1 of 46

Using Computational Methods to Design and Evaluate Catalysts for Nitrogen Reduction Reaction

Tan Jia Feng^† and Sherman Yeo Jun Yi^† (Group 5)

Principal Investigator: Asst. Prof. Li Xiaoyan^‡

SPS Mentors: Jensie Low^†‡ and Kow Zhen Ting^†

1

^†Special Programme in Science, National University of Singapore; ^‡Department of Chemistry, National University of Singapore

2 of 46

Overview

2

Conclusion

Catalyst

Metal Hydride

Introduction

3 of 46

Introduction

3

4 of 46

Importance of ammonia

4

Fertiliser

Refrigerant

(Erisman et al., 2008), (Yüzbaşıoğlu et al., 2022)

Introduction

5 of 46

Current production methods

5

N₂ + 3H₂ ⇌ 2NH₃

Haber-Bosch (90% of NH₃ synthesis)

Environmentally unsustainable

Harsh reaction conditions

(Spatolisano & Pellegrini, 2023)

Introduction

6 of 46

Current production methods

6

Electrochemical nitrogen reduction reaction (eNRR)

N

H⁺

Catalyst

(Ouyang et al., 2023)

Introduction

7 of 46

Current production methods

7

Electrochemical nitrogen reduction reaction (eNRR)

Low NH₃ yield

N

H⁺

Catalyst

HER

(Tavella et al., 2022)

Introduction

8 of 46

Drawing inspiration from current methods

8

Efficient N₂ activation with high NH₃yield

Milder conditions for N₂ activation

Catalytic system

Haber-Bosch

eNRR

Introduction

9 of 46

How can we design a catalytic system that bypasses competing reactions and enables efficient N₂ activation under milder conditions?

9

Introduction

10 of 46

Proposed set-up

10

Introduction

11 of 46

Objectives

11

Find alternative source of hydrogen

Find optimal catalyst for N₂ activation

Catalytic system

Metal hydride

Primary catalyst

Introduction

12 of 46

Workflow

12

DFT Calculations

Find Optimal MH

OC20 Dataset

Generate Catalysts

Find Optimal Catalyst

Obtain ML Model

Tandem Catalyst

to predict E_N of

Introduction

13 of 46

Metal Hydride Selection

13

14 of 46

Alternative source of hydrogen

14

DFT Calculations

Find Optimal MH

OC20 Dataset

Generate Catalysts

Find Optimal Catalyst

Obtain ML Model

Tandem Catalyst

to predict E_N of

Metal Hydride

15 of 46

Alternative source of hydrogen

15

Metal Hydride (MH)

Metal Hydride

16 of 46

Density Functional Theory (DFT) Calculations

16

The Vienna Ab initio Simulation Package (VASP) employs DFT to calculate the ground state energy of atomic structures

Metal Hydride

17 of 46

17

Metal Hydride

18 of 46

18

Metal Hydride

19 of 46

Results

19

Metal Hydride

20 of 46

Objectives

20

Find alternative source of hydrogen

Find optimal catalyst for N₂ activation

Catalytic system

KH

Primary catalyst

Metal Hydride

21 of 46

Discussion

21

Explore greater range of MH and surface orientations

Metal Hydride

22 of 46

Catalyst Selection

22

23 of 46

Optimal catalyst for N₂ activation

23

DFT Calculations

Find Optimal MH

OC20 Dataset

Generate Catalysts

Find Optimal Catalyst

Obtain ML Model

Tandem Catalyst

to predict E_N of

(Chanussot et al., 2021)

Catalyst

24 of 46

Primary workflow

24

DFT Calculations

Find Optimal MH

OC20 Dataset

Generate Catalysts

Find Optimal Catalyst

Obtain ML Model

Tandem Catalyst

to predict E_N of

Catalyst

25 of 46

Complementary workflow

25

DFT Calculations

Find Optimal MH

OC20 Dataset

Generate Catalysts

Find Optimal Catalyst

Obtain ML Model

Tandem Catalyst

to predict E_N of

Catalyst

26 of 46

Optimal catalyst for N₂ activation

26

Improves overall reaction efficiency by facilitating nitrogen adsorption

Nitrogen adsorption energy (E_N)

(Medford et al., 2015)

Catalyst

27 of 46

ML models

27

Support vector regression (SVR)

Random Forest (RF)

eXtreme gradient boosting (XGboost)

EquiformerV2

Catalyst

28 of 46

Results for ML model evaluation

28

Model	Random Forest	XGBoost	SVR	EquiformerV2
MAE / eV	0.681	0.667	0.697	0.446
R²	0.625	0.628	0.591	0.751

