Geometrical Deep Learning on 3D Models: Classification for Additive Manufacturing

3D printing has a lot of potential to the classical manufacturing process: high geometric freedom of design, functional integration, production on demand, strong and lightweight parts, etc. In automotive, additive manufacturing is mainly used to produce customizable details, prototypes and pre-series vehicles, operating equipment and motorsport parts.

This project mainly focuses on the initial step of 3D manufacturing process – estimation if certain part could be produced on the specific 3D printer. Currently this involves manual work of engineers, that check if certain part meets requirements that are specified by the 3D printer manufacturer. This process is usually very time consuming and requires expert knowledge. We would like to develop a machine learning approach that would be able to differentiate between printable and non- printable parts based on their geometrical form.

The project is organized in the following steps. First, 3D model data is normalized and converted to the voxel representation. Second step is synthetic data generation: defects, that lead to the non-printable forms, are inserted into the 3D shapes objects. Finally, we will develop and test a neural-network-based solution for classification between printable and non-printable shapes.

The project pipeline will include elements of differential geometry, geometrical deep learning, transfer learning and basics behind additive manufacturing process.