. Tosca’s run from the agent’s one-shot fix ends 17 to 21% below the original design on every measure, the fix’s own gain included.

Interpretation and scope of the comparison. In the terms of § 5.3, Tosca’s node moves are polishing, and polishing needs a sound design with room to move. The original design offers neither: its crowns bend mostly at their narrowed tips; it sits exactly on its volume and shortening limits, which are its own values; and its most strained spots lie at the end of the material data, next to the fixed patches, where a stress target gets no signal. The agent’s one-shot fix changed these proportions, with uniform struts and no narrowed tips, and so left room on the volume and shortening limits and no element beyond the end of the material data. From that start, Tosca’s polishing worked. The agent did both, redesigning and polishing, and chose which one a design needed.

The stent itself does not favour Tosca: its objective and limits are measurements of the design, which the agent reads directly and Tosca only through stand-ins; a design is a CAD model, rebuilt and meshed afresh for every proposal, while Tosca’s moves pile up in one mesh; the problem is strongly nonlinear, starts on its limits and rewards coordinated changes of the proportions; and the agent sees the whole result, not a few responses and their sensitivities.

We therefore do not claim that Tosca cannot optimise this stent: an expert who would rerun it with a changed set-up, such as a target closer to the judge’s, gentler steps or a start with room on the limits, might well succeed.

To rule out a fault in how we run and judge Tosca, we also ran its own bundled example, a plate with a hole, an elastic part with a local stress peak for which Tosca is the natural tool: run as shipped, it reproduced the vendor’s result, and with fifteen simulations instead of five it lowered the peak stress by 36% on a common judge (§ 5.5, details in Section C).

Comparison across problem sizes

Table 4 steps down in the number of values, from the stent to two problems that enter the same loop (Table 1): Tosca’s own plate with a hole, with up to 26 values, and a single crown with four. All three have budgets of thirteen to fifteen simulations, the start’s included, so the number of design values per simulation falls from over a hundred to about two and to fewer than one.

Plate with a hole: Tosca’s own example. Tosca’s bundled shape example is a plate with a hole under two load cases, linear elastic, with the peak stress at the hole’s edge as the objective (Section C). Tosca ran it exactly as its vendor ships it, with fifteen solves instead of five, and reached 36% below the circle on a common judge. Bayesian optimisation, given the hole’s two semi-axes, stopped at an ellipse 27% below the circle. The agents’ geometry parametrisation adds up to eight smooth bumps on the hole’s edge, 26 values in all, more than the fourteen designs of the budget. Four LLMs ran three times each: GPT-6 Astra (40 to 45%), Claude Fable 5.2 (46 to 47%), and the open-weight Kimi K3.1 Pro (33 to 40%) and Mistral Large 4 (27 to 39%). Nine of the twelve runs beat Tosca, all of Astra’s and Fable’s among them, and the four best runs were the four that placed all eight bumps freely. Given Tosca’s own coarse model instead of the judge’s, seven of ten agent runs still beat it. On its home ground Tosca is a real competitor, which is what makes the agents’ lead count.

Single crown: fewer values than solves. On a single crown with four variables and twelve simulations after the start (Section D and Figure 12), random search found no feasible improvement and COBYLA at most 5.8%, but Bayesian optimisation reached 8.6 to 15.4% over five seeds. The best sessions of four of six LLMs lowered the peak strain by 19 to 22%, and every session that used its full budget was level with the median seed or ahead of it; but only one LLM did better than the median seed in two of its three sessions, the criterion fixed before the baselines ran.

The reduced crown problem: one unit of a zigzag ring and plastic strain after release for the start design, best Bayesian optimisation, and best LLM session

Figure 12. The reduced crown problem. (a) The model: one unit of a zigzag ring, unrolled into the plane, from the middle of one crown down its strut to the middle of the opposite crown, with its neighbours mirrored in grey. The four values set the width at the apex, waw_a, 45∘45^\circ from it, w45w_{45}, and along the strut, wsw_s, and an outward shift ss of the apex; both crowns take the same values. Dash-dotted: the symmetry planes, one held, the other moved by 0.5 mm and released. (b)–(d) The plastic strain after release in the upper crown, the box in (a); the lower crown mirrors it. (b) The start design. (c) The best design of Bayesian optimisation over its five seeds. (d) The best LLM design, by Mistral Large 4; the best designs of GPT-6 Astra, Kimi K3.1 Pro and DeepSeek V4.1 Flash are alike, with a wider apex and shoulder and a thinner strut. Outlined: the most strained 2% of the metal, whose mean strain is the objective; below each panel, its four values. Each design solved again from its own input, which reproduces its recorded result exactly.

Interpretation of the ladder. With fewer values than solves, a good classical optimiser can search the whole design space, and the LLMs’ lead becomes a matter of the run. Once the values outnumber the solves, the classical optimisers need a person to shrink the problem and stop where the shrunken space ends, and only the shape optimiser keeps up, on its home ground. At over a hundred values per solve, only the shape optimiser can take part, and from the original stent it could not redesign. The problems differ in more than their size, so the ladder is not a scaling law; it shows where each family can take part, and how it did.

Discussion

Sources of the agent’s advantage. The agent brings what the optimisers lack: physical knowledge for its first move, the full fields of every solve, and the ability to choose a structure and a scale for its moves. Its first idea came from prior knowledge; its later, crown-level moves came from the results it read, as the shuffled-results control shows, with predictions that sharpened round by round. Unlike LLM calibration, where an early lead faded within about fifteen evaluations [47], the agent kept improving over its fifteen solves. We attribute this to the number of values: with many values there is always another informed move to make, and the agent can choose how large it should be.

Position among automated methods. To our knowledge, the agent’s loop is the only automated one that meets all three conditions; every family of automated optimisers gives up one of them (Figure 1). The automated method that would also meet all three, a gradient method on the CAD values with adjoint gradients of the judge’s own measures through plasticity and contact (§ 2), does not exist off the shelf. Grey-box Bayesian optimisation [2], which models per-crown outputs rather than the score alone, is the closest machine-learning idea to reading the results, but it still scales with the number of design values.

Classical optimisers on the agent’s structure. The agent did not turn 1 650 knobs one by one: from its third design on it set about 800 to 1 200 values, in most designs as two or three crown shapes that take only 27 to 40 values to write down, a structure a classical optimiser could be given. But choosing that structure from the results of the original design is part of what the agent did; a classical optimiser needs a person to hand it one, and the agent changed its structure between designs, working on single crowns in three of them. Where every method had the same four numbers, Bayesian optimisation was a real competitor (Section D).

Division of labour with shape optimisers. The agent redesigns and polishes, and decides which one a design needs; Tosca’s shape optimisation polishes, and polished the agent’s first redesign well. A natural division of labour follows, as in hybrids of LLMs and optimisers [17, 47]: the agent finds where to act and a shape optimiser tunes how much.

Remaining human input. The set-up is kept out of the agent’s hands on purpose. When LLMs had to write a simulation set-up themselves, for a bar whose optimum is known, five of six produced a working study in most of their sessions, but the weakest accepted set-ups that ran and measured something else, and reported success (Section D). A new problem needs its four parts, of which the geometry parametrisation is the most work. The agent helps here, because at the end of each session it lists the changes it wanted but could not express. In the pilot sessions, after crown-by-crown editing, it asked for “Localized, smoothly blended shoulder reinforcement” and “Noncircular/variable-curvature inner crown profiles”, which the edge profiles of § 4.4 then provided; at the end of the clean run it asked for “Independent leg-width/curvature control, noncircular crown centerlines, curved link centerlines or different attachment topology, and through-thickness edge fillets”. A parametrisation can thus be extended where a problem needs it, between runs and never within one.

Limitations. The main limitation is that the agent changes the CAD model only through a geometry parametrisation written for this geometry. The parametrisation decides what the agent can change, and so bounds what it can find, and every new part needs its own (§ 4.4). The loop also takes time, so only one LLM, GPT-6 Astra, ran on the stent, and only twice: the clean run took 35 hours from its first simulation to its last (Section B). How much the result varies from run to run is therefore open. Of the considered three problems, only the stent is a nonlinear three-dimensional part; the crown is a planar model with four values, and the plate is linear elastic. The original stent derives from the Open Stent Design [5], an openly published research design rather than a state-of-the-art commercial one, so whether an agent can improve a state-of-the-art design remains open; as a test of the capability itself, a standard design serves. The shuffled-results control has two continuations per variant and one round each, and with shuffled results the agent’s main reinforcement did not move to the crowns shown as most strained (§ 5.3). Finally, our finite-element model covers the expansion alone; the full-cycle check supports its gains for one design but cannot establish them in general (§ 5.2).

Conclusion

We asked whether a general-purpose LLM, working in a loop with CAD and finite-element simulation, can redesign a nonlinear three-dimensional part within a budget of fifteen simulations. On a magnesium stent it did: through about 1 650 values of the stent’s CAD model, the agent lowered the strain objective by a third for US$19 of LLM calls, a gain an order of magnitude larger than the noise of a rebuild. A design from its pilot sessions held up in an independent, higher-fidelity model of the full deployment with crimping and damage, which the agent never saw: it was 14% less strained than the original, and its most damaged spot began to fail later and carried 63% less damage at full expansion. To our knowledge, this is the first stent designed by an LLM agent, and the agent’s loop is the only automated one that combines freedom to redesign with a high parameter count and a small simulation budget.

