ml-generative-adit

You can generate novel crystal structures using a latent diffusion model that samples from the MP20 learned distribution, outputting periodic material candidates directly as CIF files for downstream atomistic simulation and screening.

Yes, you can generate non-periodic molecules by sampling from the QM9 learned distribution using the same unified diffusion model, outputting molecular candidates as XYZ files for computational chemistry workflows.

Yes, running the ADiT MCP tool for crystal and molecule generation requires a CUDA-enabled environment to execute the latent diffusion model and produce the desired CIF or XYZ output files with metadata.

The best way to expand candidate sets is using a generative diffusion model to rapidly produce novel atomistic structures, replacing hand-designed starting points with data-driven candidates for downstream stability and property analysis.

Generated crystal structures feed directly into downstream workflow steps including structural relaxation, stability checks like E_hull calculation, and property prediction to validate the novel material candidates.

Generated structures from diffusion models represent candidates from learned distributions like MP20 and QM9, requiring subsequent structure validation, relaxation, and stability analysis rather than serving as immediately finalized materials.