Keywords: Carbon capture, utilization, and storage, Process control
Abstract: This presentation will explore how decades of research in systems and control engineering can contribute to ongoing decarbonization efforts and the broader energy transition. This multifaceted transition, expected to take several decades and cost trillions of dollars, aims to shift fundamentally towards renewable energy sources. Carbon capture, utilization, and storage (CCUS) technologies are considered essential bridge solutions to reduce carbon emissions until this transition is achieved. A significant challenge involves managing the spatial and temporal dependencies associated with renewable energy generation, which requires extensive, multi-scale storage solutions. Moreover, there are considerable uncertainties regarding future technological advancements and socioeconomic conditions. My talk will concentrate on two key approaches from process systems engineering and control research: superstructure model-based optimization and stochastic optimal control. I will discuss how we assess various CCUS technological strategies, focusing on their economic viability and potential for meaningful CO2 reduction. This evaluation involves techno-economic assessment (TEA) and life cycle analysis (LCA), where material and energy flows for different CCUS options are modeled through a superstructural framework. I will introduce a tool called ArKaTAC3, specifically developed to identify and evaluate the most effective CCUS pathways. Further, I will examine how the deployment planning of CCUS and green hydrogen production, amid the variabilities of renewable energy sources and prevailing uncertainties in technological and socioeconomic aspects, leads to a complex, multi-scale, multi-stage stochastic decision-making problem. This scenario closely mirrors challenges found in stochastic optimal control. I will demonstrate how this intricate issue can be tackled using a combination of traditional optimization techniques and newer machine learning strategies, such as reinforcement learning, offering a robust framework for addressing these critical environmental challenges.

Keywords: Model predictive control, Artificial intelligence and machine learning, Process control
Abstract: Approximate model predictive control based on imitation learning methods enables real-time implementation of optimal constrained control even for large-scale systems under uncertainty. Training the underlying neural networks often requires large datasets of, which can be challenging for a single process operator to gather. A federated learning scheme, where multiple operators combine smaller datasets, could alleviate this issue. Yet, off-the-shelf federated learning may conflict with privacy requirements of the participants. In this paper, we present a collaborative federated learning scheme for robust approximate model predictive control. To protect data privacy with respect to the central computing server, we integrate homomorphic encryption, allowing for encrypted learning.

Keywords: Systems biology, synthetic biology, metabolic flux modeling, Modeling and identification, Dynamic modelling and simulation for control and operation
Abstract: Synthetic biology is a recent area of biological engineering, whose aim is to provide cells with novel functionalities. A number of important results regarding the development of control circuits in synthetic biology have been achieved during the last decade. A differential geometry approach can be used for the analysis of said systems, which are often nonlinear. Here we demonstrate the application of such tools to analyse the structural identifiability, observability, accessibility, and controllability of several biomolecular systems. We focus on a set of synthetic circuits of current interest, which can perform several tasks, both in open loop and closed loop settings. We analyse their properties with our own methods and tools; further, we describe a new open-source implementation of the techniques.

Keywords: Modeling and identification, Model predictive control, Process optimization
Abstract: This paper considers the problem of designing predictive control laws for nonlinear auto-regressive exogenous (NARX) models based on measured input-output data without explicitly identifying the model parameters. We explore the case when outputs are corrupted by additive measurement noise. An upper bound on the optimal value function for the robust case is derived, and the mismatch between the predicted and actual output is also theoretically studied. The recursive feasibility and practical stability of the robust data-driven predictive control (DDPC) scheme are guaranteed. The simulation results finally quantify the effectiveness of the proposed method with the experimental data gathered from a powder compaction process performed on a rotary tablet press.

Keywords: Model predictive control, Process optimization, Process control
Abstract: Model predictive control (MPC) is an advanced control strategy that can deal with general nonlinear systems and constraints but relies on accurate predictions given by a dynamic model. To satisfy constraints and improve performance despite imperfect models, robust MPC methods can be formulated. Multi-stage MPC is a robust MPC method based on the formulation of scenario trees. The resulting optimization problems can be large, as the number of scenarios considered in the tree results from the combinations of all possible uncertainties. For systems with many uncertainties, as it is the case in bioprocesses, the optimization problems become rapidly intractable. To solve this issue, heuristics are typically used to select the most relevant uncertain parameters and their range of uncertainty. In this paper, we propose a two-step approach to obtain a systematic design of multi-stage MPC controllers: First, the key uncertain parameters are extracted based on the parametric sensitivities. Second, Bayesian optimization is employed for tuning of the range of uncertainties. The approach is applied to a bioreactor simulation study. The proposed approach can avoid constraint violations that are otherwise obtained by standard MPC while being less conservative than a manually-tuned robust controller.

