HetSys News

Attending the CECAM workshop Uncertainty Quantification in Atomistic Modelling was a valuable opportunity to engage with current developments in the field and to better understand how uncertainties are approached within different research groups and themes. The programme covered a broad range of topics, from uncertainty-aware density functional theory to machine-learning-based interatomic potentials and highlighted both the technical challenges and the practical considerations involved in making atomistic simulations more reliable. The talks and discussions provided a clear picture of the methodologies currently available, their limitations, and where future work is heading.

I presented a poster on my research into model-misspecification-aware uncertainty quantification. This allowed me to showcase the challenges that arise when the underlying model class cannot represent all of the relevant physics — an issue that is often overlooked when using standard uncertainty estimates. Several attendees expressed interest in how such misspecification can be detected or mitigated in practice, and our conversations helped refine my thinking about how to frame the problem more clearly and how to evaluate proposed solutions. In turn, presenting the poster gave me an opportunity to make others aware of the tools and approaches I am developing, and how they might apply them in their own modelling workflows.

The workshop also helped situate my work within the broader landscape. Hearing about recent efforts in uncertainty-aware DFT, active-learning strategies, and uncertainty calibration for machine-learning models gave me a better understanding of how my ideas fit into the emerging push towards end-to-end uncertainty management. Overall, the event strengthened my confidence in the relevance of my research direction and provided helpful pointers on how to prioritise the next steps.