---
library_name: pytorch
tags:
- materials-science
- crystal-generation
- diffusion
- encoder
- multimodal
- gradio
license: mit
---

# Nexa_Mat2

`AethronPhantom/Nexa_Mat2` is the public artifact repository for the Nexa_Mat Gen Stack. It contains the frozen physics encoder, the constrained diffusion decoder, the experimental multimodal/controller pilot, and the stack manifest consumed by the public Space.

Space: https://huggingface.co/spaces/AethronPhantom/nexamat-crystal-viewer

## Artifacts

| Component | Path | Status |
|---|---|---|
| Encoder V1 | `encoder/v1/nexa_mat_V1_final.safetensors` | Frozen downstream handoff checkpoint. |
| Encoder manifest | `encoder/v1/manifest.json` | Architecture, source URI, checksum, and training metadata. |
| Diffusion V1 | `decoder/diffusion_v1/final_checkpoint.safetensors` | Production diffusion/checkpoint handoff for constrained sampling. |
| Diffusion manifest | `decoder/diffusion_v1/manifest.json` | Architecture, eval, checksum, and operating-mode metadata. |
| Controller pilot | `multimodal/controller/nexa_mat_controller_fft_pilot_20260518T234148Z/final_model_merged/model.safetensors` | Experimental Qwen-based controller pilot. |
| Cross-attention contract | `cross_attention_contract.json` | Interface contract joining encoder, decoder, controller, evidence, and task-head lanes. |
| Stack manifest | `stack_manifest.json` | Canonical manifest for the public Space and downstream tooling. |

## Intended Use

The stack is intended for materials candidate triage. Forward mode proposes constrained candidate structures from a design intent. Reverse mode ranks a candidate pool against a target use case. Generated candidates should be treated as hypotheses for DFT and downstream validation, not as confirmed stable materials.

## Component Roles

The encoder is the physics-grounding layer. It learned a periodic materials manifold from the Nexa_Mat V1 training surface and is frozen for downstream generative experiments.

The diffusion decoder is the proposal layer. It repairs and proposes structures under constraints, but should be used as a best-of-N sampler with filtering instead of a one-shot oracle.

The multimodal/controller layer is the semantic/evidence layer. It connects model outputs to use cases, evidence packets, explanations, and reverse-pool ranking. The published controller checkpoint is a pilot, not a final full fine-tune.

## Limitations

These artifacts do not replace DFT, relaxation, experimental validation, or synthesis review. The decoder is strongest in constrained sampling and weaker in unconditional high-yield generation. The controller should not be trusted for unsupported literature claims unless paired with an evidence retrieval layer.