Keywords: Systems biology, synthetic biology, metabolic flux modeling
Abstract: As biology enters the big data club, interest shifts from acquisition to analysis. The systems biology community has developed an array of tools for structuring and interrogating multi-omics datasets. Despite generating data at an unprecedent scale and rate, however, biological datasets present several challenges. Although genomic and transcriptomic datasets tend to be rich, other omics datasets tend to be sparse, as is process-level information. Additionally, data are often representative of a narrow range of systems and experimental conditions. Process systems methodologies often offer the go-to tools for the community to fill the knowledge gaps and provide industrially relevant solutions. This talk will give an overview of modelling approaches for mammalian cell factories and illustrate how systems tools can guide both cell and process design. It will further illustrate approaches for transferring models across systems and conditions, minimising associated experimentation, and will also review the particularities of biological systems that pose future challenges.

Keywords: Process and performance monitoring, Fault detection, diagnosis, supervision, and safety, Modeling and identification
Abstract: Multimode characteristics commonly exist in modern industrial processes. Previous multi-model approaches treat steady states and transitions separately. However, identifying each mode is often tedious, generally achieved through clustering, requiring operators to tune hyperparameters extensively. As practitioners prefer a concise and easily implemented approach for multimode dynamic process monitoring, we initially propose a hierarchical scheme to simplify the modeling process while enhancing monitoring performance. Our method iteratively constructs dynamic models in a hierarchical, monitoring-oriented manner without mode partition. It offers three advantages. Firstly, modeling is directly conducted following a hierarchical structure driven by monitoring indexes, which is more concise and ensures monitoring performance. Secondly, by eliminating mode partition, only three hyperparameters, such as model order and the termination condition, need to be decided by humans. This significantly reduces human labor and facilitates the applicability of the proposed method across various processes. Lastly, by focusing on dynamic characteristics rather than steady state and transitional modes, our method reduces the number of required models for a given process, resulting in a simpler multi-model structure that still ensures monitoring performance.

Keywords: Batch process modeling and control, Modeling and identification, Systems biology, synthetic biology, metabolic flux modeling
Abstract: Accurate estimation of growth rates from sparse and noisy concentration data is a significant challenge in bioprocess modeling. This paper presents a method that utilizes a modular non-linear interpolation framework combined with Bayesian parameter inference to address this issue. By incorporating prior knowledge of cell culture dynamics with differentiable basis functions, our approach generates credibility intervals for growth rates, enhancing the reliability of predictions. We validated the performance of our method using growth rate data generated in silico and demonstrated its application on two in vitro datasets, showcasing its robustness across various measurement conditions and practical applicability. Results indicate improvements in the reliability and credibility of predictions compared to traditional methods, making this framework a valuable resource for accurate growth rate estimations

Keywords: Artificial intelligence and machine learning, Process control, Integration between scheduling and control
Abstract: This paper presents a hierarchical framework for demand response optimization in air separation units (ASUs) that combines reinforcement learning (RL) with linear model predictive control (LMPC). We investigate two control architectures: a direct RL approach and a control-informed methodology where an RL agent provides setpoints to a lower-level LMPC. The proposed RL-LMPC framework demonstrates improved sample efficiency during training and better constraint satisfaction compared to direct RL control. Using an industrial ASU case study, we show that our approach successfully manages operational constraints while optimizing electricity costs under time-varying pricing. Results indicate that the RL-LMPC architecture achieves comparable economic performance to direct RL while providing better robustness and requiring fewer training samples to converge. The framework offers a practical solution for implementing flexible operation strategies in process industries, bridging the gap between data-driven methods and traditional control approaches.

Keywords: Artificial intelligence and machine learning, Process optimization, Scheduling, coordination and optimization
Abstract: In supply chain management, decision-making often involves balancing multiple conflicting objectives, such as cost reduction, service level improvement, and environmental sustainability. Traditional methods for multi-objective optimization, such as linear programming and evolutionary algorithms, have proven useful but struggle to adapt in real-time or handle the dynamic nature of supply chains. In this paper, we propose a novel approach that combines Reinforcement Learning (RL) and Multi-Objective Evolutionary Algorithms (MOEAs) to address these challenges. Our method leverages MOEAs to search the parameter space of policy neural networks, resulting in a Pareto front of policies. This equips the decision-maker with a swarm of policies that can be dynamically switched based on the current system conditions and objectives, ensuring flexibility and adaptability in real-time decision-making. We demonstrate the effectiveness of this hybrid approach through a series of case studies that showcase its ability to respond to the changing dynamics of supply chain environments. We also outperform state-of-the-art methods when benchmarking against our inventory management case study. The proposed strategy not only improves decision-making efficiency but also provides a more resilient framework for fast decision-making and handling uncertainty in supply chains.

