NexaMat Final Project Compendium
Date: 2026-07-24
Project state: archived and on ice indefinitely
Model state: shipped with qualified claims
1. Executive Summary
NexaMat began as a compact but brittle materials generator. Its atom predictions were strong, its outputs were novel, and its encoder carried useful physical signals, but the decoder placed too little probability mass on valid periodic geometry. Early external runs produced only a handful of strict CIFs from thousands of samples. This made every downstream stability benchmark appear catastrophic because the expensive scientific stack was starved before it could evaluate enough usable structures.
The alignment campaign solved that first problem. It introduced honest native accounting, explicit failure reasons, profile-conditioned chemistry, radius-aware geometry, charge-aware composition logic where appropriate, best-of-N native rollouts, and target-gated evaluation. The result is a 4.061859M-parameter stack that produces useful native candidate pools across six non-alloy inorganic profiles.
The second problem was harder. A statically valid crystal can still sit far from a stable energy basin. In one prior bounded test, NexaMat produced two M.S.U.N. candidates from 64 and individual hull values near 0.18 eV/atom, while the cohort mean remained 0.8750 eV/atom. This shows isolated low-hull capability, not a typical 0.18 eV/atom distribution or a verified production yield. Later cohorts did not fully verify recurrence of the historical 3.125% result, which was already far below the intended 25-50% yield. Several weeks of force, trajectory, conservative-energy, tensor, routed-expert, and foundation-residual experiments improved proxies without moving the external objective reliably. The project is therefore put on ice indefinitely at a clear plateau rather than being called a complete failure or being kept alive through unproductive iteration.
2. What The Model Is
NexaMat is a modular materials stack:
flowchart LR
P["Profile and generation request"] --> D["Native diffusion decoder"]
D --> C["Native CIF candidate"]
C --> S["Static profile gate"]
S --> E["Physics encoder and ranking"]
E --> R["External relaxation"]
R --> H["Hull and metastability evaluation"]
H --> O["Evidence-backed candidate set"]
The decoder jointly proposes atom identities, fractional coordinates, and a periodic lattice. The encoder supplies physics-aware representation and scoring. External relaxation and hull evaluation remain authoritative for stability.
The compact stack excludes the optional Qwen3-14B controller:
physics encoder: 2,813,714 parameters
diffusion decoder: 1,248,145 parameters
compact full stack: 4,061,859 parameters
3. The Original Failure
The initial decoder learned chemistry more readily than global crystal geometry. Atom vocabulary and novelty saturated while lattice scale, periodic pair distances, charge-compatible composition, and coordinate placement remained brittle. The conceptual mistake was treating generic denoising loss and teacher-forced validation as adequate substitutes for the density of native benchmark-eligible CIFs.
The decoder was therefore good at producing different-looking compositions but poor at instantiating them as usable crystals. The encoder became a crutch because it had to reject nearly everything.
4. The Static-Validity Recovery
The recovery succeeded because it changed the contract, not merely the loss:
- raw candidates were retained;
- native and repaired surfaces were separated;
- failure reasons became explicit;
- SSB chemistry became a profile rather than a global hardcoded identity;
- ionic and non-ionic profiles used different charge policies;
- radius-aware site assignment and lattice checks attacked periodic geometry;
- target-gated early stopping reduced wasted benchmark time;
- checkpoint selection used external native yield instead of final-step loss.
The strongest SSB lane, V90, reached:
| Metric | Result |
|---|---|
| Accepted | 595/4096 |
| Filter accept | 14.53% |
| Allowed elements | 100.00% |
| Charge neutrality | 100.00% |
| LeMat-distance validity | 14.53% |
| Strict validity | 54.20% |
| Formula uniqueness | 41.80% |
| Novelty | 100.00% |
The final six-profile suite accepted 513 SSB, 411 battery, 424 oxide, 436 halide/chalcogenide, 256 semiconductor, and 242 broad-inorganic candidates. The last two were tolerant passes against profile-specific targets. Alloys and intermetallics remained unresolved.
5. The Stability Gap
Static validity asks whether a crystal is structurally and chemically plausible enough to evaluate. Metastability asks whether relaxation places it close to a competitive energy basin. The latter is a much narrower target.
The broad 512-candidate screen showed the gap:
relaxation convergence: 58 / 512
weighted mean hull: 1.2330 eV/atom
M.S.U.N. candidates: 0 / 512
post-relax validity: high
The best bounded low-hull anchor improved the scientific picture:
relaxation convergence: 3 / 64
mean hull: 0.8750 eV/atom
M.S.U.N. candidates: 2 / 64
best individual hull rows: 0.1786 and 0.1837 eV/atom
The result matters because it proves the generator can reach the desired region. It is not enough because the probability mass remains sparse, the observed 2/64 yield is far below the 16-32/64 practical ambition, and the result did not reproduce consistently enough for product reliability.
