Abstract
Predicting the stability of crystals is one of the central problems in materials science. Today, density functional theory (DFT) calculations remain comparatively expensive and scale poorly with system size. Here we show that deep neural networks utilizing just two descriptors—the Pauling electronegativity and ionic radii—can predict the DFT formation energies of C_{3}A_{2}D_{3}O_{12} garnets and ABO_{3} perovskites with low mean absolute errors (MAEs) of 7–10 meV atom^{−1} and 20–34 meV atom^{−1}, respectively, well within the limits of DFT accuracy. Further extension to mixed garnets and perovskites with little loss in accuracy can be achieved using a binary encoding scheme, addressing a critical gap in the extension of machinelearning models from fixed stoichiometry crystals to infinite universe of mixedspecies crystals. Finally, we demonstrate the potential of these models to rapidly transverse vast chemical spaces to accurately identify stable compositions, accelerating the discovery of novel materials with potentially superior properties.
Introduction
The formation energy of a crystal is a key metric of its stability and synthesizability. It is typically defined relative to constituent unary/binary phases (E_{f}) or the stable linear combination of competing phases in the phase diagram (E_{hull}, or energy above convex hull)^{1}. In recent years, machine learning (ML) models trained on density functional theory (DFT)^{2} calculations have garnered widespread interest as a means to scale quantitative predictions of materials properties^{3,4,5,6,7}, including energies of crystals. However, most previous efforts at predicting E_{f} or E_{hull} of crystals^{5,8,9,10,11,12} using ML models have yielded mean absolute errors (MAEs) of 70–100 meV atom^{−1}, falling far short of the necessary accuracy for useful crystal stability predictions. This is because approximately 90% of the crystals in the Inorganic Crystal Structure Database (ICSD) have E_{hull} < 70 meV atom^{−1}^{13}, and the errors of DFTcalculated formation energies of ternary oxides from binary oxides relative to experiments are ~ 24 meV atom^{−1}^{14}.
We propose to approach the crystal stability prediction problem by using artificial neural networks (ANNs)^{15}, i.e., algorithms that are loosely modeled on the animal brain, to quantify wellestablished chemical intuition. The Pauling electronegativity and ionic radii guide much of our understanding about the bonding and stability of crystals today, for example, in the form of Pauling’s five rules^{16} and the Goldschmidt tolerance factor for perovskites^{17}. Though these rules are qualitative in nature, their great success points to the potential existence of a direct relationship between crystal stability and these descriptors.
To probe these relationships, we choose, as our initial model system, the garnets, a large family of crystals with widespread technological applications such as luminescent materials for solidstate lighting^{18} and lithium superionic conductors for rechargeable lithiumion batteries^{19,20}. Garnets have the general formula C_{3}A_{2}D_{3}O_{12}, where C, A and D denote the three cation sites with Wyckoff symbols 24c (dodecahedron), 16a (octahedron) and 24d (tetrahedron), respectively, in the prototypical cubic \(Ia\overline 3 d\) garnet crystal shown in Fig. 1a. The distinct coordination environments of the three sites result in different minimum ionic radii ratios (and hence, species preference) according to Pauling’s first rule. We further demonstrate the generalizability of our approach to the ABO_{3} perovskites (Fig. 1b), another broad class of technologically important crystals^{21,22,23,24,25}.
In this work, we show that ANNs using only the Pauling electronegativity^{26} and ionic radii^{27} of the constituent species as the input descriptors can achieve extremely low MAEs of 7–10 meV atom^{−1} and 20–34 meV atom^{−1} in predicting the formation energies of garnets and perovskites, respectively. We also introduce two alternative approaches to extend such ANN models beyond simple unmixed crystals to the much larger universe of mixed cation crystals—a rigorously defined averaging scheme for the electronegativity and ionic radii for modeling complete cation disorder, and a novel binary encoding scheme to account for the effect of cation orderings with minimal increase in feature dimension. Finally, we demonstrate the application of the NN models in accurately and efficiently identifying stable compositions out of thousands of garnet and perovskite candidates, greatly expanding the space for the discovery of materials with potentially superior properties.
