Key elements of the project
The project encompasses four core components.
Through DTU Nanolab, researchers gain access to an experimental laboratory capable of producing new materials in thin-film form. The laboratory is among the most advanced in the world, with custom-built equipment developed by the researchers themselves to synthesize new materials across most chemicals and combine them in novel ways.
Associate Professor Andrea Crovetto and his team will oversee high-throughput synthesis and characterization, employing advanced techniques to rapidly evaluate materials. This work draws on his expertise in optimizing deposition processes and characterization methods:
“With the facilities available in the IDOL lab at DTU Nanolab, it is now conceivable to synthesize AI-generated materials – no matter how exotic – in thin-film form.”
At DTU Physics, one research group focuses on simulating materials using quantum mechanics, running large-scale computational models to predict how materials behave at the atomic level. The computational method used is Density Functional Theory (DFT), a quantum mechanical approach applied in physics, chemistry, and materials science to investigate the electronic structure of many-particle systems – typically atoms, molecules, and solids.
Professor and Head of Section Kristian Thygesen (DTU Physics), an expert in theoretical materials simulation and the developer of a widely used DFT software package, will lead the DFT simulations and provide essential training data for the AI models:
“The possibility to run our codes on the Gefion supercomputer will allow us to perform advanced quantum mechanical calculations of materials properties with an accuracy close to the experimental precision limit. The ability to train our AI models on such high-precision computational data will be key to develop the deep generative models that in turn can guide us towards the materials with optimal properties,.”
“We will also explore amorphous and non-stoichiometric materials (materials lacking an ordered periodic arrangement of their atoms). Such materials are very challenging to simulate using traditional computational methods”, says Kristian Thygesen.
By integrating experimental data for amorphous and non-stoichiometric materials into their AI-models, the team hopes to be able to make predictions for this class of materials as well.
“This would greatly expand the space of materials that can be considered for solar light-trapping.”
The second research group at DTU Physics focuses on the nanostructured surfaces of materials. Their work includes simulating how light propagates through and transforms within a material, as well as investigating how surface structures can be engineered – for example, by printing patterns onto the surface.
Associate Professor Søren Raza (DTU Physics), an expert in designing and simulating nanostructured materials for light–matter interaction, will generate data to optimize optical properties and experimentally realize the optimal light-trapping layer. The project team will hold regular cross-group meetings to ensure close collaboration and data sharing:
“There is a shortage of materials in the field of optics, and we hope to help change that. We are searching for new materials that are both transparent and have a high refractive index, but they could also be used for many other optical applications beyond just solar cells, So, it is exciting to explore how AI can accelerate the discovery of new optical materials."
"This project has the potential to reshape how we design light-trapping structures in solar cells and to uncover new and better materials that could enable advances across a wide range of photonic applications," says Søren Raza.
AI scours the materials database to find the needle in the haystack
The Novo Nordisk Foundation’s special research programme, the ‘Grand AI Challenge’, addresses major challenges in artificial intelligence in a novel way – by enabling researchers to generate their own training data and to train models on a supercomputer, in this case Gefion, which is specifically designed for AI computations and was not available just a few years ago.
But, working with the AI supercomputer is not like using standard cloud platforms, where everything is mostly ‘plug and play’. Gefion provides a kind of raw, powerful infrastructure, which means DTU Compute must develop algorithms that enable the utilization of its computational power.
At DTU Compute, Mikkel N. Schmidt and his team are tasked with scaling up the AI and algorithms to effectively search the vast space of possible materials – something that is currently beyond reach.
Their calculations will primarily focus on simulating the properties of very large classes of different materials, so they can identify which ones might be of interest. There are billions of possible materials. The challenge is simply to find out which ones are promising.
“It takes a very long time to compute the properties of a single material, so we need to accelerate the process using artificial intelligence to identify interesting candidates among the billions of possibilities. We can then perform more detailed calculations and eventually synthesize the materials in the lab,” says Mikkel N. Schmidt.
“I’m very excited to get started. In the past, we typically trained AI models on existing datasets and demonstrated that they could discover something new. But the problem is that it’s cumbersome to return to the real world and verify whether the findings are actually useful – and whether the material can be produced. We want to close the loop so that the AI can feed information back to the lab, and the lab can in turn provide new data to the AI, allowing it to continuously improve. It’s a long-term effort – like searching for a needle in a haystack.”
The text continues after the slideshow.
The potential of AI superpowers
In the past, materials scientists might have proposed that combining two elements would yield a material with specific properties. That is, of course, still possible. But today, the number of ways to combine materials in novel configurations is so vast that you must explore how to prompt artificial intelligence to suggest materials with desired characteristics.
In this case, the focus is on optical properties – how light refracts through materials and how it is reflected. By analyzing a large dataset of known materials, either measured in the laboratory or computed using traditional, resource-intensive methods, AI can be trained to emulate these results and predict how a previously unseen material might behave when exposed to light.
This knowledge can then be refined using conventional methods and ultimately tested in the laboratory.
A continuous process from day one
The researchers are launching all the activities in parallel. They simultaneously simulate materials using quantum mechanics and photonic modelling and synthesize materials in the laboratory to generate data for training their AI models.
This integrated approach ensures that the AI is trained on both quantum simulations and real experimental data. In turn, AI provides suggestions for new materials worth exploring further.
“This should ideally create a self-reinforcing effect, enabling the AI to learn more and more about where to look in order to develop promising materials – ultimately leading us to a suitable candidate that can make thin-film solar cells more efficient,” says Mikkel N. Schmidt.
"We are not just improving materials – we are redefining how they are discovered and engineered. Leading a team that blends AI, physics, nanotechnology, and lab synthesis is an exciting opportunity to make a real impact on the future of clean energy."
