What was the problem?
Polypropylene (PP) is everywhere—from appliance housings to automotive trim—but tuning its mechanical performance with eco-friendly reinforcements is still a balancing act. We explored a hybrid system of long flax fiber bundles (LFF), basalt fibers (BF), and rice husk powder (RHP). Each ingredient pulls properties in different directions: BF boosts stiffness, flax can raise strength at modest loadings, and RHP helps processability and sustainability but can soften the matrix if overused. Running a full factorial campaign would be slow and costly, so we asked: could machine learning help us map the design space with only a compact Box–Behnken plan?

Why ensembles, not a single model?
Small materials datasets are noisy and rarely linear. Single learners often overfit, especially when variables interact. We therefore stacked two strong but complementary base learners and tuned capacity carefully with cross-validation. Stacking reduces variance by letting models “vote,” while still capturing non-linear trends. We also kept a strict separation of training, validation, and a 20% hold-out set so our reported skill reflects true generalization inside the tested route.

How did we keep it honest?
Before modeling, we treated the data like we would treat a specimen: inspect it. Histograms, KDEs and boxplots ensured every factor level was represented and flagged one clear modulus outlier (>10 GPa). Pair plots helped us see raw tendencies (%BF and %RHP versus tensile strength/modulus) without hiding behind a model. During training, we used k-fold cross-validation and capped complexity (depth, estimators, regularization) to avoid optimistic results.

Why SHAP (and friends)?
A prediction is only useful if engineers can act on it. SHAP values provide per-feature, per-sample attributions for tree models, telling us how much each variable nudged a prediction up or down. We paired SHAP with permutation importance (model-agnostic), partial dependence (PDP) and accumulated local effects (ALE) to see both global and local behavior. The takeaway was intuitive and actionable: BF dominates modulus gains, but strength benefits from a balanced trio—too much of any one component quickly flattens returns. These insights align with micromechanics expectations about stiffness transfer and embrittlement at high rigid-filler content.

What did we learn for design?
The ensemble reproduced measured trends on cross-validation and on the unseen hold-out set. From the SHAP/PDP landscape we highlighted composition windows where tensile strength improves ~2× over neat PP while modulus climbs strongly: moderate BF with supportive flax plies and controlled RHP. In practice, this narrows down trial-and-error—teams can start within these windows, then fine-tune around processing realities such as fiber length retention or porosity.

Limits you should care about
Our goal wasn’t a universal model for “any” PP composite. We fixed processing (extrusion → hot press → injection molding) to isolate composition effects. That means the model explains variance due to %BF, %RHP, and flax ply count within this route; jumping to a very different line or tool will change the microstructure, and models should be retrained or updated. Also, small materials datasets make uncertainty quantification valuable; future work could add conformal prediction intervals so designers see both a forecast and its confidence band.

Why this matters beyond our system
The workflow—compact DOE, capacity-controlled ensembles, and transparent explanations—generalizes to many data-limited materials problems: bio-fillers in thermoplastics, multi-phase binders, or even cementitious systems. The key is to pair domain priors with honest diagnostics and keep explanations close to mechanisms engineers trust.

A note on sustainability
Flax and rice-husk powder are renewable or waste-derived. Being able to forecast properties with fewer physical trials lowers energy, labor, and scrap. In other words, better data practices are also greener practices.

Where to read the paper
The research article is open access in Discover Materials: DOI 10.1007/s43939-025-00406-4. A shareable link is available via Springer Nature’s initiative: rdcu.be/eQUMB.