Research ArticleMATERIALS SCIENCE

Unfolding adsorption on metal nanoparticles: Connecting stability with catalysis

See allHide authors and affiliations

Science Advances  13 Sep 2019:
Vol. 5, no. 9, eaax5101
DOI: 10.1126/sciadv.aax5101
  • Fig. 1 Demonstration of CElocal as a descriptor for adsorption energy.

    (A) The BE of CO on various sites of Au NPs as a function of CElocal: 172-atom cube (rectangles), 147-atom icosahedron (hexagons), and 147-atom cuboctahedron (rhombus). Heat map of different sites on the NPs with respect to their BE of CO (B to D) and to their CElocal (E to G). The color scheme follows the range of strongest CO binding to weakest CElocal (violet) and of weakest binding to strongest CElocal (red).

  • Fig. 2 Parity plot of the model-predicted BE of adsorbates (OH, CO, and CH3) on various metal systems versus the DFT BE (eV).

    (A) The model both trained and tested on PBE DFT data for NPs (Au/Ag/Cu, 55 to 172 atoms), which includes the CENP term. This model corresponds to case (i) of Table 1. (B) The model both trained and tested on PBE DFT data for NPs (Au/Ag/Cu, 55 to 172 atoms), which does not include the CENP term. This model corresponds to case (ii) of Table 1. (C) The model trained on PBE DFT data for NPs (Au/Ag/Cu, 55 to 172 atoms) and tested against RPBE DFT data for top-site adsorptions on metal surfaces (Au/Ag/Cu) from the literature slab dataset (29). This model corresponds to case (ii) of Table 1. (D) The model both trained and tested on RPBE DFT data for top-site adsorptions on metal surfaces (Au/Ag/Cu) from the slab dataset. This model corresponds to case (iii) of Table 1. In all cases, error bars are determined from the 10-fold cross-validated RMSE on the training set. Our DFT-calculated BEs of the different adsorbates on the various sites of the metal NPs are shown in table S6.

  • Fig. 3 Parity plot between our developed model and DFT calculations on icosahedral bimetallic (Cu55−xAgx, x = 24, 33) NPs.

    The model is trained on CH3, CO, and OH adsorbing on monometallic Ag, Cu, and Au NPs and is able to capture adsorption on bimetallic NPs. Images of the two NPs are shown as inset, with copper and silver atoms colored in brown and gray, respectively.

  • Fig. 4 Our three-descriptor model extended to slab dataset.

    (A) The model trained on the slab dataset (29) on Cu, Ag, and Au surfaces and tested against the Rh, Ir, Ni, Pd, Pt, Cu, Ag, and Au surfaces from the slab dataset. (B) The equivalent model when trained separately for each column of the d-block, still using the slab dataset. Error bars in every case are the 10-fold cross-validated RMSE of the training set.

  • Fig. 5 Extension of our model to Rh and NH3.

    (A) The model parameterized on our Ag, Cu, and Au NPs adsorbing CH3, CO, and OH and tested against Rh and NH3. (B) The equivalent model with empirical (constant) corrections for Rh and NH3. In the case of NH3 bound to Rh, both corrections are simultaneously applied and indicated by two-colored dots. (C) The model trained on CH3, CO, OH, and NH3 adsorbing on icosahedral/cuboctahedral Rh55.

  • Table 1 OLS regression information for (i) four-descriptor model that includes CElocal, IPEA, MADs, and CENP; (ii) three-descriptor model that excludes CENP; and (iii) equivalent three-descriptor model using the slab dataset found in the literature.

    All cases are trained using datasets where CH3, CO, or OH adsorb to Cu, Ag, or, Au (29). The maximum error corresponds to the largest deviation of a single data point. MAE, mean absolute error; RMSE, root mean square error; DOF, degrees of freedom.