The agent did not search blindly. It chose its own structure, giving whole families of crowns the same shape and correcting single crowns only at the end, and the scale of each move, with different concepts first and polishing last. Its first move came from prior knowledge; its later, crown-level moves came from the results it read, as the shuffled-results control shows, and its predictions sharpened as the run went on. The comparisons place this capability. Tosca, the industrial shape optimiser that can work at this level of detail and budget, polished well but could not redesign: from the original design it gained a tenth and then stopped improving, while from a design the agent had already redesigned it reached 20% below the original. On Tosca’s own bundled example, run as shipped, agents built on four LLMs beat it in 9 of 12 runs, and on a single crown with fewer values than simulations a good Bayesian optimiser caught up with them. Agents thus lead where the design values outnumber the simulations, and they combine naturally with shape optimisers: the agent redesigns, the optimiser polishes.

Two limits remain. A person still writes the simulation set-up, because weaker LLMs accept wrong set-ups and report success, and every new part needs a geometry parametrisation through which the agent can change its CAD model; the agent eases the latter by naming the changes it could not express. Whether the capability carries over to other parts, to other LLMs on the stent and to designs that are already state of the art remains to be tested. Engineering CAD, with its hundreds to thousands of values, may be where reading the simulation pays most.

Acknowledgements

We thank Dirk Steglich (Helmholtz-Zentrum Hereon) for the stent that this work starts from, and for running the agent’s design and the original design through his full-cycle model of crimping and expansion [41] and sharing its results. We gratefully acknowledge the computing time on Strand, the HPC cluster of Helmholtz-Zentrum Hereon, operated by its department Logistics and Organization Coastal Research. This work was supported by the Digital Twins of Complex Systems (DToCS) Platform at Helmholtz-Zentrum Hereon, which is jointly funded by the German Federal Government and the state governments of Schleswig-Holstein, Hamburg, Lower Saxony, and Brandenburg.

Code and data availability

The geometry parametrisation and its checks, the agent’s tools and prompts, the scripts that score and plot the results, and the session records are available from the authors on reasonable request. The simulations need the commercial codes Abaqus and Tosca.

Protocols, prompts and tools

Table 1 lists, for each of the three problems, what the agent receives and writes, how its designs are simulated and judged, and its budget. This appendix gives the settings that each run’s protocol fixed, the system prompt, opening message and tools of the clean run on the stent, a round of its transcript, how the prompts and tools of the other runs differ, and what the records allow to be reproduced.

Protocols. The record of each run states what was fixed before its first LLM call: the LLM and its settings, the opening message, the tools, the budget, the cost caps, and the questions the run is to answer, with how each answer will be read. Table 5 lists the settings of every run of the agent in this paper. All LLMs were called through OpenRouter. On the stent and the plate, a session ends before any call whose worst case, its whole prompt and the longest reply allowed, could exceed the cap per call or what is left of the session’s cap; a turn is one call. Two thirds of the clean run’s cost went on the six calls that followed its rounds of simulations: the prompt cache does not last through hours of simulation, so each of these calls paid in full for reading the whole conversation. Changes made after a run had started are recorded with their dates and reasons. On the plate, during the first run of Claude Fable 5.2, the cap per call was raised from US$12 to US$25 and the turn limit from 150 to 300, for that session and every later one: the run needed about twelve turns per design and would otherwise have been stopped before its fourteenth design, and the budget compared between methods is the designs, not the turns. That run’s first 77 calls went without the prompt cache, which its provider applies only on request, so the run got a cap of US$260. Claude Fable’s second run first stopped at its third call, which the provider refused as overloaded on every retry; that attempt is kept in the records, and the run was repeated from the start. The crown’s amendments are described in Section D.

System prompt and opening message. Every session with tools on the stent and the plate had the same system prompt: “You are working on an engineering design task with the tools provided. Use them as you see fit and give your final answer with the finish tool.” The opening message states the problem in seven sections: how the design is generated, how a design file changes it, the simulation set-up, the start design with its results, the objective and the limits, the tools and the budget, and what to report with each design. It states facts only, with no hypothesis about the results and, in the clean run, nothing from the pilot sessions. The clean run’s opening message follows as it was sent, except that its table of the 60 crowns is cut to its first four rows. It gives the limits of the main reading (§ 4.3): the agent saw the verdicts of that reading only, not those of the strict or the head-to-head reading.

The clean run’s opening message
# A magnesium stent: edit the design crown by crown, with tools and 14 simulations

You get a stent design problem, its original design D0 and D0's recorded deployment. You can change the shared settings, single crowns and single links, including the shape of each crown's edges, and you work with tools: a Python sandbox with the full recorded fields of every simulated design, a free design check, and a budget of 14 deployment simulations, run in rounds of up to three at once. Section 6 describes the tools.

The design generator

A balloon-expandable stent is laser-cut from a magnesium-alloy tube: outer radius R=7.995125R = 7.995125 mm, wall thickness t=0.59285t = 0.59285 mm (inner radius 7.402275 mm). The cut pattern is drawn on the outer tube surface, unrolled (axial coordinate zz, arc length ss), and wrapped onto the tube with radial walls (every wall points at the tube axis).

  • 5 rings. Each ring is a serpentine of 12 semicircular crowns (6 at each ring end), joined by straight axial legs. Crowns alternate between the two ring ends every 360/12=30360/12 = 30 degrees around the tube.

  • Crown period along the surface: half = 2πR/12=4.1862382 \pi R / 12 = 4.186238 mm. For strut width ww, the crown inner radius is rin=(half−w)/2r_\mathrm{in} = (\text{half} - w)/2 and the outer radius rout=(half+w)/2r_\mathrm{out} = (\text{half} + w)/2.

  • crown_offset dd: each crown's outer arc is shifted by dd towards the crown opening, so the crown tip is narrower than its sides (tip width w−dw - d).

  • leg_length LL: straight leg between consecutive crowns, width ww.

  • Ring width (axial) = L−d+2routL - d + 2 r_\mathrm{out}. The ring_pitch pp sets the distance between ring centres; the axial gap between rings is pp minus the ring width.

  • Neighbouring rings are shifted by half a crown period, so crown tips face each other across each gap. Hourglass links join facing crown tips: link_crowns = ((2, 5), (0, 3)) lists the link-carrying crowns of the left ring per gap, cycled over the gaps; this gives two links per gap, 180 degrees apart, rotated by 90 degrees from one gap to the next. Link waist width wl=0.74315w_l = 0.74315 mm with a straight waist flat f=0.280963f = 0.280963 mm; each link flank is an arc tangent to the facing crown's outer arc and to the waist, with radius rlink=(dx2+(wl/2)2−rout2)2(rout−wl/2)r_\mathrm{link} = \frac{(d_x^2 + (w_l/2)^2 - r_\mathrm{out}^2)}{2(r_\mathrm{out} - w_l/2)}, where dx=p/2−(L/2−d/2)−f/2d_x = p/2 - (L/2 - d/2) - f/2.

  • The minus-zz crown tips of ring 0 are fixed at z=−27.792043z = -27.792043 mm; the first ring centre is at z=−27.792043+rout+L/2−d/2z = -27.792043 + r_\mathrm{out} + L/2 - d/2 and ring kk at that plus kpk p. As-cut length = 4p+4 p + ring width.

  • Metal volume V=A(R2−Rin2)/(2R)=0.570870∗AV = A (R^2 - R_\mathrm{in}^2)/(2R) = 0.570870 * A for the unrolled pattern area AA.

A design is a JSON file (schema "astra-direct-002/design-1"): shared settings plus optional entries for single crowns and links. A crown or link without an entry, and any field left out, takes the shared value. D0, the original design, is

{"schema": "astra-direct-002/design-1", "settings": {"strut_width_mm": 0.96314, "crown_offset_mm": 0.178238, "leg_length_mm": 0.773662, "ring_pitch_mm": 8.2498}, "crowns": [], "links": []}
  • Crowns are named (ring, theta_deg) and links (gap, theta_deg) exactly as in the result tables below (gap jj joins rings gg and g+1g+1; a link's theta is that of its crowns).

  • Crown fields: width_mm (strut width of the crown and its shoulders; the legs taper between neighbouring crowns of different width), tip_narrowing_mm (the crown's own offset dd: tip width = width - dd; negative values widen the tip; shoulders and legs keep the crown's width), extension_mm (the whole crown moved outwards along the axis, lengthening its two legs; negative shortens them), and profile (below).

  • profile: a list of smooth bumps on the crown's edges, each {"edge": "outer" or "inner", "center_deg": angle from the crown's apex (+ towards larger theta), "half_width_deg": reach to either side, "amount_mm": material added at the centre (negative removes). The outer edge (on the crown's tip side) moves out and the inner edge (on its opening side) moves in by amount * (1 - t^2)^3, t=(angle−center)/half widtht = (\text{angle} - \text{center})/\text{half width}, so the edge returns smoothly to the arc. Bumps must end within +-85 degrees of the apex and 2 degrees before the legs; on a linked crown's outer edge they must stay 5 degrees clear of the link attachment (about +-28 degrees in D0). Up to 6 bumps per crown; center_deg -80 to 80, half_width_deg 5 to 60, amount_mm -0.3 to 0.3; the local width along the crown must stay >= 0.57 mm.

  • Links fields: waist_width_mm, waist_flat_mm (a longer flat gives smaller flank radii, i.e. a narrower attachment at the crowns) and shift_mm (the link's axis moved along the circumference, + towards larger theta). Each flank is fitted to its own crown, so the two crowns of a link may differ.

