A Methodological Study on Data Representation for Machine Learning Modelling of Thermal Conductivity of Rare-Earth Oxides
Amiya Chowdhury, Acacio Rincón Romero, Eduardo Aguilar-Bejarano, Halar Memon, Grazziela Figueredo, Tanvir Hussain · 2025
Read A Methodological Study on Data Representation for Machine Learning Modelling of Thermal Conductivity of Rare-Earth Oxides by Amiya Chowdhury, Acacio Rincón Romero, Eduardo Aguilar-Bejarano, Halar Memon, Grazziela Figueredo, Tanvir Hussain (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-18951.txt.
Introduction
Amiya Chowdhury and five co-authors ask how best to feed material data into machine learning models. Their test case is the thermal conductivity of rare-earth oxides used in heat-shield coatings for engines. Predicting such traits with data models can cut years off lab work. But a model's skill depends on how each material is described to it: as a list of hand-picked traits, or as a graph of atoms and bonds.
The team compares two classic models fed hand-made trait lists with a graph neural network that reads the crystal structure itself. They also test two ways to stretch their small dataset. The graph model wins by a wide margin over both classic models. The result supports graph-based descriptions for this kind of materials search.
Catalog record
| CREATOR | Amiya Chowdhury, Acacio Rincón Romero, Eduardo Aguilar-Bejarano, Halar Memon, Grazziela Figueredo, Tanvir Hussain |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 42 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2509-18951
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
A Methodological Study on Data Representation for Machine Learning Modelling of Thermal Conductivity of Rare-Earth Oxides Amiya Chowdhury, Acacio Rincón Romero, Eduardo Aguilar-Bejarano, Halar Memon, Grazziela Figueredo, Tanvir Hussain Abstract Quantitative structure-activity relationship (QSAR) modelling is widely employed in materials science to predict properties of interest and extract useful descriptors for measured properties. In thermal barrier coatings (TBC), QSAR can significantly shorten the experimental discovery cycle, which can take years. Although machine learning methods are commonly employed for QSAR, their performance depends on the data quality and how instances are represented. Traditional, hand-crafted descriptors based on known material properties are limited to represent materials that share the same basic crystal structure, limited the size of the dataset. By contrast, graph neural networks offer a more expressive representation, encoding atomic positions and bonds in the crystal lattice. In this study, we compare Random Forest (RF) and Gaussian Process (GP) models trained on hand-crafted descriptors from the literature with graph-based representations for high-entropy, rare-earth pyrochlore oxides using the Crystal Graph Convolutional Neural Network (CGCNN). Two different types of augmentation methods are also explored to account for the limited data size, one of which is only applicable to graph-based representations. Our findings show that the CGCNN model substantially outperforms the RF and GP models, underscoring the potential of graph-based representations for enhanced QSAR modelling in TBC research. {graphicalabstract} † † highlights: Use of graph neural networks to model and predict thermal conductivity for high-entropy rare-earth oxides Comparison of graph data representation of compositions with handcrafted descriptors found in the literature Experiments show improved performance and generalisation with graph representations, demonstrating its suitability for modelling the problem keywords graph neural networks, thermal conductivity, data representation, rare-earth oxides, machine learning † † credit: Conceptualisation, Methodology, Software, Formal Analysis, Data curation, Investigation, Writing - Original Draft Preparation † † credit: Conceptualisation, Methodology, Data curation, Supervision, Writing - Reviewing and Editing † † credit: Methodology, Software † † credit: Writing - Reviewing and Editing † † credit: Conceptualisation, Methodology, Software, Supervision, Writing - Reviewing and Editing † † credit: Conceptualisation, Fund acquisition, Supervision, Writing - Reviewing and Editing † † corresponding: Corresponding author 1 Introduction Operating temperatures for nickel based superalloys are between 800-1000 {}^{o}C ( Mouritz 2012 ). Turbine inlet temperatures in gas-turbine engines can exceed 1500 {}^{o}C ( Centrich et al. 2014 ).Thermal barrier coatings (TBC) provide temperature reductions on the nickel-based superalloy of 100-200 {}^{o}C ( Darolia 2013 ), allowing turbine engines to operate effectively at higher temperatures. Higher engine operating temperatures lead to increased thermodynamic efficiency and lower