Keywords: Process control, Model predictive control, Artificial intelligence and machine learning
Abstract: In this paper, we propose a convex data-based economic predictive control method within the framework of data-enabled predictive control (DeePC). Specifically, we use a neural network to transform the system output into a new state space, where the nonlinear economic cost function of the underlying nonlinear system is approximated using a quadratic function expressed by the transformed output in the new state space. Both the neural network parameters and the coefficients of the quadratic function are learned from open-loop data of the system. Additionally, we reconstruct constrained output variables from the transformed output through learning an output reconstruction matrix; this way, the proposed economic DeePC can handle output constraints explicitly. The performance of the proposed method is evaluated via a case study in a simulated chemical process.

Keywords: Process control, Artificial intelligence and machine learning
Abstract: A zonotope-based big data-driven predictive control (BDPC) approach is developed to partition the nonlinear process behaviour (represented by an input-output trajectory set) into multiple linear sub-behaviours using a two-step hierarchical clustering: Euclidean distance-based clustering and linear subspace distance-based clustering. By approximating every linear sub-behaviour as a zonotope, a data-driven interpolation is developed based on the convex combination of zonotopes. During online control, a BDPC controller is designed by determining an interpolated zonotope where its center trajectory is closest to the online trajectory and computing the control action subject to an optimisation problem. The proposed BDPC approach is illustrated using a case study on controlling an aluminium smelting process.

Keywords: Model predictive control, Process control, Modeling and identification
Abstract: Exact discretization of continuous linear time-delay systems with quadratic objective functions,under piecewise constant manipulated variables, is used to design and implement a novel linear model predictive controller, termed the continuous-time linear-quadratic model predictive controller (CT-LMPC). The key novelty in the paper is the exact numerical discretization of CT time-delay linear-quadratic systems. The control model of the CT-LMPC is parameterized using transfer functions with delays. We formulate linear-quadratic optimal control problems (LQOCPs) with time delays for the CT-LMPC and derive their discretization under the assumption of piecewise constant inputs. Time-delay systems are ubiquitous in process industries, such as the cement industry, where conveyor belts introduce delays, making continuous-time modeling advantageous. We illustrate the CT-LMPC with both SISO and MIMO examples inspired by the cement industry. The results demonstrate that, for fixed parameters, the CT-LMPC outperforms the conventional discrete-time LMPC as the sampling time increases.

Keywords: Artificial intelligence and machine learning, Model predictive control, Dynamic modelling and simulation for control and operation
Abstract: We present a method for end-to-end learning of Koopman surrogate models for optimal performance in a specific control task. In contrast to previous contributions that employ standard reinforcement learning (RL) algorithms, we use a training algorithm that exploits the potential differentiability of environments based on mechanistic simulation models to aid the policy optimization. We evaluate the performance of our method by comparing it to that of other training algorithms on an existing economic nonlinear model predictive control (eNMPC) case study of a continuous stirred-tank reactor (CSTR) model. Compared to the benchmark methods, our method produces similar economic performance while eliminating constraint violations. Thus, for this case study, our method outperforms the others and offers a promising path toward more performant controllers that employ dynamic surrogate models.

Keywords: Artificial intelligence and machine learning, Batch process modeling and control, Bio-applications
Abstract: Hybrid modelling utilizes advantageous aspects of both mechanistic (white box) and data-driven (black box) modelling. Combining the physical interpretability of kinetic modelling with the power of a data-driven Artificial Neural Network (ANN) yields a hybrid (grey box) model with superior accuracy when compared to a traditional mechanistic model, while requiring less data than a purely data-driven model. This study aims to construct a hybrid model for the predictive modelling of a high-cell-density microalgal fermentation process for lutein production under uncertainty. In addition, transfer learning is combined with the hybrid model to simulate new fed-batches utilizing alternative substrates operated under a different reactor scale. By comparing with experimental data, the hybrid transfer model was found to be able to simulate the new fed-batch processes that achieve heightened cell densities and higher product quantities. Overall, this work presents a novel digital twin construction strategy that can be easily adapted to general bioprocesses for model predictive control and process optimization under uncertainty.