  • Legs are common tangents of neighbouring crowns, and every outline stays tangent-continuous. The legs are short, so neighbouring crowns can differ in width only by about 0.14 mm at L=0.77L = 0.77 mm before a leg vanishes; the check reports it.

  • Shared settings: strut_width_mm 0.76 to 1.1, crown_offset_mm 0.0 to 0.4, leg_length_mm 0.4 to 1.4, ring_pitch_mm 7.6 to 8.75, link_waist_width_mm 0.57 to 1.2, link_waist_flat_mm 0.0 to 2.0; ring_pitch_mm may be one value or four, one per gap.

  • Per crown: width_mm 0.76 to 1.2, tip_narrowing_mm -0.3 to 0.63, extension_mm -0.3 to 0.4. Per link: waist_width_mm 0.57 to 1.2, waist_flat_mm 0.0 to 3.0, shift_mm -0.5 to 0.5.

  • Fixed: the tube, 5 rings, 12 crowns per ring, which crowns carry links, the zz of ring 0's minus-zz crown tips in the shared construction. Rules: crown tip width >= 0.57 mm, every link flank radius >= 0.3 mm, every leg's straight part >= 0.2 mm, facing unlinked crown tips >= 0.5 mm apart, link attachments inside the crown arcs (>= 3 degrees from the legs), the profile rules above, as-cut length <= 41.0 mm, volume <= 229.193 mm^3.

  • /astra_lib/local_stent.py builds the outline from a design and runs these checks; read it for the exact construction (resolve, check, measures, crown_table, plot).

The simulation (set-up, the same for every design)

  • Abaqus/Standard, quasi-static, large deformation. C3D8 hexahedra swept through the wall (3 layers, graded 1:2:1), in-plane element size 0.19 mm (about 37,000 elements).

  • Material: elastic-plastic, von Mises yield with isotropic hardening, E=43E = 43 GPa, ν=0.3\nu= 0.3. The hardening curve is calibrated up to an equivalent plastic strain (PEEQ) of 0.1766; above that the model is outside its calibration (it is not a fracture criterion).

  • Analysis frame: stent axis = zz, angle theta = atan2(yy, xx) in degrees.

  • Balloon: a cylindrical surface, radius 3.80 mm, length 42 mm, centred on the stent; its nodes are moved radially. Steps: Approach (balloon brought to 0.05 mm from the stent's inner surface), Expansion (balloon radius increased by 7.0 mm, i.e. balloon diameter 21.6 mm), Recoil (balloon returned to its initial radius; the stent springs back freely).

  • Contact: hard normal contact; friction 0.05 between balloon and stent; frictionless stent self-contact.

  • Rigid-body motion is removed by three single nodes on the outer surface of ring 2, near target angles 0, 120 and 240 degrees on the accepted node plane nearest to 0.09 mm below the ring centre. Their positions depend on each design's mesh (listed below). At each, the tangential and axial displacements are fixed, the radial one is free.

  • The swept mesh needs a cut through the tube along the plane y=0y = 0 (theta = 0 and 180 degrees); it passes through the crown apices at theta 0 and 180 in every ring. A few slightly distorted elements lie next to this cut.

The original design D0

D0 is the design of section 2: shared settings only, no crown or link entries. It was simulated with this set-up; its design file and recorded fields are in /data/designs/D0.

Geometry of D0:

id,origin,strut_width_mm,crown_offset_mm,leg_length_mm,ring_pitch_mm,tip_width_mm,ring_width_mm,
  ring_gap_mm,link_flank_radius_mm,length_as_cut_mm,volume_mm3
D0,original design,0.963140,0.178238,0.773662,8.249800,0.7849,5.7448,2.5050,1.6115,38.744,229.193

Results (after recoil). Crown strain = mean of the 5 highest element PEEQ values in a crown (a crown here is the arc plus the adjacent halves of its two legs); Q = 90th percentile of the crown strain over all 60 crowns; shortening = as-cut length minus length after recoil. As-cut inner diameter 14.805 mm for every design. feasible: under the limits of section 5.

id,Q,median_crown_strain,max_crown_strain,max_element_peeq,elements_above_0.1766,inner_D_after_recoil_mm,
  recoil_pct,shortening_mm,feasible
D0,0.1599,0.1511,0.1802,0.1857,3,21.2591,2.090,1.3294,yes

Support nodes:

id,support_theta_deg,support_z_mm
D0,348.45 / 131.55 / 228.45,-8.509 / -8.509 / -8.509

Per crown: crown strain, max and mean element PEEQ, and tip_inward_mm = axial motion of the crown tip towards its own ring centre after recoil. linked = a link is attached at the crown tip.

ring,theta_deg,tip_end,linked,crown_strain_D0,D0_max_element_peeq,D0_mean_peeq,D0_tip_inward_mm
0,0,-z,no,0.1454,0.1470,0.0263,+1.031
0,30,+z,yes,0.1449,0.1484,0.0246,-0.022
0,60,-z,no,0.1450,0.1474,0.0261,+1.048
0,90,+z,no,0.1567,0.1600,0.0290,+0.723
[… 56 further rows, one per crown …]

D0 in more detail. Inner mean diameter 21.713 mm at the end of expansion; length 38.742 mm as cut, 36.736 mm at the end of expansion, 37.413 mm after recoil; max von Mises stress 247.5 MPa at the end of expansion, 131.4 MPa (residual) after recoil. PEEQ does not change during recoil.

Per link (D0):

gap_between_rings,theta_deg,max_element_peeq
0–1,30,0.0160
0–1,210,0.0441
1–2,120,0.0152
1–2,300,0.0504
2–3,30,0.0154
2–3,210,0.0236
3–4,120,0.0195
3–4,300,0.0415

Per ring (D0):

ring,centre_z_mm,inner_D_expansion_end_mm,inner_D_after_recoil_mm,crown_tip_to_tip_length_as_cut_mm,
  after_recoil_mm
0,-24.920,21.713,21.258,5.598,4.147
1,-16.670,21.710,21.259,5.615,4.165
2,-8.420,21.706,21.254,5.618,4.182
3,-0.170,21.710,21.261,5.615,4.164
4,8.080,21.713,21.260,5.582,4.118

The full recorded fields of D0 (every node and element, every exported frame) are in the sandbox under
/data/designs (see /data/designs/README.md); each design you simulate is added there.
## 5. Objective and rules

Objective: the lowest Q of a feasible design. D0 (Q 0.1599) is the design to beat; the best feasible
design of this session counts.

Constraints (feasible = all met):
- inner mean diameter after recoil >= 21.249 mm;
- shortening <= 1.3605 mm;
- volume <= 229.193 mm^3 (D0's volume); as-cut length <= 41.0 mm;
- the design rules of section 2.

A design that cannot be built, meshed or solved counts as Q = 0.25 and infeasible.
## 6. Tools and budget

- python: runs one Python program in an isolated sandbox (numpy, scipy, matplotlib, shapely; no network;
up to 10 minutes per run). /data and /astra_lib are read-only; /work is yours and persists between runs.
- view_image: shows you a PNG or JPEG you saved under /work.
- check_design: the free checks of section 2 plus the CAD build of a design file under /work; it
reports the volume, length, link flank radii, profiles and what changed relative to D0. Up to 60 checks.
- submit_designs: simulates one to 3 designs at once (about 2.5 to 3 hours per round; the conversation
waits) and returns all their results; their full fields then appear under /data/designs. Give your
predictions and reasons with each design. 14 simulations in total; every design must pass check_design
first and must not repeat a simulated one.
- finish: ends the session with your best design and a summary.

Between turns you keep your messages, tool calls and their results, and short summaries of your
reasoning, not your full reasoning; keep notes under /work if you want to.
## 7. What to report

With every design you submit, predict Q, the inner mean diameter after recoil, the shortening, the max
crown strain and the three hottest crowns, and give your reasons. When you are done, at the latest after
the results of your last simulations, call finish: name your best design, summarise what you learned,
and list any change you wanted to make but could not express with these controls.

Tools. Table 6 lists the clean run’s five tools as the LLM received them. With every call, the LLM gets back its earlier messages, tool calls and results, and summaries of its earlier reasoning, as its opening message says. The sandbox is an isolated Linux system without network access: /data holds the fields of every design simulated so far, /astra_lib the design code and a small library to read the fields, and /work whatever the agent writes. A tool’s result is cut in the middle beyond 30 000 characters, the result of a round beyond 60 000. At most three times in a session, a reply without a tool call is answered with “Please continue with the tools, or call finish when you are done.” The clean run needed none.

ToolDescriptionArguments
python“Run one Python program in the sandbox and return what it prints. /data and /astra_lib are read-only; /work is writable and persists between runs. No network; up to 10 minutes per run.”code: “Python source”
view_image“Show a PNG or JPEG file saved under /work; it appears in the next message.”path: “e.g. /work/outline.png”
check_design“Free checks and CAD build of a design file under /work (schema astra-direct-002/design-1). Reports pass or fail, the failures, volume, length, link flank radii, profiles, what changed relative to D0 and whether it was already simulated. Up to 60 checks.”path: “e.g. /work/d8.json”
submit_designs“Simulate one to 3 designs at once, in parallel (about 2.5 to 3 hours per round; the conversation waits), and return all their results. Each must pass check_design and must not repeat a simulated design. 14 simulations in total.”designs: one to three, each with path (“design file under /work”), reasons (“why this design”) and predictions of Q, inner_d_after_recoil_mm, shortening_mm, max_crown_strain and hottest_crowns (“the three hottest crowns, hottest first, as ‘ring@theta’”)
finish“End the session.”best_design (“id of your best design, as named in the results”), summary and wanted_but_not_expressible (“changes you wanted but could not express with these controls”)

Table 6. The clean run’s tools, with the descriptions the LLM received, verbatim. The plate’s sessions had the same tools, with the descriptions and predictions of their problem.

