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

Alexandra Starostina, Alexander Hermann, Mahish Guru, Kartik Bali, Roland Aydin, Christian C. CyronarXiv 2026 · arXiv

TL;DR

Researchers have demonstrated that general-purpose Large Language Models (LLMs), specifically GPT-6 Astra, can act as highly effective engineering design agents. By combining their internal physical intuition with feedback from nonlinear simulations, these agents can optimize complex parts with over 1,600 CAD parameters using a budget of only 15 simulations—outperforming specialized industrial shape optimizers.

What Problem Does This Paper Address?

Engineering design often faces a "trilemma" where automated tools can usually only satisfy two of three critical conditions:

  1. High Parameter Count: Real CAD models have hundreds or thousands of variables.
  2. Small Simulation Budget: High-fidelity nonlinear simulations (like 3D finite-element analysis) take hours, limiting studies to dozens of runs.
  3. Freedom to Redesign: True improvement often requires coordinated changes to a part's proportions (redesigning), not just local surface "polishing."

As shown in Figure 1, classical Bayesian optimizers handle redesigns well but fail as parameter counts explode. Conversely, industrial shape optimizers like Tosca handle thousands of surface nodes but are limited to "polishing"—making tiny local adjustments that cannot fundamentally change a part's proportions. This paper explores whether an LLM agent can bridge this gap by "reading" simulation results like an engineer and "writing" new CAD designs.

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.

Methods

The authors developed an LLM-driven optimization loop (see Figure 2). The process is entirely automated once the initial problem is defined:

  • The Agent: GPT-6 Astra acts as the central intelligence. It receives a text-based "digest" of simulation results, including strain tables and feasibility scores.
  • The Sandbox: The agent can write and execute Python code in a secure sandbox to analyze full-field simulation data (every node and element) and generate plots to inform its next move.
  • The Output: The agent writes JSON design files that define the geometry.
  • The Pipeline: A separate pipeline takes these JSON files, builds the 3D CAD model (using FreeCAD), generates a mesh, runs a nonlinear simulation (using Abaqus or FEBio), and scores the result.
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.

The core novelty is the LLM's ability to propose coordinated changes. Instead of tweaking 1,650 values individually, the agent might decide to give all "unlinked crowns" in a stent the same new shape, effectively imposing its own structure on the high-dimensional design space.

Datasets

The paper evaluates the agent across a "ladder" of three increasingly complex engineering problems:

  1. Magnesium Stent (Primary Benchmark): A 3D nonlinear problem involving large-deformation plasticity and contact. The agent manages ~1,650 CAD parameters with a budget of 15 simulations.
  2. Plate with a Hole: A linear elastic problem based on a standard industrial example for the Tosca optimizer, involving up to 26 parameters.
  3. Single Crown: A simplified 2D version of the stent problem with only 4 parameters, used to compare LLMs against classical optimizers in low-dimensional spaces.

Key Insights and Ideas

  • Intuition + Data: The LLM doesn't search blindly. Its first move usually draws on prior knowledge (e.g., "uniform struts are better than narrowed tips"). Subsequent moves are driven by the specific simulation data it "reads."
  • Redesign vs. Polishing: The agent successfully performed "redesigns"—changing leg lengths and crown proportions—whereas the industrial tool Tosca was stuck "polishing" the existing surface.
  • The "Aha" Moment on Shuffling: To prove the LLM was actually using the simulation data, the researchers performed a control test where they shuffled the results (assigning the strain of one crown to another). The agent immediately followed the "fake" data, proving it was actively learning from the feedback rather than just relying on pre-trained biases.
  • Cost Efficiency: The entire optimization run for the complex stent cost only US$19 in LLM API calls, a negligible fraction of the engineering and compute time involved.

Key Results

  • Stent Optimization: GPT-6 Astra reduced the strain objective (the 90th percentile of crown strains) by 32% within its 15-simulation budget. In contrast, the industrial optimizer Tosca improved the design by only 10% before stalling.
  • High-Fidelity Validation: A design from the agent's pilot runs was tested in an independent, much more complex model (including crimping and damage). It showed 14% less strain and 63% less damage at full expansion compared to the original design.
  • Scaling: On the 26-parameter plate problem, LLM agents beat the industrial baseline in 9 out of 12 runs. However, on the 4-parameter "Single Crown" problem, classical Bayesian optimizers were able to catch up, suggesting LLMs provide the most value in high-dimensional spaces where simulations are expensive.

Limitations and Future Work

  • Human-in-the-Loop Setup: A human still needs to write the initial simulation setup and define the CAD parameterization. The authors noted that weaker LLMs often struggle with this setup phase, accepting incorrect configurations.
  • Geometry Bounds: The agent can only change what the CAD model allows. If a better design requires a topology change not covered by the parameters, the agent is stuck.
  • Temporal Cost: While the simulation budget is small, the wall-clock time is significant (the stent run took 35 hours).
  • Generalization: While successful on stents and plates, the approach needs testing on a wider variety of engineering components to prove its universality.

Summary updated

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.
BibTeX
@misc{starostina2026ai,
  author = {Alexandra Starostina and Alexander Hermann and Mahish Guru and Kartik Bali and Roland Aydin and Christian C. Cyron},
  title = {{AI agents can redesign engineering parts with more than a thousand CAD parameters and nonlinear simulations}},
  year = {2026},
  eprint = {2611.26003},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.26003},
}
Figure