Keywords: Artificial intelligence and machine learning, Batch process modeling and control, Process control
Abstract: Process optimisation and quality control are crucial in process industries to minimise waste and enhance economics. However, common uncertainties ranging from feedstock variability to human error can cause significant deviations in product quality leading to batch discards. This study introduces a novel framework combining machine learning with optimisation strategies to identify optimal operational spaces under uncertainty. Using a process model, the framework screens a broad operational space, isolating promising sub-regions and control trajectories. Machine learning techniques are used to cluster these sub-regions by displayed control patterns, and a dynamic optimisation framework identifies the maximum operable design space, ensuring constraints are met under uncertainty. Two case studies, involving a fermentation process and a formulation manufacturing process, were conducted to demonstrate the high efficiency of the proposed framework and to showcase its strong potential for industrial applications.

Keywords: Interaction between design and control, Artificial intelligence and machine learning, Modeling and identification
Abstract: This work presents a methodology for the simultaneous generation, design, and control of chemical process flowsheets (CPF) using RL, starting from an inlet flowrate and a set of unit operations (UOs) involving reaction-separation systems, each equipped with an embedded decentralized control system. The key innovation lies in embedding neural network surrogate models, which approximate the dynamic behaviour of complex UOs within the RL environment. The proposed framework was validated through a case study focused on the reaction and separation of products at varying purities. Results demonstrate the agent’s ability to generate economically attractive CPFs that can maintain the dynamic operation of the systems in closed-loop in the presence of external disturbances.

Keywords: Renewable energy system, Interaction between design and control, Integration between scheduling and control
Abstract: This work examines the impact of process design decisions on the ability of a process to participate in Demand Response (DR) activities. Focusing on load shifting capabilities, the DR particibility potential of a given process design is evaluated using the notion of levelized cost of the load-shifting capacity, which is taken as a measure of an overall cost of the design decisions over the lifetime of the process. Initially, the concept of levelized cost of energy (LCOE), or levelized cost of electricity, is introduced and adapted to capture the levelized cost of load-shifting (LCOL). Then, a bottom-up approach to calculate the load-shifting capacity for a given process design, and the associated levelized cost, is developed based on an MILP scheduling model. The implementation of the proposed approach is demonstrated using a conceptual case study involving a reactor-storage process with a discretized scheduling model. The case study investigates the differences in load-shifting capacities when considering DR participation in a day-ahead electricity market versus participation in a five-minute market.

Keywords: Integration between scheduling and control, Artificial intelligence and machine learning, Dynamic modelling and simulation for control and operation
Abstract: This study presents a framework for the implementation of a Deep Reinforcement Learning (DRL) agent for optimal scheduling and control integration on flow-shop batch plants with input variability. The agent is designed to take multiple decisions at every time interval which allows for the integration of scheduling and control. A hybrid agent with multiple decision outputs is used to perform online scheduling and control. To account for the short-term history of the process, the agent approaches the optimization problem as a Partially Observable Markov Decision Process (POMDP). The agent makes use of a set of Long Short-Term Memory cells (LSTM) to correlate sequential states from the environment to be aware of its evolution when taking decisions. To demonstrate the advantages and limitations of the hybrid agent, the method is implemented on a batch plant under variability in the inputs. Results showed that the agent’s policy reacted to the fluctuations in concentration from raw materials. To validate the proposed method, a comparison with an agent trained on an environment with fixed inputs was performed to demonstrate the adaptive behavior of the agent developed with the presented framework.

Keywords: Biopharmaceutical processes, Systems biology, synthetic biology, metabolic flux modeling, Batch process modeling and control
Abstract: In biopharmaceutical manufacturing of monoclonal antibodies (mAbs), asparagine (N)-linked glycosylation profile of these proteins is a critical quality attribute. We introduce improved Glycosylation Flux Analysis (iGFA), enhancing our previous GFA by: (1) reformulating constraint-based modeling using enzymatic kinetics to obtain biologically interpretable factors; and (2) implementing the analysis using Python's Pyomo modeling language, which not only reduces computational costs significantly compared to the MATLAB-based GFA, but also makes the iGFA an open-source package. When applied to data from Chinese Hamster Ovary (CHO) cell culture production of mAb under varying pH conditions, the analysis revealed both common and distinct dynamic trends across different pH. We identified galactosylation as the most impacted glycosylation processing by pH. Further, the estimated enzyme-related factors correlated more strongly with gene expression levels than with nucleotide sugar availability, suggesting that glycosylation regulation is predominantly controlled at the transcriptional and/or translational level. Overall, the iGFA is a powerful tool for analyzing glycosylation dynamics in mAb production.

Keywords: Systems biology, synthetic biology, metabolic flux modeling, Biopharmaceutical processes
Abstract: Metabolic network models are powerful tools for understanding cellular functions and guiding biotechnological applications. Yet, the complexity of these models poses challenges in accurately predicting intracellular flux distributions, considering limited measurement availability. To address this, we propose a novel bilevel optimisation framework to metabolic network reduction using Bayesian optimisation. Our method assigns continuous probability values to reactions and iteratively refines a reduced model that balances network simplification with predictive performance. The upper-level Bayesian optimisation process selects reaction removal probabilities, while the lower level evaluates model feasibility and performance through flux sampling. A Gaussian Process surrogate is trained to approximate the impact of reaction removals on model accuracy, guiding the optimisation toward a minimal yet representative network. We applied our methodology to a Chinese Hamster Ovary (CHO) cell metabolic model using multiple datasets, demonstrating its ability to adapt to different datasets and suggest targeted measurements. By unifying lumping and sensitivity analysis concepts in a data-driven framework, our approach systematically simplifies metabolic models, increasing their applicability in both development and manufacturing processes.

