Page 25 - Read Online

Page 2 of 18 Liu et al. J. Mater. Inf. 2025, 5, 27 I http://dx.doi.org/10.20517/jmi.2024.105

abundant formation of [Fe 4C] squares on several reconstructed surfaces. These insights highlight the potential of
utilizing LAM for atomistic simulation for iron-based FTS processes and other complex catalytic reactions.

Keywords: Large atomic model, machine learning potentials, fine-tuning, Fischer-Tropsch synthesis, transition state
optimization, surface reconstruction

INTRODUCTION
As a key aspect of the chemical industry, heterogeneous catalysis has played a crucial role in the large-scale
production of commodity chemicals such as ammonia, alcohol, and synthetic fuels. Nowadays, computational
simulationsbasedonabinitiodensity-functionaltheory(DFT)calculationshaveofferedunprecedentedoppor-
tunities for the rational design of novel solid catalysts by providing a deep atomistic analysis of the structures
[1]
and reaction properties combined with theories in heterogeneous catalysis . However, accurate and efficient
computational simulations of complex heterogeneous catalytic processes remain highly challenging because of
the demanding computational cost. One of the typical examples of complex heterogeneous catalytic systems is
theFischer-Tropschsynthesis(FTS),whichconvertssyngas(amixtureofCOandH )intofuelsandothervalu-
2
able chemicals, holding a special position in the energy industry [2,3] . Among many possible candidates, the
iron-based catalyst has gained much attention due to its low cost, high sulfur tolerance, and low methane selec-
tivity [4,5] . The active phases of iron-based FTS are believed to be in situ formed iron carbides such as -Fe 5C 2,
-Fe 2C, -Fe 2C/ -Fe 2.2C, and -Fe 3C, which can be verified with spectroscopic and electron-microscopy ex-
periments [6–10] . However, iron-based FTS is recognized as a structure-sensitive reaction [11] , and many factors
of Fe-catalysts such as particle size, chemical composition, and promoters are found to have a great impact on
their catalytic activity [11–13] , which poses great challenges for both theoretical and experimental research.

To achieve the goal of rational catalyst design, extensive research has been conducted to elucidate the relation-
shipbetweenthesurfacestructuresofironcarbidesandtheirreactivitiesintheFTSprocess [14–18] . Forinstance,
Chen et al. found that the CO dissociation barrier on different -Fe 5C 2 surfaces can be effectively predicted
usinglocaldescriptorssuchasatomiccharges [16] . TheyalsoshowedthattheirderivedBrønsted-Evans-Polanyi
(BEP) relationship remains applicable in cases including non-stoichiometric terminations, carbon vacancies,
and potassium promotion. Li et al. studied the CO activation processes on -Fe 3C(010) surfaces, and revealed
that on the Fe/C-terminated surface, direct CO dissociation is not favored due to the high concentration of
surface carbon atoms, and the participation of hydrogen is essential for CO dissociation, while on the Fe-
terminated -Fe 3C(010) surface the direct CO dissociation is preferred [17] . Yin et al. investigated the CH 4
formation and C-C coupling reactions on multiple -Fe 5C 2 surfaces based on a Wulff structure [18] . They
demonstrated that certain “active facets” only account for a small fraction of the total exposed surface area,
but could play a significant role in the overall FTS activity. Although these studies have provided valuable
insights into the FTS-related properties of iron carbides under various chemical environments, there is still a
lack of comprehensive understanding of the pattern of catalytic sites and reaction mechanisms under realistic
FTS conditions. This is primarily due to the high computational cost of DFT calculations, particularly when
applied to systems with large temporal or spatial scales, which are common in iron-based FTS reactions and
other complex heterogeneous catalytic processes.

The past several years have witnessed enormous advances in artificial intelligence (AI) methods, fueling a
new paradigm shift of discoveries in natural sciences and giving rise to a new area of research, known as AI
for science [19] . Especially, with the rapid development of machine learning potentials (MLPs) with both the
accuracy of DFT and the efficiency of classical force fields, the detailed atomistic simulation for complex het-
erogeneous catalysis systems on a large temporal and spatial scale has become feasible [20] , so that there emerge
a large number of so-called AI-driven atomistic simulation platforms incorporating MLPs sharing the same
basic architecture that using structural descriptors to represent the atomic structures and a fitting model (usu-