Periodic Reporting for period 1 - PROSAFE (Smart Integration of Process Systems Engineering & Machine Learning for Improved Process Safety in Process Industries)
Période du rapport: 2024-01-01 au 2025-12-31
The process industries – chemical, petrochemical, pharmaceutical, and energy – are cornerstones of our modern society. However, they operate under conditions that present inherent safety risks. We're seeing increasing plant complexity, aging infrastructure, and the potential for subtle, hard-to-detect failures that can cascade. Furthermore, human operators often face information overload. Simultaneously, the Industry 4.0 revolution provides an unprecedented opportunity. We now have access to vast amounts of data from sensors and control systems, coupled with significant advances in computational power. This sets the stage for AI and Machine Learning to provide new, predictive insights that can transform safety management. ProSafe is a strategic response to these challenges and opportunities. At its heart, ProSafe is a collaborative research and doctoral training program, bringing together expertise in Process Safety, Process Systems Engineering, and Machine Learning. Our vision is to improve safety in the process industries. We aim to do this by creating data-driven methodologies, crucially informed by process domain knowledge, that allow for real-time risk assessment and better operational decision support. Our mission is twofold: firstly, to train 12 doctoral candidates who will become the future leaders in this interdisciplinary field. Secondly, to foster strong, synergistic collaborations between top universities and key industrial players across Europe, ensuring our research is relevant and impactful. The ProSafe research program is meticulously structured to tackle key challenges in process safety. At the foundation, we have three core methodological Work Packages
WP2: This focuses on 'Model-based foundations for improved risk assessment and process safety.' Here, we leverage Process Systems Engineering principles to build robust models that help us understand and quantify risks.
WP3: This is dedicated to 'Artificial Intelligence and Machine Learning for risk monitoring and safe process operation.' This WP explores how AI can analyze vast amounts of data for early warnings, anomaly detection, and improved operational safety.
WP4: This crucial work package focuses on 'Hybrid approaches and tools integration.' The goal here is to synergistically combine the strengths of the model-based approaches from WP2 with the data-driven techniques from WP3, creating powerful new tools. Each DC contributes to the specific objectives of these WPs.
Our research is strongly application-driven, which is covered in WP5: 'Domain applications to selected high-hazard multisector process industries.' This ensures that the methodologies developed in WP2, 3, and 4 are relevant and tested in real-world contexts.
WP2: This focuses on 'Model-based foundations for improved risk assessment and process safety.' Here, we leverage Process Systems Engineering principles to build robust models that help us understand and quantify risks.
WP3: This is dedicated to 'Artificial Intelligence and Machine Learning for risk monitoring and safe process operation.' This WP explores how AI can analyze vast amounts of data for early warnings, anomaly detection, and improved operational safety.
WP4: This crucial work package focuses on 'Hybrid approaches and tools integration.' The goal here is to synergistically combine the strengths of the model-based approaches from WP2 with the data-driven techniques from WP3, creating powerful new tools. Each DC contributes to the specific objectives of these WPs.
Our research is strongly application-driven, which is covered in WP5: 'Domain applications to selected high-hazard multisector process industries.' This ensures that the methodologies developed in WP2, 3, and 4 are relevant and tested in real-world contexts.
The technical and scientific work in the ProSafe project was carried out in Work Packages 2–5, with the following activities and achievements delivered during the reporting period (M1–M24)
Work Package 2:
o Developed a comprehensive methodological workflow for global uncertainty and sensitivity analysis in Quantitative Risk Assessment (QRA).
o Selected and adapted open-source CFD tools (primarily OpenFOAM and FireFOAM) for loss-of-containment and escalation scenarios relevant to QRA.
o Developed an original CFD case study of a liquid hydrogen (LH2) tank exposed to fire (initial phase of a BLEVE).
Main achievements/outcomes: Standardised uncertainty workflow that reduces fluctuations in risk estimates; coherent CFD framework with new benchmarking case; clear guidance on when high-fidelity CFD is justified versus simplified approaches in QRA. These outputs are fully documented and ready for integration into industrial QRA practice.
Work Package 3:
o Developed and validated a Multi-Scale Convolutional Neural Network (MSCNN) framework with automated hyperparameter optimisation (Tree-structured Parzen Estimator) and ensemble learning for fault diagnosis and early warning.
o Established a hybrid semi-empirical + deep-neural-network model for proton exchange membrane fuel cells (PEMFCs) and electrolysers, enabling prediction of power density, efficiency, and hydrogen crossover with explicit uncertainty handling, subsequently embedded in a mixed-integer linear programming (MILP) optimisation.
Main achievements/outcomes: MSCNN framework with uncertainty quantification delivered and benchmark-validated; hybrid SD–ABM and PEMFC–MILP models providing interpretable, physics-constrained predictive analytics for online risk monitoring.
Work Package 4:
o Introduced a new dynamic benchmark simulation model of green ammonia production (Haber–Bosch coupled with fluctuating renewable hydrogen from water electrolysis), fully dimensioned for feasibility and control under variable supply.
o Developed a novel hybrid learning architecture based on q–Kronecker algebra and q-deformed Nonlinear Vector Autoregression (q–NVAR) for health-aware operation and maintenance.
o Implemented and tested integration of machine-learning long-term degradation models with short-term model predictive control (MPC), including robust MPC variants that incorporate safety-critical disturbances and faults while reducing computational complexity.
Main achievements/outcomes: Green ammonia benchmark model and q–NVAR framework delivered; risk-informed, degradation-aware control methods that simultaneously optimise short-term operation and long-term equipment health.