Results
Model construction and definitions
We start with the hypothesis that the formation energy E_{f} of a C_{3}A_{2}D_{3}O_{12} garnet is some unknown function f of the Pauling electronegativities (χ) and Shannon ionic radii (r) of the species in the C, A, and D sites, i.e.,
Here, we define E_{f} as the change in energy in forming the garnet from binary oxides with elements in the same oxidation states, i.e., \(E_f^{\mathrm{oxide}}\) as opposed to the more commonly used formation energy from the elements \(E_f^{\mathrm{element}}\) in previous works^{8,9,10,11}. Using the Ca_{3}Al_{2}Si_{3}O_{12} garnet (grossular) as an example, \(E_f^{\mathrm{oxide}}\) is given by the energy of the reaction: 3CaO + Al_{2}O_{3} + 3SiO_{2} → Ca_{3}Al_{2}Si_{3}O_{12.} This choice of definition of E_{f} is motivated by two reasons. First, binary oxides are frequently used as synthesis precursors. Second, our definition ensures that garnets that share elements in the same oxidation states have E_{f} that are referenced to the same binary oxides, minimizing wellknown DFT errors. In contrast, \(E_f^{\mathrm{element}}\) and E_{hull} are both poor target metrics for a ML model. \(E_f^{\mathrm{element}}\) suffers from nonsystematic DFT errors associated with the incomplete cancellation of the selfinteraction error in redox reactions^{28}, while E_{hull} is defined with respect to the linear combination of stable phases at the C_{3}A_{2}D_{3}O_{12} composition in the CADO phase diagram, which can vary unpredictably even for highly similar chemistries. Henceforth, the notation E_{f} in this work refers to \(E_f^{\mathrm{oxide}}\) unless otherwise stated. The binary oxides used to calculate the E_{f} for garnets and perovskites are listed in Supplementary Table 1 and 2, respectively.
Based on the universal approximation theorem^{29}, we may model the unknown function f(χ_{C},r_{C},χ_{A},r_{A},χ_{D},r_{D}), which is clearly nonlinear (see Supplementary Fig. 1), using a feedforward ANN, as depicted in Fig. 2. The loss function and evaluation metric are chosen to be the mean squared error (MSE) and MAE, respectively. We will denote the architecture of the ANN using n^{i}−n^{[1]}−n^{[2]}−···−1, where n^{i} and n^{[l]} are the number of neurons in the input and l^{th} hidden layer, respectively.
Neural network model for unmixed garnets
We developed an initial ANN model for unmixed garnets, i.e., garnets with only one type of species each in C, A, and D. A data set comprising 635 unmixed garnets was generated by performing full DFT relaxation and energy calculations (see Methods) on all chargeneural combinations of allowed species (Supplementary Table 3) on the C, A, and D sites^{30}. This dataset was randomly divided into training, validation, and test data in the ratio of 64:16:20. Using 50 repeated random subsampling cross validation, we find that a 6241 ANN architecture yields a small root mean square error (RMSE) of 12 meV atom^{−1}, as well as the smallest standard deviation in the RMSE among the 50 subsamples (Supplementary Fig. 2a). The training, validation and test MAEs for the optimized 6241 model are ~7–10 meV atom^{−1} (Fig. 3a), an order of magnitude lower than the ~100 meV atom^{−1} achieved in previous ML models^{5,8,9,10}. For comparison, the error in the DFT E_{f} of garnets relative to experimental values is around 14 meV atom^{−1} (Supplementary Table 4). Similar RMSEs are obtained for deep neural network (DNN) architectures containing two hidden layers (Supplementary Fig. 2b), indicating that a singlehiddenlayer architecture is sufficient to model the relationship E_{f} and the descriptors.
