back to MMS Days 2023 Main Page »

6th Leibniz MMS Days
April 17 - April 19, 2023
Potsdam

Focus topic
Computational Material Science

At the MMS days 2023 contributions from eight institutes in the field of materials science were presented. This was so far the best participation in this area. It included topics from solid state physics (Engineering of materials for CMOS microelectronics, C. M. Manganelli, IHP Frankfurt/Oder; High-throughput band-structure calculations and machine learning for spin models of quantum magnets, O. Janson, IWF Dresden) and optics (Extending the emission wavelengths of GaAs-based diode lasers from red towards orange, F. Mauerhoff, FBH Berlin), on growth kinetics (A Kinetic Monte Carlo Tool for Computation of Epitaxial Processes Programmed in Julia, W. Miller, IKZ Berlin) and related topics (Lean plasma models for deposition processes, M. Rudolph, IOM Leipzig). Free interfaces and contact angles were a subject of two talks (Calculation of the shape of melt free surface using the finite volume method and interface tracking for crystal growth applications, I. Tsiapkinis, IKZ Berlin; Discrete Differential Geometric Formulations for Mesh-Independent Three-Phase Dynamic Contact Angle Modelling in Multiphase Flows, S. Endres, IWT Bremen). The topic of machine learning (ML) was addressed in Machine Learning Assisted Multiscale Simulation for Liquid Composite Molding, St. Cassola, IVW Kaiserslautern). ML was also used for obtaining realistic force fields for molcular dymamic computations (Ab initio modelling of liquids with ionic compounds, St. Zahn, IOM Leipzig). The IPF Dresden contributed by a poster on Scaling Properties of Tree-like Self-Similar Polymers (Ron Dockhorn).
The story in computational material science was completed by two talks from WIAS about Symmetries in TEM imaging of semiconductor nanostructures with strain (A. Maltsi) and an overview on FBH - WIAS collaborations on semiconductor lasers (Mindaugas Radziunas).

Regardless of the application and the numerical methods used in computational material science ML is a hot topic. Consequently, one topic in the World Cafe was "Linking ML tools to simulation".

There was not a deep experience among the participants at this topic table. However when tools were used, it was in python and tensor flow, scikit, or pytorch was used. JAX was also mentioned as an interesting alternative but no one has used it yet.

ML is seen to be useful for the following:

  • get good initial guess
  • use for subtopics in classical simulations, where other methods are too cpu intensive or too inaccurate
    (e.g. convection in crystal growth processes or others, microphysics of droplets in geophysical computations)
  • inverse computing by ML for parts, where data is missing in classical computations
    (e.g. using experimental in flow, when boundary condtion is unknown)

In engineering problems one has to make a case-dependent decision if ML is superior to classical approaches: computational time versus accuracy.

Nevertheless, the overall opinion was that on long term ML will win against classical simulation in many fields considering the increase in computer power and in particular, the outcome of new computer architectures.

However, there were certain concerns with ML as a black box, where the validity might be questionable. This means that all implementations of ML should be carefully documented and its validity should be clear.

The discussion can be concluded as follows:

  • There are great challenges both in mathematical issues of ML and application, to be addressed and solved in the near future.
  • This requires a close communication and joint work between mathematicians and domain specialists in material science.
  • The MMS network is the ideal tool to bring the relevant researchers together to eventually start projects.