Keywords: Artificial intelligence and machine learning, Waste water treatment processes, Modeling and identification
Abstract: In the pursuit of a resource-efficient economy, purple non-sulfur bacteria (PNSB) represent a promising solution due to their capacity to convert waste from various sources into valuable products, including biomass. However, scaling up PNSB technology remains challenging, and developing reliable dynamic models for monitoring and control is essential to facilitate this transition. Despite recent efforts dedicated to PNSB modeling, existing gaps in process understanding and difficulties in data collection still limit their development. In this regard, physics-informed neural networks (PINNs) emerge as a natural candidate, considering their ability to integrate partial physical information. This paper presents a PINN-based model developed by combining an existing first-principles model---whose performance declines under new conditions---with additional data representing these conditions. To assess the performance of the PINN-PNSB model, we compare it with other modeling alternatives, including the updated parametric model obtained by classical parameter identification and a pure artificial neural network (ANN). A PINN-derived model, obtained by updating the physical model parameters during PINN training, is also evaluated. Results using training and test data demonstrate the superior performance of the PINN-based model for PNSB applications.

Keywords: Process control, Artificial intelligence and machine learning, Bio-applications
Abstract: Efficient multiple setpoint tracking can enable advanced biotechnological applications, such as maintaining desired population levels in co-cultures for optimal metabolic division of labor. In this study, we employ reinforcement learning as a control method for population setpoint tracking in co-cultures, focusing on policy-gradient techniques where the control policy is parameterized by neural networks. However, achieving accurate tracking across multiple setpoints is a significant challenge in reinforcement learning, as the agent must effectively balance the contributions of various setpoints to maximize the expected system performance. Traditional return functions, such as those based on a quadratic cost, often yield suboptimal performance due to their inability to efficiently guide the agent toward the simultaneous satisfaction of all setpoints. To overcome this, we propose a novel return function that rewards the simultaneous satisfaction of multiple setpoints and diminishes overall reward gains otherwise, accounting for both stage and terminal system performance. This return function includes parameters to fine-tune the desired smoothness and steepness of the learning process. We demonstrate our approach considering an Escherichia coli co-culture in a chemostat with optogenetic control over amino acid synthesis pathways, leveraging auxotrophies to modulate growth.

Keywords: Batch process modeling and control, Artificial intelligence and machine learning, Process control
Abstract: The development of optimal control strategies for bioprocesses has become essential in pharmaceutical and industrial applications due to the growing demand for sustainable bioproducts. Traditional model-based optimization methods rely heavily on the accuracy of the system model, requiring frequent recalibration and experimentation to sustain performance. In contrast, advanced model-free control techniques, such as Reinforcement Learning (RL), are widely researched. However, training RL controllers online is constrained by the need for extensive online interactions with the biosystem environment, which can be costly and present safety risks. To overcome these limitations, we propose leveraging offline Reinforcement Learning algorithms to train control agents using historical data collected from previous bioprocess operations. These agents can subsequently be fine-tuned, improving current control strategies by utilizing past data without extensive real-time interactions with the system. The effectiveness of this approach is validated in an in-silico semi-batch bioprocess case study, outperforming alternative machine learning techniques that leverage historical data, including behavioral cloning.

Keywords: Modeling and identification, Dynamic modelling and simulation for control and operation, Bio-applications
Abstract: Fractionation of lignocellulosic biomass is a crucial step to provide cellulose, lignin, and hemicellulose for further processing. This paper is concerned with modelling biomass fractionation using the ionoSolv process, which employs low-cost ionic liquid water mixtures, with a special focus on describing the effect of acid:base ratio of the mixture on process performance. We build on an existing semi-mechanistic modelling framework describing the solvent extraction of three main biopolymers from woody biomass for varying fractionation temperature, time, and solids loading. Since the effect of acidity is poorly understood from a mechanistic standpoint, we use sparse regression with lasso regularisation to incorporate it in the semi-mechanistic model. We investigate both polynomial and exponential functional forms and find that the latter yields more physically-consistent results. This enabled us to recalibrate the parameters of the combined semi-mechanistic and sparse data-driven models simultaneously to accurately predict the effect of varying acid:base ratio. This hybrid modelling framework opens new opportunities for further analysis and optimisation of ionic liquid-based biomass fractionation processes.

