This paper introduces a prototype for constructing an edge AI system utilizing the contemporary Machine Learning Operations (MLOps) concept. By employing microcontrollers such as the Raspberry Pi as hardware, our methodology includes data scrubbing and machine learning model deployment on edge devices. Crucially, the MLOps pipeline is fully developed within the ecoKI platform, a research platform for ML/AI applications. In this study, we thoroughly investigate the performance of our ecoKI platform by comparing it with the established Edge Impulse platform. We deployed the ML model with different weight quantization methods, such as FP32 and INT8, to compare accuracy variations and inference speed between these two platforms and quantization strategies on edge devices. In our experiments, we identified that the average accuracy performance of the ecoKI platform is 3.61% better than the edge impulse. Moreover, real-time AI processing on edge devices enables microcontrollers, even those w... [more]

Deep learning faces significant challenges in efficiently training large-scale models. These issues are closely linked, as efficient training often depends on precise and computationally feasible gradient calculations. This work introduces innovative methodologies to improve deep learning network (DLN) training in complex systems. A novel approach to DLN training is proposed by adapting the block coordinate descent (BCD) method, which optimizes individual layers sequentially. This is combined with traditional batch-based training to create a hybrid method that harnesses the strengths of both techniques. Additionally, the study explores Iterated Control Random Search (ICRS) for initializing parameters and applies quasi-Newton methods like L-BFGS with restricted iterations to enhance optimization. By tackling DLN training efficiency, this contribution offers a comprehensive framework to address key challenges in modern machine learning. The proposed methods improve scalability and effect... [more]

Catalysis design and reaction condition optimization are considered the heart of many chemical and petrochemical processes and industries; however, there are still significant challenges in these fields. Advances in machine learning (ML) have provided researchers with new tools to address some of these obstacles, offering the ability to predict catalyst behaviour, optimal reaction conditions, and product distributions without the need for extensive laboratory experimentation. In this contribution, the potential applications of ML in heterogeneous catalysis and photocatalysis are explored by analysing datasets from different reactions, including Fischer-Tropsch synthesis and photocatalytic pollutant degradation. First, datasets were collected from literature. After cleaning and preparing the datasets, they were employed to train and test several models. The best model for each dataset was selected and applied for optimization.

This research addresses the challenges associated with data-driven soft sensors in industrial applications, where successful implementations remain limited. The scarcity of practical applications can be attributed to variable operating conditions and frequent disturbances in real-time processes. Industrial data are often nonlinear, dynamic, and highly unbalanced, complicating efforts to capture the essential characteristics of underlying processes. To tackle these issues, we propose a comprehensive solution for industrial application, that encompasses feature selection, feature extraction, and model updating. Feature selection aims to pinpoint the independent variables that have a substantial impact on key performance indicators, including quality, safety, efficiency, reliability, and sustainability. By doing so, it simplifies the model and boosts its predictive accuracy. The process begins with screening variables based on process knowledge, followed by a thorough analysis of correlat... [more]

The efficient separation of carbon dioxide (CO2) and methane (CH4) is crucial for chemical processes, including biogas upgrading and natural gas purification. Metal-organic frameworks (MOFs) have gained significant attention as promising adsorbents for these processes due to their high porosity and tunable structures. Estimating the adsorption capacity of MOFs is essential for screening high performing adsorbents. While molecular simulations are commonly used to estimate the adsorption capacities, their computational intensity acts as a bottleneck in screening MOF adsorbents. In this study, we propose a machine learning (ML)-based framework for the high-throughput prediction of adsorption isotherms for CO2 and CH4 in MOFs. A graph neural network (GNN) model was developed to predict adsorption capacities, effectively replacing the time-consuming molecular simulations. The GNN model processes the structural graphs of MOFs, capturing their spatial configurations, such as surface structure... [more]

Compartmental models are widely used to simplify the analysis of complex fluid dynamics systems, yet subjective compartment definitions and computational constraints often limit their applicability. The CompArt algorithm introduces an AI-driven framework that automates compartmentalization in Computational Fluid Dynamics (CFD) simulations, optimizing both accuracy and efficiency. By leveraging unsupervised clustering techniques such as Agglomerative Clustering, CompArt identifies coherent flow regions based on velocity and turbulent kinetic energy dissipation rate, ensuring a data-driven, physically consistent segmentation. The methodology integrates a connectivity-based clustering strategy, where compartments are dynamically optimized using the Silhouette score and adjacency matrix. This approach enables the reduction of high-resolution 3D CFD simulations into a network of interconnected sub-systems, significantly lowering computational costs while preserving system heterogeneity. The... [more]

