The world of materials is getting a revamp from AI
From nail polish to solar cells, materials are everywhere. How can AI make them better? University of Turku associate professor Milica Todorović shares the trends in AI-driven design of materials and what FCAI researchers are working on in this area.
Image: Mikko Raskinen/Aalto University
In the past decade, the use of data science and AI tools has accelerated materials research, helping to discover new materials and improve technologies. These days, new horizons are opening up. Mining information from text data has benefitted social sciences and humanities, but natural language processing (NLP) is not traditionally used in materials science. As our field wakes up to the possibility of using text as data, there is much to discover on how best to apply NLP in natural sciences. This will allow us to gain further information from materials data which can be heterogeneous, with numbers, spectra, images and text. We have already made advances in AI for spectra and microscopy image interpretation, where Aalto University’s Adam Foster and Paavo Penttilä are successfully contributing. With the rapid evolution of cutting-edge language models, we are working closely with computer science colleagues to develop methodology and relevant datasets.
Multi-fidelity and multi-objective optimization are increasingly deployed in the field. Materials design is an optimization problem, and this is where AI can help. Usually, we are limited to optimizing material properties one-by-one, but for advanced functional materials we need to improve multiple properties simultaneously, making this a multi-objective problem.
AI models that incorporate information from different data sources at the same time, so-called multi-fidelity frameworks, are an emerging AI application in materials science. We can use experiments or simulation to study a material, but traditionally we are tied to one level of accuracy. Now we can exploit less accurate but more plentiful data to help refine our understanding of materials at a higher level of fidelity, where data may be costly to acquire. Patrick Rinke and I have been working on this with FCAI’s Jukka Corander and Ulpu Remes. The holy grail would be to integrate experimental and computational data in the same model. In the future, with multi-fidelity frameworks, we can imagine cases where we use simulation data cheap to learn from, and supportive of, very costly experiments. This would be a big boost to science, because historically it has been hard to merge data from experiments and simulation. Here, FCAI is instrumental in facilitating the transfer of knowledge from computer science to materials science.
“The interfaces of inorganic and organic small molecules are everywhere, from nail polish to lubricants to solar cells and batteries, as well as in atmospheric chemistry.”
People don’t realize how related atmospheric research is to climate and health. The Virtual Laboratory for Molecular-Level Atmospheric Transformations (VILMA) is an example of virtual laboratories, AI-driven environments that combine traditional and simulated measurements with human researchers. Propagating this idea to other fields of science is one part of FCAI’s wider research agenda. In the VILMA Center of Excellence, FCAI’s Kai Puolamäki and Patrick Rinke have teamed up with Arto Klami to apply it in study of atmospheric compounds.
The VILMA project has also been instrumental in generating datasets for atmospheric science. A lack of specialized datasets can slow down the adoption of AI in different research fields. Each published dataset could kickstart a research wave, which is especially important for multi-fidelity and NLP work. Some examples are our datasets on atmospheric molecules and optical properties of organic crystals. Publishing materials data is a concrete thing that FCAI’s Highlight E can do to boost AI research on materials. Our collaboration with CSC, Finland’s IT Center for Science, enabled us to perform calculations and produce these datasets in a high-throughput manner. Open science benefits everyone, and we want to encourage our colleagues to follow similar procedures, curate datasets and share their data.
Lowering barriers for the use of AI in our community is also an important part of our work, so all our colleagues could use AI without coding or specialized expertise. Democratizing AI and facilitating its use in experimental work will make AI part of the standard scientific toolkit. AI-driven design of experiments is championed by Mikko Mäkelä at VTT—Mikko and I are now coordinating FCAI’s AI-driven design of materials research highlight. AI-guided experimentation will replace trial-and-error approaches to discover optimal solutions with less time and investment. In an example of FCAI’s joint work with the FinnCERES flagship, AI was used to optimize lignin extraction from wood. This advanced the development of sustainable materials through eco-friendly processes. Such bio-derived materials could replace plastics or petroleum derivatives in future products.
Finally, education and knowledge transfer are central to our efforts to strengthen AI research in materials science. The exchange between computer scientists and materials scientists through FCAI allows Finland to punch above its weight in global research and education. We’ve been offering the “Machine learning for materials science” course at Aalto University since 2017. Education on AI and data literacy for domain experts from physics or chemistry is important for a new generation who will develop new digitalisation trends in industry. International schools and workshops on Machine Learning for Materials Science are taking place in Finland every other year. As we move forward, a priority for me is identifying synergies between FCAI and the wider community. Users of applied AI: what are your trends and needs? What can FCAI do for you, in terms of AI-driven design of materials?