The structure generation (SG) task tackles the inverse-design challenge of creating entirely new crystal structures that satisfy stability and functional constraints without exhaustive enumeration. Models first embed known crystals into symmetry-aware latent spaces—fractional-coordinate graphs, Wyckoff-sequence tokens, or E(3)-equivariant voxel fields. Generators—diffusion models, graph-autoregressive Transformers, or symmetry-equivariant GANs—sample this space. Running on a single GPU, the framework can propose over a thousand candidate crystal structures per minute, dramatically lowering the trial-and-error cost of discovering scintillators, solid-state electrolytes, and high-entropy compounds. Combined with a rapid, tiered screening funnel—machine-learning potential relaxation, energy threshold filtering, and final DFT refinement—this keeps computation affordable and tightly couples theory with experiment.
| Supported Functions | DiffCSP | MatterGen |
|---|---|---|
| Support Material Types | ||
| Inorganic Materials | ✅ | ✅ |
| Structure Generation | ||
| Random Sample | ✅ | ✅ |
| Condition Sample | ✅ | ✅ |
| ML Capabilities · Training | ||
| Single-GPU | ✅ | ✅ |
| Distributed Train | ✅ | ✅ |
| Mixed Precision | - | - |
| Fine-tuning | ✅ | ✅ |
| Uncertainty / Active-Learning | - | - |
| Dynamic→Static | - | - |
| Compiler CINN | - | - |
| ML Capabilities · Predict | ||
| Distillation / Pruning | - | - |
| Standard inference | ✅ | ✅ |
| Distributed inference | - | - |
| Compiler CINN | - | - |
| Dataset | ||
| Materials Project | ||
| MP20 | ✅ | ✅ |
| Hrbrid | ||
| ALEX MP20 | - | ✅ |
| ML2DDB🌟 | - | ✅ |
Notice:🌟 represent originate research work published from paddlematerials toolkit