This study tackles the challenge of accurate yield prediction in fluid catalytic cracking (FCC) units by comparing conventional supervised regression with time series forecasting methods using industrial data collected from the distributed control system (DCS) of an FCC plant. We introduce a shifted forecast paradigm that preserves temporal relationships between predictors and targets. Our preprocessing pipeline, which employs trimmed mean smoothing, addresses common industrial data challenges. Results demonstrate that the forecasting approach significantly outperforms supervised regression, achieving a mean absolute percentage error (MAPE) of 1.56% for 3-hour shifted predictions compared to 6.20% for supervised regression. The model maintains robust performance even with extended shifts during predictions, showing an MAPE of 3.55% for 14-day forecasts. This research provides valuable insights for implementing predictive analytics in industrial FCC operations, demonstrating the superio... [more]

Fault diagnosis is critical for ensuring the safety and efficiency of chemical processes, as undetected faults can lead to catastrophic consequences. While deep learning-based methods have shown promise in this field, they often require manual hyperparameter tuning, which is not efficient since they heavily rely on expert knowledge and need iterative trial-and-error. This work introduces a novel approach combining a Multiscale Convolutional Neural Network (MSCNN) with Tree-Structured Parzen Estimator (TPE) for automated hyperparameter optimization to enhance the performance of fault diagnosis for chemical processes. The Multi-Scale Module is to capture complex nonlinear features from the fault data, while the TPE efficiently searches for optimal hyperparameters for MSCNN. An experimental study was carried out on the Tennessee Eastman Process (TE Process) dataset, where the proposed method was benchmarked against state-of-the-art models. The results indicate that the MSCNN-TPE method de... [more]

In an era defined by economic competitiveness and environmental awareness, engineering solutions must maximize profitability, efficiency and sustainability, underscoring the relevance of process optimization and the societal impact any contribution in this research field would bring. In chemical reactor engineering, optimization tasks pose significant challenges due to the highly non-linear and non-convex nature of reactor models, often involving differential equations. While conventional approaches have proven to be reliable strategies for solving these complex problems, their application becomes impractical as problem size and complexity increase. This work introduces a novel application of Physics-Informed Neural Networks (PINNs) to address constrained optimization problems in reactor engineering and demonstrates the proposed methodology through two illustrative case studies in chemical reactor design and control. In doing so, we highlight the capability of PINNs to efficiently lear... [more]

Process synthesis is a fundamental step in process design. The aim is to determine the optimal configuration of unit operations and stream flows to enhance key performance metrics. Traditional methods provide just one optimal solution and are strongly dependent on user-defined technologies, stream connections, and initial guesses for unknown variables. Usually, a single solution is not sufficient for adequate decision-making, especially, when properties such as flexibility or reliability are considered in addition to the process economics. Wastewater Treatment network synthesis and design is a complex problem that demands innovative approaches in design, retrofits, and maintenance strategies. Considering this, an enhanced framework for improving reliability in wastewater transportation networks based on graph theory and machine learning is presented. Machine learning models were developed to predict failure probability, where the XGBoost model provided the best predictions. To select t... [more]

Pressure swing adsorption (PSA) is an energy-efficient technology for gas separation, while the multi-objective optimization of PSA is a challenging task. To tackle this, we propose a hybrid optimization framework, which integrates three steps. In the first step, we establish surrogate models for the constraints using Gaussian processes (GPs) and employ multi-objective Bayesian optimization to search for feasible points that satisfy the constraints. In the second step, we establish surrogate models for the objective function and constraints using GPs and utilize constrained multi-objective Bayesian optimization to search for an approximate Pareto front. In the third step, we perform a local search based on the approximate Pareto front. By employing the trust region filter method, we construct quadratic models for each constraint and objective function and refine the Pareto front to achieve local optimality. This framework demonstrates the efficiency of Bayesian optimization and the loc... [more]

Batch chemical processes face significant challenges due to frequent operational shifts and varying conditions, requiring models to be retrained for each new scenario. This high retraining demand limits the scalability of traditional process monitoring systems, making them unsuitable for dynamic batch operations. To address this, we propose a transfer learning-based framework that enhances adaptability by reusing learned features across different batch conditions, reducing the need for extensive retraining. Proposed method integrates Temporal Convolutional Networks (TCNs) for capturing temporal dependencies in batch data and predicting Quality-Indicative Variables (QIVs) to identify deviations. The core innovation lies in transfer learning, enabling the model to adapt to new process variations with minimal updates. This approach ensures robust, accurate monitoring even under evolving conditions. This framework is validated using the IndPenSim penicillin fermentation dataset, which simu... [more]

