<p>The burgeoning global olive oil industry generates vast quantities of olive oil extraction waste, posing significant environmental challenges while representing an underexploited source of valuable bioactive phenolic compounds. Traditional extraction and control methods are often empirical and fail to manage the complex, non-linear dynamics of the process, leading to suboptimal economic and operational outcomes. This study introduces a novel, two-stage AI-driven framework for the optimization and control of polyphenol extraction. The first stage addresses the critical challenge of data scarcity by developing a high-fidelity data-driven surrogate model of the process, used as a simulation-based process representation for control design. We demonstrate that for a severely limited initial dataset, a simple Gaussian noise-based data augmentation technique is significantly more effective than complex generative models (GANs, VAEs), enabling the training of a LightGBM-based process surrogate model with high predictive accuracy within the studied parameter space (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(R^2 > 0.99\)</EquationSource> </InlineEquation>). In the second stage, this digital twin is exploited within a high-fidelity simulation environment to conduct a rigorous comparative analysis of two intelligent control strategies: a classic Model Predictive Control and a modern Bayesian Optimization (BO) approach. The objective was to track dynamic (triangular) and static (step) setpoints for TPC while minimizing a comprehensive economic cost function, under both ideal and process-delay conditions. Although MPC performance can be improved through appropriate tuning, the sensitivity analysis performed in this study shows that Bayesian Optimization maintains superior tracking performance and lower economic cost over the range of MPC configurations investigated. The BO strategy’s aggressive policy, which actively modulates process parameters to prioritize tracking fidelity, proved significantly more economically efficient than the MPC’s overly conservative approach. This work yields a critical insight: for complex processes where performance deviation incurs a high economic penalty, an aggressive control strategy that meets production targets is globally more cost-effective, even at the expense of higher immediate resource consumption. This research provides a robust methodology for creating digital twins from sparse data and offers a clear verdict on control strategy selection, paving the way for the intelligent, economically optimized automation of agro-industrial waste valorization.</p>

From sparse samples to optimal control: A Data-driven surrogate modeling framework for polyphenol extraction using gaussian-augmented data and bayesian optimization

期刊论文
发表日期：

2026年7月6日

查看全文

Imen Aidi,
Manel Taktak,
Mongi Besbes

摘要

The burgeoning global olive oil industry generates vast quantities of olive oil extraction waste, posing significant environmental challenges while representing an underexploited source of valuable bioactive phenolic compounds. Traditional extraction and control methods are often empirical and fail to manage the complex, non-linear dynamics of the process, leading to suboptimal economic and operational outcomes. This study introduces a novel, two-stage AI-driven framework for the optimization and control of polyphenol extraction. The first stage addresses the critical challenge of data scarcity by developing a high-fidelity data-driven surrogate model of the process, used as a simulation-based process representation for control design. We demonstrate that for a severely limited initial dataset, a simple Gaussian noise-based data augmentation technique is significantly more effective than complex generative models (GANs, VAEs), enabling the training of a LightGBM-based process surrogate model with high predictive accuracy within the studied parameter space ( \(R^2 > 0.99\) ). In the second stage, this digital twin is exploited within a high-fidelity simulation environment to conduct a rigorous comparative analysis of two intelligent control strategies: a classic Model Predictive Control and a modern Bayesian Optimization (BO) approach. The objective was to track dynamic (triangular) and static (step) setpoints for TPC while minimizing a comprehensive economic cost function, under both ideal and process-delay conditions. Although MPC performance can be improved through appropriate tuning, the sensitivity analysis performed in this study shows that Bayesian Optimization maintains superior tracking performance and lower economic cost over the range of MPC configurations investigated. The BO strategy’s aggressive policy, which actively modulates process parameters to prioritize tracking fidelity, proved significantly more economically efficient than the MPC’s overly conservative approach. This work yields a critical insight: for complex processes where performance deviation incurs a high economic penalty, an aggressive control strategy that meets production targets is globally more cost-effective, even at the expense of higher immediate resource consumption. This research provides a robust methodology for creating digital twins from sparse data and offers a clear verdict on control strategy selection, paving the way for the intelligent, economically optimized automation of agro-industrial waste valorization.