Supplementary Materials

Enabling catalyst discovery through machine learning and high-throughput

experimentation

Travis Williams

_{, Katherine McCullough}

_{, and Jochen A. Lauterbach}

1

_{Department of Chemical Engineering, University of South Carolina, Columbia, SC 29208, USA.}

*Correspondence to: [email protected]

† Authors contributed equally to this work

Supplementary Materials

Calculations

The Ru-normalized rate was calculated using the following formula: 3 2𝑄𝑁𝐻3

(

)

where 𝑄𝑁𝐻3 is the flow rate of ammonia, 𝜌𝑁𝐻3 is the density of ammonia, 𝑀𝑁𝐻3 is the molar mass of ammonia, 𝑋𝑁𝐻3 is the measured ammonia conversion, 𝑤𝑅𝑢 is the weight loading of ruthenium on the catalyst, 𝑚𝑐𝑎𝑡 is the mass of catalyst loaded into the reactor, and 𝑀𝑅𝑢 is the molar mass of ruthenium.

Machine learning algorithm selection

Fig. S1

data set (all secondary elements across all weight loadings). A full hyperparameter search was performed for each algorithm using the gridsearch algorithm from sklearn. A 3-fold crossvalidation methodology was used rather than the leave-one-out crossvalidation method to decrease the total time of the parameter search. The best combination of hyperparameters for each algorithm were used to calculate the model error.

3

grid search method to probe all possible combinations. Unlisted parameters for each algorithm were kept at default values.

Algorithm Hyperparameter Optimized _Value Algorithm Hyperparameter Optimized _Value Kernel Ridge Regression Alpha Coef0 Degree 1.5 0 2 Decision Tree Criterion Max_depth Max_features Min_samples_split splitter Mse 2 Auto 5 Random Support Vector Machine Max_iter Kernel Gamma Epsilon Degree Coef0 200 Rbf 1 0.01 2 100.0 Extremely Randomized Trees N_estimators Min_samples_split Min_samples_leaf Min_impurity_decrease Max_leaf_nodes Max_features Max_depth Criterion 25 5 2 0 20 0.1 None mae Neural Net Tol Solver Momentum Max_iter Learning_rate_init Learning_rate Hidden_layer_sizes Early_stopping Alpha Activation 1e-05 Lbfgs 0.95 500 0.001 Invscaling 10 True 0.001 Tanh Adaboost Base Estimator Learning Rate Loss N_estimators None 2 linear 50 Lasso Regressor Alpha Fit_intercept Max_iter Normalize 0.9 False 100 True Gradient Tree Boosting Subsample N_estimators Min_samples_split Min_samples_leaf Min_impurity_decrease Max_leaf_nodes Max_features Max_depth Loss Learning_rate Criterion 1 100 10 1 0 None Sqrt None Ls 0.1 Friedman_mse Ridge

Regressor Alphasolver sag1 Random Forest

N_estimators Min_samples_split Min_samples_leaf Min_impurity_decrease Max_leaf_nodes Max_features Max_depth 100 10 1 0 50 Auto 5 k-Nearest Neighbor Regressor Weights P N_neighbors Leaf_size Algorithm Uniform 1 7 2 Ball_tree

Table S2. Training set summary for machine learning models. Summary of training sets used in study, including catalyst

compositions, target predictions, and mean absolute error.

Training Set _{Training Set}Catalysts in Prediction _Target Mean Absolute _Error Squared ErrorRoot Mean Catalysts within Percent 5% Error 3-catalyst

training set RuCaK, RuMnK, 3,1,12 RuInK Secondary Elements for substitution into K-promoted Ru catalyst 0.157 0.198 23% 22-catalyst

training set RuCaK, RuMnK, 3,1,12 RuInK, RuCuK, RuYK,RuMgK, RuNiK, RuCrK, RuWK, RuHfK, RuScK, RuZnK, RuSrK, RuBiK, RuPdK, RuMoK, RuRhK, RuOsK, RuPtK, RuAuK, RuNbK, RuFeK 1,3,12 and 2,2,12 weight loadings for secondary metals substituted into K-promoted Ru catalyst 0.127 0.169 20% 28-catalyst

training set RuCaK, RuMnK, 3,1,12 RuInK, RuCuK, RuYK,RuMgK, RuNiK, RuCrK, RuWK, RuHfK, RuScK, RuZnK, RuSrK, RuBiK, RuPdK, RuMoK, RuRhK, RuOsK, RuPtK, RuAuK, RuNbK, RuFeK 2,2,12 RuCaK, RuMnK, RuInK 1,3,12 RuCaK, RuMnK, RuInK 1,3,12 and 2,2,12 weight loadings for secondary metals substituted into K-promoted Ru catalyst 0.115 0.164 37%

5

in literature at similar conditions for NH3 decomposition.

a _{Data from this work.} Catalyst Conversion NH3 (300°C) NH3 Conversion (400°C) NH3 % GHSV (mLNH3 gcat-1 h-1) Ru-normalized rate @ 350℃ (molH2 molRu-1 h-1) Ref. 3,1,12 RuYK/Al2O3 43% 94% 100 5200 826.4 a 3,1,12 RuMgK/Al2O3 44% 98% 100 5200 919.8 a 3,1,12 RuSrK/Al2O3 46% 98% 100 5200 879.3 a 3,1,12 RuScK/Al2O3 46% 97% 100 5200 885.4 a 1,3,12 RuYK/Al2O3 47% 97% 100 5200 2702.3 a

7, 20 RuCs/CNT 28% 92% _listedNot 5200 337.9 (1)

3,4 RuCs/CNT 3% 33% _listedNot 5200 112.6 (1)

7,7 RuNa/CNT 10% 100% 30 6000 316.7 (2)

7,7 RuNa/CeO2 NR 5% 100% 30 6000 165.6 (2)

Table S4. Catalyst features used in machine learning algorithm. Elemental properties were taken from (66). Temperature

and space velocity values were measured during the experiment. Element loadings are the nominal mol fraction during synthesis. Precursor chlorine mols were calculated from the weight loading of the precursor and the number of chlorine ligands.

Feature Name

Atomic Number Ru Loading

Atomic Volume K Loading

Atomic Weight Mg Loading

Boiling Temperature Ca Loading

Conductivity Sc Loading

Covalent Radius Ti Loading

Density V Loading

Electron Affinity Cr Loading

Electronegativity Mn Loading

Fusion Enthalpy Fe Loading

Ground State Energy Co Loading

Heat Capacity (Mass) Ni Loading

Heat Capacity (Molar) Cu Loading

1st _{Ionization Energy Zn Loading}

2nd _{Ionization Energy Sr Loading}

3rd _{Ionization Energy Y Loading}

4th _{Ionization Energy Zr Loading}

5th _{Ionization Energy Nb Loading}

6th _{Ionization Energy Mo Loading}

7th _{Ionization Energy Rh Loading}

8th _{Ionization Energy Pd Loading}

Melting Temperature Ag Loading

Polarizability Cd Loading

Adjusted Work Function In Loading

Number of Valence Electrons Sn Loading

Number of s-shell Valence Electrons Hf Loading

Number of p-shell Valence Electrons Ta Loading

Number of d-shell Valence Electrons W Loading

Number of f-shell Valence Electrons Re Loading

Temperature Os Loading

Space Velocity Ir Loading

Precursor Chlorine Mols Pt Loading

Au Loading Pb Loading Bi Loading