EPICCAD benchmark gives AI the rules behind editable engineering designs

Artificial intelligence can now generate plausible mechanical parts from a text prompt, but the models it produces often fail at the step that matters most in real engineering: being edited. Change one dimension of an AI-generated component and the surrounding features frequently refuse to follow, because the system learned only what a part looks like, not the rules that hold it together. A team of researchers — Zehe Yin, Weixin Lin and Xiangbiao Kong — reports in the journal Discover Artificial Intelligence a new benchmark called EPICCAD, built to close that gap. It pairs more than 160,000 CAD models with the procedural histories, explicit geometric constraints and intent-aware language descriptions that AI systems need in order to generate designs that remain valid, editable and faithful to engineering intent after their parameters change.

The problem the benchmark addresses is rooted in what a CAD model actually is. Unlike a mesh or a point cloud, a parametric CAD model is a recipe: an ordered sequence of sketches, extrusions and Boolean operations, each carrying dependencies that must survive every future modification. The engineering-design community has long held that high-quality solid models must be usable, reusable and semantically rich rather than merely geometrically correct. Meeting that standard depends on a part’s modeling history, its geometric constraints and its feature interactions — precisely the structure that most learning-based CAD corpora fail to expose. As a result, even systems built on large language models and modern generative methods struggle to produce outputs that are at once valid, editable and faithful to the designer’s intent.

The authors identify four specific shortcomings in existing datasets. Most sequence-based resources omit explicit geometric constraints such as tangency, parallelism or concentricity, so learned models cannot recover the relational structure that governs editability. Widely used corpora are dominated by simplified academic parts and underrepresent the modeling depth, feature diversity and transformation patterns of industrial practice. Their text labels are semantically sparse, capturing coarse appearance while rarely expressing functional intent or finer geometric detail. And their hierarchical sketch encodings introduce redundant nesting that inflates token length and makes sequence learning harder than necessary. EPICCAD was designed to attack all four weaknesses at once.

The benchmark combines 152,360 academic samples with 8,141 industrial Siemens NX parts, yielding 160,501 models. Every sample aligns five modalities: an executable modeling sequence, multi-view renderings, STEP boundary-representation files, native parametric files, and textual annotations. The academic split draws on three public corpora — DeepCAD, SketchGraphs and Fusion 360 Gallery — which describe design state at different levels and had to be bridged by a conversion pipeline. DeepCAD stores feature histories in hierarchical 3D form but omits sketch constraints; SketchGraphs exposes rich 2D constraints but lacks extrusion parameters and links from sketches to 3D features; Fusion 360 Gallery contains real design histories whose sketch-feature correspondences must be reconstructed before alignment. A cross-dataset constraint-propagation pipeline flattens DeepCAD profiles through a symmetric-difference formulation and then transfers SketchGraphs constraints onto the recovered primitives, producing constraint-aware 3D histories that none of the source datasets contain individually.

The industrial subset contributes something the academic data cannot: professionally built models read directly from Siemens NX files, parsed programmatically into the shared EPICCAD schema. These parts carry longer feature chains, more tightly coupled constraints and a wider repertoire of transformations, including rotated extrusions fixed by an axis and a pair of angular limits — a construct that is rare in the largely upright academic data. Although the industrial subset accounts for only about five percent of all samples, its deeper histories carry a disproportionate share of the procedural difficulty, and the authors show that this added complexity translates directly into better generalization rather than merely harder training data.

The benchmark’s central technical decision concerns how sequences should be written down. EPICCAD flattens only the redundant sketch-loop nesting inside individual profiles — the verbose face-and-loop containers that wrap primitives — while retaining the ordered feature history of sketches, extrusions and Boolean operations as a procedural sequence. Loop membership is not discarded but reconstructed deterministically from primitive connectivity at execution time, and the boundary-representation topology of surfaces, edges and vertices is preserved in the aligned STEP files. The geometric relationships that nesting implicitly encoded are re-expressed explicitly through a constraint set of ten types: coincident, parallel, perpendicular, horizontal, vertical, tangent, equal, concentric, fix and normal. The authors are careful to note that this choice targets a specific machine-learning obstacle — order-sensitive loop containers that force autoregressive models to learn bracketing conventions carrying no geometric information — rather than a claim that flatter trees are universally preferable for human authoring.

