Tamara
Husch
and
Martin
Korth
*
Institute for Theoretical Chemistry, Ulm University, Albert-Einstein-Allee 11, 89069 Ulm, Germany. E-mail: martin.korth@uni-ulm.de
First published on 27th July 2015
Li–air batteries are very promising candidates for powering future mobility, but finding a suitable electrolyte solvent for this technology turned out to be a major problem. We present a systematic computational investigation of the known chemical space for possible Li–air electrolyte solvents. It is shown that the problem of finding better Li–air electrolyte solvents is not only – as previously suggested – about maximizing Li+ and O2− solubilities, but also about finding the optimal balance of these solubilities with the viscosity of the solvent. As our results also show that trial-and-error experiments on known chemicals are unlikely to succeed, full chemical sub-spaces for the most promising compound classes are investigated, and suggestions are made for further experiments. The proposed screening approach is transferable and robust and can readily be applied to optimize electrolytes for other electrochemical devices. It goes beyond the current state-of-the-art both in width (considering the number of compounds screened and the way they are selected), as well as depth (considering the number and complexity of properties included).
The identification of new Li–air electrolyte solvents was targeted in the past (see below for details and references). These studies investigated compounds or compound families that were hand-picked based on previously reported desirable properties or chemical intuition. As the chemical space of possible polar-aprotic, organic liquids is tremendously large, a rational decision-making model which investigates compounds experimentally is highly desirable. The mere vastness of the space under consideration does make it seem very likely, that improving upon the current solvents is possible. At least today it seems impossible to test compounds experimentally in a magnitude that allows systematic investigations in this sense, but we will show in the following that computational high-throughput screening now offers a way to probe the full known chemical space and make systematic investigations of full sub-spaces possible. Screening should be seen as complementary to detailed experimental and computational investigations, as it first requires a detailed insight into the relevant processes to identify suitable screening parameters, but offers then a way to transfer insight into innovation by reducing effort on trial-and-error procedures through rational pre-selection.
Large-scale computational screening in battery research was first applied to identify new inorganic materials for cathodes by Ceder and coworkers within the Materials Project.8 Other fields of renewable energy research have seen similar investigations. A prominent example is the Harvard Clean Energy Project that strives to identify organic molecules for photovoltaics.9 The scope of electronic structure theory based screening projects in renewable energy research reaches from a few thousands to a few millions (Materials Project: 60K,8 Harvard Clean Energy Project: 2.3 Mio9). The utilization of screening techniques to optimize battery electrolytes is still in its early stages. Several exploratory studies with a strong focus on redox stabilities were published in the past.10 Only this year the groups of Korth and shortly afterwards Curtiss published larger scale studies based on more properties than just redox stabilities.11,12
Another noticeable feature of all published studies is the choice of structural pool. The structural pool consists of known electrolyte molecules or the candidates that are derived from a given motif, which is identified from experimental insight or chemical intuition. The chemical space under investigation thus suffers from a ‘selection bias’,13 and the question of how to navigate chemical space needs to be addressed to alleviate the effects of this bias. First steps in this direction were made by us when evaluating computational methods at different theoretical levels for the identification of new battery electrolyte solvents.11,14 Here we chart the known chemical space represented by the largest publicly available database, to identify promising candidates and relevant structural motifs for new Li–air battery electrolyte solvents. As a second step we systematically investigate full sub-spaces for the most promising compound classes.
Our screening methodology itself goes beyond the current state-of-the-art by including computational estimates for all properties reported as relevant so far. By evaluating our data with respect to multiple properties at the same time, many false predictions are avoided, which is of utmost importance when making suggestions for subsequent experimental work. A reasonably large amount of knowledge on Li–air electrolyte solvents is available from both experimental4 and theoretical15–21 investigations. This allows us to validate our screening results for the known chemical space in the first part, thereby giving support to our suggestions for new compounds in the second part.
Very recent results by Johnson et al. highlight the importance of good Li+ solubilities, as they are related to changes in the morphology of the Li2O2 discharge product:5 solvents with poor Li+ solubility lead to Li2O2 film growth that is associated with low capacity, decaying rates and early cell death. In contrast, solvents with good Li+ solubility lead to particle growth, a higher capacity and sustained discharge. Shortly afterwards Luntz and coworkers reached the same conclusion, additionally emphasizing the importance of the O2− solubility.6 According to these studies, the problem is thus (to first approximation) two-dimensional: solvents with high solubilities for both Li+ and O2− should allow for high-capacity Li–air batteries. Both studies are based on quantifying ion solubility with Gutman donor (for cations) and/or acceptor (for anions) numbers (AN and DN). The COSMOtherm model allows us to compute solubilities based on input from quantum chemical calculations, thus providing an alternative that is suitable for large scale computational screenings. As high solubilities are indicated by zero by the COSMOtherm model, we turn to the chemical potential as defined in the COSMO-RS theory23 for quantification. We found the correlation between the donor number and the Li+ chemical potential to be high (R = 0.85), with substantial deviations only for very high donor numbers, which in turn do not correlate well with experimental data (cf. ESI† Section S1).