Averaged neural network models for mixed garnets
To extend our model to mixed garnets, i.e., garnets with more than one type of species in the C, A, and D sites, we explored two alternative approaches—one based on averaging of descriptors, and another based on expanding the number of descriptors to account for the effect or species ordering. The data set for mixed garnets were created using the same species pool, but allowing two species to occupy one of the sites. Mixing on the A sites was set at a 1:1 ratio, and that on the C and D sites was set at a 2:1 ratio, generating garnets of the form C_{3}A’A”D_{3}O_{12} (211 compositions), C’C’’_{2}A_{2}D_{3}O_{12} (445 compositions), and C_{3}A_{2}D’D’’_{2}O_{12} (116 compositions)_{.} For each composition, we calculated the energies of all symmetrically distinct orderings within a single primitive unit cell of the garnet. All orderings must belong to a subgroup of the \(Ia\overline 3 d\) garnet space group.
In the first approach, we characterized each C, A, or D site using weighted averages of the ionic radii and electronegativities of the species present in each site, given by the following expressions (see Methods):
where X and Y are the species present in a site with fraction x and (1−x), respectively, and O refers to the element oxygen. The implicit assumption in this “averaged” ANN model is that species X and Y are completely disordered, i.e., different orderings of X and Y result in negligible DFT energy differences.
Using the same 6241 ANN architecture, we fitted an “averaged” model using the energy of the ground state ordering of the 635 unmixed and 772 mixed garnets. We find that the training, validation, and test MAEs of the optimized model are 22, 26, and 26 meV atom^{−1}, respectively (Supplementary Fig. 3a). These MAEs are about double that of the unmixed ANN model, but still comparable to the error of the DFT E_{f} relative to experiments. The larger MAEs may be attributed to the fact that the effect of species orderings on the crystal energy is not accounted for in this “averaged” model.
Ordered neural network model for mixed garnets
In the second approach, we undertook a more ambitious effort to account for the effect of species orderings on crystal energy. Here, we discuss the results for species mixing on the C site only, for which the largest number of computed compositions and orderings is available. For 2:1 mixing, there are 20 symmetrically distinct orderings within the primitive garnet cell, which can be encoded using a 5bit binary array [b_{0},b_{1},b_{2},b_{3},b_{4}]. This binary encoding scheme is significantly more compact that the commonly used onehot encoding scheme, and hence, minimizes the increase in the descriptor dimensionality. We may then modify Eq. 1 as follows:
where the electronegativities and ionic radii of both species on the C sites are explicitly represented. In contrast to the “averaged” model, we now treat the 20 orderingE_{f} pairs at each composition as distinct data points. Each unmixed composition was also included as 20 data points with the same descriptor values and E_{f}, but different binary encodings.
We find that a twohiddenlayer DNN is necessary to model this more complex compositionorderingenergy relationship. The final optimized 132281 model exhibits overall training, validation and test MAEs of ~11–12 meV atom^{−1} on the entire unmixed and mixed dataset (Supplementary Fig. 3b). The comparable MAEs between this extended DNN model and the unmixed ANN model is clear evidence that the DNN model has successfully captured the additional effect of orderings on E_{f}. We note that the average standard deviation of the predicted E_{f} of different orderings of unmixed compositions using this extended DNN model is only 2.8 meV atom^{−1}, indicating that the DNN has also learned the fact that orderings of the same species on a particular site have little effect on the energy. Finally, similar MAEs can be achieved for A and D site mixing (Supplementary Fig. 3c and 3d) using the same approach.
