Research in AI-Driven Computational Materials Science
The computational material science group led by Milica Todorović combines algorithms from artificial intelligence with materials simulations and design. In collaboration with computer science partners and experimental colleagues, we are developing new methodologies for optimising functional materials and their performance in devices.
Active machine learning methods for atmospheric science (ActiveAtmos) project employs novel data curation techniques to build high-quality transferable AI models. Given atmospheric molecules as input, the model will make instant predictions on their affinity to condense into aerosol particles, expanding on our understanding on aerosol formation and impacting climate change models. This Grand Challenge project on CSC’s Mahti supercomputer is carried out jointly with Prof. Patrick Rinke (Aalto University), Dr. Theo Kurtén and Prof. Hanna Vehkamäki (University of Helsinki).
This cross-disciplinary study opens the route towards green, cost-efficient biorefinery protocols to convert wood-derived lignin from waste into valuable materials for everyday use. We apply AI to infer how experimental processing conditions for lignin extraction correlate with the structural and functional properties of resulting lignin, so we can customize lignin types towards their target applications. The AI for Wood-based Functional Materials (AI-WOOD) collaboration with Prof. Mikhail Balakshin, Prof. Patrick Rinke and the FinnCERES CoE (Aalto University) is supported by the Aalto Platforms Seed Fund (2020).
Hybrid perovskite materials are prime candidates for next-generation solar cells, but proposed devices suffer from instability and toxicity. With data-driven compositional engineering of halide perovskite materials on the atomic level, we pursue a combination of material stability and favourable material properties while eliminating toxicity and undesirable traits. The LearnSolar collaboration with Prof. Patrick Rinke (Aalto University) is supported by the Academy of Finland (2020-2023).
In this data-driven project, we train AI models to map molecular structure to their electronic levels and spectra, allowing fast inference of their light-absorbing and emitting properties. The trained models can be used to screen millions of candidate organic molecules for those with optimal properties in optoelectronic devices. This joint work with the groups of Prof. Patrick Rinke and Prof. Aki Vehtari (Aalto University) promises to accelerate the development of new technologies by cutting down on trial-and-error approaches in materials design.
Bayesian optimization (BO) is a key AI tool for autonomous complex optimisation tasks. The BOSS code incorporates basic and advanced BO functionality into a free python tool for solving optimisation problems in materials science. BOSS has been applied to atomistic structure search in complex functional materials, optimal parameter search, multi-target materials design, and for guiding experimental data collection. BOSS code development with groups of Prof. Patrick Rinke (Aalto University) and Prof. Jukka Corander (University of Oslo) is supported by the Academy of Finland through the AI in Physical Sciences and Engineering AIPSE call (2018-2021).
Devices for future technologies are based on advanced functional heterostructures, which typically feature blends of organic molecules and inorganic crystals. Their performance critically depend on the binding, structure and properties at the organic/inorganic contact, which is difficult to probe experimentally and computationally. We use the BOSS code to conduct global structure search at the organic/inorganic interfaces, and extract all possible contact structures to study their relative stabilities and properties. In collaboration with global partners, our objective is to understand and optimise device performance.
Materials science abounds with approximated models for more tractable research, with computational frameworks fitted to experiment, or force fields fitted to costly quantum mechanical simulations. Transfer learning exploits the high correlation between different fidelity levels to learn primarily from cheap low-fidelity data, then refine the accuracy of predictions with a few costly data points. These innovative AI tools are introduced into materials science in joint work with Prof. Patrick Rinke (Aalto University) and Prof. Jukka Corander (University of Oslo).
HITL builds upon data-driven AI applications in materials science by additionally sampling the scientist to collect domain knowledge. This information can be employed to guide AI data collection, make learning more data-efficient, correct data biases, and resolve issues where data is inconclusive. Research on this upcoming trend is carried out in collaboration with the group of Prof. Samuel Kaski (Aalto University and FCAI).