Open models
Synthetic-data modelOpen · synthetic

Rare Earth Recovery Predictor

Predicts element recovery across a solvent-extraction separation train from stage conditions and feed assay. Reference implementation trained on synthetic separation data.

Rare earthsRare Earth Separation IntelligenceRecovery predictionv0.3.1Apache-2.0
8.4k 214 46

This model is a reference implementation. Industrial use requires validation on plant-specific data. Customer-specific fine-tuning remains private unless explicitly approved for publication.

Input schema

FieldTypeDescription
stage_phfloat[]pH by extraction / scrub / strip stage
org_aq_ratiofloatOrganic-to-aqueous phase ratio
feed_assayfloat[] · wt%Feed composition by element
feed_flowfloat · m³/hFeed volumetric flow

Output schema

FieldTypeDescription
recoveryfloat · %Predicted element recovery
recovery_cifloat[]Prediction interval

Training data & evaluation

Training data type

Synthetic separation dataset

Evaluation metrics

MAE 1.8 % 0.91

Example usage

from matereal import load_model
m = load_model('matereal/ree-recovery-predictor')
m.predict({'stage_ph': [...], 'org_aq_ratio': 1.2, ...})

Limitations

  • Trained on synthetic data with idealized stage behaviour.
  • Does not model reagent degradation over long campaigns.

Discussion

  • s.

    s.iyer

    ML researcher · 3d

    Validation note

    Reproduced the reported metric on the public split with a fixed seed. Numbers line up within noise, so it's a good baseline to build on.

    18
  • l.

    l.zhang

    Process engineer · 1w

    Improvement

    Adding an uncertainty head would make this far more useful for prioritization. Happy to open a PR against the reference repo.

    11
Reference, not production-proven

Validate on your data, keep the result private

Reference models are a starting point. We can fine-tune privately on plant-specific data, with customer-owned IP and no public disclosure.