DeepMind's new crystal structure prediction model outperforms CGCNN

The new model combines message passing with symmetry-aware tokenization and reports stronger performance on formation energy and elastic tensor prediction benchmarks than CGCNN and MEGNet baselines. Their ablation suggests most gains come from enforcing space-group constraints during training rather than raw parameter count.

I would still like to see broader transfer tests. Many benchmark sets overlap heavily with public repositories used in pretraining, so true out-of-domain generalization is hard to judge. If anyone has tried this model on low-symmetry organic-inorganic hybrids, please share failure cases.