ChemComm

View Article Online

COMMUNICATION View Journal | View Issue

Active learning and neural network potentials

This article is licensed under a Creative Commons Attribution 3.0 Unported Licence.

accelerate molecular screening of ether-based

Cite this: Chem. Commun., 2020,
solvate ionic liquids†
Open Access Article. Published on 16 June 2020. Downloaded on 7/4/2023 5:59:58 AM.

56, 8920

Received 16th May 2020,

Accepted 16th June 2020 Wujie Wang, Tzuhsiung Yang, William H. Harris and Rafael Gómez-Bombarelli *

DOI: 10.1039/d0cc03512b

rsc.li/chemcomm

Solvate ionic liquids (SIL) have promising applications as electrolyte 103 ether molecules seeking to maximize ligand-ion binding energy,
materials. Despite the broad design space of oligoether ligands, the electrochemical stability of the chelate, and the size of the
most reported SILs are based on simple tri- and tetraglyme. Here, complex. Neural-network-accelerated atomistic simulations allow
we describe a computational search for complex ethers that can identifying a number of synthetically accessible ligands with stronger
better stabilize SILs. Through active learning, a neural network affinities than the baselines and mapping the chemical space of
interatomic potential is trained from density functional theory data. oligoether–ion interactions for SILs (Scheme 1).
The learned potential fulfills two key requirements: transferability An optimal novel SIL ligand would form strongly-coordinated
across composition space, and high speed and accuracy to find chelates with small hydrodynamic radius and weak ionic inter-
low-energy ligand-ion poses across configurational space. Candi- action to allow high cation transference number, low viscosity,
date ether ligands for Li+, Mg2+ and Na+ SILs with higher binding and low melting temperature.2,11 Common salts are typically not
affinity and electrochemical stability than the reference com- dissociative enough to form SILs, which results in the need to use
pounds are identified. Lastly, their properties are related to the complex and expensive anions such as TFSI.12 New ligands with
geometry of the coordination sphere. higher binding affinities than current glymes can enable SILs with
less costly – but also less dissociative – anions than TFSI by
Solvate ionic liquids (SILs) consist of chelated metal cations strongly solubilizing ions. The chemical space of ligands for SILs
and small molecular anions. They are synthesized by mixing a is vast, even if only ether derivatives are considered, since small
salt and a ligand, commonly an oligoether. Prepared in an variations in alkyl spacers can have large effects on the binding
equimolar ratio, the cation (e.g., Li+, Na+, or Mg2+) and the energy. Charting this design space experimentally is slow and
oligoether form a stable coordination complex where oxygen
donors in the ligand coordinate the cation. This results in SILs
with better thermal and oxidative stability than the two com-
ponents alone and high ionic conductivity like a conventional
IL. The properties of these SILs can be tuned by adjusting the
ligand chemistry.1–4 SILs synthesized with tri- and tetraethylene
glycol dimethyl ethers (triglyme (G3) and tetraglyme (G4),
respectively) and lithium bis(trifluoromethylsulfonyl)imide (LiTFSI),
have been extensively studied in experiments1,3,5,6 and molecular
dynamics simulations,7–10 and demonstrate the potential of SILs as a
promising alternative to conventional ILs for electrolyte applications.
Hence, finding SILs with desired melting temperature, oxidative
stability, transport properties and cation-ligand association can
advance battery technologies. We screen a combinatorial library of
Scheme 1 An automated machine learning workflow to explore the
chemical space of ion–ether complexes. We train neural network poten-
Department of Materials Science and Engineering Massachusetts Institute of tials that take 3D molecular graphs as inputs. The loop iterates through
Technology, 77 Massachusetts Avenue, Cambridge, MA 02319, USA. data generation, validation and model training to actively explore the
E-mail: [email protected] configurational spaces. The trained neural network potentials calculate
† Electronic supplementary information (ESI) available. See DOI: 10.1039/ cation-ligand binding poses by running short molecular dynamics simula-
d0cc03512b tions followed by geometry optimization.

8920 | Chem. Commun., 2020, 56, 8920--8923 This journal is © The Royal Society of Chemistry 2020
 View Article Online

Communication ChemComm

costly. Virtual screening allows quantification of the ion-ligand

binding energy with accurate ab initio simulations by identifying
the ligand-ion pose with the lowest energy and comparing with
the energies of free ligand and ion.13
The exhaustive configurational sampling needed to evaluate
every ion and ligand pair is computationally demanding,
because the ion-ligand potential energy surface (PES) has many
local minima. Algorithms such as simulated annealing or
seeding gradient-based optimization from different random
This article is licensed under a Creative Commons Attribution 3.0 Unported Licence.

guesses can perform such global optimization but are typically

too costly to be combined with quantum chemical methods.
Open Access Article. Published on 16 June 2020. Downloaded on 7/4/2023 5:59:58 AM.

