SELTO: 3D Topology Optimization Problems and Solutions

SELTO is a benchmark dataset for deep learning in 3D topology optimization, containing voxelized problems and solutions. The dataset consists of four subsets (disc simple, disc complex, sphere simple, sphere complex), each split into training and validation sets. It was created by Sören Dittmer of the University of Bremen in cooperation with Ariane Group and Synera, using the Altair OptiStruct implementation of SIMP.

Use Cases

• Training neural networks for 3D topology optimization based on voxel-wise design space and boundary condition data.
• Benchmarking sample efficiency of learned optimization methods using the provided training and validation splits.
• Developing models to predict optimal material distribution given Young's modulus, Poisson's ratio, and force constraints.
• Researching the application of SIMP-based topology optimization solutions generated via commercial software.

Strengths

• Dataset is explicitly designed as a benchmark for deep learning, with defined training and validation subsets.
• Solutions were generated in cooperation with industry partners Ariane Group and Synera using established commercial software (Altair OptiStruct).
• Includes four distinct 3D problem types (disc simple, disc complex, sphere simple, sphere complex).
• Data structure is shape-consistent, with spatially varying quantities defined at voxel centers.

Limitations

• Row count and total dataset size are unknown, which may limit suitability assessment for large-scale training.
• Column-level documentation is absent for the main dataset; field semantics must be inferred from the description of sample files.
• Last update date is unknown; freshness unverified.

Provenance

Source

University of Bremen, created in cooperation with Ariane Group and Synera.

Collection Method

Solutions generated using the Altair OptiStruct implementation of the SIMP method within Synera software.