Iterative Approach of Experiment–Machine Learning for Efficient Optimization of Environmental Catalysts: An Example of NO_x Selective Reduction Catalysts

Abstract

An iterative approach between machine learning (ML) and laboratory experiments was developed to accelerate the design and synthesis of environmental catalysts (ECs) using selective catalytic reduction (SCR) of nitrogen oxides (NO_x) as an example. The main steps in the approach include training a ML model using the relevant data collected from the literature, screening candidate catalysts from the trained model, experimentally synthesizing and characterizing the candidates, updating the ML model by incorporating the new experimental results, and screening promising catalysts again with the updated model. This process is iterated with a goal to obtain an optimized catalyst. Using the iterative approach in this study, a novel SCR NO_x catalyst with low cost, high activity, and a wide range of application temperatures was found and successfully synthesized after four iterations. The approach is general enough that it can be readily extended for screening and optimizing the design of other environmental catalysts and has strong implications for the discovery of other environmental materials.

Short abstract

Minimal research works on the application of artificial intelligence technology in environmental catalyst development. This study reports an iteration approach of a data-driven model with lab experiments to develop a novel catalyst of atmospheric pollutants rapidly.

Introduction

Catalysis-based approaches have been increasingly used in various areas, including pollution control and contaminant treatments.¹⁻³ However, the discovery of novel catalysts is often labor-intensive and time-consuming.^4,5 The trial-and-error approach remains the most common way to design new catalysts, which relies heavily upon the intuitive knowledge and experience of researchers.⁶⁻⁸ On the other hand, a large number of experimental data are available in the literature that may be used to provide insights into the design and optimization of new catalysts.

Modern big data and machine learning (ML) approaches provide a way to accelerate the screening and discovery of novel materials and have already made impacts in areas such as biomaterials, electrocatalysis, and batteries.⁹⁻¹³ For example, Falivene et al. used a ML approach to derive a correlation relationship between the reactivity and structure features, such as the Sterimol steric parameters and average steric occupancy in biomaterials, which were then used to accelerate the screening, modification, and optimization of the metalloproteins applied in biocatalysis.¹⁴ Similarly, Tran et al. used a ML approach and the adsorption data from an established database (Materials Project¹⁵) as fingerprints to successfully screen the candidates for CO₂ reduction.¹⁶ The effectiveness of the ML-based approaches, however, strongly relies on the availability and abundance of data.¹⁷ In practice, the relevant data for a novel catalyst to be developed are inherently limited, because the catalyst does not yet exist. When the relevant data are limited, significant uncertainties would exist in the screening process as demonstrated in a previous study,¹⁸ posing a significant challenge for the practical application of the ML-based approaches for screening novel catalysts. A new approach is therefore required that can not only integrate previous data but also minimize the uncertainty in the screening process for the acceleration of novel catalyst development.

An iterative approach of machine learning and laboratory experiments¹⁹ is therefore proposed for screening and optimizing potential new catalysts for environmental applications. The approach consists of the following steps: (1) training a ML model using experimental data collected from relevant publications, (2) screening candidate catalysts with the trained model, (3) synthesizing and characterizing the candidate catalysts experimentally, (4) updating the ML model by incorporating the new experimental results, and (5) screening candidate catalysts again with the updated model. This process is iterated until a new optimal catalyst with ideal properties is identified and synthesized. The approach was demonstrated using the selective catalytic reduction (SCR) of nitrogen oxides (NO_x) with ammonia as an example. Using the approach, a low-cost, high-performance SCR NO_x catalyst with a wide range of applicable temperatures was successfully identified and experimentally synthesized. The developed approach is general enough that it can be used for screening and optimizing other environmental catalysts and has strong implications for the discovery of other environmental materials.

Iterative Approach

ML-Based Screening Step

The ML part of the iterative approach includes database development, ML model training, and the application of the trained ML model for screening candidate catalysts (Scheme 1). The database development starts with collecting relevant data from the literature as described previously.¹⁸ In this study, the database was iteratively updated using the experimental data generated during this research. For the case of SCR NO_x catalysts, five categories of data comprising composition, structure, morphology, preparation method, and reaction condition that are commonly provided in the literature were collected to establish a raw database. A total of 2748 sets of data from 49 research articles were collected.¹⁸ The collected data sets were organized in a form of a high-dimensional matrix with the five categories of data arranged in columns. The data sets of these five categories consist of 62 independent variables, which were hereafter termed feature variables, and 1 dependent variable that is the NO_x conversion, which was hereafter termed the target variable. The detailed category information and variables with their value ranges are provided in Table S1 of the Supporting Information.