Classical force fields are possible alternatives to the expensive

ab initio methods for sampling. However, they describe electro-
static interactions poorly and struggle to capture the interac-
tions between ions and lone pairs. This results in inaccurate
structure guesses which, in the best case, require extensive Fig. 1 (A) The validation accuracy improves with each round of active
refinement with, and in the worse case, do not correspond with learning for Li–ether clusters with a 2D t-SNE projection of the SOAP
fingerprints of geometries obtained from active learning. The dispersity of
the global minimum of the quantum chemical PES. In order to
the training data increases from the first generation (purple) to the ninth
combine higher accuracy and fast sampling, we trained neural generation (yellow). (B) Distribution of true force components of NN
network (NN) interatomic potentials based on graph convolu- optimized geometries. (C) Root Mean Squared Distances (RMSD) between
tions (GC)14,15 using density functional theory (DFT) energies DFT converged structures and NN optimized structures. The average
and atomic forces for a set of reference chemistries. An active RMSD for trained species and untrained species is 0.09 Å2 per atom and
0.19 Å2 per atom respectively.
learning (AL) strategy was used to acquire training data effi-
ciently by carrying out DFT only on areas with high model
uncertainty.16,17
AL augments the data space and automatically improves the the end of learning cycle the training set was composed of
model by iteratively acquiring training data and deploying mole- 22 763 geometries. To understand how AL helps explore con-
cular dynamics (MD) simulations with the learned potential. Addi- figurational space, we performed t-distributed stochastic neigh-
tional DFT simulations are carried out on points in the trajectory bor embedding (t-SNE) analysis20 on SOAP fingerprints21 of NN-
for which the model is uncertain. CGNN models trained with this generated configurations and colored the data by generations
approach showed high accuracy both in terms of energies and (Fig. 1A). The configurations in the 1st generation do not cover
forces at the end of the AL loop, and also in predicting converged enough configurational space and hence showed poor perfor-
ion-binding structures that agree with the DFT refined ones (Fig. 1). mance when the learned potential is used to run MD (Fig. 1A).
The screened library enumerates over 1000 linear and As the evolution progresses, the diversity in configurational
branched oligoethers with methyl and/or ethyl capping groups, space was systematically improved in keeping with the valida-
alkyl spacers of length one to five, and up to six ether oxygens. A tion accuracy of the model.
reference dataset of 44 oligoethers was used for training. Starting We then applied the trained NN models to sample and
geometries were obtained using RDKit,18 followed by DFT opti- optimize the conformations of more than 1000 oligoethers with
mizations at BP86-D3/def2-SVP level of theory using ORCA,19 and each of the three cations. The NN-optimized geometries are
augmented with randomly-generated ion–molecule poses. Normal shown to be close to equilibrium geometries which are
mode sampling of each DFT-optimized configuration was per- obtained with DFT refinements at BP86-D3/def2-SVP level of
formed at 500 K to obtain non-equilibrium geometries, upon theory, as shown in small forces and RMSD (Fig. 1B). The low-
which energy and gradient calculations were performed. This energy geometries were refined using a long-range-corrected
produced an initial training set of around 5000 geometries. The hybrid DFT functional, wB97X-D3/def2-TZVP.21 We first applied
first generation of GCNN was trained on these energy and our workflow on well-studied crown ether binding with Li+ and
gradient values. Constant-energy NN MD runs were then per- Na+. It was found that the calculated binding energies correlate
formed for 10–200 ps to explore the configurational space. The well with experimental data measured in gas-phase and correlate
predicted energies and forces at a fraction of frames were then reasonably with solvent data (see Table S2, ESI†). The [Li-18C6]+
validated by DFT calculations and added to the training data for crown ether–lithium complexes form SILs2 with [TFSI] ,22 but
a new model. The procedure was repeated until the accuracy of show varied binding behaviors across solvents.23,24 DFT calcula-
forces and energies converged. Fig. 1 shows the improvements tions of the binding energy of [Li-18C6] in organic solvents predict
in force and energy accuracy for all generations of NN models weak binding, despite its known ability to form SILs. Hence, we
trained on Li–oligoethers chemical space. After several rounds primarily utilize gas-phase DFT results as a proxy for the binding
of training, the resulting NN model for Li–ether complexes affinity of oligo–ethers in SILs.
shows accurate energy (0.83 kcal mol 1) and force predictions Through this NN-accelerated virtual screening workflow for
(0.50 kcal mol 1 Å 1) on newly-sampled poses from NN MD. At ion-ligand complexes, we identified synthetic targets that have

