Still predicting how things yield

Like

Memory of the past

Two years ago I had a paper about how to predict the stress-strain curves of metals that yield. The work was published in Nature Communications 9, 5307 (2018) and involved me, Lasse Laurson (now a professor at Tampere University, Finland), and Henri Salmenjoki as the guy who did the work, our joint PhD student. The main point of what we did is that we showed that the dislocation assemblies that are in materials the engine of plastic deformation present a memory of how the sample was created. This imprint is then at the heart of such predictability, an example of a wide class of such problems in physics and materials science (see Keim et al. Rev. Mod. Phys. 91, 035002 (2019) for a very recent review arising from a Kavli Institute for Theoretical Physics program). To exploit that memory we tried various Machine Learning tools such as Support Vector Machines and Neural Networks and of course a part of the game was that we tried various measures to quantify such memories ("descriptors" is the usual term).

Since then

We have expanded on this work in various directions like the prediction of how soft matter systems or foams yield (L Viitanen et al., Phys. Rev. Research 2, 023338 (2020)), how to guess how more realistic 3D dislocation systems yield depending on their microstructure (H. Salmenjoki et al., Materials Theory 4, 5 (2020)) or how dislocation pileup yield or depin (M. Sarvilahti et al., APL Materials 8, 101109 (2020)) and other work is in progress in several directions. We are part of an exponentially increasing effort in the world on the applications of Machine Learning to materials and condensed matter physics and statistical mechanics problems. For instance, such methods allow to find the key features in amorphous, disordered materials (D. Richard et al., Phys. Rev. Materials 4, 113609 (2020)) or in foams (L. Viitanen et al. unpublished). or classify the disorder of a material from measured data (S. Papanikolaou, npj Computational Materials 4, 27 (2018)). The two key physical ideas are here the presence of a memory of the past of the material - the Process-Structure-Property paradigm of materials science - and that this leads to the possibility to classify materials and even individual samples and tell them from each other. At the best this is since there are hidden different behaviors that originate from separate phases.

What next

Of my co-authors Henri is wrapping soon up his PhD at Aalto University in Finland and he has already published a number of papers on the subject. Lasse became a tenure-track professor at Tampere, Finland and we are still collaborating. Myself, I started to share my time between my appointment at Aalto University in Finland and as the director of a new Center of Excellence in Poland. NOMATEN is at the National Center for Nuclear Research close to Warsaw and is funded among others by the EU H2020 Teaming program and the Foundation for Polish Science and we have as partners CEA from France and VTT from Finland. My interest in this effort stems to a great degree from the possibility of applying material informatics and ML to real materials problems. With Lasse, Stefan Sandfeld (now at Julich, Germany) and Damien Vandembroucq (ESPCI, Paris), we shall organize a CECAM workshop on the symbiosis of plastic deformation and ML in early 2022 in the, hopefully, post-pandemic era (Understanding plastic deformation via Artificial Intelligence).

 

 

Please sign in or register for FREE

If you are a registered user on Electrical and Electronic Engineering Community, please sign in