Machine learning (ML) techniques have rapidly found applications in many domains of materials chemistry and physics where large data sets are available. Aiming to accelerate the discovery of materials for battery applications, in this work, we develop a tool (http://se.cmich.edu/batteries) based on ML models to predict voltages of electrode materials for metal-ion batteries. To this end, we use deep neural network, support vector machine, and kernel ridge regression as ML algorithms in combination with data taken from the Materials Project Database, as well as feature vectors from properties of chemical compounds and elemental properties of their constituents. We show that our ML models have predictive capabilities for different reference test sets and, as an example, we utilize them to generate a voltage profile diagram and compare it to density functional theory calculations. In addition, using our models, we propose nearly 5,000 candidate electrode materials for Na-and K-ion batteries. We also make arXiv:1903.06813v2 [cond-mat.mtrl-sci] 9 May 2019 available a web-accessible tool that, within a minute, can be used to estimate the voltage of any bulk electrode material for a number of metal-ions. These results show that ML is a promising alternative for computationally demanding calculations as a first screening tool of novel materials for battery applications.

We have investigated the stability, maximum intercalation capacity, and voltage profile of alkali metal intercalated hexagonal BC3 (MxBC3), for 0 < x ≤ 2 and M = Li, Na, and K. Our calculations, based on dispersion-corrected density functional theory, show that these intercalation compounds are stable with respect to BC3 and their bulk metal counterparts. Moreover, we found that among all MxBC3 considered, the maximum stable capacity corresponds to an x value of 1.5, 1, and 1.5 for Li, Na, and K, respectively. These values are associated with large gravimetric capacities of 572 mA h/g for Na and 858 mA h/g for Li and K. Importantly, we show that metal intercalated hexagonal BC3 has the advantage of a small open-circuit voltage variation of approximately 0.49, 0.12, and 0.16 V for Li, Na, and K, respectively. Our results suggest that BC3 can become a robust alternative to graphitic electrodes in metal ion batteries, thus encouraging further experimental work.

Rechargeable batteries provide crucial energy storage systems for renewable energy sources, as well as consumer electronics and electrical vehicles. There are a number of important parameters that determine the suitability of electrode materials for battery applications, such as the average voltage and the maximum specific capacity which contribute to the overall energy density. Another important performance criterion for battery electrode materials is their volume change upon charging and discharging, which contributes to determine the cyclability, Coulombic efficiency, and safety of a battery. In this work, we present deep neural network regression machine learning models (ML), trained on data obtained from the Materials Project database, for predicting average voltages and volume change upon charging and discharging of electrode materials for metal-ion batteries. Our models exhibit good performance as measured by the average mean absolute error obtained from a 10-fold cross-validation, as well as on independent test sets. We further assess the robustness of our ML models by investigating their screening potential beyond the training database. We produce Na-ion electrodes by systematically replacing Li-ions in the original database by Na-ions and, then, selecting a set of 22 electrodes that exhibit a good performance in energy density, as well as small volume variations upon charging and discharging, as predicted by the machine learning model. The ML predictions for these materials are then compared to quantum-mechanics based calculations. Our results reaffirm the significant role of machine learning techniques in the exploration of materials for battery applications.

No abstract

We analyze the effect of removing self-interaction error on magnetic exchange couplings using the Fermi-Löwdin orbital self-interaction correction (FLOSIC) method in the framework of density functional theory (DFT). We compare magnetic exchange couplings obtained from self-interaction-free FLOSIC calculations with the local spin density approximation (LSDA) with several widely used DFT realizations and wave function based methods. To this end, we employ the linear H–He–H model system, six organic radical molecules, and [Cu2Cl6]2− as representatives of different types of magnetic interactions. We show that the simple self-interaction-free version of LSDA improves calculated couplings with respect to LSDA in all cases, even though the nature of the exchange interaction varies across the test set, and in most cases, it yields results comparable to modern hybrids and range-separated approximate functionals.

The static electric dipole polarizability of a system is a measure of the binding of its electrons.In density functional theory (DFT) calculations, this binding is weakened by the presence of unphysical self-interaction in the density functional approximation (DFA), leading to overestimates of polarizabilities. To investigate this systematically we compare polarizabilities for the atoms from H to Ar and their anions and cations calculated in several DFA's and the corresponding self-interaction corrected (SIC) DFAs with experiment and with high-level quantum chemistry reference values. The SIC results are obtained using the Fermi-Löwdin orbital self interaction correction (FLO-SIC) method. Removing self-interaction generally leads to smaller polarizabilities that agree significantly better with reference values. We find that FLO-SIC improves the performance of the local spin density approximation and the generalized gradient approximation (GGA) for polarizabilities to a quality that is comparable to so-called rung 4 functionals, but slightly degrades the performance of the strongly constrained and appropriately normed (SCAN) meta-GGA functional.

The self-interaction error (SIE) is one of the major drawbacks of practical exchange-correlation functionals for Kohn–Sham density functional theory. Despite this, the use of methods that explicitly remove SIE from approximate density functionals is scarce in the literature due to their relatively high computational cost and lack of consistent improvement over standard modern functionals. In this article we assess the performance of a novel approach recently proposed by Pederson, Ruzsinszky, and Perdew [J. Chem. Phys. 2014, 140, 121103] for performing self-interaction free calculations in density functional theory based on Fermi orbitals. To this end, we employ test sets consisting of reaction energies that are considered particularly sensitive to SIE. We found that the parameter-free Fermi–Löwdin orbital self-interaction correction method combined with the standard local spin density approximation (LSDA) and Perdew–Burke–Ernzerhof (PBE) functionals gives a much better estimate of reaction energies compared to their parent LSDA and PBE functionals for most of the reactions in these two sets. They also perform on par with the global PBE0 and range-separated LC-ωPBE hybrids, which partially eliminate the SIE by including Hartree–Fock exchange. This shows the potential of the Fermi–Löwdin orbital self-interaction correction (FLOSIC) method for practical density functional calculations without SIE.

The prerequisite of therapeutic drug design and discovery is to identify novel molecules and developing lead candidates with desired biophysical and biochemical properties. Deep generative models have demonstrated their ability to find such molecules by exploring a huge chemical space efficiently. An effective way to generate new molecules with desired target properties is by constraining the critical fucntional groups or the core scaffolds in the generation process. To this end, we developed a domain aware generative framework called 3D-Scaffold that takes 3D coordinates of the desired scaffold as an input and generates 3D coordinates of novel therapeutic candidates as an output while always preserving the desired scaffolds in generated structures. We demonstrated that our framework generates predominantly valid, unique, novel, and experimentally synthesizable molecules that have drug-like properties similar to the molecules in the training set. Using domain specific data sets, we generate covalent and noncovalent antiviral inhibitors targeting viral proteins. To measure the success of our framework in generating therapeutic candidates, generated structures were subjected to high throughput virtual screening via docking simulations, which shows favorable interaction against SARS-CoV-2 main protease (Mpro) and nonstructural protein endoribonuclease (NSP15) targets. Most importantly, our deep learning model performs well with relatively small 3D structural training data and quickly learns to generalize to new scaffolds, highlighting its potential application to other domains for generating target specific candidates.