Server Infos
Server:
Server IP:
Environment:
Authorized User:
cm-p115160-e1228952-aem-publish-5d55478c88-5s4bt
172.24.18.7
publish,local

On Good Terms with AI

Reading Time 8 min
October 08, 2025

Sometimes researchers come up against their limits. Experiments are too complex—there are too many possibilities, and that prevents them from reaching the goal quickly and efficiently. In cases like these, artificial intelligence can get a project moving again. At Evonik, AIChemBuddy supports its human colleagues with advice and provides space for new inspiration

Karsten Lemm
By Karsten Lemm

Karsten Lemm is a freelance writer based in Berlin, specializing in business, technology, and innovation.

The task ahead of them was trickier than expected. Sabine Kanbach and her team were looking for a production process for a new lipid. This chemical building block can be used to produce lipid nanoparticles, which form a protective shell around active ingredient molecules, for example in vaccines. This protection should be stable, a high degree of purity is mandatory, and the lipids must be easy to process. The expert team’s goal was clear, but the path to it lay hidden in a fog of unknown factors.

“We had a lot of different parameters,” recalls Kanbach, a chemist working in the Research and Development department of Evonik’s Health Care division in Hanau. The starting materials could still be narrowed down without computer assistance—but after that it became difficult. Which solvent should they choose and what temperature? How could the desired reaction be set going, and how long should it run? 

Portrait Sabine Kanbach

»It wouldn’t have been possible for us to calculate by ourselves everything we needed to vary«

Sabine Kanbach Chemist in Research and Development at Evonik Health Care

Normally, Kanbach would have approached the result with the help of empirical values and experiments—a lengthy and cost-intensive procedure. But this time, the 33-year-old turned to AIChemBuddy, an expert system that combines artificial intelligence with specialist knowledge. Evonik has developed it precisely for cases where the sheer abundance of possible combinations pushes the human imagination to its limits. Fed with all the important data and initial test results from the project, the software narrowed down the list of reagents, helped to optimize the synthesis and found a way to double the purity. “It wouldn’t have been possible for us to calculate by ourselves everything we needed to vary,” says Kanbach. With the help of computers, her team was finally able to “adjust the parameters to decimal places” and reached their goal just a few weeks after the start. “We got to the sweet spot,” she concludes.

A view over the shoulder of a computer user looking at a graphic.

The support provided by AIChemBuddy enabled the development phase to progress extremely quickly. This gave the team of experts scope to focus at an early stage on other parameters that are important in process development, for example.

Inspiration and support

AIChemBuddy is a prime example of how AI helps to make complex processes more efficient and develop new products faster. One in three companies in the chemical and pharmaceutical industry in Germany is already using the smart algorithms for these purposes, according to a current study from the German Economic Institute. In the authors’ view, the most innovative companies are those that do not buy AI off the shelf, but instead invest in their own systems that are tailored to their needs.

Evonik has several AI tools, including the company’s internal chatbot EvonikGPT, which can generate texts and process files. AIChemBuddy is aimed at employees in research. “The classic user that we always have in mind gets stuck at some point during the lab work,” says Thomas Asche. “AIChemBuddy is there to provide inspiration or calculate an optimum.”

As a materials researcher, the 37-year-old is familiar with the challenge of juggling a large number of open parameters. In 2021, his interest in digital topics led him to the digitalization department of Research, Development & Innovation at Evonik in Marl. RD&I initially provided inspiration and ideas and then implemented AIChemBuddy together with Process Engineering and Evonik IT.

Thomas Asche is sitting at a computer workstation and smiling.

When Asche joined RD&I, the first attempts to use machine learning to develop new materials were already under way. The results were encouraging, but the partners—external software companies—demanded a revenue share or access to sensitive data. This was unacceptable for Evonik. “So we asked ourselves: Can we do this on our own?” Asche recalls. “Johannes has been working on this topic ever since.”

Periodic Table and Algorithms

Johannes Dürholt is sitting next to Asche in a conference room at Creavis in Marl. In the world of Evonik, Creavis is the strategic innovation unit and business incubator. Dürholt, a 34-year-old expert in theoretical chemistry, feels just as comfortable in the world of algorithms as he does on excursions into the periodic table. As a data scientist at Process Engineering, he often works together with researchers in the laboratory. “It’s often about planning experiments,” says Dürholt. “And that’s exactly what AIChemBuddy takes care of now.”