Scheme 1. Schematic Diagram of the Iterative Approach for Environmental Catalyst Screening and Optimization.

Collected Data Sets Used To Train a ML Model

In this study, an artificial neural network (ANN) model was selected that can handle data sets with a large scale and high dimensionality with the goal to establish the correlation relationships linking 62 feature variables with the target variable. The number of hidden layers and neurons was adjusted to optimize the performance of the ANN model based on the values of the correlation coefficient R and root means square error (RMSE) (eqs 1 and 2), respectively. As demonstrated previously,¹⁸ a three hidden-layer structure with 6, 4, and 2 neurons for each layer (Figure S1 of the Supporting Information) was an optimal combination for this problem

1

where P_mesu is the measured value and P_pred is the predicted value. N is the total amount of data

2

where P_mesu, P_pred, P̅_mesu, and P̅_pred are the measured, predicted, average of measured, and average of predicted values, respectively. N is the total amount of data.

The trained ML model was used to search for a set of specific values for the feature variables that can yield candidate catalysts with desired performance. Because a large number of the feature variables were involved in identifying the candidate catalysts, a genetic algorithm (GA)²⁰ was used to search for a set of feature variables that give the desired performance of the corresponding catalysts. The threshold for the desired performance was set to 90% conversion efficiency in this study, so that the pool of the identified candidate catalysts was large enough for further screening. The procedure is described below:

(1) A temperature range relevant to the practical application of the target catalyst was input. For the SCR NO_x catalysts, it is desirable to have a high conversion efficiency over a wide temperature range. A range of 100–300 °C was then selected in this research. (2) The temperature within the range of 100–300 °C was sampled with a step size of 1 °C, and the total number of samples was recorded as N_total. (3) The NO_x conversion at each temperature sample was predicted with the trained ANN model by varying other features. Note that a threshold Thr (90% in this work) was predefined, and the number of the samples that greater than Thr is counted as N_s. (4) The objective function was defined. A parameter-dependent probability P(X_i) for the samples greater than Thr was defined (eq 3) and used as the objective function for the GA-based maximization. (5) The optimization results can be expressed as eq 4

3

where X_i values are the ANN input parameters that need to be optimized

4

where function argmax is an operation that finds the argument that gives the maximum value from the target function P(X_i).

The identified catalysts were further screened for experimental synthesis and characterization. In this step, the appearance of each metal element in all of the candidate catalysts was counted, and those candidate catalysts containing the elements with the highest frequency of appearance were selected as the final list of candidate catalysts for experimental preparation and characterization.

Synthesis and Characterization

The trained ML model as described in the previous section provides a list of candidate catalysts that have the potential for the desired performance in SCR NO_x. The candidates in the list primarily consist of Fe–Mn–Ni catalysts, as described in the Results and Discussion, and consequently, in this experimental procedure description, the synthesis of Fe–Mn–Ni catalysts was provided. The procedure will require modification if other types of catalysts were selected by the trained ML model.

The Fe–Mn–Ni type of catalysts were synthesized by first dispersing Na₂CO₃ (24 mmol) in 50 mL of deionized (DI) water in a three-neck round-bottom flask. Another aqueous solution (50 mL, 1 M in total) of Fe(NO₃)₃·9H₂O, Mn(NO₃)₂, and Ni(NO₃)₂·6H₂O was gradually added (2 mL/min) in proportion to the aqueous Na₂CO₃ solution (50 mL, 0.5 M) to form a solution mixture, which was consistent with the result predicted from the ML model (see Table 3). The mixture solution was stirred (500 rpm) at room temperature and maintained at pH 11.5 (adjusted using 4 M aqueous NaOH solution) for 17 h. The precipitates from the mixture solutions were collected by filtration, washed with DI water until the supernatant had pH 7, and then dried in a vacuum oven at 30 °C overnight. The products were calcined at 500 °C at a ramp of 10 °C min^–1 for 2 h for catalyst activation.

