The accurate prediction of pure component properties has consistently been a critical issue in fields such as chemical engineering, biomedicine, and environmental science. In recent years, end-to-end deep learning methods have shown significant improvement over traditional machine learning approaches. This is due to their ability to automatically learn task-relevant representations from raw molecular data. In addition to accurate property prediction, researchers have increasingly focused on how specific fragment structures influence molecular properties. However, existing fragmentation methods based on predefined rules and group libraries struggle to capture novel molecular structures, which hampers the development of new materials and drugs. To address these challenges, this work proposes a hierarchical molecular fragmentation method. This method can automatically segment molecules into multiple fragments containing key functional groups. Then a three-branch graph attention network wa... [more]

Accurate life cycle assessment (LCA) depends on robust characterization factors (CFs), which quantify impacts such as ecotoxicity through the integration of fate (FF), exposure (XF), and effect (EF) factors. While databases such as USEtox and Ecoinvent provide essential CFs, significant data gaps remain, particularly in ecotoxicity endpoints like hazardous concentration 50% (HC_50), which directly inform effect factor calculations. Existing machine learning models can predict such values, but they often lack interpretability, which limits trust and transparency in environmental modeling. To address this, a machine learning framework is applied that utilizes symbolic regression (SR) and genetic programming (GP) to predict missing HC_50 values from physicochemical descriptors. A dataset with 14 descriptors was used to train SR models capable of generating interpretable mathematical expressions that link chemical properties to HC_50 values. SR models were benchmarked against prominent bla... [more]

Addressing the urgent transition to low-carbon energy systems requires microgrids capable of locally coordinating electricity generation, storage, and flexible consumption. Their efficient integration calls for control strategies that are scalable, privacy-preserving, and robust to uncertainty. To address such a challenging control problem, this work proposes a decentralised Multi-Agent Reinforcement Learning (MARL) approach based on the Cross-Entropy Method (CEM) for the coordination of prosumers, equipped with renewable generation and vehicle-to-grid capabilities. To improve sample efficiency and robustness, the policy is warm-started using Behaviour Cloning (BC) from a classical Proportional-Integral-Derivative (PID) controller, resulting in a hybrid BC-CEM framework. The proposed method is evaluated in a realistic microgrid simulation with stochastic demand and real weather and generation profiles. Results show that BC-CEM accelerates convergence and achieves lower energy costs com... [more]

Digitalization, artificial intelligence, and autonomous experimentation are reshaping chemical process development by enabling data-driven system identification and model-based optimization. Despite these advances, mechanistic models remain a cornerstone for predicting chemical reaction behavior and supporting optimization. However, purely mechanistic models often exhibit limited predictive accuracy when key phenomena affecting kinetics, mass and energy transfer are not fully captured. To address limitations on kinetic modelling, a hybrid modelling framework is proposed in this work that integrates a lumped power-law kinetic model with a Gaussian Process (GP) residual model to predict the reaction rate across the experimental design space while quantifying the uncertainty of the predicted rate. The hybrid model is then coupled with multi-objective Bayesian optimization (MOBO) by employing a weighted-sum approach and an upper confidence bound acquisition function to guide experimental d... [more]

Biological systems are governed by coupled interactions between intracellular metabolism and bioreactor operation that span multiple time scales. Constraint-based metabolic models are widely used to describe intracellular metabolism, but repeatedly solving the optimization problem at each time step in dynamic models introduces numerical challenges related to infeasibility and computational efficiency. This work presents a multi-scale modeling framework that integrates genome-scale, constraint-based metabolic models with dynamic bioreactor simulations. Intracellular metabolism is described using positive flux variables in a parsimonious flux balance analysis, and the resulting embedded optimization problem is replaced by a neural network surrogate. The surrogate provides a smooth approximation of the embedded optimization mapping and eliminates repeated linear program solves during simulation. The approach is demonstrated for fed-batch fermentation of Escherichia coli, in which the surr... [more]

