Page 57 - Read Online

Page 12 of 14 Li et al. J Mater Inf 2024;4:4 I http://dx.doi.org/10.20517/jmi.2023.41

ing sizes. It is particularly noteworthy for addressing the longstanding challenge in the materials science do-
main of applying DL techniques to large molecules or 2D materials. As illustrated in Figure 6, our model
maintains exceptional precision for metallic nanoparticles containing around 80 atoms. Beyond that, our
model can be further generalized to surfaces. In comparison, the benchmark model only achieves commend-
able results on crystalline materials. Moreover, the training time of LERN is faster than that of all other neural
networks by an order of magnitude. Notably, all the aforementioned models were trained on a single RTX3080.
When iterating 200 times, with the exception of LERN, the training time for all models exceeded two hours,
while LERN required only approximately 26 min. This is considerably faster than other conventional graph
neural networks while the accuracy is greatly optimized. The above results show that our model solves the
current challenges of graph neural networks faced with the computational complexity and information cap-
ture problems present in complex systems. At the same time, because we only focus on the local characteristic
information of the adsorption site, our model can be adapted to systems with wider system diversity. It helps
develop catalysis research in different dimensions.

CONCLUSIONS
In summary, we have spearheaded a novel and universal ML framework utilizing the local environment in-
teraction to enhance feature input for effectively predicting molecular adsorption energy. Within this frame-
work, descriptors are constructed by modified graph-based VT representation (geometric information) and
improved fingerprint feature engineering (chemical information). These descriptors can be input into conven-
tional ML models as weighted features or deep neural networks as a graph form. We took conventional ML,
CGCNN, and ResNet as examples across diverse systems, including 0D, 2D and 3D catalysts, to demonstrate
the robustness and generalization of this framework. Such local environment interaction-based descriptors
make a lightweight advantage over other models, showcasing a significant boost in computational speed. This
universal and robust LEI-framework can be expanded to broader applications, such as catalysis, sensors, and
drug design.

DECLARATIONS
Authors’ contributions
Made substantial contributions to the conception and design of the study and performed data analysis and
interpretation: Li Y, Wu Y, Han Y, Shen L
Conducted data acquisition and contributed to the feature engineering part and model training: Lyu Q, Wu
H, Zhang X

Availability of data and materials
The source code of our work is freely accessible at https://github.com/mpeshel/LEI-framework_LERN.

Financial support and sponsorship
This work was supported by Singapore MOE Tier 1 (No. A-8001194-00-00) and Singapore MOE Tier 2 (No.
A-8001872-00-00).

Conflicts of interest
All authors declared that there are no conflicts of interest.

Ethical approval and consent to participate
Not applicable.

Consent for publication
Not applicable.