Nexa_Mat Gen Stack

Generating candidate0%

Preparing model stack

La7Ge7Te7

stack 0.782

min d 1.104 Å

drag rotate · fixed zoom

Forward Generation

Generate a plausible crystal candidate from constraints, then trace encoder, diffusion, and multimodal agreement.

Forward use case

Allowed elements Samples

Forward constraints Novelty temperature 0.42 Lattice strictness 0.72

Reverse Pool Search

Upload or paste a CSV, JSON, or JSONL candidate pool. The stack goes in reverse: parse candidates, score them against the use case, then rank what should be inspected first.

Reverse use case Pool file (.csv, .json, .jsonl) Candidate pool Reverse constraints

Viewer Knobs

Realtime Inference Pass

1. Physics EncoderValidates atoms, local distances, lattice volume, and manifold score.

2. Diffusion GeneratorRepairs noisy coordinates and proposes candidate variants under constraints.

3. Multimodal ControllerAdds application context and evidence while flagging unsupported claims.

4. Decision LayerSelects candidates for DFT follow-up or reverse ranking.

Adapter Contract

This app is wired so local demo functions can be replaced with real endpoints: `encode(structure)`, `diffuse(prompt, constraints)`, `ground(candidate, context)`, and `rank(pool, target)`.

How To Use This App

Use Forward Generation when you want the stack to propose a new material from a design intent and element constraints. Use Reverse Pool Search when you already have candidates and want the stack to rank them against a target application.

The 3D view is inspection-first: drag to rotate, scroll to zoom, hover atoms for species details, and toggle bonds or the unit cell when the structure gets dense.

What It Can Do

It can sketch candidate crystals, show basic geometry health, expose candidate distributions, rank a pool, and explain which model lane is responsible for each part of the decision.

Limits

This is a UI/controller render. It is not a DFT result and should not be treated as proof of stability. Production use should plug in the frozen encoder, diffusion checkpoint, multimodal evidence model, and downstream DFT triage.