The identification of predictive mathematical model structures (i.e. set of model equations) is essential for the development of digital twin models of chemical reactor systems. Recent work demonstrated the use of artificial neural networks (ANNs) for kinetic model recognition in a conceptual batch reaction experimental system. In practical chemical processes, however, system behaviour is governed not only by reaction kinetics but also by reactor hydrodynamics and system thermodynamics. While a very recent study incorporated hydrodynamic effects, this work integrates the three aspects: reaction kinetics, reactor hydrodynamics, and system thermodynamics, to develop a general reactor modelling recognition framework. The framework, which comprises three modules: 1) model generator module; 2) data generation module; and 3) ANN classifier module, was applied to a case study of benzoic acid esterification in a Taylor vortex flow reactor system. Analysing the framework's sensitivity, results... [more]

Activity coefficients are key thermodynamic quantities for describing phase equilibria, but their experimental determination entails laborious and costly phase-equilibrium measurements, making predictive approaches highly desirable. The potential of machine learning for such predictions has received growing attention as an alternative to physics-based models that require experimental data or expensive calculations for parameterization. We propose a physics-informed edge-enhanced graph attention network (PEGAT) to predict activity coefficients in multicomponent mixtures, where each molecule is encoded as a graph in which the nodes correspond to atoms and the edges to chemical bonds. The excess Gibbs free energy of the mixture is predicted using the proposed model, including a nonlinear transformation in the final layer to ensure that the excess Gibbs free energy vanishes for pure components. To further enforce thermodynamic consistency, the relevant activity coefficients are obtained vi... [more]

This work presents a data-driven surrogate modeling framework for predicting distillation behavior assuming an infinite number of stages and distillation limits informed by residue-curve topology and pinch-point feasibility analysis. The framework provides a direct mapping from feed composition and distillate-to-feed ratio (D/F) to distillate and bottom product compositions, making it suitable for flowsheet synthesis and optimization applications. The approach combines three components: a classifier that identifies feasible singular-point splits, a boundary regression model that predicts D/F limits separating pure- and mixed-product operating regimes, and a neural network that interpolates product compositions in the intermediate regime. The method is demonstrated for the ternary system ethanol, benzene, and water at 1 atm using data generated from rigorous vapor-liquid-liquid equilibrium analysis. Results show that the framework provides reliable predictions for pure splits while reta... [more]

Near-infrared (NIR) sorting systems in waste sorting plants operate under multiple settings, creating distinct input-output relationships that challenge predictive modeling. Conventional neural networks, such as multilayer perceptron (MLP), often suffer from catastrophic forgetting under continual training, limiting reliability across settings. This study evaluates Kolmogorov-Arnold Networks (KAN) for continual regression modeling of multi-setting NIR systems. KAN assign nonlinear transformations to network edges using localized spline grids, enabling structural isolation between input regions. We introduce controlled input-space manipulations (shifting successive settings to adjacent or non-overlapping grid regions) and compare KAN performance with MLPs of comparable parameter count. We also examine single-input versus multi-input configurations to assess dimensionality effects. Results show that KANs with sufficient input-space separation maintain previously learned knowledge with pe... [more]

Manufacturing kinetically stable products remains a challenge in the agrochemical industry. Current agrochemical formulation design relies on semi-empirical and trial-and-error methods. The inconsistency is caused by the lack of a mechanistic understanding of the formulation, making the design a black-box optimisation problem. In addition, validating the ground truth of the high-dimensional design space is expensive, driving chemists to explore possible solutions using data-driven methods. We proposed a Bayesian optimisation framework employing a Gaussian process as the surrogate model to intelligently guide the screening of the design space. The uniqueness of our framework is the application to the classification task to increase the number of hits of stable formulation recipes. The framework was tested on a provided industry dataset with a focus on emulsifiable concentrates. The performance reached a comparable accuracy with only ~25% of the data being sampled and hit more stable for... [more]

The availability of temperature-dependent physicochemical property data forms the cornerstone of process simulation, optimization, and sustainable molecular and product design. However, a critical data gap persists, as experimental measurements are accessible for only a small subset of known chemicals. This renders experimental characterization resource-prohibitive, often compelling reliance on empirical estimation methods. Moreover, although many models offer single-point predictions at fixed temperatures, accurately modeling continuous temperature-dependent behavior remains challenging. Conventional methods frequently overlook intermediate variations, resulting in limited extrapolation capability. To overcome these limitations, we introduce a mechanism-guided hybrid modeling framework that integrates physical insights into data-driven models. This framework is built on two strategies. Strategy I targets trend correction by generating a continuous representation from discrete single-p... [more]

