Machine Learning for Discrete Granular Media

Granular material behavior is fundamental to many problems we encounter in society: dikes can fail because sand grains lose stability and form failure zones under rising water levels; foundations may become unstable due to the shaking of the soil; the quality of construction materials such as concrete requires the perfect mixing of the ingredients. All these problems are related to understanding how grains move and interact with each other. Computational methods, such as discrete element methods, can be used to simulate millions of grains individually and investigate how they influence the processes at the macroscale. However, these simulations are expensive to run, hard to interpret, and therefore difficult to use for practitioners from industry.

Machine learning (ML) can help interpret and make fast and accurate predictions after being trained with data generated from physics-based simulations and laboratory experiments. These ML surrogates make digital twinning affordable for real-life applications, including uncertainty quantification and optimization of industrial/natural processes involving grains. This is a new field of research that requires a collaboration between modelers, data analysts, experimentalists, and practitioners from various engineering disciplines.

The workshop will bring together experts in numerical modeling and machine learning to answer the following questions.

• To what extent can ML help discover the constitutive laws of granular materials from discrete particle data and granular experiments?
• What are the machine learning techniques best suited to emulate the temporal and spatial response of granular materials?
• How to integrate physics-based models and ML surrogates such that uncertainty quantification and optimization can be made affordable?
• What are the steps to take towards an open data infrastructure where data from different software packages are shared and repurposed?