LAB NOTICE
- [AI-HAZOP] Vision based HAZOP Safety checkup AI-Automation
- [AI-HAZOP] Vision based HAZOP Safety checkup AI-Automation Safer by Design: How AI agents are rethinking hazard analysis https://www.linkedin.com/safety/go/?url=https%3A%2F%2Flnkd%2Ein%2Fg7Ju7AN5&urlhash=uYeS&mt=rKNZjv3l3wkYfhl7iIFEC5E74xVqcpD5i9ibN-yhYqhXmXbMkL17J5Wo5i--0gxmGrkDyhNs9hw6bWlGcsMxu_FChHTqKPSnJBQevKL2BPwKMMe9BHVl52W8Rg&isSdui=true https://edu.nl/vtrq4Safer by Design: How AI agents are rethinking hazard analysisA thousand pages of blueprints. Reactor units, separation columns, storage tanks, boilers, furnaces; all interconnected through a maze of pipes, pumps, and valves. Gathered around these drawings is a team of process engineers and plant experts, painstakingly examining every connection and every possible failure mode. This is a typical Hazard and Operability (HAZOP) study: a structured risk assessment used to identify process hazards and operability problems in industrial plants. While essential, the process is notoriously time-consuming, and thus very costly. According to Artur Schweidtmann, artificial intelligence agents could make HAZOP studies not only more efficient, but also safer. We interviewed him and his PhD student Angga Alimin about their project to use multi-agent AI systems to improve the safety analysis of chemical processes. Hazard analysis for safety assuranceChemical plants are industrial facilities where chemical reactions are carried out on a large scale to manufacture a wide range of products. A thorough understanding of chemical plant operations is essential for ensuring both the safety and efficiency of these processes. Chemical plants are inherently hazardous environments due to the presence of flammable and reactive substances, toxic chemicals, high pressures, and elevated temperatures. Failure to maintain adequate safety standards can lead to catastrophic consequences, as demonstrated by, for example, the Bhopal disaster, which resulted in thousands of fatalities and remains one of the deadliest industrial accidents ever recorded. <section class="special has-entered" style="position: relative; margin-top: -3.75rem;">Hazard analysis for safety assuranceChemical plants are industrial facilities where chemical reactions are carried out on a large scale to manufacture a wide range of products. A thorough understanding of chemical plant operations is essential for ensuring both the safety and efficiency of these processes. Chemical plants are inherently hazardous environments due to the presence of flammable and reactive substances, toxic chemicals, high pressures, and elevated temperatures. Failure to maintain adequate safety standards can lead to catastrophic consequences, as demonstrated by, for example, the Bhopal disaster, which resulted in thousands of fatalities and remains one of the deadliest industrial accidents ever recorded. It’s tedious work, and it requires a team of plant experts and process engineers. The company can spend 10,000 hours on updating and maintaining HAZOPs.<cite style="color: rgb(155, 155, 155); margin-top: 0rem; display: block; font-size: 1rem; margin-left: 7.188em; overflow: hidden; margin-bottom: 10px;">Artur Schweidtmann</cite>The purpose of a HAZOP study is to identify potential problems and issues within complex systems. A HAZOP study is centred around the piping and instrumentation diagram (P&ID), a schematic that provides an overview of the process equipment together with the instrumentation and control devices. During the study, the P&ID is divided into smaller sections, known as nodes, and per section you ask the team: “What is the design intent and what can go wrong in this section?" According to Schweidtmann, this is tedious work, which requires a team of plant experts and process engineers. He states that “the company can spend 10,000 hours on updating and maintaining HAZOPs”. Artur Schweidtmann and his team are developing a model called HAZOPCopilot. This model uses as input a P&ID from the entire industrial process and then produces a HAZOP study. </section><section class="special has-entered" style="position: relative; margin-top: -3.75rem;">Building transparent AI the industry can trustAccording to Artur Schweidtmann, the chemical industry can be sceptical about integrating machine learning and AI into process engineering, and can underestimate the things that are possible. Everybody knows about general LLM agents, such as ChatGPT or Gemini. However, these generalized agents often fail at simple tasks in process engineering context. Schweidtmann states: “This is what scares the industry. So, what commonly happen is, I asked ChatGPT something stupid or incomplete and it doesn't work. And then they're like, this will never work.” The problem with letting generalized AI read a P&ID as an image, is that images themselves lack structured engineering context, and they don’t have access to relevant engineering databases. Added to this, generalized AI remains largely black-box, making it difficult to verify, which poses a problem in process engineering. This is what scares the industry. So, what commonly happen is, I asked ChatGPT something stupid or incomplete and it doesn't work. And then they're like, this will never work.