PhyCAD trains AI-generated CAD designs against geometry and physics checks

From descriptions to constrained designs

In a preprint posted on November 6, 2026, Yuxin Cui, Weijie Zheng, Zhengyang Li, Han Wang, Wenjie Feng and Qiyu Kang of the University of Science and Technology of China describe PhyCAD. The work is not established peer-reviewed evidence.

The system generates programs in CadQuery, a Python-based tool for creating editable, parametric computer-aided design models. Instead of reproducing one precisely specified object, it receives a description of required features, allowable dimensions, material properties, loads and physical limits. Its task is to satisfy those requirements while minimizing mass.

That distinction matters: several different geometries can be acceptable solutions. Matching the appearance of a reference component is not the goal.

Building a simulation-based dataset

The researchers started with the Text-to-CadQuery corpus, containing approximately 110,000 programs. They retained 77,067 designs that successfully completed their CAD and engineering-analysis pipeline. Reference designs had to produce exactly one valid connected solid.

For each reference, an automated procedure identified opposite terminal surface groups. One group was fixed, while the other received a distributed load totaling 100 kilonewtons. The simulations used three-dimensional, small-strain linear elasticity and one material setting: isotropic structural steel.

The researchers then removed exact dimensions from the descriptions and modeling-operation information. Prompts retained structural features and relationships, but provided only coarse dimensional ranges. Stress, displacement and mass limits were derived from reference measurements by adding 10% and rounding upward. This kept the reference feasible without disclosing its exact measurements.

These limits are therefore benchmark requirements derived from reference designs, not independently established safety ratings.

The dataset supplied 11,024 examples for supervised fine-tuning, in which models learned from reference programs. Another 62,670 prompts supported reinforcement learning, with 3,373 held-out tasks excluded from policy updates and used for evaluation.

Making feasibility take priority

PhyCAD checks generated programs in three stages: successful CAD construction, geometric compliance and finite-element analysis. The last stage evaluates displacement, mass and a volume-weighted 99th-percentile von Mises stress measure—not the absolute maximum stress.

Candidates fall into four levels: construction failure; successful construction but failed geometry; passed geometry but failed simulation or physical requirements; and full compliance.

The training method builds on Group Relative Policy Optimization, or GRPO. For each prompt, the model produces four candidates. PhyCAD compares their verification levels, then separately ranks candidates within the same level on relevant objectives.

This structure addresses a weakness of simply adding rewards together: an improvement in one measurement could otherwise compensate for a requirement that remains unmet. PhyCAD weights progress between levels strongly enough that a higher-level candidate always receives a larger training advantage than a lower-level candidate in the same group.

That guarantee concerns the ordering of training feedback, not guaranteed success on future designs. Among fully compliant candidates, feedback rewards lower mass alongside stress and displacement margins, with margin rewards capped at 10%.

After training, the model generates a program directly, without an iterative simulation-guided refinement loop at inference time.

What the evaluation found

Across five open-source models, joint success—meeting both geometric and physical requirements—increased from 21.76–37.36% after supervised fine-tuning to 39.49–58.05% after PhyCAD’s reinforcement learning.

Ministral-3-3B achieved the highest reported joint success, rising from 37.36% to 58.05%, a gain of 20.69 percentage points. Its valid-CAD rate reached 98.99%, illustrating the remaining gap between executable modeling code and a design that satisfies every requirement.

An ablation using Qwen2.5-3B helped isolate the training method’s contribution. Ordinary, non-hierarchical GRPO raised CAD validity but reduced joint success from 21.76% to 18.80%. Hierarchical GRPO increased joint success to 39.49%. A more costly, staged training sequence reached 42.01%.

The trained models also outperformed three proprietary models under direct prompting on this benchmark. That comparison evaluates specialized training against direct prompting; it does not establish superiority across other engineering tasks or workflows.

What remains unproven

Even the best reported model failed to meet all requirements on more than two-fifths of tasks. High CAD validity should therefore not be mistaken for reliable engineering compliance.

The authors acknowledge that broader engineering intent is weakly represented, material variation is not systematically covered, and critical local dimensions need stronger constraints. The reported evaluation also remains within the specified linear-elastic steel simulations and automatically constructed loading conditions.

PhyCAD provides evidence that structured engineering feedback can improve text-to-CAD generation. It does not establish that generated components are ready for manufacture or safe in service. The authors provide a code repository for the framework.