Table 3. Chemical Compositions of the Catalysts Determined with ICP–MS^a.

catalyst	Fe (mol)	Mn (mol)	Ni (mol)
Fe0.4Mn0.15Ni0.45	0.5520	0.1644	0.2835
Fe0.2Mn0.2Ni0.6	0.3114	0.2637	0.4248
Fe0.2Mn0.6Ni0.2	0.2772	0.5951	0.1275
Fe0.319Mn0.206Ni0.475	0.4335	0.269	0.2967

^aNote that catalyst Fe_0.319Mn_0.206Ni_0.475 was identified from the second round of iteration, and the other three catalysts were identified from the first round of iteration.

The catalytic materials were characterized using X-ray powder diffraction (XRD) on an X-ray diffractometer (Rigaku SmartLab 9kW, Japan) using Cu Kα radiation (λ = 1.5418 Å). The accelerating voltage was set at 40 kV with a current of 40 mA and a set scan speed of 5°/min. The catalytic materials were also analyzed using transmission electron microscopy (TEM, FEI Talos F200X) with an accelerating voltage of 300 kV. The contents of Fe, Mn, and Ni in each catalyst sample were quantified using inductively coupled plasma mass spectrometry (ICP–MS, ICAP RQ, Thermo Fisher Scientific).

Nitrogen adsorption–desorption was analyzed using a Micromeritics ASAP 2020 analyzer to obtain the information on the surface area, pore size, and pore volume. The sample (about 150 mg) was degassed for 8 h to remove impurities and moisture and then analyzed and tested at a low temperature of −196 °C using liquid nitrogen as the adsorbate.

NH₃ temperature-programmed desorption (TPD) was performed on a FINESORB-3010C (FANTAI Co., Ltd., China). For each measurement, the catalyst (ca. 40 mg) was pretreated in pure He at 300 °C for 1 h and cooled to 50 °C, and then 5% NH₃/He was injected into the reactor for 30 min, followed by the purge with pure He for 1 h to remove physically adsorbed NH₃. In the analysis, the TCD signal was recorded from 50 to 700 °C at a rate of 10 °C/min.

X-ray photoelectron spectroscopy (XPS) was collected using a spectrometer (PHI 5000 Versaprobe III, Japan) equipped with a hemispherical electron analyzer and an Al Kα radiation source. The binding energies were calibrated internally using the carbon deposit C 1s binding energy of 284.6 eV.

Catalytic Reaction and Performance

The catalytic performance of the synthesized catalyst materials was evaluated in a fixed-bed quartz flow reactor using a simulated flue gas containing 400 ppm of NO_x, 400 ppm of NH₃, and 5% O₂, with N₂ as the balance gas. A fixed bed quartz tube with a diameter of 6 mm was used to place 100 mg (40–60 mesh) samples with the total gas flow rate of 100 mL/min, corresponding to gas hourly space velocity (GHSV) of 60 000 h^–1. The evaluation temperature range was set to 150–450 °C. The NO_x analyzer (CLD60, ECO PHYSICS) was used to measure the outlet concentration of NO, and the steady-state activity data were collected after the reaction stabilized. The NO_x conversion percentage was calculated as follows:

5

The measured performance of the synthesized candidate catalysts was used to update the database together with the feature variables determined from the characterization. The updated database was used to screen candidate catalysts again for experimental characterization and performance measurements. This process is iterated until an ideal catalyst was found.

Results and Discussion

ML Model Training and Testing

MATLAB was used as the computation platform to train the ANN model. The results (Figure 1) show that the correlation coefficients between the measured and predicted values are 0.98 and 0.95 for training and testing steps, respectively, with a normal distribution of residual errors. The result indicated that the model captured the data well (Figure 1A). Figure 1C shows that the RMSE quickly decreased at the beginning from 60 to 20 for the training set, followed by a slow decrease to 7 at about 160 epochs, and then stabilized. The best performance has a RMSE value of 6.32 at 989 epochs.

Figure 1.

ANN model training results: (A) training and test results, (B) error histogram, and (C) training performance.

Screening Candidate Catalysts

The application of the trained ML model was demonstrated in screening NO_x SCR candidate catalysts with a desired composition of chemical elements. The database for NO_x SCR catalysts contains 25 elements among 62 variables. Some elements were excluded, such as alkali metals (Na, K, etc.) and alkaline earth metals (Mg, Ca, etc.), in the screening process in this study because of their toxic effects on the catalysts. Noble metals, such as Pt, were also excluded because of their high cost. Uncommon elements (Nb, Sm, etc.) were not considered because they are only present in a few catalyst data sets. After the above consideration, eight elements (Ti, V, Mn, Fe, Ni, Cu, Ce, and W) were selected as the potential elemental composition to find candidate catalysts. The range of each element value was set as 0–1 as the molar ratio of the element in the catalyst with a constraint that their addition is unit.

