Publications

Journal Paper

Large language model agent for user-friendly chemical process simulations

Jingkang Liang, Niklas Groll, Gürkan Sin

Abstract

Modern process simulators enable detailed process design, simulation, and optimization; however, constructing and interpreting simulations is time-consuming and requires expert knowledge. This limits early exploration by inexperienced users. To address this, a large language model (LLM) agent is integrated with AVEVA Process Simulation (APS) via Model Context Protocol (MCP), allowing natural language interaction with rigorous process simulations. An MCP server toolset enables the LLM to communicate programmatically with APS using Python, allowing it to execute complex simulation tasks from plain-language instructions. Three case studies assess the framework across different task complexities and interaction modes. The first shows the agent autonomously analyzing flowsheets, finding improvement opportunities, and iteratively optimizing, extracting data, and presenting results clearly. The framework benefits both novice users (for educational purposes), by translating technical concepts and demonstrating workflows, and experienced practitioners, by automating data extraction, speeding routine tasks, and supporting brainstorming. The next case study assesses autonomous flowsheet synthesis through both a step-by-step dialogue and a single prompt, demonstrating its potential for novices and experts alike. The step-by-step mode gives reliable, guided construction suitable for educational contexts; the single-prompt mode constructs fast baseline flowsheets for later refinement. Further, the agent’s answers proved stable across variations in LLM versions, prompting types, and process flowsheet complexities. While current limitations mainly involve the accurate interpretation of the physical results seen in minor reasoning mistakes, such as oversimplification or misleading suggestions, mean expert oversight is still needed. However, the framework’s capabilities in analysis, optimization, and guided construction suggest LLM-based agents can become valuable collaborators for process simulation tasks in future.

View PDF

Journal Paper

Applying Extreme Value Theory to a Century of Ammonium Nitrate Disasters: Persistent Safety Risks in Chemical Supply Chains

Gürkan Sin

Abstract

We have analyzed 39 major ammonium nitrate (AN) accidents (1916–2022) that reveal systemic catastrophic risks have not diminished over the past century. This challenges the prevailing narrative of improved process safety of AN on account of iterative guidelines and technology advances. Applying extreme value theory (EVT), we model temporal occurrence as a homogeneous Poisson process (maximum likelihood estimate λ ≈ 0.23 events/year) and casualty severity as heavy-tailed (generalized Pareto distribution, shape parameter ξ ≈ 1.13). These findings indicate that there are persistent systemic patterns. This is shown by the implied “shadow mean” that exceeds the sample mean by 1.5–19-fold, and societal risk profiles occupy the “intolerable” region per frequency–number curve criteria. To address this problem, we establish an EVT-based systemic monitoring framework with quantitative benchmarks, identifying 50-year (∼315 fatalities) and 100-year (∼715 fatalities) return levels as a data-driven baseline for assessing safety performance. This ensemble framework provides a necessary tool for stakeholders to monitor realized systemic risks over mechanism-driven simulations or narrative expectations.

View PDF

Journal Paper

On Neuro-Symbolic AI for Abnormal Event Detection in Process Safety

Julien Amblard, Alessandra Russo, Gürkan Sin

Abstract

In recent years, Artificial Intelligence (AI) techniques have led to numerous applications, from self-driving cars, to chatbots and crop monitoring. However, due to the high-risk environments in which they work, many safety-critical fields have fallen back on adopting these new methods. To be considered for use in such hazardous settings, an AI system would need to be capable of reasoning while fully taking into account relevant knowledge of the domain it is applied to, as well as being able to provide a clear explanation for each conclusion it makes. With this in mind, Neuro-Symbolic Learning — a hybrid AI approach combining Neural Networks with logic-based reasoning — shows much potential for use in Process Safety. This paper aims to provide the reader with a survey of the state-of-the-art Neuro-Symbolic approaches, with an emphasis on abnormal events. Additional topics of interest will be AI safety and Explainable AI. The paper concludes by setting the scene for future research focused specifically on abnormal event detection in chemical processes, suggesting novel frameworks that could be developed and used in real-time applications. By the end of this paper, the reader should have a rough idea of what Neuro-Symbolic Learning entails, and how it can be used to model complex systems in the context of Process Safety.