Optimization under uncertainty is inherent to many PSE applications ranging from process design to RTO. Reaching process true optima often involves learning from experimentation, but actual experiments involve a cost (economic, resources, time) and a budget limit usually exists. Finding the best trade-off on cumulative process performance and experimental cost over a finite budget is a Partially Observable Markov Decision Process (POMDP), known to be computationally intractable. This paper follows the nonmyopic Bayesian optimization (BO) approximation to POMDPs developed by the machine-learning community, that naturally enables the use of hybrid plant surrogate models formed by fundamental laws and Gaussian processes (GP). Although nonmyopic BO using GPs may look more tractable, evaluating multi-step decision trees to find the best first-stage candidate action to apply is still expensive with evolutionary or NLP optimizers. Hence, we propose modelling the value function of the first-st... [more]

Solvent selection is a critical decision-making process that balances economic, environmental, and societal factors. The vast chemical space makes evaluating all potential solvents impractical, necessitating pre-selection strategies to identify promising candidates. Predictive thermodynamic models, such as the UNIFAC model, are commonly used for this purpose. Recent advancements in deep learning have led to models like the Gibbs-Helmholtz Graph Neural Network (GH-GNN), which overall offers higher accuracy in predicting infinite dilution activity coefficients over a broader chemical space than UNIFAC. This study presents a systematic comparison of solvent pre-selection using GH-GNN and UNIFAC-IL in the context of liquid-liquid extraction. The original GH-GNN model is extended to simultaneously predict organic and ionic systems. This extended GH-GNN model predicts more than 92 % of the logarithmic IDACs with an absolute error of less than 0.3. By comparison, UNIFAC-based models only achi... [more]

Heuristic decomposition schemes like moving horizon schemes are a common approach to approximately solve large-scale mixed-integer programs. The authors propose Bayesian optimization as a methodological approach to systematically tune parameters of decomposition schemes for mixed-integer programs. This paper discusses detailed results of three studies of the Bayesian optimization-based approach using hoist scheduling as a case study: Firstly, two objectives of the tuning problem are examined considering sequences of incumbent solutions found by the Bayesian optimization. Secondly, the Bayesian optimization is applied to a set of test instances of the hoist scheduling problem using four types of acquisition functions; they are compared with respect to the convergence of the tuning problem solutions. Thirdly, the scaling behaviour of the Bayesian optimization is studied with respect to the dimension of the space of tuning parameters. The results of the three studies show that the solutio... [more]

Industrial process monitoring can benefit from utilizing historical data, providing insights for decision-making and operational efficiency. This study develops a soft-sensor-based approach leveraging multi-fidelity modeling to correct discrepancies between online sensors and laboratory analyses. A Gaussian process-based strategy is used to predict deviations between high-frequency low-fidelity sensor data and less frequent high-fidelity laboratory measurements. By exploring static and dynamic modeling frameworks, we assess their suitability for capturing process dynamics and addressing time-dependent variability. The multi-fidelity soft sensor noticeably improves predictive accuracy, outperforming high-fidelity and low-fidelity methods. This approach demonstrates applicability across various industrial settings where integrating diverse data sources enhances real-time process control and monitoring, reducing reliance on costly laboratory sampling.

In process systems engineering (PSE), obtaining accurate process models for optimization can be expensive and time-consuming. Black-box Bayesian Optimization (BO) with Gaussian process (GP) surrogates offers a promising approach. However, full black-box optimization neglects valuable prior knowledge, which could otherwise improve the optimization process. This work explores methods of integrating prior knowledge in the form of low-fidelity data into BO by evaluating these methods on synthetic multi-fidelity test functions. Our results highlight possibilities for improved convergence of the BO optimization. However, our work further highlights potential pitfalls of these multi-fidelity models, such as bias, convergence to local optima, and overfitting on low-fidelity data. Hence, leveraging low-fidelity data in multi-fidelity models can improve BO convergence, but there are instances where the algorithms are more susceptible to failure.

The definition of strategies for operation of process networks is a key research focus in process systems engineering. This challenge is commonly formulated as a numerical constraint satisfaction problem, where most practical algorithms are limited to identifying inner approximations to the feasible operational envelope. Sampling-based approaches so far have only been developed for formulations that required coordinated operation of the units within the network. We propose a decomposition approach that enables decentralized operation for acyclic muti-unit processes by sampling. Our methodology leverages problem structure to decompose unit-wise and deploys surrogate models to couple the resultant subproblems. We demonstrate it on a serial, batch chemical reactor network. In future research, we will extend this framework to consider the presence of uncertain unit parameters robustly.