Keywords: Dynamic modelling and simulation for control and operation, Bio-applications, Artificial intelligence and machine learning
Abstract: General anesthesia, typically induced using a combination of hypnotic (Propofol) and analgesic (Remifentanil) drugs, is crucial for the success of surgical procedures, but it can cause dangerous cardiovascular side effects. In this context, models and simulations offer new opportunities to address the intrinsic complexity of the process, accelerating advances and innovation in the technology of anesthesia. This study aims to improve the modeling of hemodynamic effects under general anesthesia by expanding the applicability of a recent mechanistic model in combination with data-driven modules. In particular, we use a dataset related to plastic surgery for both model calibration and testing, preserving the physical interpretability of the mechanistic model while integrating it with data-driven components to enhance its predictive capabilities. The results demonstrate a significant improvement in the model ability to simulate hemodynamic variables under surgical conditions, offering potential applications for anesthesia monitoring and control systems design that consider the patient's cardiovascular safety. This enhanced hybrid model provides a more accurate representation of the complex interactions between anesthetic drugs and cardiovascular dynamics in real surgical settings.

Keywords: Process control, Plantwide control, Dynamic modelling and simulation for control and operation
Abstract: Ratio control and bidirectional inventory control are simple and powerful data-based strategies for feedforward control and coordination, respectively. By “data-based” it is meant that no explicit process models is needed, which simplifies implementation. The paper demonstrates the power of these simple architectures when applied to distillation columns.

Keywords: Model predictive control, Plantwide control, Process control
Abstract: Electric submersible pump (ESP) systems are essential in the oil industry. These systems allow operation with high flow rates and efficiency, even in mature and deep wells. This paper compares the practical implementation of Model Predictive Controllers (MPC) in an ESP system in the Artificial Lift Laboratory at UFBA. The first controller is the traditional MPC, and the second is a target MPC with targets at the input. The zone controller is a more advantageous option for the scenarios tested since tuning is more straightforward, has an easy operating point for the operator to understand, and operates naturally in the maximum production region.

Keywords: Process control, Modeling and identification
Abstract: Abstract: This paper investigates the application of two distinct control structures conventional single-layer control and cascade control in a Core Annular Flow (CAF) system simulated through Computational Fluid Dynamics (CFD). Both control strategies were tested and carried out open-loop tests to tune the controllers following the SIMC rules. Results demonstrate that both structures, one I controller for the oil fraction and one cascade controller PI-I for the velocity ratio and the oil fraction, successfully controlled the system, each exhibiting unique behaviors and performance characteristics. The analysis highlights the strengths and limitations of each approach, where the single-layer structure with an I controller was faster to reach the setpoint and was efficient to reject disturbances.

Keywords: Bio-applications, Modeling and identification, Industrial biotechnology
Abstract: The fast pace of urbanization, population growth, and fossil fuel dependency have brought environmental challenges such as global warming and water contamination, pressing for solutions in greenhouse gas reduction and wastewater treatment. Microalgae cultivation offers promising results by assimilating CO₂ and purifying wastewater. This study focuses on the identification of pH models in microalgae raceway reactors, essential for accurate control and optimization of growth conditions. A novel multisine-based persistent excitation approach combined with a range controller is proposed to enhance data quality and coverage across varying operating points, without violating output constraints. This method demonstrates improved operational stability and enriched dataset acquisition for model identification. Experimental results, conducted at the University of Almeria and IFAPA research center, confirm the method’s effectiveness in generating reliable excitation data, supporting accurate pH model identification for industrial-scale applications.

Keywords: Model predictive control, Process control, Dynamic modelling and simulation for control and operation
Abstract: This paper presents a nonlinear model predictive control (NMPC) strategy using Gaussian Processes (GPs) to control a froth flotation process under partial observability. The GP state-space model predicts future states for both observable and latent variables, using available data while incorporating the probability distribution of these predictions into an optimization problem. This improves robustness against measurement noise and process disturbances and evaluates the impact of feed particle size, a typical process disturbance. We assessed the framework's ability to maintain optimal process performance across varying operating conditions. The results demonstrate that the proposed GP-MPC framework improves process efficiency, even with frequent changes in particle size and measurement noise, confirming its potential for online control of partially observable systems.