View PDF

Conference Paper

Identification of critical inputs in QRA studies using Monte Carlo-based sensitivity analysis

Mercedes Belda-Ley,  Gürkan Sin

Abstract

Quantitative Risk Assessment (QRA) is a well-established methodology used in different fields to identify, quantify and evaluate the risks associated with human and industrial activities. It provides a structured and extensive approach to calculate risk values, providing as a result the ability to identify major risk contributors and to assist with decision-making among others. Owed to its extensiveness and complexity, numerous decisions and assumptions are to be made throughout its execution. With the intention of harmonizing and facilitating the execution of QRAs, multiple guidelines and methodologies have been developed. However, these diverge depending on the country, region, and can further be translated differently by each risk analyst, aggravating the uncertainty in the estimation of QRA results. Consequently, QRA reliability has frequently been questioned, reiterating and analysing its – inherent – uncertainties and related implications (e.g., Rae et al., 2014). This scrutiny has led to proposing and developing diverse strategies for treating its uncertainty (e.g., Abrahamsson, 2002; Xu et al., 2023) as well as discussing the role of sensitivity analysis in QRA (e.g., Flage and Aven, 2009). In this context, the Monte Carlo (MC) methods provide a suitable framework for performing uncertainty analysis to complex problems (Sin and Espuña, 2020) where the description of the context, e.g., possible inputs, can be highly uncertain, as is the case for QRA (Abrahamsson, 2002; Li et al., 2022). Instinctively, these methods also serve as an effective framework for sensitivity analysis, since it is closely related to uncertainty analysis. MCbased sensitivity analysis has been applied to specific sections composing QRA in Pandya et al. (2012), however, the focus of this work was set on analysing the influence of model parameters in the calculated output. This study presents an initial framework for applying MC-based Global Sensitivity Analysis (GSA) to QRA with the goal of pinpointing the most critical input parameters driving uncertainty in risk estimates. The aim is to quantify the contribution of individual input uncertainties to the variance of the overall risk outputs, i.e., the impact that the assumptions and decisions made throughout the QRA studies may have in the calculated output.

View PDF

Conference Paper

ProSafe: Smart Integration of Process Systems Engineering & Machine Learning for Improved Process Safety

Gürkan Sin, Merlin Alvarado Morales, Eulàlia Planas, Elsa Pastor, Johannes Jäschke, Idelfonso Nogueira, Miguel Muñoz, Çan Erkey, Erdal Aydin, Alessandra Russo

Abstract

ProSafe is a novel interdisciplinary initiative (Figure 1) aiming to make a step change and reinforce the process safety effectiveness with new methods and skills exploiting emerging digital transformation opportunities (Big Data, ML, AI) in alignment with the EU digitalization roadmap of the European manufacturing industry initiative. Successful process safety enhancement in high-hazard industries calls for the development of new high-level process safety research. In this regard, the synergy between process safety, process systems engineering, machine learning, and artificial intelligence, which is usurping the new era of industry, has not yet been exploited. The underlying reasons are, among others, segregated research, and development efforts at different academic and non-academic centers in EU combined with the complexity of the problem, which makes it too big to tackle alone with a single discipline/research center. Thus, there is a strong need for a European doctoral training program bringing together complementary disciplines in research and training, which sets the motivation for ProSafe.

View PDF

Conference Paper

Enhancing Fault diagnosis for Chemical Processes via MSCNN with Hyperparameters Optimization

Jingkang Liang, Gürkan Sin

Abstract

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 demonstrated improved performance in terms of precision and recall, achieving 5.26% and 5.63% higher values, respectively, compared to the CNN model. Comparisons of the MSCNN with default hyperparameters further confirmed the effectiveness of these techniques in improving fault diagnosis performance. Additionally, model ensemble technique was explored to further enhance the performance of the model and provide uncertainty estimations. In conclusion, this approach offers a robust and reliable solution for fault diagnosis in the chemical industry, enhancing process safety and efficiency.

View PDF

Conference Paper

A Benchmark Simulation Model of Ammonia Production: Enabling Safe Innovation in the Emerging Renewable Hydrogen Economy

Niklas Groll,  Gürkan Sin

Abstract

The green transition accelerates innovations and developments targeting the integration of green hydrogen in the chemical industry. However, all new hydrogen pathways and process designs must be tested on operability and safety. A big challenge is the typical fluctuating characteristic of green hydrogen supply that contrasts the steady-state operation of most conventional chemical processes. Therefore, to adequately assess control and monitoring techniques, a benchmark model tailored to the relevant aspects of the hydrogen economy is required. We introduce a benchmark model based on the production of green ammonia using the Haber-Bosch process that remains operable when coupled to a fluctuating hydrogen supply from water electrolysis. The main section of the process model is an adiabatic indirect cooled reactor system that provides realistic modeling of industrial applications. Like the ammonia reactor, all process units and the underlying control structure are precisely dimensioned to ensure feasible operation across various hydrogen supply rates. Eventually, the flexible operating ammonia benchmark model can serve as a new benchmark for analyzing process safety and control aspects.

View PDF