Stability classification of garnets using ANN models
While E_{f} is a good target metric for a predictive ANN model, the stability of a crystal is ultimately characterized by its E_{hull}. Using the predicted E_{f} from our DNN models and precalculated DFT data from the Materials Project^{31}, we have computed E_{hull} by constructing the 0 K CADO phase diagrams. From Fig. 4a, we may observe that the extended Cmixed DNN model can achieve a >90% accuracy in classifying stable/unstable unmixed garnets at a strict E_{hull} threshold of 0 meV atom^{−1} and rises rapidly with increasing threshold. Similarly, high classification accuracies of greater than 90% are achieved for all three types of mixed garnets. Given the great flexibility of the garnet prototype in accommodating different species, there are potentially millions of undiscovered compositions. Even using our restrictive protocol of singlesite mixing in specified ratios, 8427 mixed garnet compositions can be generated, of which 2307 are predicted to have E_{hull} of 0 meV atom^{−1}, i.e., potentially synthesizable (Supplementary Fig. 4a). A web application that computes E_{f} and E_{hull} for any garnet composition using the optimized DNNs has been made publicly available for researchers at http://crystals.ai.
Neural network models for unmixed and mixed perovskites
To demonstrate that our proposed approach is generalizable and not specific to the garnet crystal prototype, we have constructed similar neural network models using a dataset of 240 unmixed, 222 Amixed and 80 Bmixed ABO_{3} perovskites generated using the species in Supplementary Table 5. We find that a 4121 singlehiddenlayer neural network is able to achieve MAEs of 21–34 meV atom^{−1} in the predicted E_{f} for unmixed perovskites (Fig. 3c), while two 10241 neural networks are able to achieve MAEs of 22–39 meV atom^{−1} in the E_{f} of the mixed perovskites (Supplementary Fig. 5). These MAEs are far lower than those of prior ML models of unmixed perovskites, which generally have MAEs of close to 100 meV atom^{−1} or higher^{9,16}. As shown in Fig. 3b, the accuracy of classifying stable versus unstable perovskites exceeds 80% at a strict E_{hull} threshold of 0 meV atom^{−1} and maintains at above 70% at a loosened E_{hull} threshold of 30 meV atom^{−1}. During the review of this work, a new work by Li et al.^{32} reported achieving comparable MAEs of ~28 meV atom^{−1} in predicting the E_{hull} of perovskites using a kernel ridge regression model. However, this performance was achieved using a set of 70 descriptors, with model performance sharply dropping with less than 70 descriptors. Furthermore, Li et al.’s model is restricted to perovskites with E_{hull} < 400 meV atom^{−1} and only a single ordering for each mixed perovskite, while in this work, the highest E_{hull} is 747 meV atom^{−1} for the perovskite dataset and all symmetrically distinct orderings on the A and B sites within a √2×√2×1 orthorhombic conventional perovskite unit cell (ten structures each) are considered.
Discussion
To summarize, we have shown that NN models can quantify the relationship between traditionally chemically intuitive descriptors, such as the Pauling electronegativity and ionic radii, and the energy of a given crystal prototype. A key advantage of our proposed NN models is that they rely only on an extremely small number (two) of sitebased descriptors, i.e., no structural degrees of freedom are considered beyond the ionic radii of a particular species in a site and the ordering of the cations in the mixed oxides. This is in stark contrast to most machinelearning models in the literature utilizing a large number of correlated descriptors, which render such models highly susceptible to overfitting, or machinelearning forcefields, which can incorporate structural and atomic degrees of freedom but at a significant loss of transferability to different compositions. Most importantly, we derive two alternative approaches—a rigorously defined averaging scheme to model complete cation disorder and a binary encoding scheme to account for the effect of orderings—to extend highperforming unmixed deep learning models to mixed cation crystals with little/no loss in error performance and minimal increase in descriptor dimensionality. It should be noted that our NN models are still restricted to the garnet and perovskite compositions (with or without cation mixing) with no vacancies, though further extensions to other common crystal structure prototypes and to account for vacancies should in principle be possible. Finally, we show how predictive models of E_{f} can be combined with existing large public databases of DFT computed energies to predict E_{hull} and hence, phase stability. These capabilities can be used to efficiently traverse large chemical spaces of unmixed and mixed crystals to identify stable compositions and orderings, greatly accelerating the potential for novel materials discovery.