This journal is © The Royal Society of Chemistry 2020 Chem. Commun., 2020, 56, 8920--8923 | 8921
 View Article Online

ChemComm Communication
This article is licensed under a Creative Commons Attribution 3.0 Unported Licence.
Open Access Article. Published on 16 June 2020. Downloaded on 7/4/2023 5:59:58 AM.

Fig. 3 The chemical space of Li–ether binding mapped by the number of

ether groups: O: ion (color bar) and carbon and oxygen ratio (C : O) with G3, G4
and G5 marked by red square, pentagon and hexagon respectively. (A) Ebind
of the Li–ether complexes. Weakly-binding outliers correspond ethers
whose alkyl substituents crowd out oxygen-ion contact. (B) HOMO level of
Li–ether complexes. (C) Scatter plots HOMO and Ebind overlaid with the
Fig. 2 (A) Binding energy distributions and (B) selected symmetrical Pareto front.
structures from DFT optimization (w-B97xD3/def2-TZVP) that demon-
strate stronger binding energy than baselines for each glyme families of
interests with different ion. Some asymmetrical oligoethers may have
superior properties, but are likely less synthetically accessible, and thus Because the hydrodynamic radius of SIL complexes is related
reported in the ESI.† In the X:X:X:X short notation,: represents the oxygen to their self-diffusion,27 and hence to their conductivity, we have
and X is the number of carbons spacers or caps. Dashed vertical lines mark estimated the solvent-excluded volume of the converged ion-
the energy of baseline glyme molecules (G3, G4 or G5).
ligand complexes using a spherical probe with radius 1.9 A.
Fig. S6 in ESI† reports how the larger ethers, which usually have
higher affinities, are bulkier and might be more viscous.
stronger binding strength than and comparable oxidative sta- The HOMO level of the complex was used as a surrogate for
bility to the state-of-the-art G3 and G4 ligands. A few selected oxidative stability (ESI,† Fig. S7 reports the correlation between
complexes from our screened molecules, shown in Fig. 2, are ionization potential and HOMO levels). As shown in Fig. 3B and
predicted to be more stable than Li(G3)+ and Li(G4)+ baseline C the predicted stability follows the opposite trend to binding
molecules by 0.2–0.3 eV while maintaining a similar EH level. energy: as the length of the spacers increases, HOMO values
Li+ binding can be improved by increasing the number of become less negative (less oxidatively stable). This compensa-
oxygen atoms and by increasing spacer length (Fig. 3). The tion effect between binding strength and oxidative stability is
longer carbon spacers facilitate the arrangement of oxygen reflected in Fig. 2C. The Pareto front of HOMO and Ebind traces
atoms around the ion and therefore to the formation of favor- the line connecting all the optimal points, so that one property
able oxygen coordination structures. In particular, having 3-C cannot be improved without sacrificing the other. Interestingly
spacers is beneficial because it contributes to the formation of a G3 and G4 form SILs but are not on the Pareto for Li+ which
six membered-ring with low strain energy.25 Linear energy decom- suggests that it is possible to improve over them in both
position analysis suggests that larger rings contribute more to the binding and stability. G5 lies on the Pareto front, but is known
binding energy (Fig. S5 in ESI†). This hypothesis is further not form a SIL in experiments.
supported by geometrical analysis using t, an index ranging from Batteries based on sodium and magnesium are promising and
0 to 1 that captures the deviation from an ideal coordination earth-abundant alternatives for lithium, but they lack effective and
structure, a value of 1 corresponding to the lowest-energy coordi- stable electrolytes, so a similar analysis was performed for Mg2+
nation structure (tetrahedral and bi-pyramidal for 4- and 5- and Na+ complexes. Fig. S4 (ESI†) reports the Ebind and HOMO level
coordinated structures respectively).26 Strongly-binding ethers for Mg2+ and Na+ complexes. The trends about spacers are similar
have high t4 and t5. In comparison, the baseline G3 and G4 to Li+, although in the case of Mg2+, the number of ether oxygen
complexes are closer to square planar (4-coordination) and square atoms is a more dominating factor regardless of spacer. Interest-
pyramidal (5-coordination) geometries, resulting form the ingly, for both Mg2+ and Na+, G3, G4 and G5 lie on the Pareto front
constraints imposed by the short spacers. of HOMO and Ebind, implying that increased binding would come

