---
title: PHBV Mass-Loss Reliability Explorer
emoji: 🧪
colorFrom: blue
colorTo: green
sdk: static
app_file: index.html
pinned: false
license: mit
short_description: Study-aware PHBV mass-loss prediction
---

# PHBV Mass-Loss Reliability Explorer

This public research-use Space predicts degradation-induced specimen mass/weight loss (%) for PHBV profiles represented by the released modelling corpus. It provides single-time-point and trajectory modes, 0–100% bounds, a positive monotonic constraint on degradation time, marginal range/category warnings, and a conservative post-selection residual band. Inference runs locally in the browser from an exported XGBoost tree ensemble; no prediction server is required.

## Scientific boundary

The output is **gravimetric mass loss**, not ultimate biodegradation, mineralisation, carbon conversion, environmental safety, or compliance with a biodegradability standard. Mass loss can include disintegration, recovery loss, leaching, and biological conversion.

## Validation

- Evidence base: 1,467 observations from 129 curves in 16 source studies.
- Primary research assessment: nested leave-one-study-out validation with study-grouped inner tuning.
- Primary macro-study MAE: 23.554 percentage points (95% study-bootstrap CI 17.421–30.251).
- Point MAE/RMSE/$R^2$: 22.061 / 30.483 / 0.145.
- Random-row XGBoost $R^2$ was 0.915, but this is an optimistic interpolation benchmark because related time points cross the split.
- Nominal 90% post-selection grouped residual intervals covered 97.6% overall, but their mean width was 91.9 percentage points and one held-out study had only 37.5% coverage. The residual-producing inner folds also informed hyperparameter selection; this is not independent conformal calibration.

These results do not support high-stakes or study-independent certification use. The interface reports the strict grouped evidence rather than deployment-training fit.

## Intended use

Use the tool to explore hypotheses, inspect plausible bounded trajectory shapes, identify marginal out-of-range inputs, and prioritise experiments. Marginal checks do not establish joint plausibility; unvalidated combinations are exploratory. Do not use the tool to replace laboratory testing, make regulatory or safety decisions, or infer causal effects of a formulation or environment.

## Model

The model is an XGBoost regressor trained on a bounded-logit target with explicit 0–100% clipping, a positive monotonic constraint on degradation time, and equal total training weight per study. Three raw training responses slightly above 100% were clipped for bounded fitting. Categorical variables are one-hot encoded with unknown levels ignored; numeric missingness is median-imputed inside the fitted preprocessing pipeline. The interactive model is fitted on all 16 studies after the research evaluation was completed. The source table is additive-enriched and inherited global temperature completion for 210 rows, category coarsening, and zero filling of unreported additive amounts.

## Version and licensing scope

- Model export: `phbv-bm-xgb-2026-07-26-v1`
- Release date: 26 July 2026
- The Space's MIT declaration covers the interface code and project-authored packaging. It does not relicense the source article, workbook, CSV, or notebooks in `origin_data/`; source rights remain with their holders and source artifacts are not uploaded to this Space.

## Citation

The source data were compiled by Kotzabasaki et al. (2026), *Polymers* 18, 897: https://doi.org/10.3390/polym18070897. Cite the companion reliability-analysis manuscript and archived code/data release when their final bibliographic records are available.