The dividing wall column (DWC) has significant energy saving potential compared to conventional column sequences. However, to reach these savings in practice, it is essential that the control structures can track the optimal operation point despite inevitable changes in feed properties, performance characteristics and other uncertainties. Otherwise, the energy consumption may rise significantly or, more commonly, the DWC becomes unable to produce pure products even at its maximum reboiler duty. Extremum seeking control (ESC) is a model-free optimisation technique that may mitigate off-optimal operation in this environment. By active perturbation of selected manipulative variables, the algorithm infers gradient properties of the measured cost function and, by that, enables tracking of a moving optimum. Extremum seeking control can be used also in combination with other approaches, e.g. self-optimising control. Applied to the DWC, the presented perturb-and-observe algorithm, which can be... [more]

In chemical process control, decision-making often involves balancing multiple conflicting objectives, such as maximizing production, minimizing energy consumption, and ensuring process safety. Traditional approaches for multi-objective optimization, such as linear programming and evolutionary algorithms, have proven effective but struggle to adapt in real-time to the dynamic and nonlinear nature of chemical processes. In this paper, we propose a framework that combines Reinforcement Learning (RL) with Multi-Objective Evolutionary Algorithms (MOEAs) to address these challenges. Specifically, we utilize MOEAs, such as NSGA-II, to optimize the parameter space of policy neural networks, resulting in a Pareto front of policies. This Pareto front provides a diverse set of policies that enable operators to dynamically switch control strategies based on real-time system conditions and prioritized objectives. Our proposed methodology is applied to a Controlled Stirred Tank Reactor (CSTR) case... [more]

Anthropogenic carbon-dioxide emissions, exceeding 51 billion tons annually, are a major driver of global climate impacts. Aqueous amine scrubbing offers an effective carbon-capture solution, but the energy-intensive thermal regeneration step of the process significantly increases costs, limiting large-scale adoption. To address these challenges, computational optimization of process and molecular design is promising but often too resource-intensive, emphasizing the need for efficient surrogate models. Specifically, we develop a surrogate model based on an artificial neural network (ANN) that is employed to replace rigorous phase-equilibrium computations performed with the SAFT-? Mie group contribution method within a steady-state aqueous amine carbon-capture process model. Our ANN is trained on 32,768 vapour–liquid equilibrium data points of a quaternary mixture of water, monoethanolamine, carbon dioxide, and nitrogen over industrially relevant temperature, pressure, and composition ra... [more]

Process optimisation and quality control are crucial in process industries for minimising product waste and improving plant economics. Identifying robust operational regions that ensure both product quality and performance is particularly valued in industries. However, this task is complicated by operational uncertainties, which can lead to violations of product quality constraints and significant batch discards. We propose a novel robust optimisation strategy that integrates advanced machine learning and process systems engineering to systematically identify optimal operational regions under uncertainty. Our approach begins by using a process model to screen a broad operational space across various uncertainty scenarios, pinpointing promising control trajectories to satisfy process constraints and product quality. Machine learning is then employed to cluster these trajectories into sub-regions. Finally, a two-layer dynamic optimisation framework is employed to determine the optimal co... [more]

Asserting the phase stability of multi-component mixtures is an important task in computer-aided mixture/blend design (CAMbD), but it is often hindered by the lack of reliable and tractable models. In this paper, we propose a group-contribution machine-learning (GC-ML) method to predict phase coexistence for a large set of ternary mixtures consisting of two solvents and one (fixed) solute. Each solvent is represented by a vector of functional group numbers, encoded by integer values. The solvent vectors are combined with mixture composition and temperature to form the input features to a GC-ML surrogate classifier, which distinguishes between four types of stable phase configurations as possible outputs: liquid (L), solid-liquid (SL), liquid-liquid (LL) or solid-liquid-liquid (SLL). To explore the performance of the trained GC-ML multi-classifier, it is embedded as a surrogate phase-stability constraint in the optimisation of an ibuprofen crystallisation process. A two-step solution s... [more]

Recent advancements in machine learning and time series analysis have opened new avenues for improving predictive control in complex systems such as mineral flotation. Techniques leveraging multivariate predictive control in mineral flotation have seen significant progress in recent years. However, challenges in developing an accurate dynamic model that encapsulates both the pulp and froth phases have hindered further advancements. Now, with a readily available model containing equations that describe the physics of flotation froths, an opportunity for novel control strategies presents itself. In this study, a Gaussian Process (GP) Model Predictive Control (MPC) strategy is proposed to integrate uncertainty quantification directly into the control framework. By leveraging the probabilistic nature of GP models, this approach captures process variability and adapts dynamically to new data, ensuring continuous refinement of the GP model within the MPC strategy. Unlike previous implementat... [more]

This study focuses on optimizing production scheduling in multi-product plants with shared resources and costly changeover operations. Specifically, two main challenges are addressed, the unknown changeover behavior of new products and the need for rapid schedule generation after unforeseen events. An innovative framework integrating Machine Learning (ML) techniques with Mixed-Integer Linear Programming (MILP) is proposed for single-stage production processes. Initially, a regression model predicts unknown changeover times based on key product attributes. Then, a representation where distances correlate with changeover times is compiled through multidimensional scaling, allowing constrained clustering to group production orders according to available packing lines. Ultimately, the MILP model generates the production schedule within a constrained solution space, utilizing optimal product-to-line allocation from cluster segmentation. A case study inspired by a Greek construction material... [more]

