Beyond prediction, machine learning can propose entirely new molecules and crystal structures with desired properties, inverting the structure–property relationship to enable de novo design. This workshop explores the full spectrum of generative approaches applied to chemical and materials space, including variational autoen- coders (VAEs), generative adversarial networks (GANs), diffusion models, and flow-based methods.

The session examines how latent space representations of chemical structure are constructed, focusing on the Junction Tree VAE for molecular graphs and the Crystal Diffusion Variational Autoencoder (CDVAE) for periodic crystal structures. Conditioning strategies that enable the generation of structures with specified target properties are developed alongside the critical question of validation: how to assess whether a generated structure is thermodynamically stable, synthesisable, and genuinely novel. The landmark GNoME result from DeepMind — the prediction of 2.2 million stable inorganic crystal structures — illustrates both the ambition and the current state of the field, and the workshop examines what experimental follow-through and validation actually require to move from predicted structures to synthesised compounds.

Participants will gain a clear understanding of the capabilities and practical limitations of current generative approaches, the validation requirements for scientific credibility, and the prospects for generative design as a component of the materials discovery pipeline.

Extended Key Takeaways & Outputs

• Understanding of VAE, GAN, diffusion, and flow-based generative approaches for chemical design

• Knowledge of latent space construction and conditioning strategies for target properties

• Critical assessment of validation requirements: stability, synthesisability, and novelty

• Contextualisation of the GNoME result and its implications for materials discovery

• Realistic evaluation of the path from generated structure to synthesised material