A round in the transcript. The clean run’s transcript holds the system prompt, the opening message, 30 replies of the LLM and the results of its 38 tool calls: 16 Python programs, one image, 14 checks, six rounds and the finish. In its eighth reply, after five Python programs, one image and three checks, the LLM opened its first round with the call below, whose three designs are the three concepts of Section 5.3. The tool’s reply follows, cut like the opening message. Its verdicts are those of the main reading; under the head-to-head reading, only the first design, which breaks the lumen limit, is invalid. The run ended with the LLM’s call of finish, the last listing.

──────────────── Reply 8: the call that opened the first round, with line breaks added ────────────────
{
  "designs": [
{
  "path": "/work/longlegs.json",
  "predictions": {
    "Q": 0.146,
    "inner_d_after_recoil_mm": 21.222,
    "shortening_mm": 1.24,
    "max_crown_strain": 0.168,
    "hottest_crowns": ["2@210", "3@150", "2@180"]
  },
  "reasons": "Increase straight-leg length to 1.15 mm to reduce the crown rotation required by hoop expansion; reduce shared width to 0.922944 mm to hold volume at 228.90 mm3. This probes the strain/recoil tradeoff, with anticipated diameter risk."
},
{
  "path": "/work/less_taper.json",
  "predictions": {
    "Q": 0.148,
    "inner_d_after_recoil_mm": 21.251,
    "shortening_mm": 1.3,
    "max_crown_strain": 0.173,
    "hottest_crowns": ["2@210", "1@210", "3@150"]
  },
  "reasons": "Reduce the tip narrowing from 0.178238 to 0.04 mm, making the tip relatively stronger and shoulders thinner at fixed volume. Shorten legs to 0.58 mm to offset extra compliance and test whether a broader bending region lowers the apex peak while maintaining recoil diameter."
}
},
{
  "path": "/work/redistribute.json",
  "predictions": {
    "Q": 0.141,
    "inner_d_after_recoil_mm": 21.239,
    "shortening_mm": 1.33,
    "max_crown_strain": 0.158,
    "hottest_crowns": ["2@210", "3@150", "1@210"]
  },
  "reasons": "Use the D0 element fields to redistribute material: unlinked crowns receive a broad outer apex addition and outer shoulder relief; linked crowns receive small inner additions at +/-30 degrees and outer shoulder relief. Longer link waist flats narrow the attachments. This directly tests spreading plastic curvature away from the measured hot zones."
}
]
}

The tool’s reply

Round of D1 (d8e49580ef84), D2 (9e89cacbac8f), D3 (1675e130361e):

Results (after recoil):

id,Q,median_crown_strain,max_crown_strain,max_element_peeq,elements_above_0.1766,inner_D_after_recoil_mm,| recoil_pct,shortening_mm,feasible
D1,0.1359,0.1295,0.1465,0.1520,0,21.1847,2.433,1.1460,no
D2,0.1667,0.1589,0.1725,0.1764,0,21.2958,1.934,1.4296,no
D3,0.1212,0.1153,0.1412,0.1465,0,21.2719,2.143,1.4397,no

D1: infeasible: inner diameter after recoil 21.1847 mm < 21.2491 mm. Predicted against actual: Q 0.146 / 0.1359; inner D after recoil 21.222 / 21.1847 mm; shortening 1.24 / 1.1460 mm; max crown strain 0.168 / 0.1465; hottest crowns 2@210, 3@150, 2@180 / 2@30, 3@150, 2@60.

D2: infeasible: shortening 1.4296 mm > 1.3605 mm. Predicted against actual: Q 0.148 / 0.1667; inner D after recoil 21.251 / 21.2958 mm; shortening 1.3 / 1.4296 mm; max crown strain 0.173 / 0.1725; hottest crowns 2@210, 1@210, 3@150 / 2@330, 2@270, 2@300.

D3: infeasible: shortening 1.4397 mm > 1.3605 mm. Predicted against actual: Q 0.141 / 0.1212; inner D after recoil 21.239 / 21.2719 mm; shortening 1.33 / 1.4397 mm; max crown strain 0.158 / 0.1412; hottest crowns 2@210, 3@150, 1@210 / 2@210, 2@180, 3@180.

Support nodes:

id,support_theta_deg,support_z_mm
D1,11.69 / 108.31 / 251.69,-8.341 / -8.341 / -8.341
D2,18.24 / 134.24 / 221.76,-8.554 / -8.546 / -8.554
D3,348.45 / 131.55 / 228.45,-8.509 / -8.509 / -8.509

Per crown (crown strain of D0 and of this round; the rest for D3):

ring,theta_deg,tip_end,linked,crown_strain_D0,crown_strain_D1,crown_strain_D2,crown_strain_D3,| D3_max_element_peeq,D3_mean_peeq,D3_tip_inward_mm
0,0,-z,no,0.1454,0.1245,0.1549,0.1110,0.1132,0.0276,+1.097
0,30,+z,yes,0.1449,0.1243,0.1517,0.1138,0.1153,0.0259,-0.097
0,60,-z,no,0.1450,0.1226,0.1498,0.1081,0.1096,0.0286,+1.105
0,90,+z,no,0.1567,0.1229,0.1558,0.1173,0.1216,0.0320,+0.483
[... 56 further rows, one per crown ...]

[... the tables per link and per ring of D1, D2 and D3, as for D0 in the opening message ...]

Their full fields are now under /data/designs (D1, D2, D3).

Simulations left: 11.


Reply 30: the call that ended the run, its three arguments

best_design: D13

summary:

D13 is the best feasible design from all 14 simulations. Design file: /work/tail_refine.json; key: → d38c6420a4b5.

Verified results: Q = 0.108489, a 32.2% reduction from D0; inner mean diameter after recoil = 21.2521 mm; → shortening = 1.2015 mm; volume = 228.5469 mm³; as-cut length = 39.1309 mm. Maximum crown strain is → 0.112744 and maximum element PEEQ is 0.114017, with no elements above 0.1766. All geometry and → feasibility checks passed.

The design uses 0.965666 mm shared width, 0.940 mm legs, D0's crown offset and ring pitch, and 1.500 mm
↪ link waist flats. Crown-specific profiles combine outer-apex reinforcement and shoulder relief on
↪ unlinked crowns, broader inner-edge reinforcement on linked crowns, and small asymmetric corrections
↪ informed by the element fields. Terminal extensions were removed.

The largest improvement came from spreading crown bending, not simply thinning struts. Longer legs reduced
↪ strain but consumed recoil-diameter margin; excessive shoulder relief created new strain peaks.
↪ Support-node changes also shifted the hottest crowns, limiting the predictability of fine local
↪ adjustments. D14 remained feasible but had slightly worse Q, so D13 is retained. Its diameter margin is
↪ only about 0.003 mm; these simulation results are not a fracture, fatigue, or clinical validation.

wanted_but_not_expressible:
Independent leg-width/curvature control, noncircular crown centerlines, curved link centerlines or
↪ different attachment topology, and through-thickness edge fillets were not available. For validation, I
↪ would also want mesh refinement and controlled support/seam sensitivity studies with consistent support
↪ mapping; those simulation controls were not exposed.

Other runs. The plate’s sessions had the same system prompt, tools and records, and an opening message with the same seven sections for the plate: the hole’s two semi-axes and up to eight bumps, the simulation set-up, the original design’s peaks and its stress along the hole’s edge every 3.75° for both load cases, and the area limit. Each design is submitted with predictions of the objective and of both load cases’ peaks and their angles. One rule is added: “Do not solve the plate problem numerically yourself: no finite-element, boundary-element, finite-difference or similar numerical solution of the plate, with your own code or with a library’s. The simulations of your budget are the only solutions of the plate in this session.” A scan of every Python program for code that could solve the plate flagged 8 of 505 programs on the judge’s model and 5 of 241 on Tosca’s; read one by one, none solves it. On Tosca’s coarse model, the opening message also describes the judge’s fine model and gives the start design’s value there; the agent saw no other value of the judge during its session.

The crown’s sessions received their framework’s own instructions as the system prompt and a task in four steps: have a critic sub-agent review the fixed protocol and file its verdict; submit designs in rounds of up to three, each with a reason and predictions of the peak strain, the opening and the shortening; read every design’s results before the next round, among them where along the crown the plastic strain sits; and finish with the best design, a summary and what could not be expressed with the four values. Every crown session but the first was also told that the protocol is fixed for the whole session and must not be edited or replaced. The task ended: “Nobody will answer questions during this session: decide yourself.” The agent had a design-study tool, through which it binds the review, submits rounds, reads the results and finishes, a tool that puts the critic’s review on record, and file tools confined to the session’s folder; it had no shell and no Python sandbox. The set-up sessions got the bar of Section D in plain words and, besides these tools, three more for looking up the solver’s documentation, retrieving example inputs and checking an input before it runs.

The pilot sessions led up to the clean run. The single call and the round on the shared settings had no tools; their system prompt asked for one JSON object, and their opening message gave the same kind of facts, with questions to answer and values to predict. Of the single call’s three variants, one left out the tables per crown, link and ring and kept the global results; one kept every number and table but replaced every physical description by neutral names, such as parameters p1 to p7, “units” and “fixed points”; and one repeated the original call. The crown-by-crown session added the tools, with one design per simulation, and the edge-profile session added the profiles and rounds of up to three. The shuffled-results control continued the clean run’s own conversation from the results of a round, with the same tools, sandbox and checks, and ended each continuation at its first accepted round.