Methods
DFT calculations
All DFT calculations were performed using Vienna ab initio simulation package (VASP) within the projector augmentedwave approach^{33,34}. Calculation parameters were chosen to be consistent with those used in the Materials Project, an open database of precomputed energies for all known inorganic materials^{31}. The PerdewBurkeErnzehof generalized gradient approximation exchangecorrelation functional^{35} and a planewave energy cutoff of 520 eV were used. Energies were converged to within 5 × 10^{−5} eV atom^{1}, and all structures were fully relaxed. For mixed compositions, symmetrically distinct orderings within the 80atom primitive garnet unit cell and the 40atom √2×√2×1 orthorhombic perovskite supercell were generated using the enumlib library^{36} via the Python Materials Genomics package.^{37}
Training of ANNs
Training of the ANNs was carried out using the Adam optimizer^{38} at a learning rate of 0.2, with the mean square error of E_{f} as the loss metric. For each architecture, we ran with a random 64:16:20 split of training, validation and test data, i.e., random subsampling cross validation.
Electronegativity averaging
Pauling’s definition of electronegativity is based on an “additional stabilization” of a heteronuclear bond X–O compared to average of X–X and O–O bonds, as follows.
where χ_{X} and χ_{O} are the electronegativities of species X and O, respectively, and E_{d} is the dissociation energy of the bond in parentheses. Here, O refers to oxygen.
For a disordered site containing species X and Y in the fractions x and (1−x), respectively, we obtain the following:
We then obtain the effective electronegativity for the disordered site as follows:
Data availability
The datasets generated during and/or analysed during the current study are available in the GitHub repository https://github.com/materialsvirtuallab/garnetdnn as well as the Dryad Digital Repository (doi: 10.5061/dryad.760r5b6). A web application that estimates E_{f} and E_{hull} for any given garnet or perovskite composition using the optimized DNNs is available at http://crystals.ai/.
References
 1.
Ong, S. P., Wang, L., Kang, B. & Ceder, G. Li−Fe−P−O2 phase diagram from first principles calculations. Chem. Mater. 20, 1798–1807 (2008).
 2.
Hohenberg, P. & Kohn, W. Inhomogeneous electron gas. Phys. Rev. 136, B864 (1964).
 3.
Pilania, G., Wang, C., Jiang, X., Rajasekaran, S. & Ramprasad, R. Accelerating materials property predictions using machine learning. Sci. Rep. 3, 2810 (2013).
 4.
Lee, J., Seko, A., Shitara, K., Nakayama, K. & Tanaka, I. Prediction model of band gap for inorganic compounds by combination of density functional theory calculations and machine learning techniques. Phys. Rev. B 93, 115104 (2016).
 5.
Schmidt, J. et al. Predicting the thermodynamic stability of solids combining density functional theory and machine learning. Chem. Mater. 29, 5090–5103 (2017).
 6.
Pilania, G. et al. Machine learning bandgaps of double perovskites. Sci. Rep. 6, 19375 (2016).
 7.
Isayev, O. et al. Universal fragment descriptors for predicting properties of inorganic crystals. Nat. Commun. 8, 15679 (2017).
 8.
Meredig, B. et al. Combinatorial screening for new materials in unconstrained composition space with machine learning. Phys. Rev. B 89, 094104 (2014).
 9.
Faber, F. A., Lindmaa, A., Von Lilienfeld, O. A. & Armiento, R. Machine learning energies of 2 million elpasolite (ABC_{2}D_{6}). Phys. Rev. Lett. 117, 135502 (2016).
 10.
Ward, L., Agrawal, A., Choudhary, A. & Wolverton, C. A generalpurpose machine learning framework for predicting properties of inorganic materials. npj Comput. Mater. 2, 16028 (2016).
