CoolFace
Apppublic

pavement-peer-review/asphalt-va-predictor

sourceHugging Faceupdated 4mo agoView on Hugging Face
0likes
App README

Asphalt Air Void (Va%) Predictor

XGBoost model predicting air void content in compacted asphalt mixes from mix design inputs alone — no destructive testing required. Trained using Leave-One-Domain-Out (LODO) cross-validation across 9 published studies (356 specimens, 13 composite domains).

LODO R² = 0.850 · SET_FINAL features: BC, CA, Gse, F, Gmm

Live demo

[Open in browser — no installation needed] https://huggingface.co/spaces/pavement-peer-review/asphalt-va-predictor

Cite this software

[Authors] et al. (2026). Asphalt Air Void Predictor [Software]. Submitted for peer review.

Run locally

bash
pip install -r requirements.txt
streamlit run app_streamlit.py

Inputs

FieldSymbolUnitTypical range
Bitumen contentBC% by mix weight4.0 – 7.0
Coarse aggregateCA% retained on No.8 (2.36 mm)33 – 81
FillerF% passing No.2001.0 – 10.0
Bulk sp. gravity (aggreg.)Gsb—2.4 – 2.7
Water absorption (aggreg.)Abs%0.0 – 3.5

The app computes Gse and Gmm internally (AASHTO formulas).

Output

  • —Predicted Va (%) — full model (trained on all 13 domains)
  • —Domain status — IN DOMAIN / BORDERLINE / OUT OF DOMAIN (based on P5–P95 of training data)
  • —LODO ensemble plot — 13 predictions, one per leave-one-domain-out model; spread quantifies generalisation uncertainty
  • —Batch mode — upload Excel, download results with colour-coded domain status

Model limitations

Not valid for: porous asphalt, SMA, gap-graded mixes, Va > ~11%, mixes with no filler. Extrapolation outside the training feature space is flagged automatically.