Keywords: Dynamic modelling and simulation for control and operation, Modeling and identification, Biopharmaceutical processes
Abstract: Oligonucleotides show great promise for therapeutic applications. While traditional solid-phase oligonucleotide synthesis presents manufacturing challenges due to low scalability, lack of real-time monitoring and high process mass-intensity, liquid-phase synthesis (LPOS) combined with soluble anchors and membrane diafiltration has emerged as a viable alternative. Herein, we formulate a multi-stage dynamic model of the LPOS process using a multi-branched homostar support hub. We exploit this model to analyse the interplay between the durations of the reaction and diafiltration steps on the oligonucleotide yield and purity through the solution of an inverse feasibility problem.

Keywords: Modeling and identification, Dynamic modelling and simulation for control and operation
Abstract: This paper proposes a new generic mass balance model that allows simulating biological cultures in bioreactors at a macroscopic scale. A multi-layer perceptron (MLP) describes the kinetic and stoichiometric parts of the model with one input layer (made of the concentrations of the different components, as well as their inverse TReLU – Thresholded Rectified Linear Unit – transforms), one hidden layer (each neuron output corresponding to one specific reaction rate and being activated by a reciprocal function 1 / x ) and one output layer (each neuron output being the sum of all the reaction contributions of a specific component to its mass balance). The parameters to be identified are split into two subsets: one for the kinetic parameters (weights on the links between input layer and hidden layer) and one for the stoichiometric parameters (weights on the links between hidden layer and output layer). This MLP structure exhibits several advantages, among which its versatility, the biological interpretation of the parameters, and an easy and efficient first estimation of the kinetic and stoichiometric parameters based on measurements of the component concentrations and estimations of their time derivatives. The first parameter estimation can subsequently be used for model reduction and as initial guess for a final nonlinear parameter estimation of the set of ODEs describing the mass balances. The performances of the new generic model are illustrated with a simulated case study.

Keywords: Dynamic modelling and simulation for control and operation, Model predictive control, Systems biology, synthetic biology, metabolic flux modeling
Abstract: Oncolytic virus therapy (OVT) is emerging as a potent alternative to conventional cancer treatments by employing engineered viruses that selectively infect and lyse tumor cells while sparing normal tissues. Although mathematical models have been developed to elucidate the dynamics of OVT and inform personalized therapies, they are often specific to certain organisms. Mathematical models tailored to more recently developed animal models of OVT, such as zebrafish, are not yet available. Here, we introduce the first mathematical model of OVT trained on zebrafish data from published studies to bridge the gap. We explore a variety of mathematical model structures and perform parameter estimation and model selection. The selected model effectively captures the observed tumor dynamics, i.e. delayed tumor shrinkage, and provides valuable insights into the underlying mechanisms of OVT in zebrafish. Our work establishes the groundwork for advancing experimental studies in zebrafish, contributing to the design of more effective cancer treatment strategies in the future.

Keywords: Food engineering, Process analytical technology (PAT), Sensors and soft sensors
Abstract: The problem of estimating the protein content in dough during kneading in a farinograph is addressed exploiting underyling structural observability of a suitable model describing the biochemical reaction network of different protein fractions involved in the formation of the gluten network in the dough and its deterioration due to excessive kneading. Based on an observability analysis for the model a reduced order geometric observer and an extended Kalman Filter are designed and tested with experimental data from dough kneading experiments. The results highlight the potential of incorporating such model-based process analytic tools for improved monitoring of the underlying biochemical mechanisms involved in dough formation, and for providing an additional means for estimation of protein content in flour during standard kneading processes.

Keywords: Artificial intelligence and machine learning
Abstract: This work explores the application of symbolic regression to learn symbolic solutions to Burgers' equation without data. We demonstrate a stepwise symbolic regression strategy that explores models that provide tractable logic from coordinates to state estimates. The first step is to learn a model representing part of the system's physics. This partial model is then used to help discover a model capturing the entire physics of the system. The method was able to learn a model of the solution of the diffusion equation with an R-squared value of 0.99 and produced models for Burgers' equation with different values of the convection coefficient, all with R-squared values greater than 0.98. These solutions to Burgers' equation, represented as transformations of the solution to the diffusion equation, demonstrate the potential of leveraging domain knowledge to simplify the symbol space and build useful primitives for symbolic regression. This work highlights how domain knowledge, expert intuition, and symbolic regression can complement each other to create more interpretable solutions to dynamical system models.

