| Popis: |
Machine learning interatomic potentials (MLIPs) have seen significant advances as efficient replacement of expensive quantum chemical calculations. Uncertainty estimations for MLIPs are crucial to quantify the additional model error they introduce and to leverage this information in active learning strategies. MLIPs that are based on Gaussian process regression provide a standard deviation as a possible uncertainty measure. An alternative approach are ensemble-based uncertainties. Although these uncertainty measures have been applied to active learning, it has rarely been studied how they correlate with the error, and it is not always clear whether active learning actually outperforms random sampling strategies. We consider GPR models with Coulomb and SOAP representations as inputs to predict potential energy surfaces and excitation energies of molecules. We evaluate, how the GPR variance and ensemble-based uncertainties relate to the error and whether model performance improves by selecting the most uncertain samples from a fixed configuration space. For the ensemble based uncertainty estimations, we find that they often do not provide any information about the error. For the GPR standard deviation, we find that often predictions with an increasing standard deviation also have an increasing systematical bias, which is not captured by the uncertainty. In these cases, selecting training samples with the highest uncertainty leads to a model with a worse test error compared to random sampling. We conclude that confidence intervals, which are derived from the predictive standard deviation, can be highly overconfident. Selecting samples with high GPR standard deviation leads to a model that overemphasizes the borders of the configuration space represented in the fixed dataset. This may result in worse performance in more densely sampled areas but better generalization for extrapolation tasks. |
