PHBV Reliability Explorer

Study-aware polymer modelling

Predict a trajectory.
Keep the uncertainty visible.

A browser-native, bounded and time-monotonic XGBoost model for literature-derived PHBV degradation-induced mass loss.

16independent source studies
23.55macro-study MAE, percentage points
0.145nested LOSO point R²
91.9mean width of post-selection 90% intervals
RESEARCH USE

Cross-study errors were large and heterogeneous. The released corpus is additive-enriched and inherited source-level preprocessing. Use this tool to explore hypotheses and plan experiments—not for certification, regulatory decisions, safety claims, causal inference, or replacement of laboratory tests.

Describe one observation

All fields mirror the released modelling table. Recorded marginal ranges are shown beneath numeric inputs.

0–100
Ready for a study-aware estimate

Enter a scenario, then calculate. The reliability band will remain visible beside the point estimate.

Trace a fixed-profile trajectory

Only time changes along the curve. Every other formulation and environmental descriptor remains fixed.

Suggested 0–360
Must exceed start
2–200
Generate a trajectory to populate the chart
PredictionPost-selection 90% residual band

Read before use

A model card inside the interface

The page reports the strict research estimate, not training fit. Browser inference keeps the method accessible without hiding its evidential limits.

What the model guarantees

Predictions remain within 0–100% through a bounded-logit transform plus explicit output clipping and are nondecreasing with degradation time for fixed descriptors. Three raw training responses slightly above 100% were clipped for bounded fitting. These are output-shape guarantees, not guarantees of accuracy.

How it was evaluated

The primary analysis used 16 outer leave-one-study-out folds with study-grouped inner tuning. Macro-study MAE was 23.554 percentage points (95% study-bootstrap CI 17.421–30.251). Random-row R² of 0.915 is shown only as an optimistic interpolation benchmark.

Why the interval is so broad

The nominal 90% post-selection grouped residual intervals covered 97.6% overall but averaged 91.9 percentage points wide. One held-out study achieved only 37.5% coverage. The same inner folds informed tuning and residual generation, so the displayed band is not independent conformal calibration or a conditional guarantee.

What applicability notes mean

Range and category checks can flag obvious marginal departures from the recorded evidence. They do not validate a joint scenario, detect every study shift, omitted protocol variable, or new material–environment interaction, and they do not certify a prediction as reliable. Unvalidated combinations are exploratory.

What population supports the model

The released evidence base contains 1,467 rows from 129 curves and 16 studies, with 1,247 rows marked additive-present. It inherited global temperature completion for 210 rows, category coarsening, and zero filling of unreported additive concentrations. It is not a probability sample of PHBV experiments.

Privacy and computation

The model, preprocessing map, and tree traversal run locally in this browser. Inputs are not sent to a prediction API. The static files are hosted by Hugging Face Spaces.