The name indicates what the AI system is supposed to do: act as a buddy to support people in the chemical laboratory. “Our aim is not to offer ‘hyperintelligence,’” Asche explains. Instead, the aim is to provide researchers with “a virtual colleague at their side who they can ask in tricky situations: ‘What would you do?’” 

Johannes Dürholt is looking at his monitor.

The developers repeatedly emphasize one thing: all the decisions are still made by humans—and if you want to elicit helpful tips from the AI, you have to know your subject pretty well. The system has nothing in common with talkative chatbots. For all its user-friendliness, it is nonetheless a specialist for specialists. When you log in, you will find an intuitive user interface that prompts the user to create a project, upload data or evaluate the results of calculations. Some things are similar to the spreadsheet program Excel, others to the presentation tool PowerPoint. The system offers a wide range of options.

Tennis balls made from end-of-life tires

Thomas Asche and Johannes Dürholt give interested parties a personal orientation session. In this onboarding, they explain what AIChemBuddy can do—and what it can’t. Some people, says Asche, have the idea that you can just throw anything into an AI—a little experimental data here, a handful of research reports there, mixed with a vague idea of what the end result should be. But that’s not the role of artificial intelligence, he emphasizes in his role as a digital manager. “AIChemBuddy is not a crystal ball.”

The smart assistant is particularly useful if results that can be significantly improved are already available from experiments, or if a screening should show which variables are relevant for a project. “We propose experiments on the basis of the desired outputs,” says Asche. “This optimization is the core element of the platform.”

The rebirth of car tires in the form of sports equipment shows how sober analysis can still result in magic. Two years ago, Dürholt learned about Vestaro, a joint venture between Evonik and the Munich-based development company Forward Engineering. The partners had set out to develop a raw material made from end-of-life tires for the production of tennis balls. Around 25 million tons of rubber are produced every year because the tread of tires is worn out—so the source material for possible recycling exists.

Or at least, it should. However, it turned out in the laboratory that it’s very difficult to find the right mixture of recycling material, additives from Evonik, and natural rubber to give the environmentally friendly tennis balls the same properties as those produced conventionally. “In addition, there was a lot of time pressure in the project,” Dürholt recalls.

The chemists and product developers consulted Dürholt and a colleague as experts in Process Engineering. The first step both of them took was what experts call “data hygiene.” They checked which values could be used by AI, removed outliers and deleted information that was not very promising. That’s because the cleaner the data, the more reliable the results delivered by AI. 

A yellow tennis ball lies on an illuminated surface.

After a few days of this preparatory work by humans, the algorithm behind AIChemBuddy only needed a few minutes to provide suggestions for several compositions of old and new. “The rubber experts were very impressed,” says Dürholt. The ball has recently gone on sale under the name “Code Planet.” With it, the manufacturer Balls Unlimited is presenting the first sustainable tennis ball. According to the company, 40 percent of the ball core is made from recycled end-of-life tires.

No more “rule of thumb”

The success impressively demonstrates how the internally developed AI solution can drive projects forward in key areas. And it reinforces the view among experts that companies that make targeted use of artificial intelligence have a clear competitive advantage in those areas where other companies focus on personal experience and intuition. Jakob Zeitler, Pioneer Fellow at the Statistics and Big Data Institute at the University of Oxford, criticizes that too many laboratories are still working according to the rule of thumb.

“When we go into the lab and run experiments, it’s a bit like baking a cake,” he explains. Here too, there are different variables, ranging from the ingredients to the mixing ratio and the temperature and time in the oven. Every change has an influence on whether the cake tastes good, the dough rises or the cake base burns. These variables may still be manageable in the kitchen, says Zeitler, but in chemistry there are many “interactions between complicated factors that are not directly visible to us.”

