Page 26 - Read Online

Liu et al. J. Mater. Inf. 2025, 5, 27 I http://dx.doi.org/10.20517/jmi.2024.105 Page 3 of 18

ally an artificial neural network) to decode the representation from descriptor to the potential energy. One
notable example is the LASP software [21] with its neural network potential constructed by the power-type
structural descriptors represented by atom-centered symmetry functions [22] developed by Huang et al., which has been
widely utilized in computational simulation for various heterogeneous catalysis topics. For instance, in the
ethene epoxidation reaction on silver [23] , they identified the O 5 phase as highly active. Moreover, the calcu-
lated selectivity and ethene conversion are consistent with experimental results. In the case of the Fe-FTS
system, Liu et al. showed that surface reconstructions of iron carbides under a typical FTS gas atmosphere
are significant and usually involve the relocation of surface carbon atoms [24] . They demonstrated that the A-
P5 site, a pentagon configuration consisting of five iron atoms and a carbon atom bonded to four of the iron
atoms, is abundant on several active surfaces of iron carbides as an active site for CO dissociation and C-C
coupling [25] . Their obtained product yield also shows good agreement with experiments [26] . Another well-
known example is the DeePMD-kit software [27–30] with its deep potential (DP) method using an end-to-end
symmetry-preserving structural descriptor consisting of smooth internal coordinates and embedding network
transformation for encoding local atomic environment [27,31] , widely used for atomistic simulation including
the dynamics of growing carbon nanotube interfaces [32] , the sintering of Au nano-particles on supports having
different metal affinities [33] , the size effects of supported Au catalysts for CO oxidation activity [34] , CO adsorp-
tion induced surface reconstruction dynamics on the Cu surfaces [35] , and structural/compositional evolution
of Pd(111) surfaces for CO oxidation under varying adsorption coverages [36] .

Despite numerous successes of MLPs in atomistic simulations, their applications often face economic and
scalability limitations. The most obvious one is that all of the data used for training and validating MLPs are
generated from scratch, namely, by ab initio molecular dynamic (AIMD) simulation or other simulation meth-
ods with ab initio calculation (typically DFT). Efficient data generation platforms through an active learning
procedure such as DP GENerator (DP-GEN) [37] or Generating DP with Python (GDPy) [38] can significantly
facilitate this process. However, a substantial amount of effort is still needed to construct DFT-labeled datasets,
especially for heterogeneous catalysis domain due to its complexity with characteristics of bulks, surfaces, and
adsorbates altogether. The publicly available large datasets, such as OC20 [39] , OC22 [40] , and MPtrj [41] , have
covered extensive physical and chemical knowledge, becoming very useful for overcoming the challenges of
data generation. However, traditional MLPs usually struggle to generalize to applications not covered by the
training data, especially when additional elements and structural configuration are included in the simulation
tasks, making them poor at combining and utilizing the knowledge from these multiple and large datasets
together due to the varying ab initio calculation methods employed in different datasets and the vast compo-
sitional space contained across all these datasets.

Recently, the rise of “universal” or “fundamental” MLPs offers opportunities for addressing the issues above
and greatly extending the application scope of MLPs, often referred to as large atomic models (LAMs). One
remarkable advancement in LAM is the second version of the deep potential with attention pre-trained model
(DPA-2)developedbyZhangetal. fromtheDeepModelingopen-sourcecommunity [42,43] . Thismodelutilizes
unified descriptors constructed by deep neural network architecture, and uses the multi-task pre-training strat-
egy to jointly pre-train a multi-head model using multiple datasets that encompass multidisciplinary knowl-
edge from a broad range of application domains. By fine-tuning for specific downstream tasks, the pre-trained
DPA-2 model can be efficiently deployed to evaluate the potential energy surface (PES) of a specific research
domain with precision and generalization.

In this work, we have developed an MLP based on the pre-trained DPA-2 model by fine-tuning methodology,
named fine-tuned Fischer-Tropsch deep potential (FT DP), which aims to describe the global chemical space
2
consisting of Fe-C-H-O compositions in an accurate and extendable way, enabling us to perform efficient
computational simulation to investigate the characteristics of Fe-FTS systems. This paper is organized as fol-
lows. In the next section, we present theoretical schemes and computational settings used in this work. In the