What I told NSF I wanted to do in 2019

Background

Ammonia (NH₃) is often credited as one of the essential compounds that has powered the last century, widely used in agriculture as fertilizer to support a growing human population and in industry as a building block of most synthetic nitrogenous compounds. [1] Thanks to its high energy density and hydrogen content, ammonia has also recently gathered attraction as a promising energy storage intermediate in the hydrogen economy. [1] The production of NH₃ is dominated industrially by the Haber-Bosch process, an energy-intensive nitrogen fixation reaction that is conducted at extremely high pressures (200-300 atm) and temperatures (~780K). [2] The extreme reaction conditions and hydrogen gas feedstock make the reaction responsible for 1-2 % of the world’s annual energy consumption. [2] Electrocatalytic nitrogen reduction reaction (NRR) has emerged as an attractive candidate that can produce NH₃ under ambient conditions. NRR avoids the high-purity H₂ input feed and instead utilizes H₂O in the following reaction:

N₂ + 3H₂O → 2NH₃ + 32O₂

In the last half-decade, there has been an effort to identify effective catalysts that have high faradaic efficiency but also suppress competing reactions such as the hydrogen evolution reaction (HER). Several noble metals and bimetallic catalysts [3] have been explored experimentally, but the poor stability, high cost, and deactivation are obstacles for their large-scale deployment. [5] Metal-organic-framework-derived (MOF) carbon catalysts have recently gathered attention for their low manufacturing cost and high relative stability. [4] Due to a large combination of heteroatom doping and vacancy ratios, experimentally exploring the full chemical space for MOF-derived carbon catalyst is infeasible. Therefore, there is a real demand for a high-throughput screening method that can prioritize promising candidates for future experimental validation.

Hypothesis

I propose a machine learning (ML) guided DFT model that can explore MOF-derived catalysts tuned with doping and oxygen/nitrogen vacancies. To automate catalyst discovery, ML models can act as surrogate models to guide the optimization towards more attractive candidates between each DFT task. The defining features of the proposal aim to 1) develop a computational model that select promising MOF-derived carbon catalyst based on key DFT calculations, 2) tune catalysts by introducing vacancies and heteroatom doping, and 3) obtain the complete energy profile of selected catalysts. If successful, this proposal will accelerate the search towards optimal, low-cost catalysts for NRR and shed light on the reaction mechanisms of promising candidates.

Aim 1 – Develop a computational model that selects promising catalysts with key DFT parameters

For most catalyst discovery or design studies, we rely on chemical intuition with preliminary experimental data to select the next viable candidate for investigation. However, to screen thousands of combinations, an optimal guide is required. By training a machine learning algorithm, we can make more intelligent choices based on experimental data and previous DFT calculations. To do so, we must represent every catalyst activity with numerical parameters, such as adsorption energy, Pauling electronegativity, and coordination number.

Catalytic activity modeled in DFT is often described in literature by two parameters: adsorption energy of nitrogen (E_{ads, N}) to describe NRR and the Gibbs free energy change of absorbed H* (ΔG_H*) to describe HER. Vienna ab initio simulation package (VASP) will be used as the periodic DFT package. The exchange-correlation function under the generalized gradient approximation with norm-conserving pseudopotentials and the Perdew–Burke–Ernzerhof function was adopted to describe the electron–electron interaction, similar to previous DFT studies. [5] Scripts will be developed to automate DFT submission using a workflow package such as Luigi or FireWorks.

Aim 2 – Tune catalysts by introducing vacancies and heteroatom doping

Once several promising catalyst structures have been identified, the catalyst structures will be modified by inserting different doping ratios and vacancies to identify characteristics that are more selective towards NRR and less selective towards HER. Structures from The Materials Project will be extracted using its API and modified before the automatic submission to the DFT-ML model from Aim 1. Doping ratios and nitrogen/oxygen vacancies will be conjoined to the ML descriptors to help guide the next set of DFT calculations.

Aim 3 – Obtain complete energy profiles of promising catalysts

After locating both the doping/vacancy ratios and the structure of promising catalysts, I will obtain the complete energy profiles based on two accepted mechanisms for NRR: distal and alternating. DFT calculations of each adsorption step will map a catalytic mechanism for promising candidates, which will aid experimentalists during future experimental validation. Finding potential-determining steps and species will be essential in experimental validation.

References

[1] Science, 2018, 360, 6391.

[2] Energy Environ. Sci., 2018, 11, 45.

[3] J. Mater. Chem. A, 2019, 7, 801.

[4] Inorg. Chem. Front., 2019, 6, 391.

[5] Nano Energy, 2018, 48, 217-226.