Structures were obtained from the PubChem Compound database.24 It comprised 67 million compounds at the time of retrieval. The PubChem database started in 2004 as a United States Government initiative and is maintained by the National Center for Biotechnology Information. Initially it was designed to collect information on (biological) activities of small molecules, but was since extended and many journals today automatically contribute to the extension. The PubChem Compound database is the closest image of the known chemical space that is publicly available. The only larger database in this field is the fee-based CAS registry (currently 91 million), which is commonly seen as a direct competitor.
For the first stage of the screening process we propose a hierarchical down-selection strategy as illustrated in Fig. 1. (Arguments for the validity of this strategy are given in ESI† Section S2.) The first steps at this stage are based on global criteria for organic, molecular electrolyte materials, i.e. compounds are discarded that are unlikely for application in any lithium battery technology. In the next steps, criteria specific for Li–air battery electrolytes are applied.
After retrieving and converting structures from the database (step 1), compounds were pre-screened based on simple rules (step 2): candidates with more than 18 heavy atoms or elements other than 1st and 2nd row elements, as well as pure hydrocarbons were excluded. The remaining 6.2 million structures were subjected to very fast COSMOfrag calculations25 to evaluate their viscosity (step 3). All compounds with a viscosity greater than 10.0 cP (at the COSMOfrag level) were discarded. The threshold for the viscosity was selected generously to not discard any compounds prematurely. The remaining 927 thousand polar, organic liquids were evaluated with respect to the chemical potential for Li+ within the bulk compound based on fast semiempirical quantum mechanical (SQM) PM6-DH+26 and SQM-based COSMOtherm calculations, leaving 188 thousand electrolyte solvent candidates after discarding all candidates with a chemical potential larger than 10 kcal mol−1 at this level of theory. These compounds were screened with respect to their ionization potential (IP), electron affinity (EA), viscosity, boiling point, flash point and the free energies of solvation and chemical potentials of Li+, O2− and O2 in the bulk compound at the SQM/COSMOtherm level. 48 thousand compounds with a viscosity below 5.0 cP, boiling points above 373 K and flash points above 343 K were optimized at the GGA-DFT level. (High boiling points are of importance also because the Li–air battery is open to ambient pressure, so that long-term stability can be problematic in the case of high vapor pressures.) All before mentioned collective properties have been re-evaluated using COSMOtherm on the basis of the BP86/TZVP calculations. Additionally, free energies of solvation and chemical potentials of water and carbon dioxide in the bulk compound were included, as the solubility of water and carbon dioxide should be very low to promote the use of air or less refined oxygen and minimize parasitic reactions.6 IP and EA values have been computed at higher levels of theory (Hybrid-DFT, CEPA27) to ensure sufficient accuracy. The final set of molecules comprises about 20 thousand potential polar, non-protic organic liquids that may be good candidates for application in Li–air batteries. Additional calculations of pKa values in DMSO using DFT/COSMOtherm were performed for this dataset, to estimate the chemical stability against nucleophilic attacks, H-abstraction reactions and autooxidation. We found DFT/COSMOtherm pKa predictions to be highly correlated (R = 0.99) with results from the best methods available, but computationally much cheaper (cf. ESI† Section S3). For each compound all possible proton abstractions were considered, and the lowest pKa was picked as the descriptor. A substantial change in molecular geometry after proton abstraction (e.g. ring opening) was taken as a sign of unsatisfactory electrochemical and chemical stability. Finally, QSPR melting point predictions were checked for selected compounds.
Our choice of screening parameters does not include estimators for every thinkable property, but every parameter previously identified as substantially important is included. Detailed follow-up investigations can be carried out subsequently for the most promising compounds. For this purpose, and to allow other researchers to try out different strategies for picking best compounds, the whole dataset will be made available on our project web page.28
Our following analysis will very much concentrate on taking into account the above-mentioned experimental results on the importance of the role of the different ion solubilities. Other experimentally working groups might want to question the significance of these results for the further development of Li–air batteries, but we will show below that also competing experimentally derived hypotheses can easily be incorporated as selection criteria within our screening approach.
As a second strategy we tested pre-filtering compounds to remove cases with poor viscosities and Li+/O2− chemical potentials. Candidates are sorted out that do not beat MeIm (which itself is beaten by DMSO) with respect to Li+ and O2− solubility. This only leaves 7 structures of which 5 are Pareto-optimal (cf. ESI,† Section S5). Among the suggestions is an amine oxide that most likely has a very high melting point. The next two hits, a dihydro-thiophene oxide and a phosphinic acid ester, are most likely reactive and cannot withstand nucleophilic attacks, which disappointingly leaves us only with DMSO and MeIm.