8922 | Chem. Commun., 2020, 56, 8920--8923 This journal is © The Royal Society of Chemistry 2020
 View Article Online

Communication ChemComm

at cost in stability, and vice versa, for any of the other Pareto-optimal Notes and references
oligoether we have identified.
1 K. Ueno, K. Yoshida, M. Tsuchiya, N. Tachikawa, K. Dokko and
Mg2+ has a similar ionic radius as Li+ and thus the proposed M. Watanabe, J. Phys. Chem. B, 2012, 116, 11323–11331.
ligands for Li+ are also good candidates for Mg2+ SILs. Top 2 T. Mandai, K. Yoshida, K. Ueno, K. Dokko and M. Watanabe, Phys.
candidates have more than 1 eV improvement in binding Chem. Chem. Phys., 2014, 16, 8761–8772.
3 K. Ueno, R. Tatara, S. Tsuzuki, S. Saito, H. Doi, K. Yoshida, T. Mandai,
energy over reference glymes. Mg2+ complexes we calculated M. Matsugami, Y. Umebayashi, K. Dokko and M. Watanabe, Phys.
to have very deep HOMO levels due to the double positive Chem. Chem. Phys., 2015, 17, 8248–8257.
4 H. M. Kwon, M. L. Thomas, R. Tatara, A. Nakanishi, K. Dokko and
charges (Fig. S5A, ESI†). However, due to stronger electrostatic
M. Watanabe, Chem. Lett., 2017, 46, 573–576.
interactions, [Mg(Glyme)](TFSI) has very high melting points
This article is licensed under a Creative Commons Attribution 3.0 Unported Licence.

5 K. Yoshida, M. Nakamura, Y. Kazue, N. Tachikawa, S. Tsuzuki,

which hinder its use at room temperature.28 Our proposed S. Seki, K. Dokko and M. Watanabe, J. Am. Chem. Soc., 2011, 133,
molecular libraries include side chain modifications and bulk- 13121–13129.
6 K. Dokko, N. Tachikawa, K. Yamauchi, M. Tsuchiya, A. Yamazaki,
Open Access Article. Published on 16 June 2020. Downloaded on 7/4/2023 5:59:58 AM.