As industries embrace the digitalization of Industry 4.0, the abundance of process data creates new opportunities to optimize industrial control systems. Traditional Proportional-Integral-Derivative (PID) controllers often require manual tuning to address changing conditions. This paper introduces an automated, adaptive PID tuning method using historical data and machine learning for a continuously evolving, data-driven approach. The method centers on training a surrogate model using historical process data to replicate real system behavior under various conditions. This enables safe exploration of control strategies without disrupting live operations. An RL (Reinforcement Learning) agent interacts with the surrogate model to learn optimal control policies, dynamically responding to the plant's state, defined by variables like operational conditions and measured disturbances. The agent adjusts PID parameters in real-time, optimizing metrics such as stability, response time, and energy... [more]

Physics-Informed Symbolic Regression (PISR) offers a pathway to discover human-interpretable solutions to partial differential equations (PDEs). This work investigates three fitness metrics within a PISR framework: PDE fitness, Bayesian Information Criterion (BIC), and a fitness metric proportional to the probability of a model given the data. Through experiments with Laplace’s equation, Burgers’ equation, and a nonlinear wave equation, we demonstrate that incorporating information theoretic criteria like BIC can yield higher fidelity models while maintaining interpretability. Our results show that BIC-based PISR achieved the best performance, identifying an exact solution to Laplace’s equation and finding solutions with R2-values of 0.998 for Burgers’ equation and 0.957 for the nonlinear wave equation. The inclusion of the Bayes D-optimality criterion in estimating model probability strongly constrained solution complexity, limiting models to 3-4 parameters and reducing accuracy. Thes... [more]

Balancing economic and environmental goals in industrial applications is critical amid challenges like climate change. Multi-objective optimization (MOO) and multi-criteria decision analysis (MCDA) are key tools for addressing conflicting objectives. MOO generates viable solutions, while MCDA selects the optimal option based on key performance indicators such as profitability, environmental impact, safety, and efficiency. However, large datasets pose a challenge in selecting the preferred solution during the MCDA process This study introduces a novel machine learning-enhanced MCDA framework and applies the method to analyze decarbonization solutions for a European refinery. A stage-wise dimensionality reduction method, combining AutoEncoders and Principal Component Analysis (PCA), is applied to simplify high-dimensional datasets while preserving key spatial features. Geometric analysis techniques, including Intrinsic Shape Signatures (ISS), are employed to refine the identification of... [more]

Operational decision-making in Process Systems Engineering (PSE) has achieved high proficiency at specific levels, such as supply chain optimization and unit-operation optimization. However, a critical challenge remains: integrating these layers of optimization into a cohesive, hierarchical decision-making framework that enables sustainable and automated operations. Addressing this challenge requires systems capable of coordinating multi-level decisions while maintaining interpretability and adaptability. Multi-agent frameworks based on Large Language Models (LLMs) have demonstrated significant promise in other domains, successfully simulating traditional human decision-making tasks and tackling complex, multi-stage problems. This paper explores their potential application within operational decision-making for PSE, focusing on sustainability-driven objectives. A realistic Gas-Oil Separation Plant (GOSP) network is used as a case study, mimicking a hierarchical workflow that spans from... [more]

This work presents a model for optimizing shell-and-tube heat exchanger design using Q-learning, a reinforcement learning technique. An agent is trained to interact with a simulated environment of a heat exchange model, iteratively refining design configurations to maximize a reward function. This reward function balances heat exchanger effectiveness and pressure drop, emphasizing designs that minimize pressure drop. Results showed that simpler configurations consistently achieved higher rewards, despite complex designs offering better heat transfer efficiency.

The accumulation of heavy polynuclear aromatics (HPNAs) in hydrocracking units (HCUs) poses significant challenges to catalyst performance and process efficiency. This study proposes an integrated machine learning framework that combines ridge regression, K-means, and long short-term memory (LSTM) neural networks to predict HPNA formation, enabling proactive process management. For the training phase, weighted average bed temperature (WABT), catalyst deactivation phase—clustered using unsupervised K-means clustering—and hydrocracker feed (HCU feed) parameters obtained from laboratory analyses are utilized to capture the complex nonlinear relationships influencing HPNA formation. In the simulation phase, forecasted WABT values are generated using a ridge regression model, and future HCU feed changes are derived from planned crude oil blend data provided by the planning department. These forecasted WABT values, predicted catalyst deactivation phases, and anticipated HCU feed parameters s... [more]

