At COSMO we perform bottom-up atomistic modeling of condensed matter and molecular systems, focusing on method development and statistical sampling. Our goal is to understand the mechanisms that underlie the macroscopic properties of materials and improve them through rational design.
We tackle this challenge through a combination of fundamental science and data-driven approaches. By incorporating physics and chemistry into machine learning methods we build accurate and predictive surrogate models. These models allow us to run reliable simulations that bridge different time and size scales and capture subtle physical effects and complex, large-scale structural features. We employ these simulations as a tool of discovery: by improving the fidelity of the simulations, we aim to achieve a better understanding of materials’ properties and a reliable description of physical reality. As we believe that research is a collective endeavour, we develop open-source tools for the benefit of other researchers and the larger community.
Creating a “plug and play” framework to combine machine learning with physics-based modeling
investigating the fundamental questions related to the mathematical representations of structures at the atomic scale
Aiming to blur the line between data-driven and physics-based models
Developing surrogate models to simulate materials’ properties and structure
In COSMO we develop software to make atomistic simulations via machine learning easier. In particular, we are currently working on the following projects:Equistore: a specialized data storage format for all your atomistic machine learning needs, and more. Think numpy ndarray or pytorch Tensor equipped with extra metadata for atomic — and other particles — systems. You can find (…)