Work Package 5:
o Performed Round-1 validation of all tools from WP2–4 on four high-hazard industrial case studies: green hydrogen production, hydrocarbon production, Tennessee Eastman process, and green ammonia production.
o Mapped and integrated the developed methods (Monte Carlo uncertainty QRA, adapted CFD, MSCNN fault diagnosis, q–NVAR health-aware modelling, fault-tolerant MPC, green ammonia benchmark) into the case studies
Main achievements/outcomes: All novel methods successfully applied and undergoing iterative refinement on real-world process systems, demonstrating the integrated ProSafe hybrid modelling approach for risk monitoring,operation, and maintenance.
Work Package 2:
o Developed a comprehensive methodological workflow for global uncertainty and sensitivity analysis in Quantitative Risk Assessment (QRA).
o Selected and adapted open-source CFD tools (primarily OpenFOAM and FireFOAM) for loss-of-containment and escalation scenarios relevant to QRA.
o Developed an original CFD case study of a liquid hydrogen (LH2) tank exposed to fire (initial phase of a BLEVE).
Main achievements/outcomes: Standardised uncertainty workflow that reduces fluctuations in risk estimates; coherent CFD framework with new benchmarking case; clear guidance on when high-fidelity CFD is justified versus simplified approaches in QRA. These outputs are fully documented and ready for integration into industrial QRA practice.
Work Package 3:
o Developed and validated a Multi-Scale Convolutional Neural Network (MSCNN) framework with automated hyperparameter optimisation (Tree-structured Parzen Estimator) and ensemble learning for fault diagnosis and early warning.
o Established a hybrid semi-empirical + deep-neural-network model for proton exchange membrane fuel cells (PEMFCs) and electrolysers, enabling prediction of power density, efficiency, and hydrogen crossover with explicit uncertainty handling, subsequently embedded in a mixed-integer linear programming (MILP) optimisation.
Main achievements/outcomes: MSCNN framework with uncertainty quantification delivered and benchmark-validated; hybrid SD–ABM and PEMFC–MILP models providing interpretable, physics-constrained predictive analytics for online risk monitoring.
Work Package 4:
o Introduced a new dynamic benchmark simulation model of green ammonia production (Haber–Bosch coupled with fluctuating renewable hydrogen from water electrolysis), fully dimensioned for feasibility and control under variable supply.
o Developed a novel hybrid learning architecture based on q–Kronecker algebra and q-deformed Nonlinear Vector Autoregression (q–NVAR) for health-aware operation and maintenance.
o Implemented and tested integration of machine-learning long-term degradation models with short-term model predictive control (MPC), including robust MPC variants that incorporate safety-critical disturbances and faults while reducing computational complexity.
Main achievements/outcomes: Green ammonia benchmark model and q–NVAR framework delivered; risk-informed, degradation-aware control methods that simultaneously optimise short-term operation and long-term equipment health.
Work Package 5:
o Performed Round-1 validation of all tools from WP2–4 on four high-hazard industrial case studies: green hydrogen production, hydrocarbon production, Tennessee Eastman process, and green ammonia production.
o Mapped and integrated the developed methods (Monte Carlo uncertainty QRA, adapted CFD, MSCNN fault diagnosis, q–NVAR health-aware modelling, fault-tolerant MPC, green ammonia benchmark) into the case studies
Main achievements/outcomes: All novel methods successfully applied and undergoing iterative refinement on real-world process systems, demonstrating the integrated ProSafe hybrid modelling approach for risk monitoring,operation, and maintenance.
o Introduction of a standardised global uncertainty and sensitivity analysis workflow specifically tailored for QRA, enabling systematic identification of critical assumptions and reduction of risk-estimate fluctuations.
o Adaptation of open-source CFD tools (OpenFOAM/FireFOAM) with a new LH2 tank-fire BLEVE case study and explicit risk-assessment-compatible outputs, extending high-fidelity modelling to emerging low-carbon carriers (H2, NH₃, CH₃OH).
o Development of a Multi-Scale CNN framework with ensemble-based epistemic uncertainty quantification, achieving >87 % precision/recall on the Tennessee Eastman benchmark while providing operator confidence intervals.
o Novel q–Kronecker algebra and q–NVAR hybrid architecture that embeds nonextensive statistical mechanics directly into learning operators, enabling continuous interpolation between linear/non-linear regimes and degradation-aware prediction.
o Physics-informed hybrid PEMFC/electrolyser model integrated into MILP optimisation, allowing quantitative trade-off analysis between efficiency, durability, and safety indicators under uncertainty.
o New green ammonia benchmark model capturing fluctuating renewable hydrogen supply and its safety/control implications.
o Adaptation of open-source CFD tools (OpenFOAM/FireFOAM) with a new LH2 tank-fire BLEVE case study and explicit risk-assessment-compatible outputs, extending high-fidelity modelling to emerging low-carbon carriers (H2, NH₃, CH₃OH).
o Development of a Multi-Scale CNN framework with ensemble-based epistemic uncertainty quantification, achieving >87 % precision/recall on the Tennessee Eastman benchmark while providing operator confidence intervals.
o Novel q–Kronecker algebra and q–NVAR hybrid architecture that embeds nonextensive statistical mechanics directly into learning operators, enabling continuous interpolation between linear/non-linear regimes and degradation-aware prediction.
o Physics-informed hybrid PEMFC/electrolyser model integrated into MILP optimisation, allowing quantitative trade-off analysis between efficiency, durability, and safety indicators under uncertainty.
o New green ammonia benchmark model capturing fluctuating renewable hydrogen supply and its safety/control implications.