<cite style="color: rgb(155, 155, 155); margin-top: 0rem; display: block; font-size: 1rem; margin-left: 7.188em; overflow: hidden; margin-bottom: 10px;">Artur Schweidtmann</cite>That is why the AI agents developed for this project are very transparent. They are trained on specialized data and the interactive environment in which they work allows for an easy overview. Finally, there is always a human in the loop, whose task it is to verify the results generated by the system, explains Schweidtmann. “It's important that it tracked who changed what and when, because you need to audit the entire process. [… ] Who changed what? When? Did someone just approve the AI or did they actually verify the results? What changes were made? Why? It’s safety critical to have human oversight.” </section><section class="special has-entered" style="position: relative; margin-top: -3.75rem;">Momentum: From demonstrator to communityAccording to Artur Schweidtmann, the goal of the seed funding is to invest in the AI-assisted HAZOP by hiring software developers to create a demonstrator level of the HAZOPCopilot. In the coming year, the priority is showing the industry illustratively what is possible for the future. “We’re now in the process to get out something convincing, to get people excited and to start a big community. Luckily, we have a strong team combining synergistic expertise.” The project is led by Artur Schweidtmann, who has developed the process knowledge graph and is working on multi agent systems for chemical process engineering. His PhD student Angga Alimin, who brings several years of industry experience including a year as lead process engineer, is doing most of the work on the prototype. Both of them are from the chemical engineering department. Another collaborator is Jana Weber, who has a computer science and chemistry background and whose research focuses on network science and machine learning for bioprocesses. She is working on molecular property prediction and on building and maintaining open-source reaction databases. Both projects are important for HAZOPCopilot, as industrial processes are big chemical processes, and understanding the reactions that form and the risks that are involved, is essential. An external collaborator is Tijs Koerts. He is the director of the European Process Safety Centre (EPSC), which is an institution representing different process industry sectors such as oil and gas, pharmaceuticals, agrochemicals, steel, energy, storage and chemical production. Tijs facilitates the bridge between research and industry. Other collaborators are Farzad Mousazadeh, Leon Urbas and Jürgen Schmidt. </section><section class="special is-entered" style="position: relative; margin-top: -3.75rem;">Turning blueprints into interactive dataIn a live demonstration, Angga Alimin walks through how the system works in practice.First, he explains, the engineering data needs to be contextualised. This is done by a computer vision pipeline, which converts the non-interactive PDF pages into a smart P&ID. This smart P&ID is later contextualized into knowledge graphs, the tools for this transformation has already been developed by their team and published as an open-source package, called pyDEXPI, available in GitHub. Second, once this data is contextualised in a knowledge graph, different Graph Analysis Tools can be used for P&ID interaction and analysis. Angga shows one example: finding a path within the graph, from one component to another component. This is just like an engineer tracing parts of the P&ID, but now can be done by the computer, while also visualising the intermediate components, the different relationships between the components and their respective attributes. Moreover, connecting the graph with the large language model (LLM) allows for straight-forward interaction with the graph. To illustrate, Angga gives the system a task like “List all valves with their locations”, which now becomes very easy. Not only does the agent list the valves, but it also selects and highlights them in the interactive graph, which allows for easy verification by the engineer. Ultimately, he explains, the final purpose of both the P&ID as a contextualized knowledge graph, and the many graph analysis tools, is to integrate the HAZOP workflow. This is done by using a multi-agent system. So, for example, we have one agent contextualizing the P&ID, several agents breaking down the P&ID, finding different deviations in the system, which could produce a hazard, and finding the causes and consequences. All of these agents can talk to each other, and another agent can write the reports. Finally, there is a human in the loop, which verifies the results. This complete system, which they call HAZOPCopilot, works together to create a HAZOP. </section><section class="special" style="position: relative; margin-top: -3.75rem;">We have big vision for what AI can do in safety and process engineering, in particular on HAZOPs.<cite style="color: rgb(155, 155, 155); margin-top: 0rem; display: block; font-size: 1rem; margin-left: 7.188em; overflow: hidden; margin-bottom: 10px;">Artur Schweidtmann</cite>For Schweidtmann and Alimin, HAZOPCopilot is about more than efficiency; it is about the quality and safety of HAZOPs, and about combining the domain knowledge of thousands of plants in a way a single team never could. "We have big vision for what AI can do in safety and process engineering, in particular on HAZOPs," Schweidtmann says.</section>He sees a big potential here, though he is quick to note that the harder question is still open. "Now the pressing question is" he says, "how can we do this in the best way?"