Keywords: Dynamic modelling and simulation for control and operation, Modeling and identification, Artificial intelligence and machine learning
Abstract: Heat exchangers in vapor compression cycles (VCCs) typically involve phase transitions of the internal refrigerant. While numerous studies have focused on modeling these systems, existing techniques require high computational burden or model switching depending on the refrigerant phase at the heat exchanger outlet. These kinds of problems make each model less suitable for control applications. To overcome these challenges, this study proposes a hybrid model that combines conventional modeling methods with data-driven time series analysis to obtain unified model of heat exchanger with refrigerant phase transition, especially the evaporator in the VCC. The proposed model is fundamentally based on the moving boundary (MB) method, and integrates long short-term memory (LSTM) networks to predict unknown parameters arising from the phase transition. The proposed hybrid model shows high accuracy with the Simulink-Simscape simulation result, and yields higher accuracy compared to a fully data-driven LSTM black-box model, which is another approach to make an unified model. This hybrid approach creates a model that improves accuracy compared to the black-box model and eliminates the need for model switching, ultimately facilitating the design of advanced control.

Keywords: Modeling and identification, Dynamic modelling and simulation for control and operation, Artificial intelligence and machine learning
Abstract: Digitization, which involves the adoption of various domain-relevant technologies to create a digital equivalent of physical assets, is the main principle of Industry 4.0. As most process plants such as Waster water treatment process (WWTP) and Post-carbon combustion capture (PCC) process exhibit multi-scale dynamics, identification of models in differential equation form (continuous-time) is advantageous. Further, any underlying physical understanding of the system can be easily captured in this modeling strategy, through appropriate choice of functionality resulting the model Gray-box in nature. Since models in differential equation form are considered, the accuracy of modeling depends on the estimation of derivatives from the sampled data. Therefore, the main objective of this paper is to develop an identification methodology in continuous-time (CT) framework that can capture the physical behavior of the system. To address the issue with the derivative information, the data set is fitted using functions like B-splines subjected to a model-based penalty to ensure that the data fit also satisfies the model of the process. For estimation of a parsimonious model, a sparsity constraint in terms of zero norm on the parameter vector of the model is considered. The efficacy of the method is demonstrated on a Van der Pol oscillator and a Continuous stirred tank reactor (CSTR) system and the results are compared with the existing methods.

Keywords: Artificial intelligence and machine learning, Modeling and identification
Abstract: Surrogate models are becoming an important technique in process design, control, and optimization. These models are often developed at the process level, tailored to specific process configurations. Even when unit-level models are constructed, they rarely account for all units within a flowsheet. In this work, we use a unit-based surrogate model framework for an amine scrubbing process designed to remove CO2, H2S, and other sulfur species from gas streams. Unit-level Artificial Neural Network (ANN) surrogate models are created for each process unit. The developed unit-based models remain independent of process configuration and can be connected to recreate many possible arrangements of the amine scrubbing process. To demonstrate the versatility of our approach, we present two case studies. The first involves a typical amine scrubbing process configuration, while the second considers two absorber columns operating in parallel. This unit-based surrogate approach proves scalable and modular, enabling accurate predictions across diverse process configurations. By adopting this unit-based surrogate modeling framework, we can explore various process scenarios and simulate the amine scrubbing operations across multiple plant configurations.

Keywords: Artificial intelligence and machine learning, Process analytical technology (PAT), Modeling and identification
Abstract: There is growing interest in shifting from batch manufacturing (BM) to continuous manufacturing (CM) in the pharmaceutical industry. Constructing statistical models that accurately predict critical quality attributes (CQAs) from operating conditions with minimal experiments in the new process is required to identify the optimal operating conditions and monitor the process. This study aims to demonstrate that heterogeneous transfer learning (TL) using data from the batch direct compression (BDC) process can enhance the prediction performance of CQAs in the continuous direct compression (CDC) process. We conducted 26 BDC experiments and 19 CDC experiments. Predictive models of tablet hardness were then built using partial least squares regression (PLSR), Gaussian process regression (GPR), and random forest regression (RFR). We employed frustratingly easy heterogeneous domain adaptation (FEHDA) to the two experimental datasets, treating BDC as the source domain and CDC as the target domain. We found that FEHDA achieved lower RMSE and higher R2 than those by models trained using only the CDC dataset. RFR attained the best predictive performance with an average RMSE improvement of 9.36 N. Notably, FEHDA improved the prediction performance in the region where no samples were obtained from the BDC process. These results support the effectiveness of heterogeneous TL for the shift from BDC to CDC.

Keywords: Process control, Modeling and identification, Model predictive control
Abstract: Industrial processes often exhibit complex nonlinear dynamics. Controlling such processes can be computationally intensive, making it advantageous to replace these nonlinear models with a series of linear models defined at various operating points. This approach reduces the computational burden while sufficiently preserving the system's nonlinear dynamics. To enhance the robustness of this control strategy, we focus on designing a multimodel predictive controller (mMPC). The MPC cost function considers weighted model formulation and includes state constraints from all linear models. The approach is applied to control an industrial chemical reactor model and compared with multiple-model adaptive control (mMAC) implementing weighted state constraints. As a base for comparison, a nonlinear model predictive controller (nMPC), and a linear MPC that switches to the best model (sMPC) according to predefined state regions. The results demonstrate greater robustness and reduced constraint violations of the proposed method.