A recurrent neural network (RNN) model for a CO2 methanation reactor was developed based on synthetic data generated from a validated mechanistic model of the same unit. The model was used to predict the main properties of the reactor - methane productivity and hotspot temperature - during a dynamic operation of the unit. The dynamic profile of feedstock availability was simulated taking into account the H2 flow that can be produced from PV-powered water electrolysis using solar irradiation profiles over a year in Milan, Italy. The dataset therefore consists of 366 data instances (one for each day), each composed of one datapoint per minute of sunlight. The best agreement between the predictions from the RNN and the target output values from the mechanistic model was found using a shallow RNN of 20 hidden-layer neurons, trained with a batch size of 10 and an 80/20 training-testing split. This showed that RNNs can constitute a reliable tool for dynamic surrogate modelling of energy conv... [more]

Bioprocess modelling, optimisation and scale-up are central components for improving sustainable manufacturing within pharmaceutical and chemical industries. However, developing accurate bioprocess digital twins remains a challenging process. Conventional mechanistic models are difficult to construct because of limited mechanistic understanding and large complexity of cellular metabolisms. While data-driven models have gained popularity, they often require large amounts of experimental data that is often time consuming to obtain and lack any quantitative description of the process. Hybrid modelling methods have emerged as promising alternatives however fail to provide physical insight to the root cause of model error. This work therefore presents a promising solution by developing a data-efficient symbolic regression (SR) based framework to enable the automated discovery of interpretable bioprocess models. A universal kinetic model backbone was used to capture overall process behaviour... [more]

In pharmaceutical process engineering, accurate prediction of tablet properties is crucial for ensuring product quality, optimizing manufacturing efficiency, and advancing sustainable production practices. This study presents a physics-informed neural network (PINN) framework for predicting the physical properties of pharmaceutical tablets from near-infrared (NIR) spectra. The PINN framework integrates revised Kubelka-Munk theory and physical constraints to ensure physically consistent predictions while requiring less training data than conventional artificial neural networks. Tablets were manufactured using acetaminophen and microcrystalline cellulose formulations with varying compositions and compression settings. The PINN framework successfully predicts critical quality attributes, including tensile strength, porosity, and density. It offers a data-efficient, interpretable solution for pharmaceutical tablet quality control.

Wine consumers are increasingly health- and environmentally conscious. At the same time, white wine and rosé drinkers favour freshness and varietal aromas, which requires low-temperature regimes that extend fermentation time and increase energy demand. Additionally, global warming accelerates grape ripening which increases alcohol level in wine. To reduce cost and alcohol levels while maintaining quality, predictive tools that forecast how fermentation conditions impact fermentation time, and primary and secondary metabolite concentrations, can provide practical benefits to wineries by expediting oenological decisions-making and in turn reducing energy demand. Additionally, literature highlights static models in smart manufacturing suffer from performance degradation with data drift. In light of this, we successfully developed and evaluated pipelines for the automated design and training of three ML methods - support vector regression, random forest and artificial neural networks - to... [more]

Plastic pyrolysis is widely promoted as a techno-economic industrial scale recycling strategy. Nevertheless, the fate and reactivity of plastic chemical additives during pyrolysis are mostly overlooked in product quality and environmental release assessments. Here, we present an integrated modeling framework to elucidate the role of additives in plastic pyrolysis and evaluate the implications of their transformation products and environmental releases. Using high-density polyethylene (HDPE) as a case study, chemical additives of concern are selected based on occurrence, concentration data, and potential risk to human health and the environment. Bond dissociation energies are predicted using a machine learning model to identify dominant radical species formed under pyrolytic conditions. These additive-derived radicals are incorporated into an automatic chemical reaction mechanism generator that constructs kinetic models composed of elementary chemical reaction steps. These kinetic model... [more]

This paper presents OpenAD-lib, an open-source Python framework for anaerobic digestion (AD) digital twins, unifying mechanistic models, machine learning (ML) surrogates, and model predictive control (MPC) within a modular ecosystem. OpenAD-lib addresses the critical fragmentation in AD digitalisation by bridging mechanistic and data-driven paradigms under explicit uncertainty. By integrating uncertainty-aware feedstock characterisation with robust process control, the platform enables the transition from isolated research tools to fully integrated digital twins, delivering economic and environmental value in AD systems.