Records and reproduction. Every run’s record holds its opening message as sent, every message, tool call and result, the tokens and costs of its LLM calls, the images the agent viewed and the files it wrote, and, for every simulated design, its design file, geometry, mesh and solver input, the exported fields, the digest and the score. On the stent and the plate, the driver records each step before taking the next, so a run that stops, for a pause, for solver licences or for a lost connection, resumes from its record without repeating an LLM call or a simulation. Before the clean run, its driver ran two rounds with a scripted stand-in for the LLM against a stand-in for the simulation, with a pause and a resume, and its opening message was checked to hold nothing from the pilot sessions. LLM calls cannot be repeated exactly, since the LLMs sample their replies and their providers may change them, so a rerun under the same protocol is a new run; this is why the protocols fix in advance what counts across repeated runs (§ 3.4). The records reproduce every score of the agent’s designs, which the judge computes from the stored fields, and they allow a run to be continued from a recorded point. The shuffled-results control did so: it restarted the clean run’s conversation from the results of a round, after rebuilding the agent’s working folder by running its recorded Python programs again, each of which had to print what it had printed before.

Stent finite-element model

Geometry and load case. The original design is cut from a tube of 15.99 mm outer diameter and 0.593 mm wall thickness and is 38.74 mm long; its struts are 0.963 mm wide and narrow to 0.785 mm at the crown tips. A rigid cylindrical surface acts as the balloon: it is driven radially to a diameter of 21.6 mm and back, expanding the stent from its as-cut state and then letting it recoil freely, with hard contact and friction 0.05 between balloon and stent and frictionless self-contact. The analysis is static with large deformations in Abaqus/Standard 2022.

Material. The magnesium alloy Mg-4Y-3Gd from the public data of Steglich et al. [42] is modelled with linear elasticity (E=43E = 43 GPa, ν=0.3\nu= 0.3) and von Mises plasticity with isotropic hardening from one monotonic tensile test, from 102 MPa at first yield to 247.5 MPa at an equivalent plastic strain of 0.1766, where the data end; beyond that the stress is held constant.

Supports and mesh. Three nodes on the middle ring, 120° apart and about 0.09 mm below its centre plane, are fixed axially and tangentially; they stop rigid-body motion but load the nearby crowns artificially. Moving them to another ring lowered the original’s most strained crown by 10% but QQ by only 1.4%. Abaqus/CAE meshes each design with eight-node hexahedra (C3D8), swept through the wall in three layers with an in-plane seed of 0.19 mm, about 35 000 elements.

Crown strains. The element strain is the equivalent plastic strain (PEEQ) at the end of recoil, the largest of the element’s eight integration points. A per-crown trace assigns every element to its nearest crown along the design’s own outline, so a crown includes its arc and the adjoining half-legs, and a crown’s strain is the mean of its five highest element values.

Rebuild noise. Every new design gets its own CAD solid, a new free mesh and supports placed anew by the rule above, so its score scatters in a way that re-solving the same mesh does not show. We measured the scatter by rebuilding four designs with their struts 0.001 mm thinner, a change expected to lower QQ by about 0.0002. All four rebuilds scored higher than their originals, by 0.0008 to 0.0031, as expected when the originals were the best designs of their campaigns, because the best of several noisy evaluations is optimistic. The scatter of QQ, the root mean square of the four changes divided by 2\sqrt{2}, is 0.0015; the difference between two rebuilds scatters by 2\sqrt{2} times as much, and we take twice that, 0.004, as the noise of a rebuild. The most strained crown scatters by 0.006, about four times as much as QQ, so its noise of a rebuild would be about 9% of the original’s value, against 2.5% for QQ. The lumen scatters by 0.0005 mm, well inside its tolerance. The shortening scatters by 0.06 mm, twice its tolerance, because it rests on the extreme nodes at the stent’s ends, and one rebuilt design went from meeting the shortening limit to breaking it; its noise, computed like that of QQ, is 0.17 mm, the margin of the head-to-head reading (§ 4.3).

Computing. Our simulations ran in Abaqus 2022 on Strand, the HPC cluster of Helmholtz-Zentrum Hereon. Each design of the agent was meshed and checked in about three minutes and solved on one core of a node with two 24-core Intel Xeon Platinum 8160 processors at 2.1 GHz. The clean run’s fifteen solves took 1.86 to 2.29 h each, 2.14 h in the median and 31.4 core-hours in all; they mostly ran one after another, because the solver licences, which we share with colleagues, rarely left room for two, and with the agent’s own time and these waits the run took 35 h from its first solve to its last. Tosca solved its designs on three or four cores, in 1.1 to 1.7 h each. The full-cycle model, as recorded in its solver’s job summaries, took 10.1 h for crimping and the crimper’s release and 7.0 h for the expansion and the balloon’s release of the original design, both on two cores, 17.1 h in all and 33 core-hours; for the agent’s pilot design it took 25.0 h on two cores and 14.2 h on one, 39.2 h in all and 63 core-hours. A run of the full chain thus costs 16 to 29 times the core-hours of a solve of the reduced model.

The plate with a hole

Problem. Tosca’s bundled shape example, copied unchanged from the Abaqus 2023 installation, is a quarter of a 100 × 100 mm plate, 1 mm thick, with a central hole of radius 20 mm, in plane stress (CPS4, E=210E = 210 GPa, ν=0.3\nu= 0.3) with symmetry supports on the two cut edges. Two load cases are solved separately: one pulls the right edge with 1800 N in all, the other the top edge with 900 N. Tosca’s task is to lower the largest von Mises stress on the hole’s edge with its controller-based method, moving the edge nodes of a mesh with 16 elements around the quarter hole. Our test reproduced the vendor’s run: 162.30 to 109.22 MPa in five cycles. For every method alike we added a fine mapped mesh of the same type, with 96 elements around the quarter hole, so that every design has the same mesh topology and no remeshing noise; the loads as uniform tractions with the shipped totals; and a limit on the material area, at most the original’s.

Judge and solver. The objective is Tosca’s own measure, the largest von Mises stress at the hole’s edge nodes over both load cases, read on the fine mesh: 171.6 MPa for the circle. Our reference solver implements the same element and nodal stresses as Abaqus and reproduces Tosca’s values on the shipped deck to all printed digits, and on all 15 meshes of Tosca’s run to better than one part in ten million. Every agent run’s best design was also solved on a mesh twice as fine, where it reads at most 1.6% higher and the order of the runs is unchanged. Tosca’s best design reads 1.8% higher there (111.4 MPa against 109.4), enough to fall behind one more agent run.

Geometry parametrisation and agents. The geometry parametrisation, built on the stent’s geometry code, sets the hole’s two semi-axes (10 to 35 mm) and up to eight bumps on its edge (centre 0 to 90°, half width 10 to 90°, amount −6 to 6 mm), mirrored at the symmetry lines; the outline is extruded instead of wrapped. Rules keep the edge 5 mm from the outer edges and the hole’s centre and the mesh’s element angles within 15 to 165°. The agents had the stent’s tools and budget. One rule the stent did not need: an agent must not solve the plate numerically itself, since a linear 2D solve is easy to write; an audit of every python cell found none that did.

Results. Table 7 lists every method. Given Tosca’s own coarse model, with the same mesh, loads and budget as Tosca, seven of ten agent runs still beat Tosca on the judge, every Astra run among them (Figure 11); their results moved by 1.6 to 3.4 MPa against their fine-model runs, less than the spread between one LLM’s runs (3 to 20 MPa). Read on finer and finer meshes, Tosca’s best design rises from 95.5 MPa on its own deck to 109.4 on the judge’s mesh, so its own model flatters it by 13%, because it builds a sharp step near 2525–30∘30^\circ (Figure 11d) that its 16 elements along the quarter hole cannot resolve; the coarse model flatters the agents’ designs by 4 to 10%. The agents knew how they would be judged, and one corrected for the finer mesh explicitly, which Tosca cannot be told; their bumps are smooth by construction, so part of that robustness comes from the parametrisation. The four best runs placed all eight bumps freely, tuning 24 to 26 values.

MethodDesign spaceBest per run (MPa)Below the circle
Claude Fable 5.2, three runsUp to 26 values93.0, 92.7, 91.846–47%
GPT-6 Astra, three runsUp to 26 values101.5, 102.6, 94.740–45%
Kimi K3.1 Pro, three runsUp to 26 values103.7, 110.2, 115.033–40%
Mistral Large 4, three runsUp to 26 values125.6, 107.4, 105.427–39%
Tosca, as shippedIts mesh: 17 edge nodes109.436%
Bayesian optimisationTwo semi-axes125.927%

Table 7. The plate with a hole: the best valid design within 15 solves, the start’s included, on the judge’s fine mesh (the circle: 171.6 MPa). The agents changed the hole’s two semi-axes and up to eight bumps on its edge, up to 26 values.

Figure 11. The plate on Tosca’s own coarse model, with 16 elements along the quarter hole. (a) Tosca’s best design within 15 simulations, as its own mesh with its own loads. (b) The same design on the coarse model the agents were given: the same 17 edge nodes, with the shipped mesh’s straight lines inside where Tosca smooths them, which changes the stress at the hole by at most 1.1%. (c) The best agent design on that model, by Claude Fable 5.2. Colour: von Mises stress, the larger of the two load cases; circle: the edge node with the highest stress, where each value is read; dashed: the original hole. (d) The hole shapes: Tosca’s 17 edge nodes and the curve the judge reads through them, and each LLM’s best design from its runs on the coarse model, with the judge’s value in brackets. (e) The best valid design so far against simulations for every run on the coarse model, judged on the fine mesh, and Tosca’s reading on its own mesh.