2026-10-01
- [정찬혁 박사과정][ICAE2026,구두 발표] 바이오가스생산을 위한 딥강화학습기반 2단 소화조의 pH 제어 (DRL-based pH control in two-stage AD for stable biogas production)
- [정찬혁 박사과정][ICAE2026,구두 발표] 바이오가스생산을 위한 딥강화학습기반 2단 소화조의 pH 제어 (DRL-based pH control in two-stage AD for stable biogas production)연구실 정찬혁 박사과정발표를 하였습-Conference: ChanHyeok Jeong1 , SangYoun Kim1 , TaeYong Woo, NaYoung Jeon, SeonJu Kim, SungKu Heo*,ChangKyoo Yoo*, Deep reinforcement learning-based stage-wise ph control in two-stage anaerobicdigestion for stable biogas production, 18 th International Conference on Applied Energy (ICAE2026),559, pp. 46, Oral presentation (On-site), JeJu Island, South Korea, (2026.09.25) (Ack 5개: 도약,
2026-09-26
PUBLICATIONS
- Sustainable energies and machine learning: an organized review of recent applications and challenges
- Pouya Ifaei, Morteza Nazari-Heris, Amir Saman Tayerani Charmchi, Somayeh Asadi, ChangKyoo Yoo*, Sustainable energies and machine learning: an organized review of recent applications and challenges, Energy (ISSN: 0360-5442, SCI, Top 5% journal – THERMODYNAMICS), 266(1), pp.126432 (2023.3) (Ack 3개: Brain Pool Program2019H1D3A1A02071051 & NRF중견2021R1A2C2007838 & Collabo R&D of SMEs and Startups in 2022.(Project No. S3301144)
- Deep-AI soft sensor for sustainable health risk monitoring and control of fine particulate matter at sensor devoid underground spaces: A zero-shot transfer learning approach
- Shahzeb Tariq+, Jorge Loy-Benitez+, KiJeon Nam, SangYoun Kim, MinJeong Kim*, ChangKyoo Yoo*, Deep-AI soft sensor for sustainable health risk monitoring and control of fine particulate matter at sensor devoid underground spaces: A zero-shot transfer learning approach, Tunnelling and underground space (ISSN: 0886-7798, SCIE, JCR Top 10%-ENGINEERING, CIVIL, DOI), Elsevier, 131(1), pp.104843 (Ack: 2021R1A2C2007838 & Fine Dust Reduction Technology(21QPPW-B152306-03)
- Explainable multisensor fusion-based automatic reconciliation and imputation of faulty and missing data in membrane bioreactor plants for fouling alleviation and energy saving
- Abdulrahman H. Ba-Alawi, KiJeon Nam, SungKu Heo, TaeYong Woo, Hanaa Aamer, and Chang Kyoo Yoo*, Explainable multisensor fusion-based automatic reconciliation and imputation of faulty and missing data in membrane bioreactor plants for fouling alleviation and energy saving, Chemical Engineering Journal (IF>16.4, JCR TOP 3% journal/Top 1 journal-Chemical Engineering), 452(1), pp.139220 (2023.1) (Ack 3개:NRF2021R1A2C2007838, “Prospective green technology innovation project (2020003160009), SMEs and Startups in 2022 (Project No. S3301144)