ier capping alkyl groups which can be used to tune SIL melting E. Takashima, J.-W. Park, K. Ueno, S. Seki, N. Serizawa and
temperature by increasing van der Waals interactions. Such M. Watanabe, J. Electrochem. Soc., 2013, 160, A1304–A1310.
design strategies have proven effective for tuning the melting 7 Y. Sun and I. Hamada, J. Phys. Chem. B, 2018, 122, 10014–10022.
8 A. Thum, A. Heuer, K. Shimizu and J. N. Canongia Lopes, Phys.
temperature and viscosity of conventional ionic liquids,29 Chem. Chem. Phys., 2020, 22, 525–535.
[Mg(Glyme)](TFSI)228 and [Li(Glyme)](TFSI).11 Selected geome- 9 W. Shinoda, Y. Hatanaka, M. Hirakawa, S. Okazaki, S. Tsuzuki,
K. Ueno and M. Watanabe, J. Chem. Phys., 2018, 148, 193809.
tries with side akyl groups for Mg2+ ether complexes which have
10 M. Callsen, K. Sodeyama, Z. Futera, Y. Tateyama and I. Hamada,
high binding energies in are reported in the ESI† (Fig. S2). J. Phys. Chem. B, 2017, 121, 180–188.
Experiments show relatively weaker binding in Na[Glyme]+ 11 T. Tamura, K. Yoshida, T. Hachida, M. Tsuchiya, M. Nakamura,
than the smaller lithium ion. This results in higher presence of Y. Kazue, N. Tachikawa, K. Dokko and M. Watanabe, Chem. Lett.,
2010, 39, 753–755.
uncoordinated glymes which compromise the stability of the 12 M. Huang, S. Feng, W. Zhang, J. Lopez, B. Qiao, R. Tatara,
electrolytes.30 Our simulations are in keeping, the larger radius L. Giordano, Y. Shao-Horn and J. A. Johnson, Chem. Mater., 2019,
of sodium ions results in weaker ion–dipole interactions and 31, 7558–7564.
13 B. Qiao, S. Mohapatra, J. Lopez, G. Leverick, R. Tatara, Y. Shibuya,
lower affinity than Li+ by about 1 eV. We thus focused on Y. Jiang, A. France-Lanord, J. C. Grossman, R. Gómez-Bombarelli,
potential candidates from the G4 and G5 families (Fig. 2), and J. Johnson and Y. Shao-Horn, ChemRxiv, 2020, DOI: 10.26434/
CHEMRXIV.12152820.V1.
observed relatively minor improvement over the baseline.
14 D. K. Duvenaud, D. Maclaurin, J. Aguilera-Iparraguirre, R. Gómez-
Design of optimal ligands for a desired ion requires exploring a Bombarelli, T. Hirzel, A. Aspuru-Guzik and R. P. Adams, Advances in
complex compositional and configurational design space. By Neural Information Processing Systems, 2015, pp. 2215–2223.
combining fast and transferable machine learning with accurate 15 K. T. Schütt, H. E. Sauceda, P.-J. J. Kindermans, A. Tkatchenko and
K.-R. R. Müller, J. Chem. Phys., 2018, 148, 241722.
DFT simulations through active learning, over 1000 candidate 16 J. S. Smith, B. Nebgen, N. Lubbers, O. Isayev and A. E. Roitberg,
oligoethers have been evaluated as ligands for Li+, Mg2+ and Na+ J. Chem. Phys., 2018, 148, 241733.
17 S. J. Ang, W. Wang, D. Schwalbe-Koda, S. Axelrod and R. Gomez-
SILs. The tradeoff between affinity, electrochemical stability and
Bombarelli, ChemRxiv, 2020, DOI: 10.26434/CHEMRXIV.11910948.V1.
size have been charted, and candidate ethers with better predicted 18 RDKit: Open-source cheminformatics, http://www.rdkit.org, [Online;
properties than the baseline compounds have been identified. accessed 05-June-2020].
The authors thank Bo Qiao for insightful discussions and 19 F. Neese, Wiley Interdiscip. Rev.: Comput. Mol. Sci., 2018, 8, e1327.
20 L. Van Der Maaten and G. Hinton, J. Mach. Learn. Res., 2008, 9,
providing helpful comments. W. W. thanks the Toyota Research 2579–2605.
Institute, T. Y. thanks Sumitomo Chemical, and W. H. thanks the 21 A. P. Bartók, R. Kondor and G. Csányi, Phys. Rev. B: Condens. Matter
MIT Energy Initiative for financial support. The computations in Mater. Phys., 2013, 87, 184115.
22 R. E. Dillon and D. F. Shriver, Chem. Mater., 1999, 11, 3296–3301.
this paper were executed at the Massachusetts Green High- 23 A. R. Fakhari and M. Shamsipur, J. Inclusion Phenom. Mol. Recognit.
Performance Computing Center with support from MIT Research Chem., 1996, 26, 243–251.
24 J. D. Lin and A. I. Popov, J. Am. Chem. Soc., 1981, 103, 3773–3777.
Computing.
25 T. Dudev and C. Lim, J. Am. Chem. Soc., 1998, 120, 4450–4458.
26 L. Yang, D. R. Powell and R. P. Houser, J. Chem. Soc., Dalton Trans.,
2007, 955–964.
Data availability 27 J. M. Slattery, C. Daguenet, P. J. Dyson, T. J. Schubert and I. Krossing,
Angew. Chem., Int. Ed., 2007, 46, 5384–5388.
The training data and optimized geometries are available for 28 K. Hashimoto, S. Suzuki, M. L. Thomas, T. Mandai, S. Tsuzuki,
download at https://github.com/learningmatter-mit/MLElectrolytes. K. Dokko and M. Watanabe, Phys. Chem. Chem. Phys., 2018, 20,
7998–8007.
29 S. Murray, R. O’Brien, K. Mattson, C. Ceccarelli, R. Sykora, K. West
and J. Davis, Angew. Chem., Int. Ed., 2010, 49, 2755–2758.
Conflicts of interest 30 P. Geysens, V. S. Rangasamy, S. Thayumanasundaram, K. Robeyns,
L. Van Meervelt, J. P. Locquet, J. Fransaer and K. Binnemans, J. Phys.
There are no conflicts to declare. Chem. B, 2018, 122, 275–289.

This journal is © The Royal Society of Chemistry 2020 Chem. Commun., 2020, 56, 8920--8923 | 8923