alm-gen/README.md

metadata

license: apache-2.0
tags:
  - materials
  - diffusion
  - crystal-generation
  - mattergen

ALM Gen · de-novo crystal generation

ALM Gen turns a natural-language description into a novel crystal by steering the mattergen_base diffusion decoder with classifier-free guidance. The K=8 [atoms_i] soft tokens feed a consumer-only bridge (a per-token projection whose output is read by the cross-attention consumer inside the decoder); there is no learnable-query producer.

Contents: atoms_mapper.pt (consumer-only bridge, optimizer stripped) + lora_adapter/ (r=8 LLM bridge LoRA) + projector_and_state.pt. The base decoder is fetched separately (mattergen_base, via external/setup_mattergen.sh); pass this repo subdir as --alm_checkpoint (the r8 adapter applies directly).

De novo generation against the MP-20 hull: stability S is E_hull ≤ 0.016 eV/atom, structures pre-relaxed, N=10×1000 (95% CIs in the paper):

Method	E_hull (eV)↓	U (%)↑	V_struct (%)↑	V_chem (%)↑	SUN (%)↑
CrystalTextLLM	0.61	47.40	90.01	91.59	0.38
PLAID++ Wyckoff	0.57	40.70	89.06	91.59	0.50
CrysReas-Base (SFT)	0.58	35.25	84.03	90.36	0.57
CrysReas-Thinking	0.52	38.64	91.29	91.72	0.59
CrysReas-RL	0.53	82.49	89.85	91.10	1.23
CrysReas	0.45	87.23	94.92	91.78	1.70
MatterGen (Base, g=0)	0.079	93.50	100.00	86.50	5.53
ALM Gen (g=0.5)	0.085	98.90	100.00	83.20	7.80
ALM Gen + FK-stoich	0.086	73.80	100.00	84.50	5.21

Steering the base decoder with language at g=0.5 improves SUN over the g=0 MatterGen base (5.53 → 7.80), SoTA on this protocol.

De novo generation on LeMat-GenBench (N=2500): strict stability is Ē_hull < 0, metastability E_hull < 0.1; E_f / E_hull / RMSD scored by 3 MLIPs:

Model	Valid↑	Unique↑	Novel↑	E_f↓	Ē_hull↓	RMSD↓	Stable↑	SUN↑	Meta↑	MSUN↑
MatterGen	95.7	95.1	70.5	-0.70	0.18	0.39	2.0	0.2	33.4	15.0
PLaID++	96.0	77.8	24.2	-0.50	0.09	0.13	12.4	1.0	60.7	7.6
WyFormer	93.4	93.0	66.4	-0.43	0.50	0.81	0.5	0.1	15.7	1.9
WyFormer-DFT	95.2	95.0	66.4	-0.67	0.27	0.42	3.7	0.4	24.8	7.8
MCFlow-S	97.2	96.3	52.2	-0.85	0.10	0.16	11.7	0.7	49.5	18.9
MCFlow-B	97.7	95.5	25.4	-0.91	0.05	0.08	17.6	0.7	64.3	11.9
MCFlow-L	98.6	95.2	18.6	-0.93	0.04	0.06	18.8	0.5	68.3	9.3
ALM Gen	92.2	91.3	61.5	-0.44	0.09	0.20	3.6	0.8	58.7	35.2

Tops the field on metastable yield (MSUN 35.2); second to the de novo-specialist flow models at strict SUN.

Generate structures from a description (inference):

alm-generate generate --alm_checkpoint alm-gen \
    --atoms_mapper alm-gen/atoms_mapper.pt --mattergen_pretrained mattergen_base \
    --prompt "A cubic rock-salt oxide of magnesium." --num_samples 8 \
    --guidance_factor 0.5 --out_dir gen_out

Evaluate (de-novo S/U/N/SUN/MSUN):

alm-eval-dng --alm_checkpoint alm-gen \
    --atoms_mapper alm-gen/atoms_mapper.pt --mattergen_pretrained mattergen_base \
    --guidance_factor 0.5 --num_samples 1000 --out_root out --run_id dng

Links

Paper: arXiv · HuggingFace · Code: GitHub

License

Apache-2.0.

Citation

@article{edamadaka2026atomistic,
  title   = {Atomistic Language Models Understand and Generate Materials},
  author  = {Edamadaka, Sathya and Ramesh, Krithik and Li, Ju and G\'omez-Bombarelli, Rafael},
  journal = {arXiv preprint arXiv:2606.21395},
  year    = {2026}
}