The future of energy is renewable. How the mix of solar, wind, and hydropower will ultimately be composed remains uncertain. What is certain, however, is that these forms of energy must be efficiently convertible into one another. Yet high losses occur in the process – and the interfaces between materials play a decisive role. This is precisely where the interdisciplinary scientists of e-conversion come in. TUM researcher Prof. Patrick Rinke is one of the pioneers of machine learning (ML) in materials research. In e-conversion, he now facilitates the digital transformation to accelerate the search for new energy materials with the help of AI. 

Your research combines machine learning, data science, and materials science. What first led you down this path? 

I have always been fascinated by mathematical structures and theoretical calculations – for example, what holds matter together at the atomic level and how this can be described computationally. During my master’s studies at the University of York in England, I became interested in electronic structure theory and atomistic modeling. Throughout my PhD and postdoctoral years, I continued to expand my methodological toolkit, particularly at the Fritz Haber Institute (FHI) in Berlin. There, the focus was on surfaces, organic molecules on semiconductors, and the further development of theoretical tools. The FHI was and still is an exceptional place – highly international and very well connected. Through these networks, I also spent time in California at the University of California, Santa Barbara. My time at the FHI shaped me profoundly and laid the foundation for my current research. Machine learning methods were already explored at the FHI early on. I witness these early days, found the emerging potential fascinating and carried this enthusiasm with me to Helsinki. 

You moved to Aalto University in Helsinki in 2014. How did this happen? 

I was leading a research group at the FHI and asked myself what the next step could be. Then an offer from Aalto University came along – including a suitable position for my partner. It was a classic dual-career moment. I then spent nearly ten years in Helsinki. During that time, I became a full professor and began to systematically integrate machine learning and data science into materials science. Initial efforts in that direction had already started at the FHI, but in Finland we had an ideal environment to take the development further, particularly with the Finnish Center for Artificial Intelligence and excellent collaborators in computer science. We also gradually moved beyond purely atomistic modeling and began to systematically digitizelaboratory and measurement data. With the Aalto Materials Digitalization (AMAD) Platform, we created and  established a materials science data infrastructure at Aalto University and integrated it into AI-driven materials synthesis and characterization processes. In 2023, AMAD received the Aalto Open Science Award. 

What particularly attracted you to data-driven materials science? 

In the 2010s, this was still a field that few people were paying attention to. I quickly realized the potential: many materials problems are high-dimensional and experimentally very difficult to access. At the same time, data are generated everywhere – from experiments, simulations, and measurements – but they were often not digitized. Finland has a strong tradition in wood and biomaterials research, and many research groups in those disciplines wanted to move away from Excel spreadsheets toward genuine digital workflows. This is where machine learning could make a real difference: identifying structures, revealing correlations, and accelerating research. At the same time, machine learning itself was developing rapidly. We therefore began to integrate ML models directly into the research workflow to speed up property prediction, screening, and materials optimization. Our own Bayesian optimization software, played a crucial role. In retrospect, systematic digitalization was key: many questions that seemed visionary at the time are now standard practice in data-driven materials research.

You have already mentioned the large data infrastructure you built at Aalto. What was the biggest challenge? 

It was important to understand that infrastructure does not just consist of servers. We had to create a software architecture, write interfaces, establish AMAD as an electronic lab notebook, and convince researchers to actually use it. Aalto University collaborated with a company that developed the AMAD prototype, which the university then continued to adapt – for example, when new instruments such as a scanning tunneling microscope were added. We had to program our own interfaces so that image data could be processed automatically. The project was difficult and time–consuming, but in the end, we had a fully functional AMAD infrastructure, supported by IT services and used by students and researchers alike. 

Now you are building something similar in Munich. What is your goal in e-conversion? 

On a practical level, it is again about infrastructure: standardized electronic lab notebooks, secure data storage, automatic data transfer from instruments, connection between datasets, and preparing data for machine learning – and, of course, persuasion. Machine learning is not inherently useful in every case. The key is the research question. Many groups at e-conversion work on processes in which data are central – materials development, catalysis, batteries. We are currently connecting with many groups to understand what data they generate, what problems they want to solve, where we can help in the short term, and what the long-term vision is. As motivation for a joint project, I can say this: even a dataset of just 50 to 100 samples can be sufficient to train a reliable ML model. With AI, fundamental structure-property relationships can be derived from comparatively small amounts of data, narrowing down search spaces that would otherwise remain inaccessible. 

Can you give an example of such a collaboration? 

A current example is the growth of new semiconductors with specific piezoelectric properties – a project with my e-conversion colleagues Dr. Verena Streibel and Prof. Ian Sharp. In materials growth with, e.g., atomic layer deposition, there are many adjustable parameters – temperature, gas flows, pressure, to name just a few. One quickly ends up in a seven- or eight-dimensional parameter space. Experimentally testing all combinations would be extremely time-consuming. AI, by contrast, can generate digital models and suggest: Try this parameter combination next – and feed the result back to me. This creates a ping-pong scenario between experiment and model, with the latter becoming increasingly accurate. Step by step, the materials or synthesis space is explored, like filling in the blank spots in a map or sketch. The goal is to identify optimal growth parameters, correlate them with structural data, and ultimately transfer this understanding to new materials systems. 

That sounds like a vision of a digital twin in materials research. 

Exactly. A first step is for instruments to send data directly to the infrastructure. After that, workflows can be automated, and eventually digital twins can be created. This allows experiments to be simulated before they are even conducted. For that, we need not only large quantities of data but, above all, high-quality data and standardized formats – essentially DIN standards for data. A great deal is currently happening in Germany in this regard, for example through FAIRmat, a consortium within the National Research Data Infrastructure. In the long term, both industry and basic research will benefit. My wish is for a holistic, data-driven picture that ensures interoperability with industry and allows seamless access to the data pool. At the same time, datasets can be understood as digital twins of materials and their properties – like fingerprints, only digital. And then there is the search for entirely new materials. What configurations even exist? How stable are they? What properties might they have? No one knows. Since AI is developing so rapidly, this field remains extremely exciting. 

What personally drives you in this research? 

What fascinates me is how many disciplines come together in materials research. For me, it is not about “just” developing an algorithm, but also about understanding the chemistry and physics behind it. What looks good on paper has no value if it cannot be realized. That is why dialogue and learning from one another are so important. Understanding interfaces takes time – but that is also what makes it so appealing. My goal is to obtain a holistic, data-driven picture of materials. When materials with a specific chemical formula receive a kind of “signature” through data, measurements, and simulations – the digital fingerprints I mentioned earlier – we create a data cloud that encompasses all properties and states. These fingerprints can then be compared, differences identified, and insights gained. It is precisely at this interface between AI, physics, chemistry, and measurement techniques that much is happening right now – and that is deeply motivating. 

Original article: https://www.e-conversion.de/energy-pioneers-patrick-rinke-en/