Keywords: Batch process modeling and control, Process optimization, Process control
Abstract: We explore batch-to-batch optimization of a simulated semi-batch process, which represents an important industrial process at BASF. Our primary objectives are to ensure safe batch operations, to produce within specifications, and to increase the throughput. To achieve these goals, Modifier Adaptation with Quadratic Approximation (MAWQA) is employed to optimize key operational parameters iteratively over a sequence of batches. By integrating modifier adaptation with the quadratic approximation used in derivative-free optimization, sensitivity to noise is reduced, and the speed of convergence is improved. A challenging feature of the case considered here is that the constraints involve the maximum temperature and pressure over the batch, which depends on the interaction of feedback controllers with the plant, and that the product quality can only be determined at the end of the batch.

Keywords: Batch process modeling and control, Process optimization, Dynamic modelling and simulation for control and operation
Abstract: To describe a quantitative relationship between the operating conditions of cooling crystallization process and product crystal size distribution (CSD), a surrogate modelling method based on the Gaussian process regression (GPR) is proposed by using only experimental data of batch crystallization process of β form L-glutamic acid (LGA). A modified design of experiments (DoE) is presented to reduce the number of batch crystallization experiments. Based on the surrogate model, an objective function reflecting the concentration of product CSD and desired yield is introduced to optimize these operating conditions. Experiments on the seeded cooling crystallization process of β-LGA are conducted to verify the effectiveness and advantage of the proposed method.

Keywords: Artificial intelligence and machine learning, Fault detection, diagnosis, supervision, and safety
Abstract: As chemical plants evolve towards full autonomy, the need for effective fault handling and control in dynamic, unpredictable environments becomes increasingly critical. This paper proposes an innovative approach to industrial automation, introducing validation and reprompting architectures utilizing large language model (LLM)-based autonomous control agents. The proposed agentic system—comprising of operator, validator, and reprompter agents—enables autonomous management of control tasks, adapting to unforeseen disturbances without human intervention. By utilizing validation and reprompting architectures, the framework allows agents to recover from errors and continuously improve decision-making in real-time industrial scenarios. We hypothesize that this mechanism will enhance performance and reliability across a variety of LLMs, offering a path toward fully autonomous systems capable of handling unexpected challenges, paving the way for robust, adaptive control in complex industrial environments. To demonstrate the concept's effectiveness, we created a simple case study involving a temperature control experiment embedded on a microcontroller device, validating the proposed approach.

Keywords: Batch process modeling and control, Dynamic modelling and simulation for control and operation, Process optimization
Abstract: Market demand for high-performance polymers often requires flexibility to customize product properties with a high degree of consistency. Batch processes are often used to produce high-performance polymers due to their flexibility. The product properties are related to molecular weight distribution (MWD). However, due to the nature of dynamic within batch and batch-to-batch variability, controlling MWD is challenging, leading to inconsistent product properties. Additionally, changes in polymer properties often require cumbersome formulation development, which in practice is not flexible enough to meet downstream requirements for polymer property customization. To address these issues, a mechanistic model of an industrial batch polymerization process is developed, and a dynamic optimization problem is designed on this model. Within the framework of dynamic optimization, it is demonstrated that the MWD can be adjusted by manipulating the initial concentration and flow rate of the chain transfer agent at a constant reaction temperature, which provides the basis and direction for the subsequent development of control strategies to the MWD.

Keywords: Sensors and soft sensors, Modeling and identification, Fault detection, diagnosis, supervision, and safety
Abstract: Flow rate soft sensors have become an important alternative for costly hardware flow meters, as they can estimate the flow rate with sufficient precision from easily measurable variables by using models and state estimation algorithms. This paper addresses the fundamental challenge that arises from ambiguous estimation problems, where the measured variable corresponds to two or more possible flow rate values. We develop and implement a decision algorithm that yields correct results in an industrial setup with substantial measurement noise. The results demonstrate a reliable flow rate estimation, providing a viable solution for real-time flow monitoring in centrifugal pumps with complex characteristics.

Keywords: Process optimization, Process control, Plantwide control
Abstract: This paper investigates the operation of parallel compressors with variable speed drives to deliver gas at a desired flow rate while maintaining a target pressure at a common discharge header. We examine strategies to minimize energy consumption amid discharge flow fluctuations caused by changes in gas demand. Specifically, we model the energy consumption impact of varying operating points, accounting for efficiency sensitivity to flow. Our approach employs sample averaging to estimate expected energy usage under flow variations, which informs an offline surrogate objective function reflecting energy consumption under disturbances. This surrogate is subsequently used online in a deterministic nonlinear programming framework to approximate a stochastic optimization solution, determining optimal load distributions for the compressors. Additionally, we propose an economic model predictive control (MPC) method. This approach first solves a tracking problem to stabilize header pressure, using compressor flows as manipulated variables, then redistributes the calculated control effort for the first step of the solution through an economic optimization. Both methods are implemented in a comprehensive pipeline compressor station simulation model, featuring a control hierarchy for station pressure, compressor flow, and anti-surge controllers. Simulation results, with and without flow disturbances, reveal that the stochastic load-sharing approach reduces energy consumption by 4.3% compared to a purely deterministic method, with the economic MPC further improving efficiency by an additional 2.2%

