Uncertainty quantification

Fast, propagatable uncertainty estimates for machine learning interatomic potentials.

I work actively on uncertainty quantification (UQ) schemes for machine learning interatomic potentials (MLIPs).

I have developed the shallow ensembles scheme for neural networks: a fast, last-layer ensemble-based UQ scheme for neural-network MLIPs that yields fast yet high-quality uncertainty estimates. Because the ensemble lives only in the last layer, its uncertainties can easily be propagated through arbitrarily complex simulation workflows with very little computational overhead (Kellner & Ceriotti, 2024). A follow-up work discusses how to train such ensembles well (Schäfer et al., 2026).

More recently, we developed the PET-UAFD / PET-EXP scheme, the first universal machine learning potentials of their kind, whose predictions are calibrated against experiment and which yield uncertainty estimates not with respect to DFT but with respect to experiment (Kellner et al., 2026).

UQ schemes will make physical simulations with ML surrogate models more trustworthy. They also enable a better trade-off between more accurate and faster simulations, balancing the various error sources in atomistic simulations.

The supplementary code and data for the shallow ensemble paper are available on GitHub.

References

2026

  1. JCTC
    How to Train a Shallow Ensemble
    Moritz Schäfer, Matthias Kellner, Johannes Kästner, and 1 more author
    Journal of Chemical Theory and Computation, 2026
  2. arXiv
    Errors that matter: Uncertainty-aware universal machine-learning potentials calibrated on experiments
    Matthias Kellner, Teitur Hansen, Thomas Bligaard, and 2 more authors
    arXiv preprint, 2026

2024

  1. MLST
    Uncertainty quantification by direct propagation of shallow ensembles
    Matthias Kellner and Michele Ceriotti
    Machine Learning: Science and Technology, 2024