Truace
Subscribe

Predicting multi-heavy-metal adsorption on bimetallic zeolitic imidazolate frameworks via interpretable machine learning: From latin hypercube design to web implementation

Efficient remediation of heavy metal-contaminated water requires systematic evaluation of adsorbent performance under complex environmental conditions. In this study, an interpretable machine learning approach was applied to predict multi-metal adsorption on bimetallic zeolitic imidazolate frameworks (ZIFs) and to analyze the contributions of material properties, aqueous chemistry, and metal ion characteristics. Three structurally distinct bimetallic ZIF-derived materials were synthesized and evaluated via adsor…

By the Truvace desk2 min read1 sourcePeer-reviewedTRV-2026-1628· cite & verify
Predicting multi-heavy-metal adsorption on bimetallic zeolitic imidazolate frameworks via interpretable machine learning: From latin hypercube design to web implementation
Preparation of phosphorus-doped boron nitride and its adsorption of heavy metals from flue gas by Li, Yanlong; Li, Hongxi; Li, Rundong; Su, Xin; Shen, Shengqiang. CC BY 4.0 · https://creativecommons.org/licenses/by/4.0

In brief

Researchers synthesized three structurally different bimetallic zeolitic imidazolate framework materials and ran adsorption experiments planned with Latin Hypercube Sampling across environmentally relevant conditions. They compared five machine learning models, and a tuned neural network performed best while data leakage was strictly controlled.

The work matters because screening adsorbents for contaminated water usually depends on many interacting variables, and an interpretable model can show which ones to prioritize. The reported accuracy comes from laboratory data on three materials, so how well the approach transfers to other adsorbents or real wastewater is not established in the summary.

Main points

  1. A neural network optimized by grid search reached R2 = 0.98 among five tested models, under strict data leakage control.
  2. SHAP analysis attributed 48.1 % of influence to metal ion properties, 39.9 % to experimental conditions and 12.0 % to adsorbent characteristics.
  3. Pore effects worked mainly through diffusion accessibility and pore density rather than pore size alone.
  4. A web-based prediction platform was built and showed a mean absolute percentage error of 18 % on independent validation data.

The gain

A grid-search-tuned neural network predicted heavy-metal adsorption capacity on three bimetallic ZIF materials with high accuracy and exposed which factors drive uptake, supporting more targeted adsorbent design.

The rundown

The experimental design used Latin Hypercube Sampling to cover a wide range of environmentally relevant conditions with a limited number of adsorption runs, and the data came from three structurally distinct bimetallic ZIF-derived materials.

Interpretability analysis with Shapley Additive Explanations ranked the parameter groups, finding that the characteristics of the metal ions mattered more for capacity than either the test conditions or the adsorbent itself.

The authors also turned the model into a web-based platform, and its accuracy on independent validation data was lower than on the main model evaluation, with an error of 18 %. They present the framework as a route to developing targeted water remediation technologies.

Sources

  1. Peer-reviewedJournal of Environmental Sciences2026-03-07

The debate