This work presents a hybrid modelling approach for a downdraft sewage sludge gasifier within the Shit2Power (S2P) process. The gasifier is represented in CHEMCAD NXT by a series of four standard reactors that combine stoichiometric and equilibrium models with a data-driven correction step to account for deviations from ideal Gibbs equilibrium. Reaction conversions in the correction reactor are fitted to experimental synthesis gas compositions reported by Werle (2014) [7] for 30 operating points with varying equivalence ratios and reactor inlet air temperatures. The calibrated hybrid reactor model is evaluated against these data and shows conservative agreement for the combustible gas components of the synthesis gas. To overcome the limitations of linear interpolation between fitted operating points, several machine learning approaches are evaluated to predict the reaction conversions, and boosted neural networks are selected as a compromise between prediction accuracy and smooth behavi... [more]

Included in the CHEMCAD flowchart simulation model, this VBA source code extracts the inputs for the machine learning (ML) model from the simulation, passes them to the machine ML (Python), and sends the ML model outputs back to the simulation model. This code is not provided for direct use, but rather to demonstrate the methodology and as an aid for interested parties to develop their own solutions.

Anomaly detection is essential for keeping chemical plants safe and running efficiently. Although many deep-learning methods have been proposed, most are still tested mainly on synthetic benchmarks such as the Tennessee Eastman Process (TEP). While these simulators enable fair comparisons, they do not reflect the noise, complexity, and irregular fault behavior of real industrial plants. As a result, it remains unclear how well these models generalize in practice. In this work, we extend our earlier ESCAPE study and move beyond water systems to industrially relevant chemical processes. We analyze data from two continuously operated pilot plant scenarios at the Technical University of Munich: n-butanol/water heteroazeotropic distillation and poly(oxymethylene) ether purification. We published these datasets for the first time at NeurIPS 2025. In this work, 30 anomaly detection methods, including 26 deep-learning and 4 classical approaches, are benchmarked using the open-source TimeSeAD l... [more]

Supplementary material for the article "Utilizing Machine Learning for Phenomena-based Synthesis of Intensified Process Flowsheets", submitted to The 36th European Symposium on Computer Aided Process Engineering (ESCAPE 36). The document includes information about the heurstic and samplic logic rules used in generating the initial dataset, and the grid search results for hyperparamter optimization.

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]

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 bound-aries, 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 innova-tive 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 foul... [more]

Hybrid modelling is widely employed in chemical engineering to generate highly accurate predictions. Such an approach merges first-principle modelling with machine learning techniques to identify and model the epistemic uncertainty from experimental data. Despite its advantages, this still requires cross-domain competencies that are difficult to find in the chemical industry and high human involvement. The possibility of automating the identification and training model would be significantly beneficial for the widespread adoption of hybrid modelling methodology within the chemical industry. This work presents a novel algorithm for the automatic identification of hybrid models (HMs) starting from the first-principle representation of the system, described by differential equation sets. The methodology formulates the problem as mixed-integer programming, identifying the equation running under uncertainty, identifying the machine learning model hyperparameters, and training the latter. Th... [more]

Mechanistic population balance modeling (PBM) has advanced the design of pharmaceutical crystallization processes, enabling the production of active pharmaceutical ingredient (API) crystals with desired critical quality attributes (CQAs), such as purity and crystal size distribution. However, PBM development can sometimes be resource-intensive, requiring extensive design of experiments (DoE) and high-quality process data, making it impractical under fast-paced industrial development timelines. This study proposes a machine learning (ML)-based workflow for developing ‘fit-for-purpose’ digital twins of crystallization processes, leveraging industrially available DoE data to link operating conditions with CQAs. Validated on industrial data for a commercial API with complex crystallization challenges, the workflow efficiently identifies optimal operating conditions, demonstrating the potential of data-driven digital twins to accelerate the development of pharmaceutical processes.