Reduced crown problem: four values for every method

On the stent, only the agent and Tosca can work on all of the design’s values within the budget. A reduced problem that solves in a minute is the ladder’s bottom rung (Table 4), where the agent and the classical optimisers change the same four variables. It answers three questions: whether reading the results helps then, whether the result depends on the LLM, and what happens when the LLM must also write the simulation set-up, which in the loop is written and checked by hand (Table 1). The sessions ran in a lighter implementation of the protocol of § 3, a design-study tool that we built on the open-source agent framework openPASO [21].

CrownPlateStent
Design values4Up to 26About 1 650
Proposed designs121414
Ratio of design values to proposed designsBelow oneAbout twoOver a hundred
Classical optimisersOn all four values: Bayesian optimisation 8.6–15.4% (median 10.2%), COBYLA 0–5.8%, random search nothingBayesian optimisation on the two semi-axes a person picks: 27%, an ellipseNot run as a comparison
Shape optimiser (Tosca)–As shipped, on its own mesh: 36%From the original: 10%, then drifted; from the agent’s fix: a further 14% (12% on a fresh mesh)
LLMsComplete sessions 10.1–22.3%; the best of four LLMs 19–22%; one LLM ahead of the median seed in two of three sessions27–47%; 9 of 12 runs ahead of Tosca32% (clean run), 26% (pilot sessions)
LLMs, runsSix LLMs, 20 sessionsFour LLMs, three runs eachGPT-6 Astra, two runs

Table 4. Three problems with four, 26 and about 1 650 values and similar budgets: the methods that can take part, and their results. Percentages: the best valid design’s objective below the start design’s, within the budget. Details in Sections C and D.

Problem. One crown of a zigzag ring, unrolled into the plane: half a crown, its strut and half of the opposite crown, with both cut faces on symmetry planes, opened by 0.5 mm per half period and released (Figure 12a). FEBio 4.12 [30] solves it with von Mises plasticity (E=200E = 200 GPa, ν=0.3\nu= 0.3, yield stress 300 MPa, linear hardening 2 GPa) on a structured mesh of 32, 48 and 12 elements along the arc, along the strut and across the width, in one to two minutes on one core. Four variables set the width profile: the width at the apex, waw_a, 45° from it, w45w_{45}, and along the strut, wsw_s, each between 0.06 and 0.2 mm and 0.1 mm at the start, and an outward shift ss of the apex of up to ±0.05 mm. The objective is the mean plastic strain of the most strained 2% of the metal after release, 0.1003 for the start design. In checks before any session it ranked five test designs as a mesh 1.5 times finer does, whereas a mean over the four most strained elements moved by 30–50% on a mesh twice as fine. A design is feasible if its opening after recoil is at least the start design’s and its shortening at most the start design’s, each with a tolerance of 0.001 mm fixed in advance, and if its metal volume, exact in closed form, is at most the start design’s. Every method gets twelve simulations beyond the start design. The four parts a problem supplies (§ 3.4), are, for this problem: four bounded numbers written into a script that generates the solver input; the FEBio model; a digest that tables the plastic strain along the crown, edge by edge, next to the scores; and the objective and limits above. The digest shows the agent where along the crown the strain sits, whereas the classical optimisers see only the scores.

Agent and protocol. The agent submits rounds of up to three designs, each with a reason and predictions of the peak strain, the opening and the shortening, and reads each design’s digest and solver output as files. It has no Python sandbox, no image viewer and no shell, and no simulation outside its budget; no CAD is built. Before the first simulation, a critic sub-agent of the same LLM reviews the fixed protocol. The judge follows the rules described in § 3.4: a failed simulation counts and is never feasible, every limit is the start design’s own value, and every limit is read both exactly and with its tolerance. The protocol was fixed before the first call; five dated amendments, each made before the sessions it governs, added sessions, raised caps and, after the first session, told the agent not to edit the protocol.

Sessions and baselines. Twenty crown sessions used six LLMs: GPT-6 Astra (one session), Qwen 3.5 397B (three), Mistral Large 4 (three), Qwen 3.8 (one), and Kimi K3.1 Pro and DeepSeek V4.1 Flash (three each, and three more each with a larger cap). In the round that added Mistral, Qwen 3.8, Kimi and DeepSeek, the caps were sized on input tokens; Kimi and DeepSeek, which spend most of their money on reasoning, used them up early, mostly in the opening review, so both were run again with US$2 per session. Two Mistral sessions and the Qwen 3.8 session were also stopped by their caps, after six to nine simulations, and were not run again. One Qwen 3.5 session followed its critic’s false objection to the exact volume formula, edited the protocol and left the study; it is listed but not compared. The classical baselines start from the same design with the same twelve simulations and judge: COBYLA with step sizes 0.1, 0.2 and 0.4 in scaled variables, random search with five seeds, and Bayesian optimisation with five seeds. The Bayesian optimisation is the one we built for the stent and also ran on the plate: BoTorch [4], one Gaussian process per output with a Matérn-5/2 kernel, log constrained expected improvement [1], and four Sobol points before eight surrogate-guided steps. All three skip, without solving them, designs that break a rule on the widths, which the agent’s free check also shows. Bayesian optimisation and random search also skip designs with more metal than the start design, while COBYLA treats the volume as one of its constraints and solves such designs. Table 8 lists every session and baseline.

MethodRunCapSimulationsMain readingStrict readingCost
StudyFinerStudyFiner
COBYLAstep 0.1–120.00.00.00.0–
step 0.2–125.85.60.00.0–
step 0.4–124.80.40.00.0–
Random searchseed 0–12 (1 failed)0.00.00.00.0–
seed 1–12 (1 failed)0.00.00.00.0–
seed 2–12 (1 failed)0.00.00.00.0–
seed 3–12 (2 failed)0.00.00.00.0–
seed 4–120.00.00.00.0–
Bayesian optimisationseed 0–12 (1 failed)10.210.40.00.0–
seed 1–12 (1 failed)10.09.30.00.0–
seed 2–128.66.84.66.0–
seed 3–1210.410.90.00.0–
seed 4–1215.415.35.15.3–
GPT-6 Astra13012, cap19.320.814.816.338.39
Qwen 3.5 397B13left the official study0.38
231210.610.90.00.00.40
331210.110.60.00.00.39
Mistral Large 4119, cap6.25.20.00.00.94
2112 (1 failed)22.322.00.00.00.84
316, cap0.00.00.00.00.92
Qwen 3.812.56, cap0.00.00.00.02.25
Kimi K3.1 Pro10.7512, cap10.39.95.74.80.67
20.750, cap––––0.72
30.753, cap0.00.00.00.00.67
129, cap0.00.00.00.02.31
2212, cap12.612.48.78.31.93
321219.318.419.318.41.82
DeepSeek V4.1 Flash10.250, cap––––0.36
20.250, cap––––0.28
30.250, cap––––0.24
1212 (1 failed)10.310.00.00.01.33
226, no answer0.00.00.00.01.34
3212 (3 failed), cap19.720.52.11.52.11

Table 8. Every crown session and classical baseline, each from the same start design with twelve simulations: the peak strain of the best feasible design, in percent below the start design’s, under the main reading (the tolerances fixed in advance) and the strict reading (every limit exact), on the study mesh and on a mesh 1.5 times finer, where the design is solved again and judged against the start design on that mesh; 0.0 means that no feasible design beat the start. Cap and cost: US$ per session at list prices, helper agents included; “cap”: ended by its cost cap; “no answer”: ended by empty replies. For Kimi and DeepSeek, the second group of three ran with the larger cap.

Results. With the same four variables and twelve simulations, random search found no feasible design better than the start in any of its five seeds, and COBYLA lowered the peak strain by 0 to 5.8%, depending on its step size (Table 8 and Figure 13). Bayesian optimisation was the strongest baseline: 8.6 to 15.4% over its five seeds, with a median of 10.2%. The nine LLM sessions that used all twelve simulations reached 10.1 to 22.3%, so none fell clearly behind the median seed, and the best session of each of four LLMs, Mistral Large 4 (22.3%), DeepSeek V4.1 Flash (19.7%), GPT-6 Astra (19.3%) and Kimi K3.1 Pro (19.3%), also beat the best seed. On the mesh 1.5 times finer, the sessions’ improvements changed by at most 1.5 percentage points and those of Bayesian optimisation by at most 1.8, while one COBYLA lead did not hold (4.8% fell to 0.4%).

Figure 13. The reduced crown problem: the best feasible design so far against simulations, main reading, study mesh. (a) The classical optimisers; random search and COBYLA with step size 0.1 found no feasible improvement. (b) Every LLM session under the final caps for its LLM, except the one that left the study, over the median and range of Bayesian optimisation.

The ten sessions that a cost cap or empty replies ended early reached 0 to 6.2%. The criterion fixed before the baselines ran, a larger reduction than the median Bayesian-optimisation seed on both meshes in at least two of an LLM’s three sessions, is met by Kimi; DeepSeek, Mistral and Qwen 3.5 meet it once each, and GPT-6 Astra and Qwen 3.8 ran once. Under the strict reading, with every limit met exactly, the difference grows: Kimi’s best design lowered the peak strain by 19.3% and GPT-6 Astra’s by 14.8%, against at most 5.1% for any classical baseline (6.0% on the finer mesh).