Keywords: Sensors and soft sensors, Process and performance monitoring, Renewable energy system
Abstract: Green hydrogen, produced via water electrolysis using renewable energy, is seen as a cornerstone of the energy transition. Coupling of renewable power supplies to water electrolysis processes is, however, challenging, as explosive gas mixtures (hydrogen in oxygen) might form at low loads. This has prompted research into gas purity control of such systems. While these attempts have shown to be successful in theoretical and practical studies, they are currently limited in that they only consider the gas purity at locations where composition measurements are available. As these locations are generally positioned downstream of the disturbance origin, this incurs considerable delays and can lead to undetected critical conditions. In this work, we propose the use of an Extended Kalman Filter (EKF) in combination with a simple process model to estimate and control the gas composition at locations where measurements are not available. The model uses noise-driven states for the gas impurity and is hence agnostic towards any mechanistic disturbance model. We show in simulations that this simple approach performs well under various disturbance types and can reduce the time spent in potentially hazardous conditions by up to one order of magnitude.

Keywords: Fault detection, diagnosis, supervision, and safety, Dynamic modelling and simulation for control and operation, Process and performance monitoring
Abstract: Fault detection has witnessed a rapid development and many approaches have been proposed in recent decades. However, many of them are based on heuristic solutions that do not check the applicability and optimality of the resulting fault-detection systems. This paper starts from a hypothesis test that subsumes all fault-detection problems, based on which a unified optimisation problem is formulated. The resulting optimal solution defines the deemed-normal region of the system. It is proven that dynamic information shrinks the deemed-normal region and improves detection performance in Gaussian LTI systems. The theoretical results are verified on a simulated three-tank system. Compared with the static method, the fault-detection rate (FDR) of the dynamic method based on a Kalman filter increases from 96% and 56% to 99% and 100% for two faults, respectively, while the false-alarm rate (FAR) decreases from 7.38% and 0.88% to 0.75% and 0.63%. This paper provides a theoretical foundation for understanding fault detection for Gaussian LTI systems, and avoids using any heuristic proposals and solutions for the problem of fault detection. The rigorous justification of the well-known fact that incorporating dynamic information improves fault-detection performance implies a roadmap towards advanced methods for more complex cases. In addition, the analysis including the idea of a deemed-normal region has the potential to be extended to a physically meaningful framework for performance assessment of fault diagnosis and fault-tolerant control systems.

Keywords: Fault detection, diagnosis, supervision, and safety, Process and performance monitoring, Artificial intelligence and machine learning
Abstract: Ensuring safety and efficiency in industrial systems requires effective fault detection and diagnosis, which becomes increasingly challenging in high-dimensional and complex environments. Traditional multivariate statistical process monitoring (SPM) methods, such as those based on PCA and PLS, often fall short in their ability to diagnose localized faults due to their lack of causal modeling. This paper introduces a Causal Network-based Decentralized Multivariate Statistical Process Control (CNd-MSPC) framework, which employs causal networks and community detection—specifically the Leiden algorithm—to segment large systems into functional communities and perform distributed monitoring. This structural partitioning preserves essential causal and topological information, enhancing the sensitivity of fault detection by allowing focused analysis of specific sub-networks. Through extensive testing with a graph-based data simulator, we demonstrate that CNd-MSPC consistently outperforms centralized methods across various network sizes, achieving higher fault detection sensitivity for both process perturbations and sensor biases, especially in large networks. The decentralized approach retains high sensitivity, even when data from several communities are missing due to process disruptions. Additionally, the diagnostic results were conclusive and, most of the time, unambiguous.

