Scholarlink

<p>Ultrafiltration (UF) is an effective method for clarifying polyphenol-rich juices without the need for chemical agents, but its practical use is often limited by fouling and operating conditions. In this study, <i>Aronia melanocarpa</i> juice was clarified using polyethersulfone (PES) membranes with different pore sizes under varying flow rates and pressures. Experimental data were used to train two deep learning regression models: a single-output model predicting permeate flux, and a multi-output model simultaneously predicting flux, fouling index (FI), and cleaning efficiency (CE). The models were trained on 36 independent ultrafiltration runs (1,296 time-resolved samples across 12 operating configurations) using a run-level, flux-stratified train/validation/test split. On held-out test data, the single-output model predicted permeate flux with a mean absolute error (MAE) of 0.0710&#xa0;kg·m⁻<sup>2</sup>·h⁻<sup>1</sup> (R<sup>2</sup> = 0.997), while the multi-output model achieved R<sup>2</sup> &gt; 0.96 across flux, fouling index, and cleaning efficiency within the studied operating range. SHAP (SHapley Additive exPlanations) analysis revealed that operation time and flow rate had a greater influence on performance than membrane type. Within the studied range, higher feed flow rates combined with moderate pressure provided a favorable balance between flux, fouling, and CE. Moreover, the 50&#xa0;kDa PESH membrane outperformed the 150&#xa0;kDa PES membrane across the evaluated conditions, consistent with the established role of surface hydrophilicity in mitigating phenolic-driven fouling, an effect which can outweigh nominal pore size for polyphenol-rich juice feeds. These findings demonstrate that data-driven modeling reduces the need for extensive experimental trials and provides a balanced assessment of permeate flux, fouling, and cleaning indicators, supporting rational operating-point selection for UF clarification of polyphenol-rich juices. Experimental validation at 3&#xa0;bar transmembrane pressure (TMP), beyond the 1–2&#xa0;bar training range, confirmed the model predictions at the identified optimum. The ensemble standard deviation of predicted flux increased with distance from the training domain, indicating reduced prediction reliability in the extrapolation region.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep learning-based modeling and optimization of aronia juice ultrafiltration: a multi-output interpretable approach

  • 期刊论文

  • 发表日期：

    2026年7月20日

查看全文
  • Ayse Seda Apaydin,
  • Emir Ozturk,
  • Pelin Onsekizoglu Bagci

摘要

Ultrafiltration (UF) is an effective method for clarifying polyphenol-rich juices without the need for chemical agents, but its practical use is often limited by fouling and operating conditions. In this study, Aronia melanocarpa juice was clarified using polyethersulfone (PES) membranes with different pore sizes under varying flow rates and pressures. Experimental data were used to train two deep learning regression models: a single-output model predicting permeate flux, and a multi-output model simultaneously predicting flux, fouling index (FI), and cleaning efficiency (CE). The models were trained on 36 independent ultrafiltration runs (1,296 time-resolved samples across 12 operating configurations) using a run-level, flux-stratified train/validation/test split. On held-out test data, the single-output model predicted permeate flux with a mean absolute error (MAE) of 0.0710 kg·m⁻2·h⁻1 (R2 = 0.997), while the multi-output model achieved R2 > 0.96 across flux, fouling index, and cleaning efficiency within the studied operating range. SHAP (SHapley Additive exPlanations) analysis revealed that operation time and flow rate had a greater influence on performance than membrane type. Within the studied range, higher feed flow rates combined with moderate pressure provided a favorable balance between flux, fouling, and CE. Moreover, the 50 kDa PESH membrane outperformed the 150 kDa PES membrane across the evaluated conditions, consistent with the established role of surface hydrophilicity in mitigating phenolic-driven fouling, an effect which can outweigh nominal pore size for polyphenol-rich juice feeds. These findings demonstrate that data-driven modeling reduces the need for extensive experimental trials and provides a balanced assessment of permeate flux, fouling, and cleaning indicators, supporting rational operating-point selection for UF clarification of polyphenol-rich juices. Experimental validation at 3 bar transmembrane pressure (TMP), beyond the 1–2 bar training range, confirmed the model predictions at the identified optimum. The ensemble standard deviation of predicted flux increased with distance from the training domain, indicating reduced prediction reliability in the extrapolation region.