EquiformerV2 performed the best among the 4 → benchmarking model

Catalyst

29 of 46

Primary workflow

29

DFT Calculations

Find Optimal MH

OC20 Dataset

Generate Catalysts

Find Optimal Catalyst

Obtain ML Model

Tandem Catalyst

to predict E_N of

Catalyst

30 of 46

Diffusion model

30

Add random noise

Remove noise

Input structure

Generated structure

Learns to remove noise

(Yang et al., 2026)

Catalyst

31 of 46

31

Desired E_N value = -0.9 eV

E_N << 0 → N₂ binds too tightly to catalytic surface

E_N >> 0 → inefficient N₂ activation

Catalyst

32 of 46

Results of diffusion model

32

Most of the 31 generated catalysts had predicted E_N close to -0.9 eV

Catalyst

33 of 46

Results of catalyst ranking

33

Generated catalyst	Predicted E_N value (eV)
Ag₂₅Fe₃₉	-0.913
Cu₂₄Fe₄₀	-0.729
Fe₃₂Ir₃₂	-0.699
Ag₂₄Fe₄₀	-1.127
Co₂₄Fe₄₀	-0.635

Ag₂₅Fe₃₉ had predicted E_N differing from the target by only 0.013 eV

Most optimal

Catalyst

34 of 46

Objectives

34

Find alternative source of hydrogen

Find optimal catalyst for N₂ activation

Catalytic system

KH

Ag₂₅Fe₃₉

Catalyst

35 of 46

Discussion

35

EquiformerV2 model has MAE of 0.446 eV

Existing models achieved MAE of 0.297 eV

Better ML models can be employed to predict E_N

(OC20 Leaderboard, n.d.)

Catalyst

36 of 46

Conclusion

36

37 of 46

Summary

37

Optimal MH:

KH

Optimal Catalyst:

Ag₂₅Fe₃₉

Incorporate interactions between MH and catalyst

Whats next?

Benchmarking model: EquiformerV2

Experimentally evaluate computational results

Conclusion

38 of 46

38

Acknowledgement

We would like to thank:

Asst. Prof. Li Xiaoyan

Principal Investigator

Kow Zhen Ting

SPS Mentor

Jensie Low

SPS Mentor

39 of 46

39

References

Ammonia. American Chemical Society. https://www.acs.org/molecule-of-the-week/archive/a/ammonia.html (accessed 2026-04-09).
Chanussot, L. et al. ACS Catalysis 2021, 11, 6059–6072.
Erisman, J. W.; Sutton, M. A.; Galloway, J.; Klimont, Z.; Winiwarter, W. Nature Geoscience 2008, 1, 636–639.
Medford, A. J.; Vojvodic, A.; Hummelshøj, J. S.; Voss, J.; Abild-Pedersen, F.; Studt, F.; Bligaard, T.; Nilsson, A.; Nørskov, J. K. Journal of Catalysis 2015, 328, 36–42.
OC20 Leaderboard. https://opencatalystproject.org/leaderboard.html (accessed 2026-03-31).
OLAS. http://eniac.fme.vutbr.cz/links.html (accessed 2026-04-05).
Ouyang, L.; Liang, J.; Luo, Y.; Zheng, D.; Sun, S.; Liu, Q.; Hamdy, M. S.; Sun, X.; Ying, B. Chinese Journal of Catalysis 2023, 50, 6–44.
Spatolisano, E.; Pellegrini, L. A. Chemical Engineering Research and Design 2023, 195, 651–661.
Tavella, F.; Giusi, D.; Ampelli, C. Current Opinion in Green and Sustainable Chemistry 2022, 35, 100604.
Yang, J.; Ye, K.; Xie, S.; Li, Q.; Milhans, C.; Liu, F.; Che, F. Journal of the American Chemical Society 2026, 148, 537–546.
Yüzbaşıoğlu, A. E.; Avşar, C.; Gezerman, A. O. Current Research in Green and Sustainable Chemistry 2022, 5, 100307.

40 of 46

Thank you!

Q&A

40

41 of 46

Current eNRR methodology

41

Electrochemical nitrogen reduction reaction (eNRR)

Hydrogen Evolution Reaction (HER)

Limited greenhouse gas emissions
Low NH₃ yield

2H⁺ + 2e^- → H₂

N₂ + 6H⁺ + 6e^- → 2NH₃

Kinetically preferred over eNRR

Introduction

42 of 46

42

Geometric optimisation of slab

Bulk optimisation

Geometric optimisation of slab with a missing H

M

H

Metal Hydride

43 of 46

43

Different H positions in MgH₂ → large variation

Similar H positions in

SrH₂ → small variation

Metal Hydride

44 of 46

44

Catalyst

45 of 46

45

26 Fe Iron

44 Ru Ruthenium

Found in all of the top 5 bimetallic catalysts

Known to be optimal single metal, but not among the generated

Catalyst

46 of 46

Why bimetallic catalysts?

46

Too many permutations of trimetallic catalysts

Single metal catalysts have been exhausted

Need to match (111) template in composition

Bimetallic catalysts!

Catalyst