Effect of Paired Aluminum Sites on the Diffusion of [Cu(NH₃)₂]⁺ in Cu-CHA Predicted by Machine Learning Accelerated Molecular Dynamics

The selective catalytic reduction (SCR) in which ammonia is used as reducing agent (NH3-SCR-NOx) is the current state-of-the-art technology to cope with nitrogen oxides (NOx) emissions from stationary power plants and diesel vehicles[1,2]. The materials currently employed as catalysts in transport applications are copper-exchanged zeolites, in particular those possessing small-pore structures such as the CHA [3,4]. Operando X-ray absorption and emission (XAS/XES) spectroscopic studies have shown that at low temperature (T &lt; 473 K) ammonia liberates the Cu⁺ ions from its coordination with the zeolite framework forming mobile [Cu(NH₃)₂]⁺ species. These mobile amino-copper complexes are responsible for the activation of O₂ through the formation of the transient dimeric species Cu(NH₃ )₂ -O₂ -Cu(NH₃)₂ [5-7]. The diffusion of the monomeric [Cu(NH₃)₂]⁺ species to the adjacent cavity through the 8r window is the rate-determining step at low Cu loading. Previous works have shown that the reacting molecules can have a significant impact in the activation free energy and that the diffusion of such species is limited to neighboring cavities[7,8]. The aluminum distribution is another factor that possibly affects the reaction rate but so far remains unexplored[9].<br/>The diffusion of [Cu(NH₃)₂]⁺ complexes through the 8r windows of the CHA is an activated process and therefore the theoretical study of such diffusion requires the use of enhanced sampling techniques. However, the construction of free energy profiles based on DFT is limited to the most relevant cases because of the significant computational cost of umbrella sampling simulations. Machine learning (ML) techniques have already taken special relevance in the area of materials science and heterogeneous catalysis.These techniques provide an opportunity to speed up the costly Born-Oppenheimer molecular dynamics (BOMD) simulations by several orders of magnitude [10]. NN-based interatomic potentials (NNIPs) retain the accuracy of the ab initio training data and can be executed at a computational cost comparable with that of classical force fields.<br/>In this work, we applied NNIPs-based enhanced umbrella sampling (US) simulations to obtain Gibbs free energy profiles for the diffusion of [Cu(NH₃)₂]⁺ complexes through the 8-ring windows of SSZ-13 zeolite for different Al distributions at 423 K representative of the low temperature regime. All interatomic potentials converted to sub-kcal mol⁻¹ accuracy for the energies on the corresponding validation set, with a mean absolute error of 0.34 kcal/mol, respectively. The NNIPs are able to reproduce the free energy profiles with an error of 0.4 kcal/mol relative to the DFT counterpart (see Figure 1), at least 400 times faster. The calculated free energy of activation for the diffusion of an isolated [Cu(NH₃)₂]⁺ complex through the 8-ring windows of the chabazite structure is relatively low, 4.3 kcal/mol, showing that this process is fast. An increase in the free energy of activation to 6.7 kcal/mol is observed when a second complex and therefore a second Al are placed in the adjacent cavity (Figure 2). However, when the two Al atoms are placed in the same 8r, through which the diffusion takes place, the free energy of activation decreases by ~1.7 kcal/mol (Figure 2), with which the reaction would proceed ~10 times faster. Our results show that a fine tuning of the Cu-CHA catalyst can be achieved by controlling the Al distribution, that is, the selective positioning of Al pairs in the same 8-ring window.