Paths of the best sessions. The sessions converged on the same trade-off: a stronger taper, wider at the apex and shoulder and narrower along the strut, lowers the strain but costs opening after recoil, so most of the best designs sit on the opening limit and use its tolerance. Where the start design’s most strained metal is a patch on the inner edge within 40∘40^\circ of the apex, two elements deep, these designs spread it thinner and further along the inner edge of the bend (Figure 12). GPT-6 Astra then solved the same widths with three shifts of the apex, measured that an inward shift wins back opening for a little strain, and spent that opening on a stronger taper; Kimi’s best design makes the same move and needs no tolerance. Bayesian optimisation found this region in one of its five seeds (15.4%, with the apex at its inward bound); its other seeds thinned the shoulder instead (8.6 to 10.4%). The LLMs’ predictions of the peak strain sharpened as they read the results: in seven of the nine complete sessions, their error fell from 6 to 100% of the start design’s value in the first round to 0.1 to 7% in the last, about as accurate as the Gaussian processes’ predictions in Bayesian optimisation’s later steps (2 to 9%).

Study set-up from a plain description. In twenty further sessions the LLM received a plain-language task and had to write the whole study itself: the solver input, the script that reads the results, the objective and the limits. The task was a bar of five linearly varying widths, pulled with plasticity, to strain as little as possible with no more metal and no less pulling force; its optimum, a uniform bar, is known. GPT-6 Astra, Kimi (three of three sessions), Mistral (three of three), DeepSeek (two of three) and Qwen 3.8 (one of one) set up a study that measured the actual task, and all but one of those ten sessions found the optimum. Qwen 3.5, which improved the crown when it was given the set-up, managed this in one of nine sessions. Five of the others never got their start design to run, three of them before we fixed how the solver’s error messages reached the LLM, and three accepted a set-up that ran but measured something else, an elastic material instead of the plastic one, a pulling force of zero or no strain at all, and reported success. Writing the set-up is where the weaker LLM failed, which is why, in the loop, the four parts of a problem are written and checked before any call and the agent only changes the design.

Limits. The problem is small: four variables and a known trade-off between strain and opening. GPT-6 Astra ran once. The caps differed between rounds. The implementation differs from the stent’s: no sandbox, no CAD, and a critic before the first simulation. FEBio on this fixed, structured mesh is deterministic, so the finer mesh, not a rebuild, is the yardstick for every gap.

