Tutorials¶
A hands-on walkthrough of MolCraftDiffusion, ordered as a learning path: prepare data → train → generate → evaluate. Each tutorial is self-contained, but they build on one another, so first-time users should follow them in order.
Tip
New here? Start with Installation, then work through Data Preparation → Training a Diffusion Model → Generation Overview. The rest are optional deep-dives you can reach for when you need them.
Start Here¶
No data, no training. Fetch a pretrained model from the zoo and generate 3D molecules in three commands — then come back and train your own.
1 · Prepare¶
Compile raw .xyz files into ASE databases, featurise, augment, and wire them into the DataModule.
2 · Train¶
Train a 3D molecular diffusion model from scratch with the override-only config workflow.
Train a property predictor — a standalone model and the basis for gradient guidance.
Train a time-aware regressor on noisy data to steer generation towards target properties.
Adapt a pretrained model to a new chemical space, add conditions, or specialise for outpainting.
3 · Generate¶
Unconditional sampling and the three generation modes at a glance.
Inpainting, outpainting and SILVR with 3D geometric constraints, plus a full parameter-tuning guide.
Steer generation with Classifier-Free Guidance, Gradient Guidance, or a hybrid of both.
4 · Evaluate & Scale¶
Predict properties for new molecules and benchmark a model against a labelled set.
Post-generation analysis: xTB optimisation, validity metrics, RMSD, electronic properties, featurisation.
Grid and Bayesian sweeps over controlled-generation parameters with automatic metric collection.