In the first round of the iterative screening process, the values of some feature variables, such as the pore size, pore volume, and specific surface area, that depend upon the elemental composition were fixed as the mean values of the variables in the database. They were updated in the second round of iteration based on the measured values from the catalyst materials synthesized from the first round of iteration. The average values for the pore size, specific surface area, and pore volume were 14.23 nm, 124.39 m²/g, and 0.28 cm³/g, respectively, in the original database. Some feature variables, such as the preparation method, were fixed as the co-precipitation method because it is commonly used for synthesizing NO_x SCR, and morphology was fixed as nanoparticles, which is related to the preparation method. Other feature variables, such as those experimental conditions for evaluating the catalyst performance, were also fixed using literature information; for example, the flue gas contains 400 ppm of NO_x, 400 ppm of NH₃, and 5% of resulting O₂, balanced with N₂, and the GHSV was 60 000 h^–1. The temperature was varied from 100 to 350 °C in the screening. Given that the catalysts in the database were mainly metal oxides, 500 °C for 2 h was selected as the post-treatment condition for catalyst activation before catalytic reaction.

With the above treatments and by focusing on the elemental composition in the screening process, the trained ML model identified 52 candidate catalysts with different sets of values for 8 elements that can meet or are above a desired performance (90%) (Table 1). Table 1 also provided the appearance frequency of each element in these candidate catalysts (bottom row in Table 1). Among the candidate catalysts, ternary catalysts with three elements had the highest percentage (>40%), followed by binary catalysts (23%) (Chart 1). The result implied that, among the identified candidate catalysts, the optimal catalyst would likely be a ternary catalyst. Further examination of the elemental composition of the identified ternary catalysts indicated that Ni (22%), Mn (20%), and Fe (23%) had the highest appearance frequency (Table 2). The appearance frequencies for the three elements are 2–5 times higher than those of the other five elements (5–10%). The result implied that the promising catalysts would consist of these three elements. Therefore, three ternary candidate catalysts with the predicted molar ratios of Ni, Mn, and Fe (Fe_0.2Mn_0.6Ni_0.2, Fe_0.2Mn_0.2Ni_0.6, and Fe_0.4Mn_0.15Ni_0.45) in Table 2 were selected as the final list of candidate catalysts. These ternary catalysts were not in the original database and were then synthesized in the experimental step for characterization in the first round of ML–experimental iteration.

Table 1. Statistics of Multiple Optimization Calculation Results.

variable	X1	X2	X3	X4	X5	X6	X7	X8
element with its range	Ti [0,1]	V [0,1]	Mn [0,1]	Fe [0,1]	Ni [0,1]	Cu [0,1]	Ce [0,1]	W [0,1]	number of element types
1	0.48		0.52						2
2			0.35	0.32	0.18		0.15		4
3		0.45		0.11	0.45				3
4	0.44	0.56							2
5			0.99	0.01					2
6	0.11		0.56		0.33				3
8	0.34		0.31	0.34					3
9		0.45		0.11	0.45				3
10				0.28		0.51	0.21		3
11	0.14	0.26	0.26		0.07			0.26	5
12						0.83	0.13	0.05	3
13		0.11	0.32		0.24			0.32	4
14	0.28	0.26		0.21	0.24				4
15	0.51		0.35	0.13					3
16				0.49	0.51				2
17			0.21	0.32	0.47				3
18	0.27			0.02	0.27	0.27		0.18	5
19						0.87	0.13		2
20			0.25		0.30	0.15	0.30		4
21			0.36	0.64					2
22		0.39	0.06	0.17	0.39				4
23			0.36	0.28		0.36			3
24	0.31		0.25	0.28	0.06	0.10			5
25	0.28		0.18	0.16			0.10	0.28	5
26		0.58					0.42		2
27		0.52	0.23		0.25				3
28	0.22		0.45	0.12		0.10	0.11		5
29		0.25	0.24	0.25			0.25		4
30			0.35	0.40	0.25				3
31		0.47	0.26		0.15			0.11	4
32	0.29		0.60	0.11					3
33	0.41					0.59			2
34				0.30	0.30			0.40	3
35				0.37	0.26	0.37			3
36	0.01		0.29		0.47			0.23	4
37	0.20	0.08	0.10	0.06	0.09	0.20	0.08	0.20	8
38		0.22					0.78		2
39	0.28			0.36	0.36				3
40			0.44			0.56			2
41	0.18		0.37	0.44					3
42			1.00						1
43			0.08	0.11	0.18	0.25	0.30	0.08	6
44		0.19			0.58		0.23		3
45				0.42	0.16	0.42			3
46	0.24	0.12	0.13		0.03	0.14	0.10	0.24	7
47	0.29		0.32	0.39	0.01				4
48		0.88			0.12				2
49			0.52		0.12			0.36	3
50				0.48			0.52		2
51		0.27	0.11		0.62				3
52			0.39	0.22	0.39				3
frequency statistics	19	17	32	30	30	15	16	12