Keywords: Process and performance monitoring, Modeling and identification, Sensors and soft sensors
Abstract: In most industrial settings, soft sensors are used to measure variables that are challenging to estimate. However, due to various factors, the estimated performance may vary over time, causing it to fail in estimating key variables quickly and accurately enough, which could result in financial losses and security risks. The variation in the predictive performance of a soft sensor is referred to as the performance drift of the soft sensor. These changes occur due to the differences between the current characteristics of the process or plant and the soft sensor. This discrepancy is also defined as plant-model mismatch (PMM). Therefore, once the soft sensor is designed, a way to recover the performance change is required. This paper proposes a method that reduces the impact of PMM without updating the soft sensor itself. This method reconstructs the closed-loop controller based on the performance change index (PCI) for the soft sensor. Then, the online identification of the fault model is studied. Finally, the modification rules for the controller are given using the Youla-Kučera parameterisation. This approach is tested on a three-tank system, where it is shown that the performance changes of the soft sensor caused by PMM are recovered.

Keywords: Sensors and soft sensors, Artificial intelligence and machine learning
Abstract: Soft sensors, or inferential sensors, are crucial in quality and process control systems because they allow for efficient, online estimation of essential quantities that are otherwise difficult or expensive to measure directly. In many applications, it is common to use cost-effective measurement equipment, offering faster data collection than high-fidelity measurements, albeit at the price of reduced accuracy. These low-fidelity data can provide useful information to enhance the estimation of output quantities of interest, thereby facilitating the design of inferential control systems. In this work, we introduce an innovative approach to soft sensing by employing hierarchical, multi-fidelity surrogate models as soft sensors, integrated with Bayesian optimization for input variable selection. Our method creates a parsimonious model by identifying and organizing relevant inputs into a fidelity hierarchy, which enables a multi-fidelity neural network to sequentially refine estimations by extracting crucial information progressively. First, we showcase the effectiveness of the proposed framework on a numerical benchmark, then we use our method to create a surrogate model as soft sensor for accurately determining the atmospheric particulate matter concentration (PM2.5) using real data collected from low-cost sensors.

Keywords: Process and performance monitoring, Fault detection, diagnosis, supervision, and safety, Artificial intelligence and machine learning
Abstract: Stiction in control valves, among other problems, presents a formidable challenge in industrial control loops, often resulting in suboptimal system performance. Given its significant impact, stiction detection has become a crucial aspect of controller performance monitoring. While machine learning-based methods for stiction detection have gained traction, this paper investigates the effectiveness of Inception Networks and Inception-Residual Networks as potential enhancements to the previously proposed CNN method. The results highlight that these adjustments improve accuracy, from 75.3% to 79.45%, using the same training dataset, effectively capturing variations overlooked by other methods. The application of real industrial data highlights the improvements offered by the proposed framework

Keywords: Fault detection, diagnosis, supervision, and safety, Artificial intelligence and machine learning, Modeling and identification
Abstract: Modern day industries invest heavily on looking for various methods for timely and accurate fault detection and diagnosis. Since training a model to learn all possible faults is challenging and impractical, developing an active learning based methodology which is capable of learning about any new faults arriving in the plant in the due course of operation is the main objective of this paper. This objective is achieved through a two staged methodology where in, an unsupervised learning strategy using one-class SVM is considered in the first stage to detect the presence of a new fault. In the second stage a multi-class classifier of Wavelet Neural Network is utilized to detect the nature of fault. The efficacy of the proposed method is demonstrated on a benchmark Tennessee Eastman Process and the results are compared with the existing methods.

Keywords: Artificial intelligence and machine learning, Process optimization
Abstract: Bayesian optimization (BO) is a widely used method for data-driven optimization that generally relies on zeroth-order data of objective function to construct probabilistic surrogate models. These surrogates guide the exploration-exploitation process toward finding global optimum. While Gaussian processes (GPs) are commonly employed as surrogates of the unknown objective function, recent studies have highlighted the potential of Bayesian neural networks (BNNs) as scalable and flexible alternatives. Moreover, incorporating gradient observations into GPs, when available, has been shown to improve BO performance. However, the use of gradients within BNN surrogates remains unexplored. By leveraging automatic differentiation, gradient information can be seamlessly integrated into BNN training, resulting in more informative surrogates for BO. We propose a gradient-informed loss function for BNN training, effectively augmenting function observations with local gradient information. The effectiveness of this approach is demonstrated on well-known benchmarks in terms of improved BNN predictions and faster BO convergence as the number of decision variables increases.

