Active learning for conditional crystal inverse design
A crystal generator proposes alloys with a target stiffness, a three-stage screen checks them up to DFT, and verified structures retrain it.
The loop
- Proposes
Alloy crystal structures from Con-CDVAE conditioned on a target bulk modulus
- Judged by
Three screens: a CGCNN pre-filter, molecular dynamics with MACE-MP-0 (within 20%, stable), then DFT
- Steers
Structures passing DFT, with computed bulk moduli, join the training set and the generator is fine-tuned
Headline result
For a 350 GPa bulk modulus target, in a range sparse in the training data, three active learning iterations lowered the mean absolute percentage error of generated crystals from about 0.4 to 0.2 (CGCNN-evaluated). DFT bulk moduli of new structures averaged 343.54 GPa (standard deviation 27.20) after the first iteration and 356.52 GPa (standard deviation 18.61) after the second.
What this does not prove
The error trend is measured with the same kind of GNN used in the screen, and DFT is a calculation rather than a lab result. It is a case study on one property (bulk modulus) of alloys, and no structure was synthesized.
How it generalizes the harness
- The generator is not an LLM
- Feedback retrains the generator
- A domain authority judges
- The output outlives the model
A crystal VAE (not an LLM) is retrained on structures that a DFT-backed screening pipeline accepted.
Classification
| What it proposes | Crystal structure |
|---|---|
| Generative model | Con-CDVAE conditional crystal variational autoencoder, fine-tuned each iteration |
| Algorithmic authority | CGCNN pre-filter, MACE-MP-0 molecular dynamics, DFT (via APEX workflows) |
| Feedback | Checks at several fidelities |
| What gets updated | Model weights and training data |
| When it runs | During training |
| Validity | Approximate physics (simulation, not lab results) |
| Evidence | Experiments reported |
| Lineage | Original system |