Open Access

Aakash Naik

• Interactions between constituent atoms in crystalline materials have been shown to influence the properties of materials, such as elasticity, ionic and thermal conductivity, etc.[1–3] These interactions between constituent atoms, often quantified as bond strengths, can be extracted from crystalline materials using density-based[4], energy-based[5], and orbital-based methods. LOBSTER[6] is a software that relies on the orbital-based method to extract such bonding information by projecting the plane wave-based wave functions of modern density functional theory computations (DFT) onto a local atomic orbital basis. To garner a better understanding of how this bonding information relates to material properties on a larger scale, machine learning seems an obvious choice. However, for such data-driven studies, large quantities of data that are systematically generated, validated, and post-processed (feature engineering) in a form suitable for input in state-of-the-art ML models are oftenInteractions between constituent atoms in crystalline materials have been shown to influence the properties of materials, such as elasticity, ionic and thermal conductivity, etc.[1–3] These interactions between constituent atoms, often quantified as bond strengths, can be extracted from crystalline materials using density-based[4], energy-based[5], and orbital-based methods. LOBSTER[6] is a software that relies on the orbital-based method to extract such bonding information by projecting the plane wave-based wave functions of modern density functional theory computations (DFT) onto a local atomic orbital basis. To garner a better understanding of how this bonding information relates to material properties on a larger scale, machine learning seems an obvious choice. However, for such data-driven studies, large quantities of data that are systematically generated, validated, and post-processed (feature engineering) in a form suitable for input in state-of-the-art ML models are often needed.[7] Here, we first present a workflow implemented in atomate2[8] that can generate such bonding-related data using the LOBSTER program with minimal user input and a post-processing tool, LobsterPy[9], which can summarize and engineer features that could be directly used as input for ML studies. Lastly, we demonstrate the utility of these newly generated features by building a simple machine-learned model to predict harmonic phonon properties using the bonding dataset[10] generated by us for 1500 materials. We find a clear correlation between the bonding information and the phonon property.…