Keywords: Artificial intelligence and machine learning, Batch process modeling and control, Model predictive control
Abstract: In this work, we proposed a controller based on reinforcement learning (RL) for the unseeded batch crystallization of paracetamol in ethanol. The controller aims to achieve the crystal mean volume size and the crystal mass at five different targets by manipulating the temperature. We used the deep deterministic policy gradient (DDPG) algorithm to train the control agent. The performance of the RL controller was compared to an NMPC using a population balance model (PBM) as its internal model and tested for the five different scenarios. The controllers were also tested taking into account a 5% disturbance in the concentration measurement. Both controllers were able to reach values of the controlled variables close to the targets even accounting for disturbances. However, the RL controller was able to calculate the control action much faster than the NMPC and imposed less temperature changes, which is something better for real control applications. Therefore, the RL controller presented as a more suitable approach than an NMPC using a PBM as its internal model for controlling the paracetamol batch crystallization process.

Keywords: Modeling and identification, Artificial intelligence and machine learning, Dynamic modelling and simulation for control and operation
Abstract: We propose data-driven surrogate models to solve systems of time-dependent differential equations coupled with noise. Using a feedforward neural network, we separately learn the noise and solution, tackling approximations across regimes with bifurcations and rare events. Focusing on irregular data generated by a stochastic noise model on a one-dimensional spatial lattice coupled to a differential equation, we examine two profiles: the periodic complex Ginzburg-Landau equation and a saddle bifurcation equation exhibiting rare events. This coupling introduces conditional data, enabling solutions to reach new states while posing challenges for accurately learning the underlying dynamics.

Keywords: Artificial intelligence and machine learning, Batch process modeling and control, Process optimization
Abstract: Biochemical processes, characterized by nonlinear dynamics and uncertainties, pose significant optimization challenges. This work explores Robust Predictable Control (RPC) as a Reinforcement Learning (RL) algorithm to enhance a fed-batch penicillin production process utilizing the simulation model IndPenSim. Unlike some RL implementations that constrain exploration based on prior knowledge, the selected RPC approach allows the RL agent to explore freely and identify optimal control strategies by itself. We trained the RL agent under disturbance-free conditions and evaluated its performance against various unseen initial process conditions and disturbances. Results show that RPC significantly outperforms other process control methods, including other RL implementations, achieving higher yields with fewer necessary measurements as input for the RL agent. Analyzing two reward functions - penicillin concentration and yield - revealed that using concentration in the reward function improved agent training for maximizing yield, highlighting the importance of reward design in RL. Additionally, the trained RL agent effectively adapted to different action intervals, demonstrating robustness in dynamic environments without retraining. Our findings underscore RPC's potential for optimizing biochemical processes, especially in scenarios with few measurements, paving the way for AI-driven control systems in industrial applications.

Keywords: Metabolic engineering, Systems biology, synthetic biology, metabolic flux modeling, Modeling and identification
Abstract: Metabolic modeling is a valuable tool for studying microbial metabolism and has broad applications across fields like biotechnology, medicine, and environmental science. The construction of metabolic networks is crucial in this process, though their development presents significant challenges. While genome-scale networks offer detailed insights, they are computationally demanding, and smaller networks are often too simplified. This study discusses a methodology to derive a metabolic network of intermediate size by combining biological knowledge, experimental data, and mathematical tools to refine the network definition. The present study focuses on the modeling of photosynthetic cyanobacteria Arthrospira sp. PCC 8005 and experimental validation is achieved using cultures in continuous mode. The procedure is effective, yielding promising results, and metabolic analyses show predictive capabilities that are in agreement with existing studies while studying the impact of different nitrogen sources on the growth of cyanobacteria.

Keywords: Bio-applications, Data mining tools, Systems biology, synthetic biology, metabolic flux modeling
Abstract: In safety-critical fields, such as skin-lesion classification, interpretability of the decisions of a machine learning model is required. This can be provided through concept-based interpretability methods like testing with concept activation vectors (TCAV). TCAV quantifies how specific human-understandable concepts influence a model’s decisions. A further issue affecting the performance of ML models is generalizability, i.e., how well a model generalizes to unseen data from a different domain. It is currently unknown how the interpretability provided by TCAV is affected by domain shifts. Here we show that TCAV-based interpretability is predominantly unaffected by domain shifts. To that end, we introduce concept detection scores (CDS) as aggregated TCAV scores which are directionally unified and thus a suitable evaluation metric. The results show only small differences between CDS within domain and across domain for 48 models trained on three distinct source domains. This increases the viability of TCAV as an interpretability tool since it can be used without additional effort to manage generalizability.