Chart 1. Frequency Statistics: (A) Frequency of the Candidate Catalysts with a Different Number of Elements and (B) Frequency of Different Elements in the Ternary Catalysts.

Table 2. Type and Content Statistics of Ternary Catalyst Elements.

variable	X1	X2	X3	X4	X5	X6	X7	X8
element with its range	Ti [0,1]	V [0,1]	Mn [0,1]	Fe [0,1]	Ni [0,1]	Cu [0,1]	Ce [0,1]	W [0,1]
1		0.45		0.10	0.45
2	0.11		0.56		0.33
3	0.34		0.31	0.34
4		0.45		0.11	0.45
5				0.28		0.51	0.21
6						0.83	0.13	0.05
7	0.51		0.35	0.13
8			0.21	0.32	0.47
9			0.36	0.28		0.36
10		0.52	0.23		0.25
11			0.35	0.40	0.25
12	0.29		0.60	0.11
13				0.30	0.30			0.40
14				0.37	0.26	0.37
15	0.28			0.36	0.36
16	0.18		0.37	0.44
17		0.19			0.58		0.23
18				0.42	0.16	0.42
19			0.52		0.12			0.36
20		0.27	0.11		0.62
21			0.39	0.22	0.39
frequency statistics	6	5	12	15	14	5	3	3

Experimental Result in the First-Round Iteration

The three ternary catalysts identified from the ML model were synthesized and characterized with the methods described in the Synthesis and Characterization section. The molar ratios of Ni, Mn, and Fe elements in the synthesized catalysts were consistent with the desired values from the ML model (Table 3), indicating that the co-precipitation method used in the synthesis process can accurately control the molar ratio of the elements in the synthesized materials. XRD analysis (Figure 2A) showed that the synthesized materials had a hydrotalcite-like structure of layered double hydroxide (LDH).²¹ The other characteristic peaks appearing in the Fe_0.2Mn_0.6Ni_0.2 sample were assigned to MnCO₃ [Joint Committee on Powder Diffraction Standards (JCPDS) number 44-1472].²² After catalyst activation at 500 °C for 2 h, all typical LDH diffraction peaks disappeared (Figure 2B) and a new amorphous oxide phase (inverted triangle) was formed. Crystalline MnO_x was also observed in the Fe_0.2Mn_0.6Ni_0.2 catalyst with peaks identified as Mn₃O₄ (JCPDS number 24-0734).²³ The peaks marked by black asterisks in Fe_0.2Mn_0.2Ni_0.6 were assigned to traces of nickel oxide hydride, caused by the high doping level of nickel.²⁴

Figure 2.

XRD profiles of the candidate catalysts (A) before and (B) after the process of calcination (note that catalyst Fe_0.319Mn_0.206Ni_0.475 was identified from the second round of iteration, and other three catalysts were identified from the first round of iteration).

TEM images (Figure 3) show that the synthesized Fe–Mn–Ni catalysts were all nanoparticles, which was consistent with the morphology used in the ML model. Panels A and C of Figure 3 show that the particles had relatively uniform sizes of 20–60 nm. However, Figure 3A indicated that there were larger crystal-like particles for the high Mn content product Fe_0.2Mn_0.6Ni_0.2. In combination with the high-resolution transmission electron microscopy (HRTEM) lattice fringe image of the “101” planes (Figure S2 of the Supporting Information) and the XRD results, the segregation of Mn₃O₄ in the high Mn doping sample (Fe_0.2Mn_0.6Ni_0.2 LDH) can be inferred.²⁵

Figure 3.

