---
license: apache-2.0
library_name: pytorch
tags:
  - materials-science
  - crystal-structures
  - solid-state-batteries
  - representation-learning
  - screening
model-index:
  - name: SSB Screening Model (RTX6000x2)
    results:
      - task:
          type: text-classification
          name: Screening Proxy (3-class)
        metrics:
          - type: accuracy
            value: 0.8118937
          - type: f1
            value: 0.8060277
          - type: precision
            value: 0.7671543
          - type: recall
            value: 0.8694215
          - type: val_loss
            value: 0.2856999
---

# SSB Screening Model (RTX6000x2)

## Model Summary
This model is a lightweight MLP classifier trained on NPZ-encoded inorganic crystal structure features for solid-state battery (SSB) screening proxies. It is intended to prioritize candidate structures, not to replace DFT or experimental validation.

- **Architecture**: MLP (input_dim=144, hidden_dims=[512, 256, 128], dropout variable by sweep)
- **Output**: 3-class classification proxy for screening tasks
- **Training Regime**: supervised training on curated NPZ dataset with class-weighted loss
- **Best checkpoint**: `checkpoint_epoch45.pt` (lowest observed val_loss in logs)

## Intended Use
- **Primary**: ranking/prioritization of SSB electrolyte candidates
- **Not intended**: absolute property prediction or experimental ground truth replacement

## Training Data
- **Dataset**: `ssb_npz_v1` (curated NPZ features)
- **Split**: 80/10/10 (train/val/test)
- **Features**: composition + lattice + derived scalar statistics (144-dim)

## Evaluation
Metrics from the latest run summary:
- **Val loss**: 0.2857
- **Val accuracy**: 0.8119
- **Holdout accuracy**: 0.8096
- **F1**: 0.8060
- **Precision**: 0.7672
- **Recall**: 0.8694

## Limitations
- The model is a proxy classifier; it does not predict ground-truth physical properties.
- Performance is tied to the training distribution of `ssb_npz_v1`.
- Chemical regimes underrepresented in the training set may be poorly ranked.

## Training Configuration (abridged)
- Optimizer: AdamW
- LR: sweep (best around ~3e-4)
- Weight decay: sweep (0.005–0.02)
- Scheduler: cosine
- Batch size: sweep (128–512)
- Epochs: sweep (20–60)
- Gradient accumulation: sweep (1–4)

## Citation
If you use this model, please cite the dataset and training pipeline from the Nexa_compute repository.