EvoCAD: Evolutionary CAD & Material Design

• EvoCAD is an advanced framework that employs evolutionary algorithms to generate CAD code and design voxel-encoded materials with AI-driven optimizations.
• It integrates LLMs and VLMs in symbolic CAD generation and leverages physics-based simulations to ensure high topological and geometric fidelity.
• Demonstrated applications include precise material deformations and bio-inspired actuators, achieving superior performance in both virtual and physical design tasks.

EvoCAD is a class of systems that utilize evolutionary algorithms in conjunction with generative models or physics-based simulation to design and fabricate objects encoded at the level of computer-aided design (CAD) or material voxels. Recent instantiations include evolutionary CAD code generation guided by vision-LLMs and evolutionary design of voxel-encoded hard-magnetic soft active materials, both of which leverage evolutionary strategies to optimize non-trivial functional or structural properties (Preintner et al., 13 Oct 2025, Wu et al., 2020).

1. Underlying Problem Domains

EvoCAD methodologies address automatic construction of objects—either as CAD representations or physical structures—where the complexity of the design space or manufacturing constraints preclude exhaustive search or simple parametric optimization. In CAD, this involves deriving code (e.g., CADQuery scripts) that satisfies both syntactic constraints (compilable, watertight solids) and semantic constraints (alignment with detailed natural-language specifications). In voxel-encoded material design, the aim is to assign per-voxel properties such as magnetic density and orientation to morph a structure into a prescribed target shape or enable specific biomimetic actuation under external fields (Preintner et al., 13 Oct 2025, Wu et al., 2020).

2. EvoCAD for Evolutionary CAD Code Generation

The EvoCAD framework for symbolic CAD object generation integrates LLMs and vision LLMs (VLMs) within a population-based evolutionary optimization cycle (Preintner et al., 13 Oct 2025). The system iterates over two main phases:

• Initialization: An LLM is prompted, using $k$ few-shot CADQuery script examples, to produce $M$ candidate CAD scripts for a given natural-language prompt $p$. Compilation failures invoke a single-pass LLM-based self-debugging step.
• Evolutionary Optimization: Across $N$ generations, each CAD script is rendered to multiview images, described in natural language by a VLM, and ranked by a reasoning LLM for semantic alignment to the original prompt. Offspring are generated by LLM-guided crossover ("merge") and mutation ("refinement"). Elitism ensures the best designs persist.

Fitness evaluation bypasses ground-truth comparison during search, instead utilizing VLM and reasoning LLM outputs as surrogate reward signals. Selection leverages exponential ranking probabilities: $$p_i = \frac{\exp(-\lambda r_i)}{\sum_j \exp(-\lambda r_j)}$$ with $\lambda=0.5$. Mutation rate $p_m=0.5$ controls the likelihood of LLM-driven code refinement (Preintner et al., 13 Oct 2025).

3. EvoCAD for Voxel-Encoded hmSAM Inverse Design

In the context of hard-magnetic soft active materials (hmSAMs), EvoCAD denotes an integrated pipeline unifying direct-ink-write (DIW) voxel-level printing, per-voxel magnetization encoding, and evolutionary search (Wu et al., 2020). The design representation discretizes structures into $N$ voxels, each parameterized by normalized magnetic density $\rho_i \in [0,1]$ and magnetization direction $\mathbf{m}_i \in \mathbb{S}^2$:

$M$0

The inverse design problem is to minimize the deviation between desired curvature profiles $M$1 and FEM-simulated curvature $M$2 via:

$M$3

Subject to $M$4, $M$5 for all $M$6. The evolutionary loop employs tournament or rank-based parent selection, uniform crossover (direct swap or slerp on directions), scalar and vector mutations, and generational or $M$7 replacement, terminating on reaching an error threshold or maximum generations (Wu et al., 2020).

4. Fitness Metrics and Evaluation Protocols

CAD Code Generation

Fitness is assessed by LLM-based ranking of VLM-produced geometric descriptions against the natural language prompt, repeated three times and averaged for rank stability. Beyond standard spatial metrics—point-cloud distance (PCD), Hausdorff distance (HDD), Intersection-over-Union (IoU), and Dice similarity (DSC)—EvoCAD introduces two topological metrics based on the Euler characteristic $M$8:

Metric	Formula	Semantic Target
Topology Error ($M$9)	$p$0	Magnitude of topological mismatch
Topology Correctness ($p$1)	$p$2 if $p$3, $p$4 otherwise	Binary correctness of genus, holes, cavities

These metrics uniquely measure hole/cavity arrangements that can escape detection by purely spatial scores (Preintner et al., 13 Oct 2025).

Voxel-Encoded Material Design

Fitness is given directly by squared curvature error. Verification includes both FEM simulations and experimental realization, with metrics such as RMS curvature error and percentage shape error. In showcased applications, RMS curvature errors below $p$5 mm$p$6 and shape errors under $p$7 are attained in physical beams (Wu et al., 2020).

5. Demonstrated Applications

CAD Code Evolution

On the CADPrompt benchmark (200 cases), EvoCAD-4o achieves $p$8, $p$9, PCD $N$0, and DSC $N$1, outperforming baselines such as CADCodeVerify and 3D-Premise in topological fidelity and spatial agreement (Preintner et al., 13 Oct 2025). Optimization metrics improve monotonically across generations; failures in prior art to preserve genus are corrected by the evolutionary approach.

Voxel-Encoded Soft Actuators

Demonstrated cases include:

• Quarter-Circle, Parabolic, and Cosine Beams: EA-derived non-uniform $N$2 and $N$3 distributions yield target-shape deformations under external magnetic fields.
• Inchworm Robot: EA symmetry constraints produce bidirectional gait; 2 mm net displacement per actuation cycle.
• Dog Trot Walker: Differential leg curvature profiles, designed by EA, enable coordinated multi-limb biomimetic motion patterns (Wu et al., 2020).

6. Systemic Limitations and Scalability

Current EvoCAD instantiations in CAD code evolution utilize relatively small populations ($N$4) and generations ($N$5) due to LLM inference costs, limiting exploration of highly complex or multiobjective design tasks. A plausible implication is that future scaling of population size and evolutionary depth could further improve solution diversity and robustness, especially as computational resource constraints are eased (Preintner et al., 13 Oct 2025). In material design, while voxel-level encoding permits fine geometric and functional control, fabrication speed and physical simulation accuracy impose limits on achievable structure scale and design cycle time (Wu et al., 2020).

7. Outlook and Significance

EvoCAD frameworks demonstrate that the fusion of evolutionary search with model-based or physics-based surrogate fitness and population-level search strategies enables substantial advances over single-step or purely greedy refinement in both virtual (symbolic CAD) and physical (voxel-encoded material) design domains. Notably, the approach supports direct integration with manufacturing pipelines (e.g., DIW for hmSAMs) and opens research avenues in AI-driven parametric and functional design. Extensions to multiobjective optimization and human-in-the-loop strategies represent promising directions as underlying generative and evaluative technologies continue to evolve (Preintner et al., 13 Oct 2025, Wu et al., 2020).