(Cross-post from our report published in the Psi-k blog)
From 20th until 24th June 2022 I co-organised a workshop on the theme of Error control in first-principles modelling at the CECAM Headquarters in Lausanne (workshop website). For one week the workshop unified like-minded researchers from a range of communities, including quantum chemistry, materials sciences, scientific computing and mathematics to jointly discuss the determination of errors in atomistic modelling. The main goal was to obtain a cross-community overview of ongoing work and to establish new links between the disciplines.
Amongst others we discussed topics such as: the determination of errors in observables, which are the result of long molecular dynamics simulations, the reliability and efficiency of numerical procedures and how to go beyond benchmarking or convergence studies via a rigorous mathematical understanding of errors. We further explored interactions with the field of uncertainty quantification to link numerical and modelling errors in electronic structure calculations or to understand error propagation in interatomic potentials via statistical inference.
- Gabor Csanyi (University of Cambridge)
- Genevieve Dusson (CNRS & Université Bourgogne Franche-Comté)
- Michael Herbst (RWTH Aachen University)
- Youssef Marzouk (Massachusetts Institute of Technology)
A primary objective of the conference was to facilitate networking and exchange across communities. Thanks to the funds provided by CECAM and Psi-k we managed to get a crowd of 30 researchers, including about 15 junior researchers, to come to Lausanne in person. Moreover we made an effort to enable a virtual participation to the smoothest extent possible. For example we provided a conference-specific Slack space, which grew into a platform for discussion involving both in-person as well as virtual participants during the conference. In this way in total about 70 researchers from 18 countries could participate in the workshop. The full list of participants is available on the workshop website.
The workshop programme was split between the afternoon sessions, in which we had introductory and topic-specific lectures, as well as the morning sessions, which were focussed on informal discussion and community brainstorming.
Monday June 20th 2022
- Uncertainty quantification for atomic-scale machine learning. (Michele Ceriotti, EPFL)
- Testing the hell out of DFT codes with virtual oxides. (Stefaan Cottenier, Ghent University)
- Prediction uncertainty validation for computational chemists. (Pascal Pernot, Université Paris-Saclay)
- Uncertainty driven active learning of interatomic potentials for molecular dynamics (Boris Kozinsky, Harvard University)
- Interatomic Potentials from First Principles (Christoph Ortner, University of British Columbia)
Tuesday June 21st 2022
- Numerical integration in the Brillouin zone (Antoine Levitt, Inria Paris)
- Sensitivity analysis for assessing and controlling errors in theoretical spectroscopy and computational biochemistry (Christoph Jacob,
- Uncertainty quantification and propagation in multiscale materials modelling (James Kermode, University of Warwick)
- Uncertainty Quantification and Active Learning in Atomistic Computations
(Habib Najm, Sandia National Labs)
- Nuances in Bayesian estimation and active learning for data-driven interatomic potentials for propagation of uncertainty through molecular dynamics
(Dallas Foster, MIT)
Wednesday June 22nd 2022
- The BEEF class of xc functionals (Thomas Bligaard, DTU)
- A Bayesian Approach to Uncertainty Quantification for Density Functional Theory (Kate Fisher, MIT)
- Dielectric response with short-ranged electrostatics (Stephen Cox, Cambridge)
- Fully guaranteed and computable error bounds for clusters of eigenvalues (Genevieve Dusson, CNRS)
- Practical error bounds for properties in plane-wave electronic structure calculations (Gaspard Kemlin, Ecole des Ponts)
- The transferability limits of static benchmarks (Thomas Weymuth, ETH)
Thursday June 23rd 2022
- An information-theoretic approach to uncertainty quantification in atomistic modelling of crystalline materials (Maciej Buze, Birmingham)
- Hyperactive Learning (Cas van der Oord, Cambridge)
- Benchmarking under uncertainty (Jonny Proppe, TU Braunschweig)
- Model Error Estimation and Uncertainty Quantification of Machine Learning Interatomic Potentials (Khachik Sargsyan, Sandia National Labs)
- Committee neural network potentials control generalization errors and enable active learning (Christoph Schran, Cambridge)
Morning discussion sessions
The discussion sessions were centred around broad multi-disciplinary topics to stimulate cross-fertilisation. Key topics were active learning techniques for obtaining interatomic potentials on the fly as well as opportunities to connect numerical and statistical approaches for error estimation.
A central topic of the session on Thursday morning was the development of a common cross-community language and guidelines for error estimation. This included the question how to establish a minimal standard for error control and make the broader community aware of such techniques to ensure published results can be validated and are more reproducible. Initial ideas from this discussion are summarised in a public github repository. With this repository we invite everyone to contribute concrete examples of the error control strategies taken in their research context. In the future we hope to community guidelines for error control in first-principle modelling based on these initial ideas.
Feedback from participants
Overall we received mostly positive feedback about the event. Virtual participants enjoyed the opportunity to interact with in-person participants via the zoom sessions and Slack. For several in-person participants this meeting was the first physical meeting since the pandemic and the ample opportunities for informal interchange we allocated in the programme (discussion sessions, poster sessions, social dinner, boat trip excursion) have been much appreciated.
A challenge was to keep the meeting accessible for both researchers from foreign fields as well as junior participants entering this interdisciplinary field. With respect to the discussion sessions we got several suggestions for improvement in this regard. For example it has been suggested to (i) set and communicate the discussion subject well in advance to allow people to get prepared, (ii) motivate postdocs to coordinate the discussion, which would be responsible to curate material and formulate stimulating research questions and (iii) get these postdocs to start the session with an introductory presentation on open problems.
Conclusions and outlook
During the event it became apparent that the meaning associated to the term “error control” deviates between communities, in particular between mathematicians and application scientists. Not only did this result in a considerable language barrier and some communication problems during the workshop, but it also made communities to appear to move at different paces. On a first look this sometimes made it difficult to see the applicability of research results from another community. But the heterogeneity of participants also offered opportunities to learn from each other's viewpoint: for example during the discussion sessions we actively worked towards obtaining a joint language and cross-community standards for error control. Our initial ideas on this point are available in a public github repository, where we invite everyone to participate via opening issues and pull requests to continue the discussion.