Ternary spinel oxides, represented by the formula AB₂O₄, where A and B denote transition metal elements, are recognized for their exceptional characteristics as electrode materials in batteries and superconductors. Despite extensive research, the sheer multitude of possible element combinations and stoichiometries renders a comprehensive exploration challenging. Intriguingly, alterations in the spinel composition have been found to influence its electrical properties, specifically affecting the electrical conductivity and band gap energy. This highlights the dynamic nature of these materials and their potential for tailored applications. In this context, the rise of data-driven and machine learning (ML) methodologies in material science paves the way for the discovery of promising compositions, thereby expediting the quest for innovative materials with specific electrical characteristics.

Machine Learning (ML) algorithms can be viewed as sophisticated statistical models that can handle a vast array of parameters, a task made feasible by modern computing power. The application of data-driven techniques has become prevalent across a multitude of technological domains, notably within the realm of computational materials science. These techniques primarily serve to address issues related to regression, classification, clustering, anomaly detection, and dimensionality reduction. In our research we used ML algorithms for two tasks. First, we took advantage of the back-propagation method to fit tight-binding Hamiltonians to the band structure of each spinel in our database. Second, we trained multiple ML algorithms to predict the band gap and conductivity of a given spinel composition.

Although the fundamental formula for spinel oxides composition is AB₂O₄. However, the actual unit cell comprises eight of these basic formula units, resulting in a total of 54 ions. The general expression takes the form (AyB₁-_y)[AxB₂-_x]O₄, where A and B metals can occupy either tetrahedral (Td) or octahedral [Oh] sites. Here, 0 < y < 1 and 0 < x < 2. The intricate composition of these materials significantly impacts their electrical properties, while the complicated structure has historically posed challenges for material discovery and optimization.

To address the challenge of material conductivity prediction, we embarked on constructing a spinel oxide database—that was previously unavailable online. Our approach involved performing density functional theory (DFT) calculations for 190 distinct spinel oxide compositions, incorporating transition metals such as Fe, Ni, Co, and Mn. From these calculations, we obtained the band structure for each material. Subsequently, we fit these band structures to a tight-binding Hamiltonian, enabling us to calculate the current using Landauer-Büttiker and NEGF methods. As a result, our database now contains information on 190 spinel compositions, along with their associated current and bandgap.

We selected four prominent ML algorithms for training using our spinel oxide database: neural networks, random forests, support vector regression, and kernel ridge regression. Our goal? To train these algorithms using our spinel oxide database, enabling them to predict both the band gap of a given composition and the electrical current it generates under a 1V electric potential. The unique fingerprint that is used to describe each composition to the machine was composed of four numbers, x, y, and the atomic number of A and B. To ensure robust model performance, we divided our dataset into two distinct sets: a training set and a test set. The latter served as a critical evaluation tool, assessing the model’s ability to predict properties for new data—data it hadn’t encountered during training. Avoiding the pitfall of overfitting, we adhered to the standard practice of allocating 80% of the dataset for training and reserving the remaining 20% for testing. Additionally, we developed a separate model that leverages bandwidth information from the band structure as input. Ultimately, this model is capable of a successful and fast prediction of the conductivity associated with each spinel material composition.