AI is guided only by numbers, patterns and mathematical relationships—parameters that often remain hidden to the human eye, and whose alteration can lead to astonishing results. Zeitler remembers chemistry experiments in which the algorithm repeatedly suggested an acidity level that seemed absurd. In practice, however, the calculated recipe proved to be superior. “It suddenly worked after all, because the acid interacts with other factors in a very complicated way,” the researcher explains. “Since the statistical model understands this high-dimensional space better than any human being, it can predict reactions that contradict our intuition.”

The AI makes suggestions.

Mathematically, such successes are based on a method called Bayesian optimization. It uses a limited amount of initial data—such as results from laboratory experiments—and uses models to calculate what happens when individual parameters change. AIChemBuddy uses the available experiences to look for promising ways to achieve a result that the researchers have specified.

The AI finishes its job with a proposal for a promising experiment. The method is therefore particularly suitable for exploring a given search space in which the best conditions for the interaction of certain factors are to be determined.

AIChemBuddy relies on a software library for Bayesian optimization called BoFire. Johannes Dürholt is wearing a T-shirt with the AI’s logo in black visible through his light blue shirt. BoFire is the result of a cooperation between companies that may seem unusual at first glance: In addition to Evonik, BASF, Boehringer-Ingelheim and, more recently, Bayer are also involved in the further development of the AI system.

“They’ve all worked on similar problems,” Dürholt explains. They all had costs for the same approach to a solution. “That’s when we said, ‘Why don’t we join forces and make it open source?’”

Johannes Dürholt and Thomas Asche are sitting across from each other at a shared workstation.

This was the next step, which was an unusual one from a management perspective: The source code for BoFire is openly available on the Internet. Isn’t anyone worried that competitive advantages and company secrets are being revealed? No, answers Thomas Asche. “The intelligence ultimately comes from the data we use to train the models. This means that everything that generates value for us remains within Evonik,” he explains.

AIChemBuddy also belongs exclusively to Evonik as an application. Put simply, BoFire is the drive system that all participants develop together—but it only becomes a powerful engine in combination with internal company data. Each company also builds its own chassis—the app that determines what users can do with the AI.

Evonik has taken on a pioneering role in the development of BoFire and also contributed a large part of the code base. “Other participants are welcome, because together we can make BoFire even more powerful,” says Dürholt.

Johannes Dürholt gestures in front of his monitor.

AI does not replace human know-how

AIChemBuddy has its virtual home in a data center in Frankfurt am Main, and it’s frugal when it goes into action. “The energy consumption corresponds to that of two to three laptops,” says Asche.

At Evonik, AIChemBuddy is available to all employees. Among other things, this is based on the hope of finding as many new fans as possible for the AI system. That’s because the more often the digital assistant gets projects up and running, the more likely it is to pay off for the company. The same applies to the acceleration of development work, as in Sabine Kanbach’s project. 

Kanbach already sees the opportunities to use insights from her current project for future tasks. “We have now optimized our lipid synthesis to a very high degree,” she reports. “These findings can be easily transferred to other excipients that work in the same way.” She is so enthusiastic about the results that she wants to recruit more colleagues for AI. So far, many have been hesitant, she reports. For some, habit probably prevailed, while others were worried that AI could jeopardize their jobs.

ELEMENTS-Newsletter

Get exciting insights into Evonik's research and its societal relevance—delivered conveniently via email.

Kanbach believes the concern is unfounded: “I don’t have the feeling that AIChemBuddy is replacing me,” she says. “After all, we specify the experiments ourselves and deal with the results.” The developers also believe that the special value of their system lies in this close cooperation between man and machine. “The chemist’s knowledge is super important,” Johannes Dürholt emphasizes, adding that the more knowledge that humans provide to the AI, for example in the form of limits for the individual parameters, the easier it is for the machine to find good approaches to solutions.

“If you feed it with your knowledge, you’ll reach your goal faster,” adds Thomas Asche—and the word is spreading in the company. So far, around 300 employees from all business units are using AIChemBuddy. 

Dürholt and Asche, seen from behind, walk down a long corridor in the Creavis building in Marl., Germany

“We have many use cases from the areas of research and process engineering development. It’s also already being used for simulations and process optimization.” But there should be even more users. “This is not a race of ‘AI versus chemists,’” Asche emphasizes. “Instead, it’s a race of ‘chemists without AI’ versus ‘chemists with AI.’ And it’s a race we will win.”