Maintenance is critical for industrial plants to ensure operational reliability and worker safety. In process industries, fouling, the accumulation of solid residues in equipment, poses a significant challenge, causing inefficiencies and productivity losses. Effective modeling of fouling evolution over time is essential for maintenance planning to prevent equipment from operating under suboptimal conditions. Traditional approaches to fouling prediction include equation-based models, which offer high precision but may struggle with continuously changing process boundaries, and machine learning techniques, which are more adaptable but less effective at capturing rapidly evolving trends driven by complex underlying physics. This study introduces an innovative hybrid machine learning approach for predictive maintenance, combining the strengths of both methods. Pressure differential is modeled using an equation-based approach that links pressure data with fouling thickness, while the foulin... [more]

Piping and Instrumentation Diagrams (P&IDs) play a crucial role in the chemical industry, yet they are often stored as scanned images or computer-aided design (CAD) drawings, limiting their seamless integration into modern digital workflows. Consequently, the task of automating P&ID digitization has attracted significant attention from the computer-aided design (CAD) research community. Traditional approaches, which typically rely on conventional object detection techniques, often demand extensive manual annotations to accurately identify and classify the various symbols in P&IDs. In this paper, we proposed a novel framework for automating the recognition of P&IDs. Our method first extracts key features of geometric primitives in CAD drawings through an automated process. Subsequently, a Transformer-based model is employed to predict the layer assignment of these primitives. Following this, an unsupervised clustering method, guided by predefined rules and geometric distances, is applie... [more]

Photochemical monolith reactors offer advantages over microreactors by providing high mixing efficiency and surface area to volume ratio while being scalable. However, optimizing monolith design parameters like channel number, shape, and stacking is critical to maximizing light usage and reactor efficiency. This work proposes using Bayesian optimization and COMSOL Multiphysics simulations to automatically design translucent monoliths for photochemical reactions. The goal is to maximize both photochemical space-time yield and space-time yield. Ray tracing simulations were performed while evaluating five different channel geometries (circular, elliptical, triangular, square, and pentagonal) and optimizing parameters, including channel diameter, vertical stacking, shape rotation, and ellipse axis ratio. Results showed a clear trade-off between Space-Time Yield (STY) and Photochemical Space-Time Yield (PSTY), with optimized elliptical channels achieving up to 15.3% improvement in STY with... [more]

The analysis of large chemical reaction space sheds light on reaction patterns between molecules and can inform subsequent reaction pathway planning. With the aim to discover more sustainable production systems, it became worthwhile to explicitly model the reaction space reachable from biobased feedstocks. In particular, the space that reactions in integrated biorefineries span for optimised biorefinery planning is of interest. In this work we show a network-theoretical analysis of biorefinery reaction data. We utilise the directed all-to-all mapping between reactants and products to compare the reaction space obtained from biorefineries with the entire network of organic chemistry (NOC). In our results, we find that despite having 1000 times fewer molecules, the constructed network resembles the NOC in terms of its scale-free nature and shares similarities regarding its small-world property. Additionally, we analyse the coverage rate of the biorefinery reaction data and find that many... [more]

Modeling process systems by use of partial differential equations is often complex and computationally expensive, especially for inverse problems such as optimization, state identification, or parameter estimation. Data-driven methods typically provide efficient alternatives with lower computational cost. One such method is Graph Neural Networks (GNNs), which can be used to model dynamical systems as graphs. However, dynamic GNNs often face challenges with extrapolation and representability. Integrating mechanistic insights in surrogate models can improve both prediction accuracy and interpretability. This study compares different strategies for embedding physics-based insights into GNNs to model the dynamic behavior of a catalytic CO2 methanation reactor. The hybrid integration of physics-informed GNNs aims to improve the predictive ability and interpretability while reducing the model development time, thereby facilitating faster deployment. Results show that penalizing the predicted... [more]

Process engineers in the Chemical and Oil & Gas industries extensively use process simulation for the design, development, analysis, and optimization of complex systems. This study investigates the integration of Artificial Intelligence (AI) with AVEVATM Process Simulation (APS), a next-generation commercial simulation tool. We propose a framework for a custom chatbot application designed to assist engineers in developing and troubleshooting simulations. This chatbot application utilizes a custom-trained model to transform engineer prompts into standardized queries, facilitating access to essential information from APS. The chatbot extracts critical data regarding solvers and thermodynamic models directly from APS to help engineers develop and troubleshoot process simulations. Furthermore, we compare the performance of our custom model against OpenAI technology. Our findings indicate that this integration significantly enhances the usability of process simulation tools, promoting more... [more]

