Identification of systematic DFT-errors and uncertainty estimation in catalyst design using machine learning

Ensuring predictive accuracy in low-cost Machine Learning (ML) models for accelerated design of next-generation catalysts depends explicitly on the choice of ML approach and the accuracy of the training data. As we move towards fully autonomous materials discovery, this will often require accurate calculations at the DFT-level or higher, of catalysts in industrially relevant sizes and conditions, as relevant experimental data may not be available. The DFT-level training data should be accurate enough to ensure predictive accuracy of the ML model, but also independent of the specific exchange-correlation (XC) functional used to generate the training data.
Here, we present a computationally fast approach to identify and correct systematic DFT-errors in catalysis and provide calculational error-bars. We rely on an ensemble of ~2.000 approximate XC functionals generated in the ML-based BEEF-vdW approach, to identify the chemical bonds, which are responsible for the errors and to provide calculational error-bars based on a Bayesian ensemble standard deviation.
We illustrate the approach on two electrochemical model reactions, i.e. the oxygen reduction reaction (ORR) in fuel cells and electro-reduction of CO₂ (CO₂RR) into fuels and chemicals. For ORR, we show that the O-O bond in the OOH* intermediate is not well described by typical XC functionals, but correcting for this error yields improved scaling relations, which are independent of the choice of XC functional.
For CO2RR, we show that the main DFT-error is associated with C=O double bonds rather than the empirically determined OCO backbone. This leads to reduced error-bars in the prediction of CO2RR, which can lead to the identification of new electrocatalysts, e.g. for the production of formic acid.
Finally, we will show how such systematic errors in (electro)catalysis can depend on the specific chemical nature of the catalyst, e.g. metals and oxides.