 11.
Ward, L. et al. Including crystal structure attributes in machine learning models of formation energies via Voronoi tessellations. Phys. Rev. B 96, 024104 (2017).
 12.
Seko, A., Hayashi, H., Nakayama, K., Takahashi, A. & Tanaka, I. Representation of compounds for machinelearning prediction of physical properties. Phys. Rev. B 95, 144110 (2017).
 13.
Sun, W. et al. The thermodynamic scale of inorganic crystalline metastability. Sci. Adv. 2, e1600225–e1600225 (2016).
 14.
Hautier, G., Ong, S. P., Jain, A., Moore, C. J. & Ceder, G. Accuracy of density functional theory in predicting formation energies of ternary oxides from binary oxides and its implication on phase stability. Phys. Rev. B  Condens. Matter Mater. Phys. 85, 155208 (2012).
 15.
LeCun, Y., Bengio, Y. & Hinton, G. Deep learning. Nature 521, 436–444 (2015).
 16.
Pauling, L. The principles determining the structure of complex ionic crystals. J. Am. Chem. Soc. 51, 1010–1026 (1929).
 17.
Goldschmidt, V. M. Die Gesetze der Krystallochemie. Naturwissenschaften 14, 477–485 (1926).
 18.
Nakamura, S. Present performance of InGaNbased blue/green/yellow LEDs. Proc. SPIE 3002, 26–35 (1997).
 19.
O’Callaghan, M. P., Lynham, D. R., Cussen, E. J. & Chen, G. Z. Structure and ionictransport properties of lithiumcontaining garnets Li _{3} Ln _{3} Te _{2} O_{12} (Ln = Y, Pr, Nd, Sm−Lu). Chem. Mater. 18, 4681–4689 (2006).
 20.
Peng, H., Wu, Q. & Xiao, L. Low temperature synthesis of Li_{5}La_{3}Nb_{2}O_{12} with cubic garnettype structure by sol–gel process. J. Sol.Gel Sci. Technol. 66, 175–179 (2013).
 21.
Kobayashi, K.I., Kimura, T., Sawada, H., Terakura, K. & Tokura, Y. Roomtemperature magnetoresistance in an oxide material with an ordered doubleperovskite structure. Nature 395, 677–680 (1998).
 22.
Cava, R. J. et al. Bulk superconductivity at 91 K in singlephase oxygendeficient perovskite Ba_{2}YCu_{3}O_{9 − δ}. Phys. Rev. Lett. 58, 1676–1679 (1987).
 23.
Cohen, R. E. Origin of ferroelectricity in perovskite oxides. Nature 358, 136–138 (1992).
 24.
Grinberg, I. et al. Perovskite oxides for visiblelightabsorbing ferroelectric and photovoltaic materials. Nature 503, 509–512 (2013).
 25.
Green, M. A., HoBaillie, A. & Snaith, H. J. The emergence of perovskite solar cells. Nat. Photonics 8, 506–514 (2014).
 26.
Pauling, L. The nature of the chemical bond. IV. The energy of single bonds and the relative electronegativity of atoms. J. Am. Chem. Soc. 54, 3570–3582 (1932).
 27.
Shannon, R. D. Revised effective ionic radii and systematic studies of interatomie distances in halides and chaleogenides. Acta Cryst. A32, 751–767 (1976).
 28.
Wang, L., Maxisch, T. & Ceder, G. Oxidation energies of transition metal oxides within the GGA + U framework. Phys. Rev. B 73, 195107 (2006).
 29.
Hornik, K. Approximation capabilities of multilayer feedforward networks. Neural Netw. 4, 251–257 (1991).
 30.
Granatstruktur, D., Ubersicht, E., Kationen, D. & Ionenverteilung, D. Crystal chemistry of the garnet. Z. für Krist.  Cryst. Mater. 47, 1989 (1999).
 31.
