Presentation Information
[P2-8]Magnetic crystal reconstruction and optimisation
using Graph Neural Networks
*Gino Hrkac1, Hao Zhang1, Prathyush Menon1 (1. University of Exeter (UK))
Keywords:
XRD reconstruction,Graph theory,GNN,Machine learning,Magnetic materials
Magnetic crystals are described by complex electronic and spin interactions that govern magnetization dynamics, magnetic anisotropy, and phase transitions. These properties play an important role in spintronics, energy-efficient
memory devices, and magnetic rare-earth materials. However, accurate prediction of their macroscopic magnetic behaviour is challenging because of multi-phase interactions, complex crystal symmetry, and differences in the boundaries
of traditional computational methods.
In the first step, we focus on single magnetic crystals, using graph-theory tocreate graph representations of hence crystals where nodes represent the atoms and the edges the magnetic or chemical relations. Both of these extract local(node-level) and global (graph-level) structural features, which serve as input foran active teaching model that refines the predictions of magnetically prominent properties and topological features such as magnetization, Curie temperature,average degree, average shortest path length, clustering coefficient, maximum cluster size, and PageRank centrality. The amount of data points increases the efficiency of the model by prioritizing data points with higher connectivity based on the graph-model, reducing computational time and costs.
In the second phase, we extend this approach to poly-crystals, characterized by complex atomic arrangements and long-range interactions. This involves the construction of larger and more complex systems (mixed interfaces and compositions) and the extraction of higher-order structural characteristics that capture
the complexity of these systems. We employ machine learning techniques (such as XGboost for feature-based learning and Graph Neural Networks (GNNs)) to model dependencies between atomic interactions and magnetic properties with
the aim to identify and reconstruct these super-cell structures. The graphic analysis reveals that directional cutting changes local magnetic connectivity,leading to measurable changes in magnetization and anisotropic properties. In particular, the grouping coefficient and maximum cluster size display strong
correlations with post-cut temperature variations. In detail, we investigate the impact of different cutting planes ([100], [110], [111], etc.) on magnetic connectivity and anisotropy, providing a comprehensive understanding of structural modifications in 1-12 type materials. This approach enhances predictive accuracy but also offers deeper insights into how graph topology governs macroscopic magnetic behavior, deepening our understanding of structure-property relationships In figure a, this is a T hM n super-cell crystal with a (1,1,1) to a (1,1,0) interface. Figure b is the interface of a super-cell T hM n and α − F e structure. It is composed of multiple nodes, which are the atoms, and the black
lines represent the topological connections between atoms. The blue lines are the connection between these two different crystals at the interface. The entire graph forms a complex topological network. The graph analysis reveals that
structural modifications induce significant changes in local magnetic connectivity. Based the calculation from igraph we can get its Topological characteristics(e.g degree avg, avg path length, clustering coeff, pagerank avgpagerank avg and
max clique size) This is just one of the possibility.
Notably, clustering coefficients, maximum cluster size, and connectivity metricsstrongly correlate with post-cut magnetic variations, emphasizing the role of interfacial atomic bonding. The combined analysis of unit-cell and super-cell structures provides insights into the fundamental structure-property relationships, which offers a computationally efficient alternative to density functionaltheory (DFT) simulations. Compared with previous methods, this is more like a distillation-style ML method. The aim is to simplify the data used for training, use more targeted methods to derive an accurate result and simplify the branches during the training, and cut off unnecessary possibilities in a targeted manner, so that the noise of the answer can be reduced. In short, this is an artificial Mixture of Experts (MOE) machine learning technology. The purpose
is to derive an accurate and faster analysis of complex systems with smaller data sets.
memory devices, and magnetic rare-earth materials. However, accurate prediction of their macroscopic magnetic behaviour is challenging because of multi-phase interactions, complex crystal symmetry, and differences in the boundaries
of traditional computational methods.
In the first step, we focus on single magnetic crystals, using graph-theory tocreate graph representations of hence crystals where nodes represent the atoms and the edges the magnetic or chemical relations. Both of these extract local(node-level) and global (graph-level) structural features, which serve as input foran active teaching model that refines the predictions of magnetically prominent properties and topological features such as magnetization, Curie temperature,average degree, average shortest path length, clustering coefficient, maximum cluster size, and PageRank centrality. The amount of data points increases the efficiency of the model by prioritizing data points with higher connectivity based on the graph-model, reducing computational time and costs.
In the second phase, we extend this approach to poly-crystals, characterized by complex atomic arrangements and long-range interactions. This involves the construction of larger and more complex systems (mixed interfaces and compositions) and the extraction of higher-order structural characteristics that capture
the complexity of these systems. We employ machine learning techniques (such as XGboost for feature-based learning and Graph Neural Networks (GNNs)) to model dependencies between atomic interactions and magnetic properties with
the aim to identify and reconstruct these super-cell structures. The graphic analysis reveals that directional cutting changes local magnetic connectivity,leading to measurable changes in magnetization and anisotropic properties. In particular, the grouping coefficient and maximum cluster size display strong
correlations with post-cut temperature variations. In detail, we investigate the impact of different cutting planes ([100], [110], [111], etc.) on magnetic connectivity and anisotropy, providing a comprehensive understanding of structural modifications in 1-12 type materials. This approach enhances predictive accuracy but also offers deeper insights into how graph topology governs macroscopic magnetic behavior, deepening our understanding of structure-property relationships In figure a, this is a T hM n super-cell crystal with a (1,1,1) to a (1,1,0) interface. Figure b is the interface of a super-cell T hM n and α − F e structure. It is composed of multiple nodes, which are the atoms, and the black
lines represent the topological connections between atoms. The blue lines are the connection between these two different crystals at the interface. The entire graph forms a complex topological network. The graph analysis reveals that
structural modifications induce significant changes in local magnetic connectivity. Based the calculation from igraph we can get its Topological characteristics(e.g degree avg, avg path length, clustering coeff, pagerank avgpagerank avg and
max clique size) This is just one of the possibility.
Notably, clustering coefficients, maximum cluster size, and connectivity metricsstrongly correlate with post-cut magnetic variations, emphasizing the role of interfacial atomic bonding. The combined analysis of unit-cell and super-cell structures provides insights into the fundamental structure-property relationships, which offers a computationally efficient alternative to density functionaltheory (DFT) simulations. Compared with previous methods, this is more like a distillation-style ML method. The aim is to simplify the data used for training, use more targeted methods to derive an accurate result and simplify the branches during the training, and cut off unnecessary possibilities in a targeted manner, so that the noise of the answer can be reduced. In short, this is an artificial Mixture of Experts (MOE) machine learning technology. The purpose
is to derive an accurate and faster analysis of complex systems with smaller data sets.