Keywords: Dynamic modelling and simulation for control and operation, Bioenergy production (bioethanol, algae, anaerobic digestion), Artificial intelligence and machine learning
Abstract: Extended Dynamic Mode Decomposition (EDMD) has received increasing attention in the last decade, but neural networks remain the most popular approach to the data-driven representation of biochemical processes in the published literature. In this study explores the potential of pqEDMD—a variant of EDMD using a reduced set of orthogonal polynomials—to approximate the dynamics of a complex system, i.e., a raceway pond for the biological treatment of wastewater and the production of algal biomass. We carefully discuss the main ingredients of the method, and illustrate the performance of the method with numerical results, showing promising prospects.

Keywords: Fault detection, diagnosis, supervision, and safety, Process and performance monitoring, Artificial intelligence and machine learning
Abstract: Effective fault detection and diagnosis (FDD) in chemical process systems is critical for maintaining safe and reliable operations. While deep learning methods have improved fault classification performance, they often require long sequences of data after a fault occurs, delaying timely interventions. In this work, we propose an early fault diagnosis method that enables rapid fault diagnosis using multistep multivariable prediction. Our approach employs a transformer-based prediction model to predict future values of key process variables, enriching the current data with these predictions. An LSTM-based model then classifies the enriched data into specific fault categories, leveraging both current and predicted information for improved precision. We evaluate the performance of the proposed approach on the Tennessee Eastman Process benchmark, demonstrating its effectiveness in early fault diagnosis.

Keywords: Process and performance monitoring, Process analytical technology (PAT), Sensors and soft sensors
Abstract: The problem of model-based state estimation in absence of a reliable model for the liquid evaporation rate in a spray drying process is addressed. The process is described by a three-state model consisting of the product moisture, air humidity and temperature in the spray drying chamber with measurements of the temperature and air humidity. The state estimation problem is approached within the framework of unknown-input observer design, considering the evaporation rate as unknown time-varying bounded source term. It is shown that using an adequate integral state transformation based on energy and mass conservation mechanisms it is possible to obtain reliable estimates for the product moisture by model-based sensor fusion. The performance of the proposed scheme is evaluated using numerical simulations for a previously validated process model.

Keywords: Dynamic modelling and simulation for control and operation, Process control, Plantwide control
Abstract: This work examines the use of P-only and PI averaging level control (ALC) of a tanks-in-series recycle process, the most basic model for understanding material balance dynamics in non-reactive recycle systems. We show that P-only ALC tuning for a standalone tank effectively extends to the recycle process. However, directly applying isolated tank PI ALC tuning is not possible due to the complex dynamics and flow amplification that occur along the tank cascade. A systematic procedure for plantwide PI ALC (de)tuning is developed, ensuring acceptable flow amplification while fully utilizing the available surge capacity for the worst-case disturbance. The application of the systematic tuning method to a three-column azeotropic separation process shows that PI ALC achieves significantly higher high-frequency variability attenuation compared to P-only ALC at the expense of mild flow amplification around a small low-frequency resonance peak. These findings suggest that PI level control, which eliminates level offsets, a feature favored by operators, may be applied in plantwide non-reactive systems with systematic detuning.

Keywords: Artificial intelligence and machine learning, Process optimization, Biopharmaceutical processes
Abstract: Abstract: Chromatographic separation processes are essential for achieving high-purity products in industries such as pharmaceuticals and biotechnology, where complex mixtures such as monoclonal antibodies require precise purification. These processes, such as the twin-column Multicolumn Countercurrent Solvent Gradient Purification (MCSGP), are typically described by nonlinear partial differential and algebraic equations, leading to high computational demands that limit their feasibility for real-time optimization. In this work, we develop a hybrid modeling approach that combines artificial neural networks (ANNs) with process knowledge to capture the nonlinear dynamics of the twin-column MCSGP system efficiently. By retaining the mechanistic separation isotherm while eliminating the need for spatial discretization, the model reduces computational effort substantially, achieving cyclic steady state (CSS) predictions in a fraction of the time required by the respective high-fidelity model. The hybrid model is integrated within a Bayesian optimization (BO) framework to maximize process yield while meeting stringent product purity requirements. A comparative analysis with both data-driven and high-fidelity models demonstrates that the hybrid model provides a computationally efficient, accurate alternative suitable for real-time applications in continuous chromatography.

Keywords: Renewable energy system, Model predictive control, Process control
Abstract: This paper demonstrates how incorporating future input power information impacts the performance of a nonlinear model predictive control (NMPC) algorithm for an alkaline electrolyzer (AEL) plant. The primary objective of the NMPC is to maintain the stack temperature and number of moles of water in the AEL within operating limits, despite large variations in the input power. The NMPC combines an optimal control problem (OCP) with a continuous-discrete extended Kalman filter (CD-EKF). For both the OCP and the CD-EKF, we use a model that is different from the AEL simulation model. We present three closed-loop simulations: two where the NMPC operates at different stack temperature and water mole setpoints with only current input power information, and one where it receives information about future power changes in advance. The results show a 1.583% increase in hydrogen production when the NMPC utilizes information about future power changes.