TEM images of the synthesized catalysts: (A) Fe_0.2 Mn_0.6Ni_0.2, (B) Fe_0.4Mn_0.15Ni_0.45, (C) Fe_0.2Mn_0.2Ni_0.6, and (D) Fe_0.319Mn_0.206Ni_0.475 (note that catalyst Fe_0.319Mn_0.206Ni_0.475 was identified from the second round of iteration, and other three catalysts were identified from the first round of iteration).

The specific surface area of the synthesized catalysts was between 40 and 87 m²/g (Table 4) with a trend of Fe_0.2Mn_0.6Ni_0.2 < Fe_0.2Mn_0.2Ni_0.6 < Fe_0.4Mn_0.15Ni_0.45. The correlation between the surface area and the doping amount of Mn was negative. During the calcination, unevenly distributed Mn in the LDH may have caused the segregation and sintering of MnO_x as suggested by the TEM and XRD results in Figures 2 and 3, which led to the destruction of the LDH structure and the decrease of the specific surface area.²⁶ As described before, in the first round of iteration, the pore size and pore volume (14.23 nm and 0.28 m³/g) were set to the average values in the database. These values were reasonably close to the measured values from the synthesized catalysts (Table 4). However, the specific surface area used in the ML model (124.39 cm²/g) was higher than the measured values (Table 4). In the subsequent round of iterations, the measured values were used in the prediction using the ML model.

Table 4. Brunauer–Emmett–Teller (BET) Characterization of Fe–Mn–Ni Series Catalysts^a.

catalyst	pore size (nm)	surface area (m2/g)	pore volume (cm3/g)
Fe0.2Mn0.6Ni0.2	21	40.62	0.351
Fe0.2Mn0.2Ni0.6	10.5	63.19	0.275
Fe0.319Mn0.206Ni0.475	12	73.04	0.287
Fe0.4Mn0.15Ni0.45	10	87.23	0.387

^aNote that catalyst Fe_0.319Mn_0.206Ni_0.475 was identified from the second round of iteration, and other three catalysts were identified from the first round of iteration.

The SCR NO_x conversion efficiency catalyzed by the synthesized materials increased with the temperature, showing a rise from 100 to 150 °C, a plateau from 150 to 250 °C, and a decrease from 250 to 400 °C (Figure 4). The Fe_0.2Mn_0.6Ni_0.2 catalyst displayed a higher conversion (up to 96%) than the others at 100–150 °C but became lower as the temperature increased to 250–400 °C. A general trend is that a higher Fe and Mn molar ratio led to better performance of the catalysts at the lower temperature region. This may be due to the advantage of Fe³⁺ and/or Mn⁴⁺ in catalyzing NO_x reduction at a low temperature, as described later. Notably, 210 °C appeared to be a cutoff temperature, below which the order of the conversion efficiency was Fe_0.2Mn_0.6Ni_0.2 > Fe_0.4Mn_0.15Ni_0.45 > Fe_0.2Mn_0.2Ni_0.6 and above which the trend was opposite. These trends were accurately predicted by the ML model, indicating that the model is effective in identifying candidate catalysts together with their properties. Some deviation existed, indicating that there was some uncertainty in the predicted results.

Figure 4.

Catalyst performance tests of Fe_0.2Mn_0.6Ni_0.2, Fe_0.2Mn_0.2Ni_0.6, Fe_0.4Mn_0.15Ni_0.45, and Fe_0.319Mn_0.206Ni_0.475 (note that catalyst Fe_0.319Mn_0.206Ni_0.475 was identified from the second round of iteration, and other three catalysts were identified from the first round of iteration).

The adsorption of NH₃ on the catalyst surface was one important step for the SCR reaction. Three catalysts showed overlapped NH₃ desorption peaks around 100–150 and 200–400 °C (Figure 5), which corresponded to weak acid sites and middle acid sites on the catalyst surface, respectively.²⁷ The intensities of the peaks of Fe_0.2Mn_0.6Ni_0.2 were higher than others, suggesting that a high Mn content led to a more weak surface acid site. However, Fe_0.2Mn_0.6Ni_0.2 had almost no ammonia desorption peaks between 260 and 290 °C, indicating fewer medium–strong acid sites. This partly explained the performance of the Fe_0.2Mn_0.6Ni_0.2 catalyst that its catalytic activity was best at a low temperature but quickly decreased at a high temperature.

Figure 5.

NH₃-TPD profiles over the catalysts (note that catalyst Fe_0.319Mn_0.206Ni_0.475 was identified from the second round of iteration, and the other three catalysts were identified from the first round of iteration).

