(393i) DFT and Machine Learning Investigations on the Sorptive Removal of Trace Contaminants from Syngas

To utilize the syngas derived from the gasification of carbonaceous feedstocks for the production of valuable chemicals, the syngas has to be free from various toxic trace contaminants such as Hg, P Se, As, Cd, and Pb. These elements get released during gasification, incineration, or pyrolysis in their elemental form, hydride form (PH₃, H₂Se, AsH₃), or sulfide form (PbS). They need to be removed since they are potential poisons to the catalyst(s) employed downstream for syngas processing [1]. The commercially available carbon and sulfur-based adsorbents are inefficient for the removal of trace contaminants at high temperatures. In the present work, the sorption of various gas-phase contaminants, viz., Hg⁰, PH₃, H₂Se, AsH₃, Cd, and Pb is studied on monometallic and bimetallic surfaces, and machine leaning models are developed to predict the adsorption energy of the contaminants on bimetallic surfaces. Insights are obtained into the dissociation of AsH₃, H₂Se, and PH₃ on metals and into the segregation behaviour of Hg, As, Se, P, Cd, and Pb in alloys.

Methodology

First-principles density functional theory (DFT) calculations are performed to calculate the adsorption energy and to evaluate the most stable geometric configurations for the adsorption of gas-phase Hg⁰, PH₃, H₂Se, AsH₃, Cd, and Pb on various metals and bimetallic surfaces. The calculations are carried out using the Vienna ab initio simulation package (VASP) and by employing the Perdew-Burke-Ernzerhof (PBE), optPBE-vdW, and PBE-D3 functionals of the generalized gradient approximation. The reaction energy pathway for PH₃ dissociation on various metals is studied using the CI-NEB method. PH₃ dissociation on the (1 1 1) surface of various metals and Ru(0 0 0 1) surface is considered to occur by the sequence of the elementary reactions described by Eqs. (R1)-(R4), where the subscript ‘g’ represents the gas-phase species and the subscript ‘ad’ represents the adsorbed surface species.

PH_{3 (g)} → PH_{3 (ad)} (R1)

PH_{3 (ad)} → PH_{2 (ad)} + H _(ad) (R2)

PH_{2 (ad)} → PH _(ad) + H _(ad) (R3)

PH _(ad) → P _(ad) + H _(ad) (R4)

Similarly, calculations are performed considering the step-wise abstraction of H atoms from H₂Se and AsH₃ as well, and the reaction energy barriers and the reaction energies for the elementary reactions are calculated.

Machine learning models are developed to predict the DFT-calculated adsorption energy of trace contaminants on various bimetallic surfaces using scikit-learn in Python [2]. Linear algorithms viz., ordinary linear regression (OLR), partial least-squares (PLS) regression, and least absolute shrinkage and selection operator (LASSO) regression, and non-linear tree ensemble algorithms viz., extra tree regression (ETR), random forest regression (RFR), and gradient boosting regression (GBR) are used. A set of 29 descriptors comprising of physical properties of the constituent metals in the adsorbents, electronic properties of the adsorbent surfaces, and a few other properties of the adsorbent surface are initially considered. From this initial set of descriptors, 17 descriptors are shortlisted on the basis of their non-collinearity.

Results and Discussion

The adsorption of Hg⁰, Cd, and Pb is the most favourable on the hollow sites of most metal surfaces, whereas the adsorption of hydrides PH₃, H₂Se, and AsH₃ is favourable on the top sites of most metals considered in this study. The adsorption of gas-phase Hg, PH₃, H₂Se, AsH₃, Cd, and Pb is exothermic on various metals. Further, Rh, Ru, Pd, Pt, and Ir are found to be the most promising metals for the removal of these contaminants from gas streams. The adsorption energies of PH_x (x=1, 2, 3) on various metals scale with the adsorption energy of elemental P. The stepwise dissociation of PH₃ is evaluated on various metals by reactions R1-R4, and the reaction energy diagram is shown in Fig. 1. The analysis of the calculated reaction energies and activation barriers shows that PH₃ dissociation is more facile on Rh, Ru, Pd, Ir, and Pt as compared to Cu, Ag, and Au. Similar results are observed for the dissociation of H₂Se and AsH₃. The segregation energy for the elements present on the surface of various alloys is plotted in Fig. 2 (a). The negative values of segregation energy indicate the tendency of elements Hg, As, Se, P, Cd, and Pb to segregate towards the surface of Ru, Rh, Pt, Ir, Pd, Cu, Ag, and Au after alloy formation. The formation energy calculations indicate the possibility of the formation of alloys between metals Ag, Au, Cu, Ir, Pd, Pt, Rh, and Ru and elements Hg, As, Se, P, Cd, and Pb (Fig. 2 (b)).

DFT calculations are performed to evaluate the binding of the trace contaminants on 56 different compositions of bimetallic adsorbents. Bimetallic adsorbents with an overlayer of Ru, Rh, Ir, or Pt on Ag or Au exhibit a higher affinity for PH₃ as compared to the monometallic adsorbents (Fig. 3). Similar results are obtained for the adsorption of Hg, H₂Se, AsH₃, Cd, and Pb. The adsorption of these contaminants on the overlayered bimetallics is stronger than that on monometallic surfaces due to the ligand and strain effects.

Six ML algorithms are used to predict the adsorption energy of various trace contaminants on overlayered bimetallic surfaces using the 17 shortlisted descriptors as inputs. The tree ensemble methods, viz. ETR and RFR, result in the best prediction with a low value of prediction error (RMSE_test) (Fig. 4 (a)). Moreover, the standard deviation for RMSE_test is quite low for the tree ensemble models. The significance of each of the 17 shortlisted descriptors in predicting the adsorption energy of trace contaminants is investigated using the feature importance score in the tree-based ETR algorithm, and the top three descriptors are evaluated to be the upper edge of the d-band ( ), group number of the constituent metal present in the overlayered bimetallic surface, and d-band center ( ). Further, the average RMSE_test using the ETR model increases marginally (from 16.2 to 17.4 kJ/mol for AsH₃) when the number of descriptors is reduced from 17 to 3 (Fig. 4 (b)). Thus, the ETR model can be used for an accurate prediction of the adsorption energy of AsH₃ and other trace contaminants using only three input parameters.

Significance of the work

Metals and novel bimetallic adsorbents are identified for the removal of various trace contaminants, and the energetics involved in the dissociation of PH₃, H₂Se, and AsH₃ on metals is shown. Tree-ensemble-based machine learning models are developed to accurately predict the adsorption energy of various trace contaminants on bimetallic adsorbents.

References

[1] C.J. Howard, R.A. Dagle, V.M. Lebarbier, J.E. Rainbolt, L. Li, D.L. King, Progress toward biomass and coal-derived syngas warm cleanup: Proof-of-concept process demonstration of multicontaminant removal for biomass application, Ind. Eng. Chem. Res. 52 (2013) 8125–8138.

[2] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, É. Duchesnay, Scikit-learn: machine learning in python, J. of Machine Learn. Res. 12 (2011) 2825–2830.