References

  1. [1]Sebastian Ament, Samuel Daulton, David Eriksson, Maximilian Balandat, and Eytan Bakshy. Unexpected improvements to expected improvement for Bayesian optimization. In Advances in Neural Information Processing Systems 36 (NeurIPS), 2023. URL https://arxiv.org/abs/2310.20708.
  2. [2]Raul Astudillo and Peter I. Frazier. Bayesian optimization of composite functions. In Proceedings of the 36th International Conference on Machine Learning (ICML), volume 97 of Proceedings of Machine Learning Research, pages 354–363, 2019. URL https://arxiv.org/abs/1906.01537.
  3. [3]Nima Bakhtiary, Peter Allinger, Matthias Friedrich, Fritz Mulfinger, Jürgen Sauter, Ottmar Müller, and Martin Puchinger. A new approach for sizing, shape and topology optimization. SAE Technical Paper 960814, SAE International, 1996.DOI
  4. [4]Maximilian Balandat, Brian Karrer, Daniel R. Jiang, Samuel Daulton, Benjamin Letham, Andrew Gordon Wilson, and Eytan Bakshy. BoTorch: A framework for efficient Monte-Carlo Bayesian optimization. In Advances in Neural Information Processing Systems 33 (NeurIPS), 2020. URL https://arxiv.org/abs/1910.06403.
  5. [5]Craig Bonsignore. Open stent design. Nitinol Devices & Components (NDC), http://nitinol.com, n.d.
  6. [6]Shuyavan Brahmachary, Subodh M. Joshi, Aniruddha Panda, Kaushik Koneripalli, Arun Kumar Sagotra, Harshil Patel, Ankush Sharma, Ameya D. Jagtap, and Kaushic Kalyanaraman. Large language model-based evolutionary optimizer: Reasoning with elitism. arXiv preprint arXiv:2403.02054, 2024. URL https://arxiv.org/abs/2403.02054.
  7. [7]Paul Brunzema, Louis Tiao, Nhat Le, Kevin De Angeli, Yao Xuan, and Djordje Gligorijevic. Agentic Bayesian optimization through surrogate-augmented autoresearch. arXiv preprint arXiv:2608.00316, 2026. URL https://arxiv.org/abs/2608.00316.
  8. [8]Amedeo Buonanno, Salvatore Fabozzi, and Maria Valenti. AutoEnergyDesign: An LLM-driven agent for automatic energy system design optimization. arXiv preprint arXiv:2610.35160, 2026. URL https://arxiv.org/abs/2610.35160.
  9. [9]C. Chen, J. Chen, W. Wu, Y. Shi, L. Jin, L. Petrini, L. Shen, G. Yuan, W. Ding, J. Ge, E. R. Edelman, and F. Migliavacca. In vivo and in vitro evaluation of a biodegradable magnesium vascular stent designed by shape optimization strategy. Biomaterials, 221:119414, 2019. doi: 10.1016/j.biomaterials.2019.119414.DOI
  10. [10]Rory Clune, Denis Kelliher, James C. Robinson, and John S. Campbell. NURBS modeling and structural shape optimization of cardiovascular stents. Structural and Multidisciplinary Optimization, 50(1):159–168, 2014. doi: 10.1007/s00158-013-1038-y.DOI
  11. [11]Dassault Systèmes SIMULIA. Abaqus/CAE User’s Guide, Abaqus 2019. Dassault Systèmes Simulia Corp., 2019. Optimization module: “Configuring a shape optimization task” and, for a general topology optimization task, “Configuring advanced options”.
  12. [12]Dassault Systèmes SIMULIA. Abaqus 2022. Dassault Systèmes, Johnston, RI, 2022.
  13. [13]Dassault Systèmes SIMULIA. Tosca structure. https://www.3ds.com/products/simulia/tosca/structure-optimization, 2026. Accessed 11 November 2026.
  14. [14]Mohamed Elferaie, Janet Qian, Raina Wu, Qian Chen, Angela Dai, and Faez Ahmed. AI agents in engineering design: A multi-agent framework for aesthetic and aerodynamic car design. In ASME International Design Engineering Technical Conferences (IDETC/CIE), 2025. doi: 10.1115/DETC2025-169682. URL https://arxiv.org/abs/2503.23315.
  15. [15]David Eriksson and Martin Jankowiak. High-dimensional Bayesian optimization with sparse axis-aligned subspaces. In Conference on Uncertainty in Artificial Intelligence (UAI), 2021. URL https://arxiv.org/abs/2103.00349.
  16. [16]David Eriksson, Michael Pearce, Jacob R. Gardner, Ryan Turner, and Matthias Poloczek. Scalable global optimization via local Bayesian optimization. In Advances in Neural Information Processing Systems 32 (NeurIPS), pages 5497–5508, 2019. URL https://arxiv.org/abs/1910.01739.
  17. [17]Fabio Ferreira, Lucca Wobbe, Arjun Krishnakumar, Frank Hutter, and Arber Zela. Can LLMs beat classical hyperparameter optimization algorithms? A study on autoresearch. arXiv preprint arXiv:2603.24647, 2026. URL https://arxiv.org/abs/2603.24647.
  18. [18]Bingkun Guo, Wentian Li, Xiaojian Liu, Jiaqi Luo, Zibin Yu, Dalong Dong, Shuyou Zhang, and Yiming Zhang. A multidisciplinary design and optimization (MDO) agent driven by large language models. arXiv preprint arXiv:2511.17511, 2025. URL https://arxiv.org/abs/2511.17511.
  19. [19]Jinseong Han, Sunwoong Yang, and Namwoo Kang. A multi-agent system for motor design optimization via an FEA-AI hybrid approach. arXiv preprint arXiv:2606.09037, 2026. URL https://arxiv.org/abs/2606.09037.
  20. [20]Lothar Harzheim and Gerhard Graf. A review of optimization of cast parts using topology optimization. Structural and Multidisciplinary Optimization, 31(5):388–399, 2006. doi: 10.1007/s00158-005-0554-9.DOI
  21. [21]Alexander Hermann, Arman Shojaei, Ingo Scheider, and Christian Cyron. OASiS: an open-source multi-physics and multi-code framework for verified computer simulations. Software, version 1.1.0, Zenodo, https://doi.org/10.5281/zenodo.20543501, 2026. Developed further as openPASO, https://github.com/open-PASO/openPASO.
  22. [22]Carl Hvarfner, Erik Orm Hellsten, and Luigi Nardi. Vanilla Bayesian optimization performs great in high dimensions. In International Conference on Machine Learning (ICML), 2024. URL https://arxiv.org/abs/2402.02229.
  23. [23]Yayati Jadhav and Amir Barati Farimani. Large language model agent as a mechanical designer. arXiv preprint arXiv:2404.17525, 2024. URL https://arxiv.org/abs/2404.17525.
  24. [24]Antony Jameson. Aerodynamic design via control theory. Journal of Scientific Computing, 3(3):233–260, 1988. doi: 10.1007/BF01061285.
  25. [25]Mostafa Jamshidian, Adam Wittek, Saeideh Sekhavat, Hozan Mufty, Geert Maleux, Inge Fourneau, Elke R. Gizewski, Eva Gassner, Alexander Loizides, Maximilian Lutz, Florian K. Enzmann, Donatien Le Liepvre, Florian Bernard, Ludovic Minvielle, Antoine Fondanèche, and Karol Miller. Towards personalised assessment of abdominal aortic aneurysm structural integrity. International Journal for Numerical Methods in Biomedical Engineering, 42(2):e70140, 2026. doi: 10.1002/cnm.70140.DOI
  26. [26]Donald R. Jones, Matthias Schonlau, and William J. Welch. Efficient global optimization of expensive black-box functions. Journal of Global Optimization, 13(4):455–492, 1998. doi: 10.1023/A:1008306431147.DOI
  27. [27]Chau Le, Julian Norato, Tyler Bruns, Christopher Ha, and Daniel Tortorelli. Stress-based topology optimization for continua. Structural and Multidisciplinary Optimization, 41(4):605–620, 2010. doi: 10.1007/s00158-009-0440-y.DOI
  28. [28]Haoju Lin, Wenchang Zhang, Weipeng Xu, Xiang Li, Tian Xu, and Tianju Xue. TO-Master: an LLM-agent framework for automated topology optimization. arXiv preprint arXiv:2607.01812, 2026. URL https://arxiv.org/abs/2607.01812.
  29. [29]Tennison Liu, Nicolás Astorga, Nabeel Seedat, and Mihaela van der Schaar. Large language models to enhance Bayesian optimization. In International Conference on Learning Representations (ICLR), 2024. URL https://arxiv.org/abs/2402.03921.
  30. [30]Steve A. Maas, Benjamin J. Ellis, Gerard A. Ateshian, and Jeffrey A. Weiss. FEBio: Finite elements for biomechanics. Journal of Biomechanical Engineering, 134(1):011005, 2012. doi: 10.1115/1.4005694.DOI
  31. [31]Bo Ni and Markus J. Buehler. MechAgents: Large language model multi-agent collaborations can solve mechanics problems, generate new data, and integrate knowledge. arXiv preprint arXiv:2311.08166, 2023. URL https://arxiv.org/abs/2311.08166.
  32. [32]Alexander Novikov, Ngân Vũ, Marvin Eisenberger, Emilien Dupont, Po-Sen Huang, Adam Zsolt Wagner, Sergey Shirobokov, Borislav Kozlovskii, Francisco J. R. Ruiz, Abbas Mehrabian, M. Pawan Kumar, Abigail See, Swarat Chaudhuri, George Holland, Alex Davies, Sebastian Nowozin, Pushmeet Kohli, and Matej Balog. AlphaEvolve: A coding agent for scientific and algorithmic discovery. arXiv preprint arXiv:2506.13131, 2025. URL https://arxiv.org/abs/2506.13131.
  33. [33]Sanjay Pant, Georges Limbert, Nick P. Curzen, and Neil W. Bressloff. Multiobjective design optimisation of coronary stents. Biomaterials, 32(31):7755–7773, 2011. doi: 10.1016/j.biomaterials.2011.07.059.DOI
  34. [34]Claus B. W. Pedersen and Peter Allinger. Industrial implementation and applications of topology optimization and future needs. In IUTAM Symposium on Topological Design Optimization of Structures, Machines and Materials, volume 137 of Solid Mechanics and Its Applications, pages 229–238. Springer Netherlands, 2006. doi: 10.1007/1-4020-4752-5_23.DOI
  35. [35]M. J. D. Powell. A direct search optimization method that models the objective and constraint functions by linear interpolation. In Advances in Optimization and Numerical Analysis, pages 51–67. Springer Netherlands, 1994. doi: 10.1007/978-94-015-8330-5_4.DOI
  36. [36]Yupeng Qi, Ran Xu, and Xu Chu. FeaGPT: an end-to-end agentic-AI for finite element analysis. arXiv preprint arXiv:2510.21993, 2025. URL https://arxiv.org/abs/2510.21993.
  37. [37]Nelson S. Ribeiro, João Folgado, and Hélder C. Rodrigues. Surrogate-based multi-objective design optimization of a coronary stent: Altering geometry toward improved biomechanical performance. International Journal for Numerical Methods in Biomedical Engineering, 37(6):e3453, 2021. doi: 10.1002/cnm.3453.DOI
  38. [38]R. Tyrrell Rockafellar and Stanislav Uryasev. Optimization of conditional value-at-risk. Journal of Risk, 2(3):21–41, 2000. doi: 10.21314/JOR.2000.038.DOI
  39. [39]Bernardino Romera-Paredes, Mohammadamin Barekatain, Alexander Novikov, Matej Balog, M. Pawan Kumar, Emilien Dupont, Francisco J. R. Ruiz, Jordan S. Ellenberg, Pengming Wang, Omar Fawzi, Pushmeet Kohli, and Alhussein Fawzi. Mathematical discoveries from program search with large language models. Nature, 625:468–475, 2024. doi: 10.1038/s41586-023-06924-6. URL https://doi.org/10.1038/s41586-023-06924-6.
  40. [40]Titu Ranjan Sarker, Muhammed Jawaad Zulqernine, Ling Yue, Shaowu Pan, Chenxi Wang, and Shiyao Lin. A multi-AI-agent framework enabling end-to-end finite element analysis for solid mechanics problems. arXiv preprint arXiv:2606.00138, 2026. URL https://arxiv.org/abs/2606.00138.
  41. [41]Dirk Steglich and Jacques Besson. Load reversals and damage in magnesium alloys during stent crimping and expansion: Experiments and modelling. Journal of the Mechanical Behavior of Biomedical Materials, 181:107502, 2026. doi: 10.1016/j.jmbbm.2026.107502.DOI
  42. [42]Dirk Steglich, Jan Bohlen, and Bjoern Wiese. SMART-RIm: Simulation methods assuring reliable technology for resorbable load-bearing implants. Data set, Deutsches Elektronen-Synchrotron DESY, Hamburg, 2025.
  43. [43]Rikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab. Sparks of weight-space programming: LLMs can natively read & write neural network weights. Blog post, MIT CSAIL, November 2026. URL https://rikiyat.github.io/blog/weight-space-programming/. Published 2 November 2026.
  44. [44]D. A. Tortorelli and P. Michaleris. Design sensitivity analysis: overview and review. Inverse Problems in Engineering, 1(1):71–105, 1994. doi: 10.1080/174159794088027573.DOI
  45. [45]U.S. Food and Drug Administration. Non-clinical engineering tests and recommended labeling for intravascular stents and associated delivery systems. Guidance for Industry and FDA Staff, 2010.
  46. [46]Hongjun Wang, Li Jiao, Jie Sun, Pei Yan, Xibin Wang, and Tianyang Qiu. Multi-objective optimization of bioresorbable magnesium alloy stent by Kriging surrogate model. Cardiovascular Engineering and Technology, 13(6):829–839, 2022. doi: 10.1007/s13239-022-00619-1.DOI
  47. [47]Ruo-Qian Wang. When is a large language model a good calibration agent? Physics-labeled prompting for budgeted engineering simulator calibration. arXiv preprint arXiv:2610.22379, 2026. URL https://arxiv.org/abs/2610.22379. Submitted to Engineering Applications of Artificial Intelligence.
  48. [48]W. Wu, L. Petrini, D. Gastaldi, T. Villa, M. Vedani, E. Lesma, B. Previtali, and F. Migliavacca. Finite element shape optimization for biodegradable magnesium alloy stents. Annals of Biomedical Engineering, 38(9):2829–2840, 2010. doi: 10.1007/s10439-010-0057-8.DOI
  49. [49]Chengrun Yang, Xuezhi Wang, Yifeng Lu, Hanxiao Liu, Quoc V. Le, Denny Zhou, and Xinyun Chen. Large language models as optimizers. In International Conference on Learning Representations (ICLR), 2024. URL https://arxiv.org/abs/2309.03409.
  50. [50]Shaoliang Yang, Jun Wang, and Yunsheng Wang. Large language models as optimization controllers: Adaptive continuation for SIMP topology optimization. Advances in Engineering Software, 223:104304, 2026. doi: 10.1016/j.advengsoft.2026.104304. URL https://arxiv.org/abs/2603.25099.
  51. [51]Ling Yue, Nithin Somasekharan, and Shaowu Pan. Foam-Agent: A large language model-based multi-agent framework for automating computational fluid dynamics workflows. Computer Methods in Applied Mechanics and Engineering, 461:119271, 2026. doi: 10.1016/j.cma.2026.119271. URL https://arxiv.org/abs/2505.04997.
  52. [52]XinXin Zhang, Zhuoqun Xu, Guangpu Zhu, Chien Ming Jonathan Tay, Yongdong Cui, Boo Cheong Khoo, and Lailai Zhu. Using large language models for parametric shape optimization. arXiv preprint arXiv:2412.08072, 2024. URL https://arxiv.org/abs/2412.08072.
  53. [53]Yu Zhu, Selene Pirola, M. Yousuf Salmasi, Sumesh Sasidharan, Serena M. Fisichella, Declan P. O’Regan, James E. Moore, Jr., Thanos Athanasiou, and Xiao Yun Xu. The influence of material properties and wall thickness on predicted wall stress in ascending aortic aneurysms: A finite element study. Cardiovascular Engineering and Technology, 16(1):52–65, 2025. doi: 10.1007/s13239-024-00756-9.DOI

Paper details

vlnc:W_pmww7t7d
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 8

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents