Presentation Information
[PCP1-03]Classification of phase formation in layered perovskite compounds with Kolmogorov-Arnold networks
*Sonosuke Kono1,2, Yoichi Higashi2, Yuki Iwasa2, Izumi Hase2, Ryo Maezono3, Taichiro Nishio1, Hiraku Ogino2, Kenta Hongo4 (1. Tokyo University of Science (Japan), 2. National Institute of Advanced Industrial Science and Technology (AIST) (Japan), 3. Institute of Science Tokyo (Japan), 4. JAIST (Japan))
Keywords:
Quantum materials,Prediction of phase formation,Materials informatics,Kolmogorov-Arnold networks
We aim to predict the phase formation of layered perovskite compounds. These materials are expected to exhibit various functions including superconductivity, and they are composed of stacking of inverse fluorite layers and perovskite oxide layers [Fig. 1(a)]. This system is a typical multinary system with numerous variations due to the diversity of elements in the perovskite layers and the variable number of stacked layers. For ternary systems, the Goldschmidt tolerance factor, which is based on ionic radii, is a well-known indicator for the formation of a cubic perovskite structure with the composition of ABX3. Although this factor is still considered to be an important condition that must be met for these multinary systems, it is insufficient as a phase formation indicator.Previously, we calculated a phase formation indicator for multinary layered perovskite compounds and predicted the formation in new arsenic fluorides [1] using SISSO (Sure Independence Screening and Sparsifying Operator) [2].
In this study, to persuit the higher prediction accuracy and interpretability, we have introduced Kolmogorov-Arnold networks (KANs) [3]. This method, unlike conventional multilayer perceptrons (MLPs), has the advantage of high interpretability, as it allows for a visual understanding of how the model's predictions are formed. Furthermore, because KANs can achieve high accuracy with a smaller number of parameters compared to MLPs, they are expected to improve generalization performance even with small datasets. This makes them suitable for the present study, which uses a small materials dataset, such as experimental data related to phase formation. By leveraging the characteristics of KANs, we will also interpret the physical meaning of the phase formation index learned by the model and evaluate their generalization performance. Fig. 1(b) shows an example of classification for phase formation in layered perovskite compounds with KANs.
References
[1] Sonosuke Kono et al., J. Phys.: Conf. Ser. 3054 012004 (2025).
[2] Runhai Ouyang et al., Phys. Rev. Materials 2, 083802 (2018).
[3] Ziming Liu et al., arXiv:2404.19756.
In this study, to persuit the higher prediction accuracy and interpretability, we have introduced Kolmogorov-Arnold networks (KANs) [3]. This method, unlike conventional multilayer perceptrons (MLPs), has the advantage of high interpretability, as it allows for a visual understanding of how the model's predictions are formed. Furthermore, because KANs can achieve high accuracy with a smaller number of parameters compared to MLPs, they are expected to improve generalization performance even with small datasets. This makes them suitable for the present study, which uses a small materials dataset, such as experimental data related to phase formation. By leveraging the characteristics of KANs, we will also interpret the physical meaning of the phase formation index learned by the model and evaluate their generalization performance. Fig. 1(b) shows an example of classification for phase formation in layered perovskite compounds with KANs.
References
[1] Sonosuke Kono et al., J. Phys.: Conf. Ser. 3054 012004 (2025).
[2] Runhai Ouyang et al., Phys. Rev. Materials 2, 083802 (2018).
[3] Ziming Liu et al., arXiv:2404.19756.