Experiments quantify the payoff of this design. On a shared subset of roughly 129,000 models, token counts measured with the Qwen3-0.6B tokenizer show that EPICCAD sequences remain 74.7 percent shorter than DeepCAD’s and 15.0 percent shorter than Text2CAD’s, even after geometric constraints are included. Compactness does not come at the cost of accuracy: EPICCAD reaches a Chamfer Distance of 7.59 × 10⁻³ against ABC ground-truth meshes, slightly better than DeepCAD’s 7.76 × 10⁻³ and substantially better than Text2CAD’s 10.02 × 10⁻³. The authors caution that both EPICCAD and Text2CAD include parameter quantization and that final meshes are obtained through geometric-kernel execution, so small discrepancies reflect practical sources of approximation rather than loss of representational power.

Explicit constraints carry a measurable price — a 32.7 percent increase in sequence length, or 117.24 tokens per sequence on average — but the editability study shows why that cost is worth paying. Comparing constraint-preserving and constraint-free variants of the same sketches in FreeCAD, the authors find that without constraints, operations such as moving an endpoint or modifying an arc radius frequently destroy intended relations including tangency, equality and concentricity, even when the loop remains technically closed. With the original constraints retained, the same edits propagate through the dependent geometry and preserve the intended design semantics. In text-to-CAD generation, fine-tuned Qwen3-8B models trained on EPICCAD sequences with an auxiliary constraint-prediction objective achieve an invalidity ratio of 1.40 percent and an average Chamfer Distance of 3.00, lowering the invalidity ratio by roughly 40 percent and the average Chamfer Distance by 37 percent relative to a Text2CAD baseline, and raising execution success from 97.65 to 98.60 percent.

Language supervision comes from AM-EPICCAD, an annotation module that grounds descriptions in parsed procedural structure rather than rendered views. The module parses each modeling history into loop and hole organization, per-part oriented bounding boxes, and pairwise spatial relations classified through a Separating Axis Theorem test as separated, touching, intersecting or containing. This structured evidence is serialized into a natural-language transcription and passed to a large language model, which produces three complementary annotation types: a description of the modeling process, a summary of geometric structure, and a concise functional-type label such as a flange or a gear. The authors are candid that the functional labels are a coarse, weakly supervised signal — LLM-inferred function can be generic, occasionally incorrect and biased toward well-represented part categories — and they recommend pairing the labels with expert review for safety- or manufacturing-critical use. In controlled generation experiments, models trained with these intent-aware annotations more reliably produced shapes reflecting the intended usage category and better preserved salient shape cues and spatial layout than models trained on appearance-focused captions.

The industrial split proved to be the single largest driver of out-of-distribution robustness. In a training-composition ablation, adding Siemens NX histories cut the industrial-test Chamfer Distance by 68.2 percent and the weighted-average Chamfer Distance by 32.1 percent, while leaving performance on the academic test sets essentially unchanged — evidence, the authors write, that the industrial histories broaden procedural coverage rather than injecting harder, noisier examples. The team also acknowledges the benchmark’s limits: the academic subset inherits biases toward simple, repetitive parts; large assemblies, proprietary feature types and domain-specific standards remain underrepresented; and metrics such as token length, invalidity ratio and Chamfer Distance do not by themselves certify manufacturability, standards compliance or human-judged reusability. The original proprietary Siemens NX files cannot be redistributed because of licensing restrictions, though derived vendor-neutral sequences, STEP exports and statistics are available on request. Even with those caveats, the authors position EPICCAD as a practical foundation for AI systems that must generate, interpret and revise executable CAD models — a step toward design assistants that understand not just what objects look like, but the rules by which they are made.

Subject of Research: A multimodal benchmark for editable, constraint-preserving history-based CAD modeling

Article Title: A multimodal benchmark for editable constraint preserving history based CAD modeling

Article References: Yin, Z., Lin, W., & Kong, X. (2026). A multimodal benchmark for editable constraint preserving history based CAD modeling. Discover Artificial Intelligence, 6(1), Article 1022. https://doi.org/10.1007/s44163-026-01899-5

Image Credits: AI Generated

DOI: 10.1007/s44163-026-01899-5

Keywords: EPICCAD, CAD generation, geometric constraints, parametric modeling, text-to-CAD, Siemens NX, multimodal benchmark, machine learning, engineering design, procedural history, boundary representation, large language models