(246f) Active Machine Learning Model for the Discovery of Novel Material Systems and Its Application to Electrocatalysis

Authors: 
Flores, R., University of Kansas
Paolucci, C., University of Virginia
Bajdich, M., SLAC National Accelerator Laboratory
Jain, A., DTU
Aykol, M., TRI
Norskov, J. K., Technical University of Denmark
Bligaard, T., SLAC National Accelerator Laboratory

The design and discovery of new material systems with desired chemical and physical properties is imperative towards addressing global energy challenge. Traditional paradigms in material development are often slow, labor intensive and are not guaranteed to fully explore the expansive chemical space. Herein, we report on a robust and generalizable active learning framework for the discovery of novel and stable bulk crystal structures in the full space of crystal structural prototypes of fixed stoichiometry. Furthermore, we demonstrate the utility of this framework by the search in space of Ir-oxides, which has direct relevance to materials application as oxygen evolution reaction (OER) and oxygen reduction reaction (ORR) catalyst and support.

An active surrogate machine learning algorithm employing a Gaussian kernel was implemented to perform a rapid exploration of the Ir-oxide chemical space. The structural space was populated by all unique AB2 and AB3 structures from Materials Project and Open Quantum Materials Database (OQMD) databases using the prototype generator of Jain, et al available at https://www.catalysis-hub.org/prototypeSearch [1][2]). Since this approach can be easily generalized to rigorously enumerate a full structural space it also guarantees the completeness of our search. We show that the active learning framework can select the most valuable structures to perform DFT calculations on by implementing a simple acquisition criteria that biases towards low energy structures. The surrogate model resolves the most stable materials in the target space with a minimal number of DFT calculations (enough for the model’s accuracy in the low formation energy region to be sufficiently low). Next, the subset of most stable IrO2 and IrO3 structures was computationally tested for their efficacy towards the electrochemical OER. The results for our bulk and surface calculations show the importance of highly oxided Ir-phases and provide atomistic explanation of the observed high stability and activity of the previously reported IrOx/SrxIrO3 system [3]. This work demonstrates the importance of data science approaches to accelerate catalyst discovery.

References

[1] Jain A., Bligaard T, Physical Review B 2018, 98, 214112.

[2] Winther K., Hoffmann M., Mamun O., Boes J., Nørskov J., Bajdich M., Bligaard T., ChemRxiv 2018,

[3] Seitz L., Dickens C., Nishio K., Hikita Y., Montoya J., Doyle A., Kirk C., Vojvodic A., Hwang, H., Nørkov J., Jaramillo T., 2016, 353, 6303