CO_{2} emissions. Yttria stabilised zirconia has been the industry standard TBC since the 1970s due to its combination of thermal and mechanical properties ( Stecura 1978 ; Cao et al. 2004 ; Vaßen et al. 2022 ). However, this composition is limited to a temperature of 1200 {}^{o}C due to the formation of tetragonal and cubic phases above this temperature, leading to undesired expansion of about 3 to 5 \% ( Cao et al. 2004 ; Liu et al. 2022 ; Pakseresht et al. 2022 ). It is also susceptible to chemical attack by calcia-magnesia-alumino-silicate (CMAS). To develop a replacement that outperforms YSZ as a coating, the new desirable material would have lower thermal conductivity, comparable fracture toughness ( 1-2 MPam^{1/2} ), thermal expansion coefficient ( \sim 10-11\times 10^{-10} K^{-1} )( Vaßen et al. 2010 ) and thermal cyclic life-span, higher resistance to CMAS and better phase stability at its operating temperatures ( \sim 1200-1300 {}^{o}C )( Bakan and Vaßen 2017 ; Vaßen et al. 2022 ; Mehboob et al. 2020 ). Due to the existing limitations of YSZ coatings, there has been an increasing interest in the discovery of high-performance alternatives ( Vaßen et al. 2022 ). Pyrochlore structures are currently the most investigated group ( Bakan and Vaßen 2017 ; Vassen et al. 2004 ; Fergus 2014 ; Wu et al. 2004 ). This is partly due to their lower thermal conductivity, higher thermal stability and lower oxygen diffusivity ( Vassen et al. 2004 ; Zhu et al. 2021a ) compared to YSZ. However, these compositions have a much lower thermal expansion coefficient and mechanical toughness compared to YSZ , making them unsuitable for replacement. The pyrochlore structure can be modified by cation substitution, forming high-entropy or multi-component ceramics (HECs and MCCs). The cubic pyrochlore structure has the general formula A_{2}^{3+}B_{2}^{4+}O_{6}O^{\prime} , where A^{3+} is typically a rare-earth element and B^{4+} is usually a transition metal. By including multiple elements in different combinations in the A and B sub-lattices, both the thermal and mechanical properties of the material can be manipulated ( Wan et al. 2006 ; Ren et al. 2015 ; Liu et al. 2014 ). The introduction of elements with different ionic radii, mass and electro-negativities, provides the potential to tailor the thermal and mechanical properties of the material, allowing for an ensemble of properties that would not otherwise be achievable with single elements. Additionally, distortion in the lattice caused by different Re-O bond lengths has been shown to reduce thermal conductivity ( Wright et al. 2019 ; Tsai and Yeh 2014 ). The ability to tailor both the mechanical and thermal properties of these compositions makes it a worthwhile search space for potential new TBC materials. Due to the large number of combinations this allows, experimentally measuring the properties of these potential TBC materials would be too costly and time-consuming. Computational methods using first-principles calculations, such as density functional theory (DFT) take time to set up and run ( Seko et al. 2018 ). And third order problems, such as thermal conductivity are computationally expensive and often limited to crystals with a small number of atoms in their unit cell ( Choi et al. 2022 ; Luo et al. 2023 ). As an alternative to overcome laborious, resource intensive processes, machine learning (ML) has been increasingly used in materials science, for property and micro-structure predictions, and composition generation ( Liu et al. 2017 ). Conventional machine learning methods used include random forest, support vector machines and shallow neural networks. They often take as input hand-crafted and/or non-problem specific descriptors to represent information about the input data points representing materials. These descriptors are hypothesised to be relevant, that is, to have some degree of correlation to the target variable ( Ward et al. 2016 ; Jha et al. 2018 ). The definition and selection of these descriptors generally requires domain knowledge— although often generic chemistry descriptors are also used — and an understanding of the machine learning algorithms ( Ward et al. 2016 ; Jha et al. 2018 ; Guyon and Elisseeff 2003 ). This is an important step, as relevant descriptors can help understand materials structure-property relationships. Yang et al. conducted an investigation of the atomic parameters that mostly affect the prediction of thermal and mechanical properties of pyrochlores Yang et al. 2018 ). Parameters include the average atomic mass of A and B cations ( M_{A} and M_{B} ), period ( P_{A} and P_{B} ) in the periodic table, average cationic radius ( R_{A} and R_{B} ), cationic radius ratio ( R_{A}/R_{B} ), electronegativity ( EN_{A} and EN_{B} ), density ( \rho ) and lattice constant ( \alpha ).