FabriCAD Dataset

A Synthetic Dataset of Manufacturing Plans and corresponding CAD data

About the Dataset

FabriCAD is a fully synthetic dataset of manufacturing process plans paired with corresponding 3D CAD data. It is designed to support research in AI-assisted manufacturing approaches.

Each sample in the dataset primarily consists of a detailed process plan and a corresponding 3D-CAD model. To represent realistic manufacturing processes and conditions, the workflows are based on sequences derived from Markovchains. These were computed from 50,000 real-world process productionplans from the metalworking industry. As a result, all generated parts are related to metalworking. The currently covered process steps include: Affecting geometry:

Other steps: We are currently extending the dataset to include additional features and process steps. The process plan includes information such as individual steps, assigned workstations, estimated durations, and cost estimations. In addition, every step is enriched with detailed, feature-level metadata. For example, in drilling operations, the dataset specifies whether the hole is a fit, includes a thread, and which tools and cutting parameters can be used.

During the generation of the dataset, parametrized CAD-data is available for every modification at the feature level. However, due to storage constraints, the main dataset only contains the final versions of each model. If you require more fine-grained CAD-process data, feel free to contact us.

All values and estimates are highly simplified, but they abstractly reflect realistic processes and workflows in modern industrial environments. The level of detail is designed to support both benchmarking and pretraining of machine learning models in the manufacturing domain.

Disclaimer

As mentioned, all data is entirely synthetic and highly simplified. It is not suitable for production use and is intended exclusively for research in AI-assisted manufacturing. The parts are generated based on a ruleset that enforces general manufacturability, realistic wall thicknesses, and plausible feature placements. However, due to the volume and variability of the data, some unrealistic configurations may still occur. For additional information and code, please refer to the repo.

Preview

Below is a small subset of the data available for preview:

Sample Part A
Sample Part B
Sample Part B

Download

A few samples are available here

Download Test Data

If you're interested in the full dataset or would like to collaborate, feel free to reach out to us at: michel.kruse@fh-kiel.de