Predicting the critical micelle concentration (CMC) of surfactants is essential for optimizing their applications in various industries, including pharmaceuticals, detergents, and emulsions. In this study, we investigate the performance of graph-based machine learning models, specifically Graph Convolutional Networks (GCN), Graph Attention Networks (GAT), and a graph-transformer model, PharmHGT, for predicting CMC values. We aim to determine the most effective model for capturing the structural and physicochemical properties of surfactants. Our results provide insights into the relative strengths of each approach, highlighting the potential advantages of transformer-based architectures like PharmHGT in handling molecular graph representations compared to traditional graph neural networks. This comparative study serves as a step towards enhancing the accuracy of CMC predictions, contributing to the efficient design of surfactants for targeted applications.

Cell culture processes play a central role in the production of various therapeutic compounds. These processes are multiscale and highly complex, making them challenging to describe comprehensively using fully mechanistic models. In this study, we employ an integrated hybrid machine learning and first principles model to predict the viable cell density, product titer, and metabolite concentration profiles. We employ the concept of degree of hybridization, where we create a family of hybrid models each with increasing degree of process knowledge. Predictions from the feasible hybrid architecture were integrated with a genome scale metabolic model to evaluate the flux distribution of reactions related to the central carbon metabolism of the cell throughout the process duration. We demonstrate that the current approach not only reasonably predicts the bioprocess profile but also provides biologically relevant information that can uncover dynamics of intracellular metabolism which can open... [more]

An essential problem in precision agriculture is to accurately model and predict root-zone (top 1 m of soil) soil moisture profile given soil properties and precipitation and evapotranspiration information. This is typically achieved by solving agro-hydrological models. Nowadays, most of these models are based on the standard Richards equation (RE), a highly nonlinear, degenerate elliptic-parabolic partial differential equation that describes irrigation, precipitation, evapotranspiration, runoff, and drainage through soils. Recently, the standard RE has been generalized to time-fractional RE with any fractional order between 0 and 2. Such generalization allows the characterization of anomalous soil exhibiting non-Boltzmann behavior due to the presence of preferential flow. In this work, we focus on inverse modeling of time-fractional RE; that is, how to accurately estimate the fractional order and soil property parameters of the fractional RE given soil moisture content measurements. S... [more]

Pressure-swing distillation (PSD) is a frequently applied method to separate pressure-sensitive azeotropic mixtures; however, its energy demand is very high. In continuous mode, PSD is performed in a system consisting of a high- and a low-pressure column. If the composition of the feed is between the azeotropic compositions at the two pressures, it can be introduced into any of the columns, leading to two possible column sequences. Depending on the feed composition, one of the sequences is optimal whether in terms of energy demand or total annual cost (TAC). In the present work, surrogate model-based optimization is applied to determine the optimal TAC value as a function of the feed composition between the azeotropic ones. As a first step, the column sequence with feeding into the high-pressure column is studied here. The mixture to be separated consists of water and ethylenediamine, which form a maximum-boiling azeotrope. The columns are modeled separately and a large number of simul... [more]

Nucleophilic aromatic substitutions (SNAr) are key chemical transformations in pharmaceutical and agrochemical synthesis, yet their complex mechanisms (concerted or two-step) complicate kinetic model identification. Accurate kinetic models for SNAr are essential for scale-up, optimization, and control of the reaction process, but conventional methods struggle with mechanism uncertainty driven by substrates, nucleophiles, and reaction conditions, with data collection being difficult due to its source-intensive nature. We address this using DoE-SINDy, a data-driven framework for generative modelling without complete theoretical understanding. A benchmark study on the SNAr reaction of 2,4-difluoronitrobenzene with morpholine in ethanol was conducted, incorporating parallel and consecutive side-product formation. Ground-truth kinetic models validated in prior studies were used to generate in-silico data under varying noise levels and sampling intervals. DoE-SINDy successfully identified th... [more]

Fluid catalytic cracking is a crucial process in the refining industry, capable of converting lower-quality feedstocks into higher-value products. Due to the variability in feedstock properties and fluctuations in product market prices, timely adjustment and optimization of the FCC unit are essential. In this context, data-driven models have garnered increasing attention for their capacity to handle the complex, nonlinear reactions involved in the FCC process. However, on account of the limited operating range of the plants and the black-box nature of data-driven models, relying solely on these models for optimization may lead to contradictory decisions in optimization processes. To address these challenges, we integrate gradient information of product yields with respect to key variables derived from the mechanistic model Petro-SIM, into the training process of data-driven models. To mitigate the high computational demands of the Petro-SIM model, we propose the use of active learning... [more]