Thanks to new supercomputer architectures, classical molecular dynamics simulations will soon enter the realm of a thousand billion atoms, never before achieved, thus becoming capable of representing the plasticity of metals at the micron scale. However, such simulations generate a considerable amount of data, and the difficulty now lies in their exploitation in order to extract the statistical ingredients relevant to the scale of "mesoscopic" plasticity (the scale of continuous models).
The evolution of a material is complex, as it depends on lines of crystalline defects (dislocations) whose evolution is governed by numerous mechanisms. In order to feed models at higher scales, the quantities to be extracted are the velocities and lengths of dislocations, as well as their evolution over time. These data can be extracted using specific analysis techniques based on characterization of the local environment ('distortion score', 'local deformation'), a posteriori or in situ during simulation. Finally, machine learning tools can be used to analyze the statistics obtained and extract and synthesize (by model reduction) a minimal description of plasticity for models at higher scales.