Improving process safety with innovative AI model


Chemelot wants to be the safest, most competitive, and sustainable chemical site in Western Europe by 2025. To achieve this, several projects are underway within Brightsite’s program line 4 ‘Securing integral process safety and societal acceptance’. With ‘BLIC vooruit’ we focus on the development of a predictive model for process safety. With the conclusion of the Safety Deal in December 2022, Brightsite can start researching and collecting data from several companies. Envalior (formerly DSM Engineering Materials) is one of the participating companies.

“BLIC vooruit – Better Learning from Information in Chemistry by Looking Ahead – is an initiative we started following good results from an earlier Brightsite project on process safety, conducted with the company AnQore. There is a lot of data available from plants, unfortunately we currently use only a small part of it. That’s a shame, which is why we are looking at whether and how we can use this data to optimize process safety. Ideally, you want to be able to use this to predict and prevent incidents,” says Esta de Goede, Senior Consultant at Sitech and Program Manager of Brightsite’s program line 4.

Johan van Middelaar, Senior Advisor Safety & Environment at TNO:

“Applying AI for process safety is innovative in safety science.”

From walking backward to looking forward

Incident prevention is now mainly focused on explaining afterwards why and how an incident could have occurred; little research has been done on incident prediction models. “In our earlier project, we looked at whether it is possible to identify hidden patterns in security datasets using artificial intelligence (AI). Patterns that are not seen by humans. We now want to pull this wider and develop it further. We are starting with a consortium of ten parties to extract information from factory data and develop a predictive model,” explains De Goede.

“Being able to predict and prevent incidents is the dream we are working toward. We still do a lot of looking backwards and learning from incidents that happened in the past. Whereas we want to look ahead to processes or situations where we can expect risks so we can anticipate them. Using AI, it is possible to learn from all kinds of data, not just incidents.

A prior project demonstrated the feasibility of identifying weak signals in text, which involved one data source from one company with one AI technique. With BLIC vooruit, we are going to learn more broadly, with more types of data and at more companies. A guide on how to do that and an overview of different types of AI techniques available today will be inventoried and tried out in this project,” emphasizes Johan van Middelaar, Senior Consultant Safety & Environment at TNO.

From breakthrough to blueprint

“We have shown that modeling using AI offers opportunities to improve safety in the chemical industry, a breakthrough. Now is the time to accelerate, which is why we are working with several parties within BLIC vooruit. The goal is to create a guideline; a step-by-step plan with which other parties can get started immediately and do not have to reinvent the wheel. Important is that we take into account the people who will eventually work with the systems, such as process operators in the factories. That is why the model is being thought about whilst keeping these people in mind so that it is useful and usable,” says De Goede. “It requires a different way of looking at things and learning through experimentation. The experiences recorded in the manual are actively shared publicly and I expect that companies outside the chemical industry will also find it very interesting,” adds Van Middelaar.

Martin Dokter, Business Excellence Automation and Digitalization Operations at Envalior:

“We are engaged in pure innovation and that takes time.”

Pilot with Envalior

Envalior, (formerly DSM Engineering Materials) is one of the companies on the Chemelot site participating in BLIC vooruit. “We notice that in recent years a lot of experience has been lost due to outflow of older, very experienced operators. Of course, their knowledge about incidents is stored in reports, but it’s nice that AI allows us to find things that happened in the past more quickly. So that we can eventually predict if something can go wrong that has occurred before. A dream come true. We got into this journey as a company because at Envalior, safety is a top priority. We want to be at the forefront in this area. Moreover, our young employees are enthusiastic about computer systems, and it suits these times to explore the possibilities of predicting using AI,” states Heiko Ilgner, Chief of Service and Supervisor of Operators at Envalior. “We as a company have also already looked ourselves to see what more we could do with the information we have at hand, but that still proves difficult. We don’t have the tools for it. BLIC vooruit is a great initiative and we immediately indicated that we wanted to participate,” adds Martin Dokter, Business Excellence Automation and Digitalization Operations at Envalior.

“To reach the end goal of a good handbook, we need to feed the model with experiences. It’s great to see that the desire to put safety first and be innovative about it is really in Envalior’s DNA. In addition to Envalior, AnQore, SABIC and CSP BV (Chemelot Site Permit) also provide data. They decide for themselves which data sources they supply. In these, we look for characteristics that have added value in predicting incidents,” says De Goede.

Heiko Ilgner, Chief of Service at Envalior:

“The computer monitors the process, and we monitor the computer.”

The model needs to learn

“The individual companies provide information with potential learning potential. Using AI, a lot of documents can be read, much more than a human could ever do. It is then a quest to find exactly the right learning points in that data. We will do that by researching, innovating and learning from every step,” Van Middelaar explains. “My wish would be to use all the data sources that can provide something, but that’s not possible due to regulations. We will start with what we think makes the most sense; on-call shift records, everything operators experience while working for the past five years. Later we will add more sources,” Dokter expects. “The system will be as smart as we make it. The model has to learn, what words in the reports are normal and behind what terms or sentence structures is frustration, for example. This is about measuring data and sentiment,” Ilgner said.

The future

“This project is one of the steps in the development of an ‘early warning’ system. We are going to train that system with as much data as possible. Then it gets really exciting when we start feeding the system with real-time data, which is going to recognize patterns 24/7 and compare them with (patterns of) weak signals that led to an incident in the past. Ultimately, the knowledge we gain can contribute to the next step; a system that can predict incidents ever more reliably. An AI model will not replace humans in this process, but it will help humans. It can help prevent incidents by issuing an “early warning. Assessing an early warning, and what to do with it, is and will remain human work for now. Technology will help to better understand complex processes, but the operator will continue to make the decisions,” Van Middelaar concludes.