“Through the DAP²CHEM project we have learned how to fundamentally transform processes with AI”
Industry 4.0, artificial intelligence and digitalization are unleashing revolutions in numerous industries. The chemical industry does not want to miss this train either. As a participant in DAP²CHEM, a project under the Catalisti umbrella, Johnson & Johnson investigated how it could deploy AI and what added value it would bring. It initially focused this on solvent switches. Project manager Niels Vandervoort speaks of a success story.
Can you briefly introduce the site in Geel?
Niels Vandervoort: “Johnson & Johnson is a multinational company with a large R&D and manufacturing footprint in Belgium. The site in Geel is an important site for the chemical production of active pharmaceutical ingredients with a very close integration with the chemical development labs hosted on the Beerse site. This R&D to commercial production ecosystem helps us to drive new chemical innovative medicine to our patients.”
How did you yourself become involved in the DAP²CHEM project?
Niels Vandervoort: “As Director Digital Transformation Synthetics Development, I am responsible for the digital transformation strategy within the Synthetics Development department which represents our chemical development, product formulation development & analytical development organizations hosted at the Beerse & Geel site. A lot is happening today from a digital technology perspective. Hundreds of ‘digital technology’ trains are leaving the station at the same time, so to speak and my job is to determine which of those trains are relevant to us i.e. which new digital technologies & innovations can help us to transform the development of new medicine and accelerate the time it takes for new innovative medicine to find their way to our patients. DAP²CHEM gave us the opportunity to find out how AI can transform our way of developing new medicine and how we can introduce it into our pilot & commercial plants in a robust way.”
The project at Johnson & Johnson focused on solvent switches. Why?
Niels Vandervoort: “Chemical processes consist of several unit operations. A typical one is a solvent switch. Most chemical reaction steps work with a component dissolved in solution. That is an efficient way to trigger most chemical reactions. Then, to finally turn it into a product, a crystallization step happens. This process necessitates changing solvent systems. Since the solvent switch is a recurring step in most processes, we have developed strong expertise in handling it both in the lab and at the plant. Improving this step could significantly reduce lead time, energy consumption and drive overall process efficiency.”
A digital twin of a solvent switch was made. Why and what did it yield?
Niels Vandervoort: “Today we already use offline models. Those predict how processes will behave or are used to analyze how a batch could have been improved.I like to compare it to the world before we knew GPS. You look at the map before departure to determine the best possible route and you must hope nothing happens along the way to get to your destination on time. With this project, we essentially built a GPS for our process. Now, we can track progress in real time. If we encounter obstacles, like unexpected changes or inefficiencies, the system automatically adapts. You can also achieve this with more traditional technology, but it typically requires handling each unit operation separately. By applying AI, we’ve created a single system that can navigate through all solvent switches.”
With this project, we essentially built a GPS for our process. Now, we can track progress in real time. If we encounter obstacles, like unexpected changes or inefficiencies, the system automatically adapts.
Anyone who has used ChatGPT knows that AI can also hallucinate at times. How do you deal with this?
Niels Vandervoort: “We did two things to make sure the process behaves safely and with quality. First, we set hard limits within which the application must always adhere to. Second, we looked at explainable AI i.e. as AI traditionally behaves like a black box; we used a surrogate model to capture the expert user experience to map AI actions to process reality. For example AI not only tells our operators to raise the temperature from 50 °C to 60 °C, but also explains why we need to make this change and what the impact on the process would be. So, AI not only predicts the action to take but also provides the reason behind it.”
What did it all add up to in the end?
Niels Vandervoort: “We ran the digital twin on a number of processes. At the same time, we also performed a classical calculation on it. AI proved to increase throughput time by 20 to 30% on average, most recently by 50%. Solvent usage went down by 30 to 40%. That makes the ROI instantaneous. But we see it primarily as a way to support the operators so they can monitor the process at all times. In an environment like ours, that's at least as important. The figures show that for the same process, there can be up to eight hours of difference in throughput time. Thanks to AI, we can better control the process and let the operators do the things they really make a difference in. Think of it as a self-driving car. The operator keeps the oversight, while many processes run in a controlled and optimal way managed by the AI agent. The operators themselves were positive about this development.”
So what are the next steps?
Niels Vandervoort: “So far we have stayed within a pilot plant environment. The intention is to take what we have learned towards an industrial application in the follow-up project EL4CHEM. That's why two parallel development paths were running. KU Leuven built the digital twin, imec added the AI. It is not an obvious thing to do because in a development environment, there is little experimental data. Additionally, we looked at how to integrate that into existing plant systems robustly. And for example take anomalies into account: What if the temperature sensor breaks down? What if there's a problem with the pressure control? What if we run out of solvent? You must think about all the possible things that could go wrong and next, how you want the system to react as soon as you go into the plant, the digital twin has to work reliably and correctly.”
How was the project received internally?
Niels Vandervoort: “It created a lot of traction. We have laid the groundwork to work towards a self-optimizing plant. This can serve not only to make the operator more efficient and optimize processes, but also can help provide new tools to the scientists who develop new molecules. Or to train new people. The knowledge and experience are now secured in digital applications instead of documents and procedures. In other words, it has triggered a lot of ideas about how we can roll this out horizontally and vertically to other operations. If we apply what we have built with DAP²CHEM only on the crystallization step, the improvements would already be enormous. If you ask me, this project was really a success story. Because we have learned how to fundamentally transform processes with AI.”