Nexa_Mat2

AethronPhantom/Nexa_Mat2 is the public artifact repository for the NexaMat materials stack. It contains a frozen physics encoder, diffusion decoder checkpoints, an experimental Qwen-based controller, manifests, and retained evaluation artifacts.

Public viewer: NexaMat Crystal Viewer.

Artifacts

Component	Path	Status
Encoder V1	encoder/v1/nexa_mat_V1_final.safetensors	Frozen physics-grounding checkpoint.
Encoder manifest	encoder/v1/manifest.json	Architecture, checksum, and training metadata.
Original Diffusion V1	decoder/diffusion_v1/final_checkpoint.safetensors	Original constrained diffusion checkpoint.
Diffusion manifest	decoder/diffusion_v1/manifest.json	Architecture, evaluation, and operating-mode metadata.
Recovered native decoder V78	decoder/native_v78/model.safetensors	Recovered native decoder identified in the final closeout documentation.
Native decoder manifest	decoder/native_v78/model_manifest.json	Artifact identity and serialized-state metadata.
Controller pilot	multimodal/controller/nexa_mat_controller_fft_pilot_20260518T234148Z/final_model_merged/model.safetensors	Experimental partial fine-tune of Qwen3-14B.
Cross-attention contract	cross_attention_contract.json	Interface specification for structure, task, controller, and evidence lanes.
Stack manifest	stack_manifest.json	Published component references and stack metadata.
Final evidence summary	eval/FINAL_EVIDENCE_SUMMARY.md	Qualified scientific results and limitations.

The original diffusion checkpoint and recovered native V78 decoder represent different artifact lineages. Select and record the exact component used rather than treating every checkpoint as interchangeable.

Intended Use

NexaMat supports research into inorganic crystal candidate generation and materials candidate triage. Forward workflows propose structures under constraints. Reverse workflows rank or route existing candidates using available physical signals, task requirements, and evidence.

Generated structures are hypotheses for downstream screening. They are not confirmed stable materials and require appropriate relaxation, thermodynamic evaluation, and expert review.

Component Roles

The encoder provides physics-related representations and plausibility signals. It is frozen for downstream generative experiments.

The diffusion decoder proposes or repairs atom identities, fractional coordinates, and lattice geometry. The original diffusion release is intended for constrained best-of-N sampling with filtering rather than one-shot prediction. The recovered native decoder is documented separately in the final closeout artifacts.

The optional controller supports semantic interpretation, candidate routing, and evidence-based explanations. It is an experimental Qwen3-14B partial fine-tune and is separate from the compact encoder-decoder stack.

The cross-attention contract documents intended interfaces between components. It should not be interpreted as proof that every described integration is implemented or independently validated.

Public Viewer

The public Space provides an interactive interface for exploring forward and reverse workflows. At the inspected Space revision, 2b332807f1a7119145f9d1a376bd96fadd2c17a0, its backend uses a deterministic demonstration adapter. Viewer outputs should therefore not be presented as measured inference from the published checkpoints.

Evidence and Limitations

The archived evidence supports useful static validity and novelty across several inorganic profiles. It also records isolated low-hull candidates, but does not establish repeatable high-yield metastable generation.

Static validity, novelty, relaxation convergence, and thermodynamic stability are distinct properties. Success on one does not establish success on the others. Results from different checkpoints, selection procedures, or evaluation cohorts should not be combined into a single performance claim.

The controller pilot requires evidence grounding for literature or scientific claims. A successful routing fixture does not establish general materials reasoning ability.

These artifacts do not replace DFT, experimental validation, or synthesis review. Organic materials, molecular crystals, polymers, MOFs, and alloys/intermetallics are outside the qualified final evaluation panel.

The project is archived and on ice indefinitely. The release preserves useful artifacts and negative results without claiming a production-ready stable-material discovery system.

Reproducibility

For any reproduction, pin the model repository revision and the selected checkpoint path. Record the execution-code revision, input source, task definition, environment, evaluation procedure, and exact command separately.

Existing outputs in examples/ and eval/ are retained artifacts. Label them as retained results unless the corresponding computation has been rerun. Do not infer new runtime or memory claims from saved predictions.

The public reverse-pool example includes one record covering six candidates and a saved rule-based comparison baseline. Its archived scores are evidence from the original evaluation, not an independently reproduced benchmark. The original input fixture and execution details must be supplied before claiming a complete reproduction.

Licensing

Original NexaMat components and author-owned documentation and examples are released under the MIT License. The license: other entry in the archived final model card is an outdated metadata label, not an additional license.

The optional controller derives from Qwen3-14B and remains subject to the upstream Apache License 2.0 terms, including applicable attribution and redistribution requirements. The repository’s MIT declaration does not replace those upstream terms.

Third-party datasets and other upstream assets retain their respective licenses and access conditions. This release does not grant redistribution rights to source datasets.

Citation

Refer to the technical paper and final model card in docs/final_closeout/. When reporting results, cite the exact repository revision, checkpoint path, and evaluation protocol used.