Understanding data with complex patterns is a significant part of the journey toward accurate data prediction and interpretation. The relationships between input and output variables can unlock diverse advancement opportunities across various processes. However, most AI models attempting to uncover these patterns are not explainable or remain opaque, offering little interpretation. This paper explores an approach in explainable AI by introducing a multi-agent system (MaSR) for extracting equations between features using data. We developed a novel approach to perform symbolic regression by discovering mathematical functions using a multi-agent system of LLMs. This system addresses the traditional challenges of genetic optimization, such as random seed generation, complexity, and the explainability of the final equation. We utilize the in-context learning capabilities of LLMs trained on vast amounts of data to generate accurate equations more quickly. This study presents research on expa... [more]

The industrialization of catalytic processes requires reliable kinetic models for design, optimization, and control. While white box models are preferred for their interpretability, they demand considerable time and expertise for their construction. This research enhances the ADoK-S framework by embedding prior expert knowledge using mathematical constraints and integrating uncertainty quantification. The improved methodology consists of: (I) a genetic programming algorithm with constraints to produce physically coherent candidate models, (II) a sequential optimization algorithm for parameter estimation, (III) model selection based on the Akaike information criterion (AIC), and (IV) uncertainty quantification of the chosen model’s predictions. The refined approach not only requires less data for discovering kinetic models but also ensures physically sound proposals. With the inclusion of uncertainty quantification, the method bolsters prediction reliability, and aids in safer system de... [more]

The increasing complexity of industrial biotechnology demands advanced modeling techniques capable of capturing the intricate dynamics of bioreactors. Traditional regression-based and empirical methods often fall short when confronted with the highly nonlinear behavior and limited experimental data characteristic of bioprocesses. Addressing these challenges requires a more intelligent approach—one that leverages domain knowledge to model complex bioprocess dynamics effectively, even with sparse data, while maintaining interpretability and robustness. In this study, we introduce a process-informed, data-driven methodology for modeling the dynamics of industrial bioreactors, leveraging the capabilities of the rising field of Scientific Machine Learning (SciML). Our approach leverages Physics-Informed Neural Networks (PINNs) to seamlessly integrate domain knowledge encoded in physical laws with sparse experimental data and deep learning techniques, enabling precise simulation and modeling... [more]

The rapid advancement of computational drug discovery has enabled the generation of vast virtual libraries of promising drug candidates. However, evaluating the synthetic accessibility (SA) of these compounds remains a critical bottleneck. While computer-aided synthesis planning (CASP) tools can provide synthesis routes to the candidate, their computational demands make them impractical for large-scale screening. Existing rapid SA scoring methods, struggle to generalize to out-of-distribution molecules and do not account for economic viability. To address these challenges, we present MolPrice, an accurate and reliable price prediction tool. By introducing a novel self-supervised learning approach, MolPrice achieves robust generalization to diverse molecular structures of various complexities. Our comprehensive analysis of model architectures and molecular representations reveals that substructure-based features strongly correlate with market prices, supporting the relationship between... [more]

Accurate yet interpretable prediction of Global Warming Potential (GWP) is essential for the sustainable design of chemical products and processes. However, existing studies that utilize molecular structure and physicochemical properties for GWP prediction often suffer from low interpretability, relying on black-box models that obscure the underlying relationships between molecular descriptors and environmental impact. To address this limitation, this study employs a Kolmogorov–Arnold Network (KAN) to derive symbolic equations that establish explicit relationships between molecular properties and GWP. By extracting interpretable mathematical expressions, our approach provides a transparent foundation for decision-making in chemical processes and reaction development. Our comparative analysis of machine learning models—including Random Forest, XGBoost, Deep Neural Networks (DNN), and KAN—reveals that Mordred descriptors outperform MACCS keys in GWP prediction, emphasizing the importance... [more]

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]

Accurate mechanistic models provide valuable physical insight and are crucial for efficient process scale-up and optimisation, but their identification requires lengthy experimental data collection, model construction, validation and discrimination. Traditional black-box machine learning transfer methods leverage prior knowledge but lack interpretability and physical insights. To address this, we propose a novel approach using artificial neural network feature attribution to automatically locate corrections and symbolic regression to make structural modifications to an inaccurate or low-fidelity mechanistic model. In a comprehensive in-silico case study, the framework adapted a kinetic model from one biochemical system to a different but related one, enhancing predictive accuracy. Integrated within an iterative model-based design of experiments routine, it minimised the number of new experiments required. The study also discusses the impact of the inductive bias trade-off and alternati... [more]