Keywords: Process optimization, Carbon capture, utilization, and storage, Artificial intelligence and machine learning
Abstract: Post-Combustion CO₂ capture has been a major focus for decades in efforts to reduce global warming. In this study, CO₂ emissions from a coal power plant are analyzed taking into account an existing process available in Aspen Hysys. In this paper, the main objective is to identify the operating conditions that result in an economical operation of the process. Since the process is complex, instead of relying on a first-principle-based steady-state model, a data-driven approach via a wavelet neural network was considered because of its linearity with respect to the parametric structure. This allows for faster training and provides accurate predictions. Although the model is accurate, due to changes in operating conditions in a process plant, a mismatch between the actual plant output and the predicted model output may exist. To account for this mismatch, Bayesian optimization is employed using Gaussian process regression, which estimates both the mean value and uncertainty of the mismatch. The trust region approach is applied to balance the crucial factors of exploration and exploitation. The efficacy of the proposed method is demonstrated via a Benoit system and a PCC process.

Keywords: Bioenergy production (bioethanol, algae, anaerobic digestion), Modeling and identification, Waste water treatment processes
Abstract: A mathematical modelling approach for a hydrogenotrophic methanation process in a trickling filter bioreactor (TFB) is formulated and simulated. As other works have suggested, the proposed model partitions the physical space in an arbitrary number of vertical levels and uses the first Fickian diffusion law along its vertical axis and along the thickness of the biofilm layer attached to the inert bed material. The biological hydrogenotrophy reaction is modelled using Monod kinetics. According to the ideal gas law, to calculate gas flows among levels, the model approach considers the fixed amount of moles of gas in each TFB level. Simulations of the proposed model were compared to experimental results of ex-situ hydrogenotrophic methanation in a TFB. Model performance against the experimental results of the real reactor fitted remarkably well the effluent proportions within an average 2% error band in the reported best experimental case. It even surpassed real productivity by 20% on average, considering an ideal scenario in which the model formulation assumes that hydrogenotrophic methanation by archaea is the sole biological transformation. A qualitative analysis of important model parameters was crucial for fitting simulation results to real experimental data. Particularly, influent raw biogas proportions near the ideal stoichiometry for hydrogenotrophic methanation is detrimental to purity and productivity of the desired biomethane effluent. The numerical effort needed for simulations was remarkably lower than expected, given the large-sized models the approach may produce. Simulation results provided insight into possible model modifications for further discussion.

Keywords: Modeling and identification, Dynamic modelling and simulation for control and operation, Biopharmaceutical processes
Abstract: Macroscopic modeling of bioprocesses is a common approach to derive dynamic predictors suitable for process optimization and control. This study presents a data-driven methodology for inferring reaction rates without resorting to the classical numerical differentiation of experimental data, which is prone to errors in the presence of noise. The approach is based on the minimization of a nonlinear least square criterion, which parameterizes the rates in terms of the temporal values. The proposed method is appropriate for datasets with sparse measurements and few experimental replicates. A use case considering protein production by mammalian cells is used to validate the proposed approach.

Keywords: Batch process modeling and control, Artificial intelligence and machine learning, Process control
Abstract: Batch processes are indispensable for the production of low-volume and high-value products. However, control of a batch process is challenging due to the inherent nonlinearity, and time-varying characteristics. In this work, we consider the problem of nonlinear trajectory tracking control in batch processes. We assume that the reference trajectory is known in advance, and does not change between batches as it is common in repetitive processes. The main objective of this work is to develop a reinforcement learning approach to perform tracking control in a repetitive batch process. To this end, the batch process control is formulated as a Markov decision process with a suitable characterization of state vector, and reward design. A model-free function approximation approach based on deep Q learning is developed as a Deep Reinforcement Learning (DRL) controller. A nonlinear batch reactor system is used to demonstrate the efficacy of the proposed DRL controller. Further, we present the performance of the proposed DRL controller under feed and parametric uncertainties.

Keywords: Fault detection, diagnosis, supervision, and safety, Artificial intelligence and machine learning, Modeling and identification
Abstract: Mechanical faults, mud loss, and insufficient cuttings transport bear significant costs and can appear at unpredictable times during drilling operations. Early detection and diagnosis of such faults to support decisions is essential to avoid severe consequences and lengthy delays. We propose a novel system for Drilling Fault Diagnosis that combines principles of Physics-Informed Neural Networks (PINN) and Multitask Learning (MTL). Since measurements down-hole in the well are rarely available in real time, our proposed system uses measurements of flow and pressure at the drilling rig, only, from which type of fault and accompanying diagnostics such as depth in the well of the fault and its severity are predicted. State-of-the-art strategies of MTL with PINNs are deployed for effective Neural Network (NN) training. Generalization performance is shown to be high as evaluated using randomly generated values for the diagnostic variables. Drilling data collected during normal drilling-ahead conditions may be utilized in the training phase to identify uncertain characteristics of the well, thereby increasing the quality of the physics prior available to the PINN, and in turn improving prediction accuracy of faults. The potential usefulness of the method is illustrated in a simulation, admittedly under quite ideal conditions.