Multi-task graph neural networks for simultaneous prediction of global and atomic properties in ferromagnetic systems ^*

Abstract

We introduce a multi-tasking graph convolutional neural network, HydraGNN, to simultaneously predict both global and atomic physical properties and demonstrate with ferromagnetic materials. We train HydraGNN on an open-source ab initio density functional theory (DFT) dataset for iron-platinum (FePt) with a fixed body centered tetragonal (BCT) lattice structure and fixed volume to simultaneously predict the mixing enthalpy (a global feature of the system), the atomic charge transfer, and the atomic magnetic moment across configurations that span the entire compositional range. By taking advantage of underlying physical correlations between material properties, multi-task learning (MTL) with HydraGNN provides effective training even with modest amounts of data. Moreover, this is achieved with just one architecture instead of three, as required by single-task learning (STL). The first convolutional layers of the HydraGNN architecture are shared by all learning tasks and extract features common to all material properties. The following layers discriminate the features of the different properties, the results of which are fed to the separate heads of the final layer to produce predictions. Numerical results show that HydraGNN effectively captures the relation between the configurational entropy and the material properties over the entire compositional range. Overall, the accuracy of simultaneous MTL predictions is comparable to the accuracy of the STL predictions. In addition, the computational cost of training HydraGNN for MTL is much lower than the original DFT calculations and also lower than training separate STL models for each property.

Computational Workflow for Accelerated Molecular Design Using Quantum Chemical Simulations and Deep Learning Models

“…To train a surrogate model for HOMO-LUMO gap prediction, we use the multi-headed HydraGNN package developed earlier by some of our team [18,19]. This surrogate, allowing multi-headed output, is ideally suited for the simultaneous prediction of multiple important molecular characteristics, such as electronic properties and synthesizability scoring.…”

“…A variety of GCNNs have been developed to better represent the atomic systems, such as crystal GCNN (CGCNN) [43] for crystalline materials, MEGNet for both molecules and crystals [44], as well as ALIGNN [45] that adds the bond angles in the model. More details about GCNNs can be found in [18].…”

“…In this work, the GCNN model implemented in HydraGNN [18][19] is used as a surrogate model for HOMO-LUMO gap property. The GC layer used is principal neighborhood aggregation (PNA) [46].…”

“…As others before us, we start with an approximation of excitation energies by MO energy differences [12,13]. Besides the use of computationally efficient DFTB methods, our new inverse design algorithm has two major components: 1) Development of a multi-headed graph convolutional neural network (GCNN) approach that will allow the prediction of not only molecular orbital energy differences, but more complex properties such as multiple excitation energies and oscillator strengths for dye candidates [18,19]. The GCNNs surrogate models are sufficiently fast to supplant computationally expensive, explicit excited state calculations required for the inverse design step, which will be component 2): Development of a machine learning masked language model (MLM)-based generation of new dye candidates with subsequent evaluation against a target function (desirable optical properties).…”

Graph neural networks predict energetic and mechanical properties for models of solid solution metal alloy phases

“…More recent work started investigating the effectiveness of GCNN models for prediction of mixing enthalpy, atomic charge transfer, and atomic magnetic moment for solid solution ferromagnetic alloys, characterized by disordered atomic configurations, with non-relaxed crystal structure and fixed volume from open-source DFT data [20,21]. Results showed that the GCNN model can accurately estimate material properties as a function of the configurational entropy [22,23].…”