The surface chemical states of Fe, Mn, and Ni elements and oxygen species over the catalysts were investigated by XPS. As shown in Figure S3 of the Supporting Information, the measured binding energies of Mn 2p_3/2 (638–648 eV) could be divided into three characteristic peaks of Mn²⁺ (641.0–641.5 eV), Mn³⁺ (642.3–642.6 eV), and Mn⁴⁺ (643.9–644.5 eV) using peak-fitting deconvolution.²⁸ Previous research reported that the activity of MnO_x species followed the order of MnO₂ > Mn₅O₈ > Mn₂O₃ > Mn₃O₄ > MnO, suggesting that the higher ratio of Mn⁴⁺/(Mn²⁺ + Mn³⁺ + Mn⁴⁺) was favorable for higher NO_x conversion.^29,30 The ratio of Mn⁴⁺/(Mn²⁺ + Mn³⁺ + Mn⁴⁺) was calculated by integrating the area of the Mn 2p peaks. As shown in Table 5, the Mn⁴⁺ ratio decreased with a higher doping amount of Mn. Fe_0.2Mn_0.6Ni_0.2 had the highest Mn content, but only 14.67% of Mn was Mn⁴⁺. Therefore, the doping amount of Mn had an optimum range, and a higher or lower amount beyond that range would be unfavorable to the performance of the catalyst.³¹

Table 5. XPS Characterization^a.

catalyst	Mn4+/(Mn4+ + Mn3+) (%)	Fe3+/(Fe2+ + Fe3+) (%)	Ni3+/(Ni2+ + Ni3+) (%)	Oβ/(Oβ + Oα + Oγ) (%)
Fe0.4Mn0.15Ni0.45	36.53	46.36	40.96	26.81
Fe0.2Mn0.2Ni0.6	33.84	51.38	43.09	18.17
Fe0.2Mn0.6Ni0.2	14.67	46.04	41.93	13.62
Fe0.319Mn0.206Ni0.475	34.28	44.89	42.17	22.32

^aNote that catalyst Fe_0.319Mn_0.206Ni_0.475 was identified from the second round of iteration (Figure S4 of the Supporting Information), and other three catalysts were identified from the first round of iteration.

Surface-adsorbed oxygen (O_β) played a key role in the SCR reaction as a result of its higher mobility. The O 1s XPS spectra for the catalysts were deconvoluted into two peaks. The lower binding energy (530–531 eV) peak was assigned to the lattice oxygen (O_α); the higher binding energy peak (531.8–532.2 eV) corresponded to surface-adsorbed oxygen (O_β); and the peak at a binding energy of 5–5 eV was assigned to O_r.³² It was reported that a higher ratio of O_β/(O_β + O_α) in a catalyst might represent the existence of more surface oxygen vacancies in it, giving rise to more preferable SCR activity.³³ Fe_0.4Mn_0.15Ni_0.45 had the highest content of Mn⁴⁺ and O_β, but its activity was slightly lower than that of Fe_0.2Mn_0.2Ni_0.6. Apparently, the amounts of surface Mn⁴⁺ and O_β were not the only factors affecting the catalytic activity of SCR. It has been reported that excess Mn⁴⁺ on the catalyst surface can cause the NO oxidation and NH₃ overoxidation, resulting in lower NO_x conversion and N₂ selectivity.³¹ Ni doping has been reported to initiate a redox reaction between Ni²⁺ and Mn⁴⁺ to form Ni³⁺ and Mn³⁺, which may prevent overoxidation during the reaction.³⁴ Furthermore, studies have shown that the strong interaction between Mn and Fe/Ni has a synergistic effect on the catalytic reduction, leading to an enhanced redox circle, such as Mn³⁺ + Ni³⁺ ↔ Mn⁴⁺ + Ni²⁺ or Ni²⁺ + Fe³⁺ ↔ Ni³⁺ + Fe²⁺.³³⁻³⁵ Therefore, the more Fe³⁺ and Ni³⁺ species in Fe_0.2Mn_0.2Ni_0.6 catalysts may benefit the redox performance and play an important role in improving the catalytic activity. Therefore, the optimization of multicomponent catalysts is often complicated. The composition ratios of metals affect not only the morphology and structure of the catalysts but also the important parameters of the SCR reaction, such as the surface acidity of the catalysts and the synergistic catalysis effect between metals.