    (i) Trained on NPs (four-descriptor model)
    (RMSE: 0.179 eV, MAE: 0.145 eV, R2: 0.936, maximum error: 0.619 eV,
    and remaining DOF: 157)
    Coefficient estimateSEP value
    Intercept1.514770.15876<2 × 10−16
    CElocal−0.14500.016633.85 × 10−15
    IPEA0.331710.01280<2 × 10−16
    MADs0.678580.01522<2 × 10−16
    CENP−0.00020.053880.998
    (ii) Trained on NPs (three-descriptor model)
    (RMSE: 0.179 eV, MAE: 0.144 eV, R2: 0.933, maximum error: 0.619 eV,
    and remaining DOF: 158)
    Coefficient estimateSEP value
    Intercept1.515090.12148<2 × 10−16
    CElocal−0.145020.01410<2 × 10−16
    IPEA0.331710.01274<2 × 10−16
    MADs0.678570.01501<2 × 10−16
    (iii) Trained on slab dataset (three-descriptor model)
    (RMSE: 0.122 eV, MAE: 0.102 eV, R2: 0.979, maximum error: 0.259 eV,
    and remaining DOF: 113)
    Coefficient estimateSEP value
    Intercept1.676770.09220<2 × 10−16
    CElocal−0.145900.01079<2 × 10−16
    IPEA0.287430.01005<2 × 10−16
    MADs0.795160.01187<2 × 10−16

Supplementary Materials

  • Supplementary material for this article is available at http://advances.sciencemag.org/cgi/content/full/5/9/eaax5101/DC1

    Section S1. Thermodynamic data

    Section S2. Leave-one-in tests

    Section S3. Investigations on slabs

    Section S4. Adsorption configurations

    Table S1. Ionization energies and electron affinities for CH3, CO, OH, and NH3.

    Table S2. BC model–calculated CEs of the relevant NPs we investigated.

    Table S3. Calculated local CEs for relevant adsorption sites.

    Table S4. Regression statistics for the various leave-one-in tests we performed.

    Table S5. Regression statistics for intentionally overfit model plot of fig. S3B, with coefficients generated via OLS regression.

    Table S6. DFT-calculated BEs for all studied adsorbate-NP pairs, ordered first by adsorbate, then by morphology, then by element, and lastly by CN.

    Fig. S1. Adsorbate BEs versus BC model–calculated NP CEs.

    Fig. S2. Parity plots for the various leave-one-in tests we performed.

    Fig. S3. Characterization of all metal-adsorbate pairs in the slab dataset (29) simultaneously (e.g., there is one training set, which includes all adsorption interactions from the dataset).

    Fig. S4. Illustration of initial configurations for several DFT calculations performed.

    References (4658)

  • Supplementary Materials

    This PDF file includes:

    • Section S1. Thermodynamic data
    • Section S2. Leave-one-in tests
    • Section S3. Investigations on slabs
    • Section S4. Adsorption configurations
    • Table S1. Ionization energies and electron affinities for CH3, CO, OH, and NH3.
    • Table S2. BC model–calculated CEs of the relevant NPs we investigated.
    • Table S3. Calculated local CEs for relevant adsorption sites.
    • Table S4. Regression statistics for the various leave-one-in tests we performed.
    • Table S5. Regression statistics for intentionally overfit model plot of fig. S3B, with coefficients generated via OLS regression.
    • Table S6. DFT-calculated BEs for all studied adsorbate-NP pairs, ordered first by adsorbate, then by morphology, then by element, and lastly by CN.
    • Fig. S1. Adsorbate BEs versus BC model–calculated NP CEs.
    • Fig. S2. Parity plots for the various leave-one-in tests we performed.
    • Fig. S3. Characterization of all metal-adsorbate pairs in the slab dataset (29) simultaneously (e.g., there is one training set, which includes all adsorption interactions from the dataset).
    • Fig. S4. Illustration of initial configurations for several DFT calculations performed.
    • References (4658)

    Download PDF

    Files in this Data Supplement:

Stay Connected to Science Advances

Navigate This Article