Page 28 - IJAMD-2-1

International Journal of AI for
Materials and Design
ML molecular modeling of Ru: A KAN approach

high-entropy alloys (HEAs) and complex concentrated MD, are often computationally intensive and limited in
2,3
alloys, demonstrate potential as refractory high-entropy their ability to rapidly explore the vast parameter space of
4
superalloys (RSAs). In A2+B2 systems, the ordered Ru’s structure-property relationships. The development of
intermetallic B2 phase is particularly important, as it an ML model that can accurately and rapidly predict Ru’s
often acts as a strengthening precipitate, similar to the molecular structures and associated mechanical properties
role of γ’ in traditional superalloys. In addition, B2 phases would be a significant advancement. Such a model could
5
often exhibit excellent strength retention at elevated potentially allow for a faster and more comprehensive
temperatures and creep resistance compared to disordered exploration of Ru’s structural and mechanical behavior
phases, enhancing the alloy’s performance under sustained under various environmental conditions.
loads at high temperatures. Many B2 phases, particularly High-throughput ab initio simulations are increasingly
6
those containing Al, demonstrate superior oxidation used for exploring the basic properties of engineering
resistance by forming protective oxide scales. However, materials, such as determining the material properties.
14
some elements discovered by Naka and Khan are still not When combined with materials informatics, ML models
fully represented in RSAs that are currently synthesized, accelerate the exploration of these material properties.
15
especially Pd, Pt, and ruthenium (Ru), which all form The goal of the ML model is to supplement or replace
stable B2 phases with Al. 7 high-fidelity modeling methods at the quantum chemical
Among the elements critical to RSA development, Ru has or classical level to predict the properties of molecules
garnered significant attention due to its unique properties or materials directly from their structure or their
and potential ability to form stable ordered phases in chemical composition. ML methods usually use available
multi-component systems. Cerba et al. demonstrated that experimental and ab initio data to build accurate statistical
8
AlRu alloys melt at 2060°C and interact with Ru through models that can be used to predict the properties of
eutectic reactions at 1920°C, suggesting potential routes materials. This innovative combination can significantly
for exploring A2 + B2 microstructures. Moreover, the enhance predictive capabilities and has applications
B2 phase exists in some binary refractory metal systems in many fields of materials research, such as dielectric
16
involving Ru, such as Nb-Ru and Ru-Zr, indicating its polymers, critical temperatures of superconducting
7
18
19
17
versatility in phase formation. Despite its importance, materials, crystal structures, perovskites, and
20
there remains a significant gap in our understanding of nanostructures. Many of these properties were measured
Ru’s molecular structure and its influence on mechanical with high accuracy using ML approaches. For instance, de
21
and thermodynamic properties, especially in the context Jong et al. developed a model using supervised ML with
of computational materials science. Specifically, there is gradient-boosting regression to predict elastic properties
currently no machine learning (ML) model capable of of inorganic polycrystalline compounds, such as the bulk
accurately predicting Ru’s molecular structures along modulus K and shear modulus G. For metallic systems,
22
with their corresponding mechanical properties in a Song et al. constructed a general-purpose neural network
wide range of temperatures. This lack of comprehensive potential (NNP) for 16 elemental metals and their alloys
data and predictive tools hinders our ability to engineer that achieves superior accuracy compared to traditional
the structures and material properties of RSAs where Ru embedded-atom method potentials while maintaining
is used. The corresponding knowledge gap hinders the computational efficiency. Their potential accurately
efficient design and development of advanced RSAs and captures complex interface behavior, phase stability, and
other high-temperature materials that incorporate Ru. mechanical properties across diverse metallic systems.
Liyanage et al. developed a NNP for the Cu-W system
23
Ru is a platinum-group metal that crystallizes in using the Behler-Parrinello framework. Their potential
a hexagonal close-packed (hcp) structure at ambient accurately reproduces metallurgically relevant properties,
conditions. Experimentally, Ru displays superconductivity including elasticity, stacking faults, dislocations, and
9
with a critical temperature below 1699 K. A hcp Ru also thermodynamic behavior in both elemental Cu and W,
10
maintains its structural stability up to pressures of 150 GPa as well as Cu-W interfaces and solid solutions. This NNP
and temperatures of 960 K. Güler et al. used molecular enables large-scale atomistic simulations to investigate
12
11
dynamics (MD) simulations with an empirical interatomic phenomena such as the influence of interface stress on
potential to explore the temperature-dependent elastic, mechanical properties in Cu-W nano-multilayer systems.
mechanical, and anisotropic properties of hcp Ru. Another advantage of ML models is their generalizability
Likewise, Lu et al. used density functional theory (DFT) across different material systems, as it is easier to
13
to calculate the high-pressure phase transition of Ru. incorporate crystal data with defects. Compared with
24
However, computational approaches, such as DFT and other interatomic potentials, ML potentials are highly

Volume 2 Issue 1 (2025) 22 doi: 10.36922/ijamd.8291