Article
(
SnCl ·2H O, CoCl ·6H O, NiCl ·6H O, ZnCl , FeSO ·7H O,
2 2 2 2 2 2 2 4 2
AgNO , In(NO ) ·xH O, Pd(NO ) , Bi(NO ) , Ce(NO ) ,
3
3
3
2
3
2
3
3
3 3
HAuCl ) and/or a 10 mM concentration of organic additive in 0.1
4
M KCl aqueous solution.
Synthetic Conditions for Cu-Mixed Catalyst. Cu-mixed samples
were electrodeposited in a plating bath made from the above Cu-
mixed precursor. Carbon cloth was used as a substrate for
electrodeposition. Cu-mixed catalyst was in situ electrodeposited at
a constant potential −1.0 V vs Ag/AgCl until a final deposition charge
2
of 1.9 C/cm was reached on the carbon cloth. A graphitic sheet was
used as the counter electrode, and a Ag/AgCl electrode was placed
near the working electrode as the reference electrode.
Electrochemistry. Electrochemical measurements were performed
using a three-electrode system and an electrochemical workstation
(CHI660E). Electrolysis was performed at room temperature in a H
type cell with a Ag/AgCl reference electrode and a graphitic sheet
counter electrode. The cathode and anode compartments were
separated by a proton exchange membrane (Nafion 117). The
potentials on the working electrodes were converted to voltages with
respect to the RHE reference electrode by E (vs the RHE) = E (vs the
room temperature in a flow cell with an Ag/AgCl reference electrode
and a nickel foam counter electrode. Catalytic results with the flow
Figure 6. (a) Relationship between FE-CO RR, FE-H , and the XRD
2
2
phase of the samples. Samples in the red area contain the Cu O phase
2
while samples in the blue area do not. (b) Relationship between FE-
C2+ and the morphology of the samples. Samples in the red area
contain the Cu O cube while samples in the blue area do not. (c)
2
SEM images of sample 167 without the Cu O cube. (d) SEM images
2
of sample 170 with the Cu O cube.
2
Product Analysis. Liquid-phase products were analyzed by proton
nuclear magnetic resonance ( H NMR) spectroscopy (Bruker
AVANCE AV III 500), in which 0.5 mL of the electrolysis solution
1
higher FE-C showed more regular Cu O cubes before
lost after electrolysis of 20 min (Figure S41). Recent literature
was mixed with 0.1 mL of deuterated water (D O) for field locking
2
+
2
2
and 0.02 μL of dimethyl sulfoxide (DMSO) as an internal standard.
1
The H NMR spectrum was measured with water suppression using a
presaturation sequence. Gas-phase products were detected online by
using a gas chromatograph (GC) connected to the headspace of the
electrolysis cell. A thermal conductivity detector (TCD) was used to
quantify hydrogen, and a flame ionization detector (FID) equipped
with a methanizer was used to quantify carbon monoxide, methane,
ethane, and ethylene.
reports also showed that Cu O cubes as a precursor can
2
50
enhance the selectivity of C2+ products. Besides the Cu O
2
cubes, we investigated Cu O nanoparticles enclosed by
2
different crystal facets (nos. 170, 150, 131) for C2+ products.
and selected area electron diffraction (SAED) (Figure S42)
showed that sample 170 has exposed {001} facets, and sample
Machine Learning Methods. Feature Selection Using a
Random Intersection Tree. A “random intersection tree” (RIT) can
48
1
1
50 has some {11−2} facets, while sample 131 has some {1-1-
quickly extract interaction-based feature combinations. The target
property (denoted y) for the RIT analysis is labeled either ‘1’ (positive
sample) or ‘0’ (negative sample). Here we chose Faradaic efficiency as
the target property. We set a criteria for Faradaic efficiency, above
which y is labeled ‘1’ while below it is labeled ‘0’ to convert the data
into binary combinations that are suitable for RIT analysis. A positive
sample was then randomly chosen as the root node to construct a tree
} facets besides the {001} facets. Although their exposed
crystal faces are different, they all show similarly high FE-C2+
(
By contrast, sample 167 (Figure S43) contained fragmented
polycrystalline Cu O nanoparticles and showed low FE-C
2
2+
(167:0%). We found that catalysts with higher FE-C2+ showed
(
RIT). A child node of the root node was constructed by randomly
more regular Cu O crystals in the precursor before catalysis
2
choosing another positive sample and calculating the common
features of the root sample and the newly chosen one, which is
called an intersection operation. These common features constitute
one child node. Several other children nodes were obtained by
randomly choosing other positive samples to undergo the intersection
operation with the common root node to complete a second layer of
the tree. The third layer of the tree was constructed by repeating the
intersection operation between the second layer and randomly chosen
positive samples. The algorithm will stop if the selected features of a
node do not show up significantly more often in positive samples than
in negative samples. A leaf node thus contains a feature subset that is
important in determining the target property. Negative samples are
also used for the intersection operation but with a modified rule.
Features appearing in a negative sample will be given penalties to
decrease their likelihood to constitute positive subsets.
while it seems that there is no direct correlation between the
exposed crystal face and FE-C . This finding is different from
2+
the reported correlation between the exposed crystal planes of
51,52
Cu O and FE-C .
2
2+
EXPERIMENTAL METHODS
■
Materials and Apparatus. Reagents were commercially available
and used without further purification unless otherwise noted. Powder
X-ray diffraction (PXRD) was carried out on a Rigaku DMax-γA
rotation anode X-ray diffractometer equipped with graphite
monochromatized Cu Kα radiation (γ = 1.54 Å). IR spectra were
obtained on a Nicolet iS50 FTIR spectrometer. Scanning electron
1
microscopy images were obtained on a Zeiss sigma. H NMR spectra
were recorded on a Bruker NMR 500 DRX spectrometer at 500 MHz.
Transmission electron microscopy (TEM) images were acquired on a
JEOL 2100 high resolution transmission electron microscope. The
XPS measurements were performed using a PHI Quantum 2000
instrument.
In building the RITs, the number of branches from each nonleaf
node is set to 3, the maximum depth is set to 9, and the maximum
number of features in a returned subset is set to 15. Positive samples
are randomly collected as “root nodes” to produce 1000 RITs. All
these trees acted like a forest to yield 1495 possible subsets.
Synthetic Procedures. Synthetic Conditions for Cu-Mixed
48
Precursor. Cu-mixed precursors were made from 0.1 M CuSO ·
An “I-score” is used in the RIT analysis, which can be computed
4
5
H O with or without a 10 mM concentration of another metal salt
by the formula
2
5
759
J. Am. Chem. Soc. 2021, 143, 5755−5762