G. Machine learning assisted materials discovery


Keynote speakers:

Claudia Draxl Humboldt-Universität Berlin, Germany
Shyue Ping Ong University of California San Diego, USA
Patrick Rinke Aalto University, Finland
Yury Lysogorskiy ICAMS and Ruhr Universität, Bochum, Germany
Gareth Conduit University of Cambridge, UK
Surya Kalidindi Georgia Institute of Technology, USA

Machine learning has become an integral part of materials science wih numerous new possibilties for identifying and developing new materials. The challenges for extracting the intricate and largely unknown relationships between process, structure, and material properties range from semantic representations of materials to data handling and statistical learning of a rapidly increasing amount of data produced by theoretical and experimental methods. While such relationsships have been traditionally provided by theoretical approaches in the past, the large amount of data enables embracing the complexity of materials and offers a new paradigm to materials science. Tackling the combinatorial explosion encountered due to the high amount of degrees of freedom associated to materials composition and process parameters requires new strategies and efficient techniques of representing materials and their properties by feature engineering and of moving through large design spaces in a targeted fashion by active learning approaches.

The generation of computational materials data by physics-based materials simlation is still hampered by the numerical effort to solve the underlying equations. However, machine learning methods have proven highly promising to overcome this bottleneck by speeding-up such calculations with surrogate models. The recent method developments and rapidly increasing range of applications push the limits of atomistic simulations by using machine learning interatomic potentials. Similar advances are demonstrated for meso-scale approaches such as phase field simulations, cellular automata or finite element simulations. The common challenges for effecient and predictive surrogate models  and for scale-bridging approaches is the idenification of appropriate features and efficient descriptors. 

The goal of this symposium is to cover the method development and the recent applications in machine-learning assisted materials discovery along the complete chain from material representation to machine learning techniques to statisical inference and surrogate models at all length scales.

The symposium covers the following topics:

  • Structure-Property prediction based on machine learning 
  • Methods and applications of active learning in materials design 
  • Surrogate/scale-bridging approaches for materials design 
  • Development and application of machine learning interatomic potentials 
  • Descriptor and feature engineering for materials science
  • Uncertainty prediction and propagation in materials design 
  • Knowledge graphs and semantic representations for materials development
  • Hybrid methods combining statistical learning and physics-based modeling in materials science

Symposium organizers:

Thomas Hammerschmidt ICAMS Bochum, Germany
Lorenz Romaner Montanuniversität Leoben, Austria
Milica Todorovic University of Turku, Finland