Database Updating and Additional Catalyst Screening

The experimental data of the three catalysts were added to the database, which was used to retrain the ANN model, and the retrained ANN model was used to identify promising catalysts using the GA method again. In this updating process, the searching range was limited to catalysts with ternary elements consisting of Fe, Mn, and Ni, with a goal to find optimal values for their contents. A new catalyst with the molar ratios of Fe_0.319Mn_0.206Ni_0.475 was identified from the retrained ML model in this study. In this round of iteration, the pore size, specific surface area, and pore volume (13.83 nm, 63.68 m²/g, and 0.338 cm³/g, respectively) that were averaged from the experimental results in the first round of experiments (Table 4) were used in the ML model. Other parameters were the same as described for the first round of iteration.

The identified catalyst Fe_0.319Mn_0.206Ni_0.475 was synthesized and characterized experimentally with the same approaches as those described for the first round of iteration. The result confirmed that the synthesized material has a Fe_0.319Mn_0.206Ni_0.475 molar ratio of Fe_0.319Mn_0.206Ni_0.475 with the same morphology nanoparticles as described before. The synthesized catalyst performed well and was consistent with the prediction from the retrained ML model in the entire temperature range (Figure 6A). As a comparison, the ML model was also used to calculate the performance for this catalyst using the database without adding the experimental results from the first round of iteration (Figure 6B). The result indicated, as expected, that the ML model based on the database containing the data from the first round of iteration predicted much better. In addition, the uncertainty predicted from the ML model decreased, as shown by the narrower difference between the predicted maximum and minimum results in Figure 6A than that in Figure 6B. This is also expected because additional data would improve the prediction as shown in a previous study.¹⁸

Figure 6.

Performance results of the catalyst Fe_0.319Mn_0.206Ni_0.475 synthesized on the basis of the prediction of the trained ML model: (A and B) Predicted performance results from the ML model after and before the addition of the experimental data generated from the first round of the experiment, respectively.

This study demonstrated an iterative approach of the ML-based model and experiment to identify new catalysts that typically have spare data in the literature. In the approach, the ML-based model was first trained from relevant data in the literature and was then used to screen candidate catalysts to guide the experimental step, and the experiment step is used to synthesize the candidate catalysts to evaluate the performance predicted from the ML model and to provide new data to retrain the ML model. This process is iterated until an optimal catalyst is identified and confirmed experimentally. For the SCR NO_x system with a goal to find an optimal number of elements and their molar ratios in the catalyst, two iterations were found sufficiently to locate a ternary catalyst that had an excellent performance within a large applicable temperature range. The screening process involves eight elements. Significant error would be required if a traditional trial-and-error method was used to locate an ideal catalyst. The identified ternary catalyst was not previously reported in the literature, indicating that the ML-based model has a potential for identifying new catalysts, and the iterative approach can significantly expedite the screening process. The characterization results showed that the NO_x conversion rate was affected by many factors, including the specific surface area, element ratio, morphology, etc., and the effect of the Mn content on catalytic activity was even a tradeoff. Therefore, designing novel catalytic materials based on the experience of researchers alone can be difficult because all parameters need to be optimized simultaneously. The ML model method, however, can readily solve the problem of multiparameter optimization and accurately capture nonlinear changes, significantly shortening the time for catalyst optimization.

Various variables can be varied to screen candidate catalysts. While all of the variables can be mathematically varied during the screening process, some variables are independent and can be varied freely and other variables are passive and depend upon the values of independent variables. In this study, only elemental compositions were treated as free variables in the screening process. The variables such as the pore size, specific surface area, pore volume, and morphology were treated as passive variables because they were affected by the elemental composition and synthesis methods. In each round of the screening process using the ML-based model, the values of the passive variables had to be fixed and updated using the results determined from the last round of the experimental step. The results in this study indicated that the values of the passive variables can quickly converge during the iteration. This is another advantage of the iterative approach because the passive variables are difficult to independently vary during the screening process. The quick convergence property for these passive variables implied that it would be ideal to classify the variables in the database based on their mutual dependence and only select those independent variables to be varied during the screening process. This would decrease the degree of freedom of the optimal searching problem and, consequently, reduce the computational and experimental costs.

Supporting Information Available

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acs.est.3c00293.

• Additional details of the model, table of detailed information of collected data, characterization of XPS, and lattice analysis of the HRTEM image of samples (PDF)

The authors declare no competing financial interest.