Jain, A. et al. Commentary: the materials project: a materials genome approach to accelerating materials innovation. APL Mater. 1, 011002 (2013).
 32.
Li, W., Jacobs, R. & Morgan, D. Predicting the thermodynamic stability of perovskite oxides using machine learning models. Comput. Mater. Sci. 150, 454–463 (2018).
 33.
Kresse, G. & Furthmüller, J. Efficient iterative schemes for ab initio totalenergy calculations using a planewave basis set. Phys. Rev. B 54, 11169–11186 (1996).
 34.
Blöchl, P. E. Projector augmentedwave method. Phys. Rev. B 50, 17953–17979 (1994).
 35.
Perdew, J. P., Burke, K. & Ernzerhof, M. Generalized gradient approximation made simple. Phys. Rev. Lett. 77, 3865–3868 (1996).
 36.
Hart, G. L. W., Nelson, L. J. & Forcade, R. W. Generating derivative structures at a fixed concentration. Comput. Mater. Sci. 59, 101–107 (2012).
 37.
Ong, S. P. et al. Python materials genomics (pymatgen): a robust, opensource python library for materials analysis. Comput. Mater. Sci. 68, 314–319 (2013).
 38.
Kingma, D. P. & Jimmy Ba, Adam: a method for stochastic optimization. Preprint at https://arxiv.org/pdf/1412.6980 (2016).
 39.
Chollet, F. et al. Keras. http://keras.io (2015).
 40.
Abadi, M. M. et al. TensorFlow: largescale machine learning on heterogeneous distributed systems. http://www.tensorflow.org/ (2015).
Acknowledgements
This work is supported by the Samsung Advanced Institute of Technology (SAIT)’s Global Research Outreach (GRO) Program. The authors also acknowledge data and software resources provided by the Materials Project, funded by the U.S. Department of Energy, Office of Science, Office of Basic Energy Sciences, Materials Sciences and Engineering Division under Contract No. DEAC0205CH11231: Materials Project program KC23MP, and computational resources provided by Triton Shared Computing Cluster (TSCC) at the University of California, San Diego, the National Energy Research Scientific Computing Centre (NERSC), and the Extreme Science and Engineering Discovery Environment (XSEDE) supported by National Science Foundation under Grant No. ACI1053575. The authors would also like to express their gratitude to Professors Darren Lipomi and David Fenning from the University of California, San Diego, and Dr Anubhav Jain from Lawrence Berkeley National Laboratory for helpful comments on the manuscript.
Author information
Affiliations
Contributions
S.P.O., W.Y. and C.C. proposed the concept. W.Y. carried out the calculations and analysis with the help from C.C., Z.W. and I.C. W.Y. prepared the initial draft of the manuscript. All authors contributed to the discussions and revisions of the manuscript.
Corresponding author
Ethics declarations
Competing interests
The authors declare no competing interests.
Additional information
Publisher's note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Electronic supplementary material
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Ye, W., Chen, C., Wang, Z. et al. Deep neural networks for accurate predictions of crystal stability. Nat Commun 9, 3800 (2018). https://doi.org/10.1038/s4146701806322x
Received:
Accepted:
Published:
Further reading

Generative adversarial networks (GAN) based efficient sampling of chemical composition space for inverse design of inorganic materials
npj Computational Materials (2020)

Ab initio calculations of conduction band effective mass parameters of thermoelectric $${\hbox {Mg}}_{2} {\hbox {X}}_{1{}x} {\hbox {Y}}_x$$ (X, Y = Si, Ge, Sn) alloys
Scientific Reports (2020)

Automated design of a convolutional neural network with multiscale filters for costefficient seismic data classification
Nature Communications (2020)

Emerging materials intelligence ecosystems propelled by machine learning
Nature Reviews Materials (2020)

Acceleration of PDEBased Biological Simulation Through the Development of Neural Network Metamodels
Current Pathobiology Reports (2020)
Comments
By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.