Keywords: Dynamic modelling and simulation for control and operation, Modeling and identification
Abstract: This paper presents a hybrid optimization methodology for parameter estimation of reactive transport systems. Using reduced-order advection-diffusion-reaction (ADR) models, the computational requirements of global optimization with dynamic PDE constraints are addressed by combining metaheuristics with gradient-based optimizers. A case study in preparative liquid chromatography shows that the method achieves superior computational efficiency compared to traditional multi-start methods, demonstrating the potential of hybrid strategies to advance parameter estimation in large-scale, dynamic chemical engineering applications.

Keywords: Artificial intelligence and machine learning, Process optimization
Abstract: Cutting planes refine the feasible region of relaxed Integer Programming (IP) problems, but traditional methods relying on static heuristics often fail to generalize effectively. This study investigates Multi-Agent Reinforcement Learning (MARL) frameworks -- TD3-TD3, PPO-PPO, and TD3-PPO -- for dynamic, adaptive cut selection. MARL improves scalability and exploration by distributing decision-making across specialized agents, outperforming conventional techniques. The hybrid TD3-PPO configuration balances TD3’s sample-efficient learning with PPO’s robust exploration. PPO-PPO demonstrates superior exploration and success rates in a sensor network design problem relevant to process control, while TD3-TD3 offers greater stability. The results highlight MARL’s potential for enhancing IP solvers and solving complex optimization problems.

Keywords: Process optimization, Process control, Dynamic modelling and simulation for control and operation
Abstract: Optimal control problems are used for many tasks such as model-based control, state and parameter estimation, and experimental design for complex dynamic systems. The solution to these problems can be divided into two tasks, where the first corresponds to the enumeration of different arc sequences and the second is the computation of the optimal values of the decision variables for each arc sequence. For the latter task, this paper proposes a method to approximate the cost and constraints of the problem as polynomial functions of the decision variables via computation of partial derivatives up to second order and multivariate Hermite interpolation. This method allows reformulating the problem for an arc sequence as a polynomial optimization problem, which is expected to enable efficiently solving optimal control problems to global optimality. The method is illustrated by a simulation example of a reaction system.

Keywords: Process optimization, Model predictive control, Process control
Abstract: The goal of dynamic real-time optimization (DRTO) applications is to compute an optimal operational trajectory for a plant by generating set-points for the lower-level control algorithm to track. This approach can be further improved by directly incorporating the control algorithm (such as Model Predictive Control, MPC) into a closed-loop DRTO (CL-DRTO). By doing so, CL-DRTO can predict both the plant and controller responses to set-point adjustments, enhancing the performance of the entire system. However, CL-DRTO schemes require a mechanism to utilize plant measurements to adapt the model to the current plant conditions. Otherwise, the decisions will be based on a nominal model and are likely to be suboptimal. This study proposes a plant feedback scheme using an extended Kalman filter within a CL-DRTO framework that embeds an MPC model. In this novel model adaptation approach in the context of CL-DRTO, not only the states and parameters of the plant model are updated but also the embedded linear MPC model, which is adapted via an output disturbance scheme. Moreover, by adding input constraints to the CL-DRTO problem, this formulation allows a simplified representation of the MPC solution at the CL-DRTO level without directly accounting for input constraints at the MPC level, which reduces computation time. The efficacy of the proposed CL-DRTO approach is demonstrated through application to a multi-input multi-output CSTR where a critical parameter is not measurable.

Keywords: Fault detection, diagnosis, supervision, and safety, Process and performance monitoring, Artificial intelligence and machine learning
Abstract: The rapid advancement of Industry 4.0, artificial intelligence, and big data sensor technologies has made industrial systems highly complex and dynamic. Classical fault detection and diagnosis (FDD) techniques depend on insufficient information and variables with equivalent uncertainty. This paper introduced an advanced dynamic inner reconstruction-based contribution with global-local preservation projection (DiGLPP-RBC) for fault detection and diagnosis. Firstly, inner data statistics are extracted to develop an augmented matrix, which is used to characterize the dynamic latent variable using the DiGLPP framework. Secondly, reconstruction-based contribution (RBC) is used to determine fault contribution. The proposed method employs Hotelling’s T2 and squared prediction error (SPE) to detect and diagnose variable contributions and kernel density of faults in the ethanol-water industrial distillation system. The proposed framework’s robustness is compared with traditional baseline frameworks such as dynamic inner principal component analysis (DiPCA) and bi-directional long short-term memory-autoencoder (BiLSTM-AE). The results indicate that the DiGLPP-RBC technique detects, identifies, and diagnoses irregularities and faults more effectively and reliably than traditional approaches.