6. Metastability Campaign
The project tested a broad set of hypotheses:
- force and stress auxiliary supervision;
- real relaxation trajectories;
- normalized geometry and cell-shape context;
- scalar conservative energy surfaces;
- bounded mobility and symmetric lattice strain;
- rank-2 angular refinement;
- profile adapters and FiLM;
- routed periodic experts;
- source scheduling and profile balancing;
- MACE relation and force distillation;
- a fresh broad force foundation;
- frozen MACE-OMat plus a learned residual potential.
Several internal signals became useful. V32 reached true-label ROC-AUC 0.70845 and minimum balanced accuracy 0.60048. V30 reached force MAE 0.14260 eV/A and low-force MAE 0.07715 eV/A. The frozen MACE-OMat foundation reached force cosine 0.75450 and MAE 0.05728 eV/A.
The generative controller did not improve accordingly. V32 moved coordinates, lattice, and pair distances in the wrong direction and became non-finite. V30 had negative coordinate and pair contraction. The final residual grid stayed near its frozen foundation baseline and did not justify mobility or native generation stages.
7. Why The Redesigns Plateaued
The redesigns attacked useful intermediate mechanisms, but they did not change the native endpoint distribution strongly enough. The system repeatedly asked a generator trained primarily for denoising and static geometry to behave like a low-hull endpoint model after modest continuation.
Four deeper mismatches remained:
- training data was not consistently filtered and weighted around broad relaxed low-hull endpoints;
- the native sampler was not always the exact reverse process implied by the training corruption;
- composition, coordinates, and lattice were not coupled strongly enough by one endpoint objective;
- internal force, critic, and contraction gates delayed direct external metastability feedback.
This explains how proxy metrics improved while metastable yield stayed flat.
8. Technical Innovations
The project produced reusable engineering and scientific mechanisms:
- replay-grade candidate JSONL with raw payloads and failure reasons;
- strict separation of native, projected, scaffolded, ranked, relaxed, and hull-scored surfaces;
- profile-conditioned element, charge, radius, and benchmark contracts;
- charge-balanced composition search;
- radius-aware periodic geometry checks;
- generated hard-negative replay;
- external noisy-state training paths;
- stability evidence ingestion with provenance-safe joins;
- source-disjoint trajectory and composition-family splits;
- bounded symmetric lattice updates with positive volume;
- target-gated 64-to-128-to-512 external promotion;
- streaming and resumable benchmark operations;
- object-storage-first lineage, manifests, receipts, and W&B references.
9. Honest Product Position
NexaMat is ready to ship as a compact proposal model when the product preserves an explicit screening funnel:
native generation
-> static profile validation
-> encoder/ranker
-> external relaxation
-> hull and novelty checks
-> expert review
It should not present raw candidates as stable materials. Its value is efficient candidate generation and unusually strong static coverage for its parameter count, not SOTA stability proof.
10. Lessons
- Optimize the final scientific outcome from the beginning.
- Keep benchmark surfaces separate so repair cannot masquerade as generation.
- A good classifier or force model is not automatically a good actuator.
- Sparse success proves support, not distribution-level competence.
- Freeze evaluator, sampler, seeds, and lineage before interpreting trends.
- Use small external probes early; do not wait for a large campaign to reveal objective mismatch.
- Preserve negative results. The failed architectures sharply narrowed the remaining problem.
- Stop when repeated redesigns move proxies but not the external objective.
11. Reopening Strategy
The best future experiment is a composition-conditioned low-hull endpoint pilot:
- assemble relaxed primitive structures with reliable hull evidence;
- separate near-hull, metastable, and higher-hull slices;
- train a mathematically matched wrapped-coordinate diffusion or Riemannian flow over coordinates and lattice;
- warm-start useful chemistry embeddings but not the failed controller;
- evaluate fixed compositions first;
- run the same frozen 64-candidate relaxation/hull protocol.
If fixed-composition generation succeeds, composition is the next bottleneck. If teacher denoising succeeds but free sampling fails, the sampler is the bottleneck. If fixed-composition geometry fails, the periodic generator architecture must be replaced.
12. Final Verdict
NexaMat did not reach the reliable high-yield stable-material generator originally targeted, but it is not a failed project. It transformed a nearly unusable decoder into a compact, auditable, multi-profile inorganic candidate generator, demonstrated sparse metastable support, and established the exact scientific boundary that remained. The project is archived and placed on ice indefinitely with a useful shipped model, durable evidence, and a decisive restart contract.