Keywords: Process and performance monitoring, Artificial intelligence and machine learning
Abstract: Accurate battery health prediction is crucial for prolonging battery life and ensuring safety. Traditional methods relying on raw time-series data struggle with complex temporal patterns and sensor noise. To address these limitations, we propose a novel approach that utilizes image-transformed data to perform "knee classification" and State of Health (SOH) estimation concurrently. This integrated approach detects aging events and continuously monitors SOH, enabling preemptive interventions. We employ a Convolutional Neural Network (CNN) to simultaneously perform knee classification and SOH estimation, incorporating Gradient-weighted Class Activation Mapping (Grad-CAM) to enhance interpretability by emphasizing critical regions involved in the classification process. The proposed model achieves an 89% classification accuracy, with higher recall than the time-series-based approach, particularly in identifying the intermediate state. Additionally, the pre-trained CNN-based model attains an R² value of 0.977 in SOH prediction, demonstrating its effectiveness for battery condition monitoring. These findings highlight the benefits of an integrated multi-task learning approach, addressing the limitations of conventional time-series models.

Keywords: Modeling and identification, Process and performance monitoring, Dynamic modelling and simulation for control and operation
Abstract: We study real-time process monitoring, where employed online sensors yield inaccurate information. A multi-fidelity (MF) modeling approach is adopted that integrates dynamic information from online, low-fidelity (LF) data with infrequent, high-fidelity (HF) laboratory measurements. The proposed methodology is demonstrated on a composition monitoring problem derived from real oil refinery operations. The developed MF model exhibits a significant improvement in accuracy with respect to both LF data (online sensor) and the HF model (standard soft sensor). The results highlight the potential of MF modeling for improving process monitoring and control through the integration of diverse data sources.

Keywords: Artificial intelligence and machine learning
Abstract: Desiccant dryers play a critical role in industrial sulphonation processes by ensuring that moisture is effectively removed from the air used during SO2 to SO3 conversion. This is necessary to prevent the formation of sulphuric acid, which can harm machinery and lower product quality. This paper introduces a novel approach utilizing Multi-Feature k- nearest neighbours (MF-kNN) forecasting to optimise the drying and regeneration cycles of the dehumidification process units. A key advantage of the MF-kNN model is its ability to perform one-shot forecasts relatively early in the cycle, accurately predicting critical transitions without the need for recursive recalculations. The proposed approach was tested using data from a large-scale surfactant production facility. For forecasting the regeneration cycle endpoint, the model incorporates both the regeneration inlet and outlet air temperatures. Hyperparameter tuning results show that assigning 50% of the feature weight to the inlet temperature results in the lowest forecasting error. Two approaches for data window selection were investigated, namely a moving and an expanding window. The moving window approach outperforms the expanding window approach by 35% and 42% reduced errors for endpoint detection and time- series forecasting tasks respectively. Overall, the model is able to predict the endpoint within a 2 min accuracy with a 400 min lead time on the tested cycles.

Keywords: Process and performance monitoring, Artificial intelligence and machine learning
Abstract: Bonnet polishing is an ultra-precision polishing technique used for manufacturing components utilized in optics, electronics, and scientific instrumentation, where sub-nanometer accuracy is required. However, the process is not fully deterministic and requires multiple process-metrology iterations. In modern computer numerically controlled (CNC) machines, polishing is performed by moderating the bonnet tool dwell time at each location based on the input parameters and material removal rate (MRR). While the MRR is typically treated as constant once established, it continuously evolves due to the process’s dynamic nature and changing conditions. This variability in MRR impacts the convergence of the polishing process, necessitating repeated surface processing and resulting in increased manufacturing time and cost. In this work, we present a data-driven approach to estimate the amount of material removed during the pre-polishing routine in bonnet polishing. The estimations are based on the force exerted by the bonnet tool on a polished surface along the three dimensions. Measurements were obtained using a bespoke force table with load sensors across three axes, mounted on the Zeeko IRP600 machine table. The results demonstrate the effectiveness of this data-driven approach for estimating MRR, achieving a mean absolute error of 0.0541 μm and a mean absolute percentage error of 5.89% across the test sets.