This surprising result that the best known compounds are already optimal (or at least very close to optimal) choices within the known chemical space (as pre-selected by the filtering) clearly merits further investigation. A key issue is a low chemical potential for Li+ in combination with a low chemical potential for O2−, which is the two-dimensional problem very recently identified by Luntz and co-workers. Our screening results now show that (even in first approximation) a third dimension needs to be considered: low chemical potentials for both ions, i.e. high donor and acceptor numbers are connected to strong intermolecular interactions in the pure bulk compound, i.e. a high viscosity and a high melting point. The problem is illustrated in Fig. 2.
![]() | ||
Fig. 2 Illustration of the problem by optimizing both donor and acceptor numbers (DN/AN): chemical potentials for Li+vs. chemical potentials for O2− in the bulk candidate compound are plotted for all 927![]() |
Candidates that clearly beat DMSO and MeIm with respect to both Li+ and O2− chemical potentials are highly viscous or solid at room temperature. The search for new Li–air battery electrolyte solvents should therefore not focus on maximum DN and AN numbers, but on finding the optimum balance between the two ion solubilities and viscosity. If a lower performance for one property can be tolerated, room is given to optimize the two other ones.
Given this directive, also compounds with somewhat higher Li+ or O2− chemical potentials become interesting. As a third strategy we therefore used relaxed filter thresholds, adding 10% to the respective values of MeIm. The results can be seen in ESI,† Section S6. The biggest share is again N-heterocycles with a too high melting point. Aromatic heterocycles with more than one nitrogen in the aromatic rings are especially interesting. Also sulfoxides are again among the hits, but they again show higher reactivity than DMSO going by the pKa. The suggestions also comprise phosphine oxides and phosphinic acid esters.
Our study shows impressively how good the choice of DMSO and MeIm is and how strong chemical intuition actually is. The PubChem database covers, on the other hand, only a tiny fraction of the relevant chemical space (which we estimate to be at least 1010 times larger), though due to its relative homogenity far less diversity is available than this number suggests.
Compound classb | μ(Li+) [kcal mol−1] | μ(O2−) [kcal mol−1] | Comment | |
---|---|---|---|---|
a Values averaged over all entries for each compound class, listed are classes with μ(Li+) > MeIm + 10%, for more details see ESI Section S7. b According to Checkmol classification. | ||||
Phosphine oxide |
![]() |
5.85 | 13.43 | Selected |
Phosphine |
![]() |
6.34 | 13.55 | Assumed to be reactive |
Carboxylic acid amidine |
![]() |
8.13 | 15.44 | |
Secondary aliphatic–aromatic amine alkylarylamine | N-rings | 8.91 | 12.65 | Selected |
Guanidine |
![]() |
9.16 | 16.32 | |
Sulfoxide |
![]() |
9.41 | 12.82 | Low pKas |
Imine |
![]() |
10.27 | 15.39 | Low pKas |
Carboxylic acid hydrazide |
![]() |
10.33 | 14.75 | Low pKas |
Lactam |
![]() |
10.42 | 14.80 | |
Secondary amine |
![]() |
10.46 | 15.29 | Selected |
Urea |
![]() |
10.46 | 15.42 | |
Oxohetarene |
![]() |
10.48 | 12.96 | High viscosities |
Secondary aliphatic amine dialkylamine | N-rings | 10.61 | 15.53 | |
Tertiary aliphatic–aromatic amine alkylarylamine | N-rings | 10.65 | 13.76 | High viscosities |
Azide | R–N![]() ![]() |
10.65 | 14.85 | Low pKas |
Tertiary carboxylic acid amide |
![]() |
10.84 | 14.98 |
Phosphine oxides show a very good overall performance, with excellent Li+ solubilities, but are known to have high melting points. This compound class was selected for further investigations to look for a candidate with a good balance between the Li+/O2− solubility and viscosity. Amidines, guanidines and imines are interesting suggestions, but though they show a good Li+ solubility, the balance with O2− is not very promising and the non-aromatic CN bond is likely reactive. The second compound class we chose for further investigations was heterocycles, especially those containing nitrogen. Various functional groups according to Haider fall in this category, for example numerous amines, lactames, and oxohetarenes. Their properties are overall very promising aside from somewhat high viscosities. Sulfoxides show as well a very good balance for the chemical potentials, but the average pKa is worrying. As within the PubChem no sulfoxide better than DMSO could be found and a large amount of sulfoxides was already included in the PubChem database, this compound class was not chosen for further studies here. (The properties of DMSO may nevertheless be succeeded by sulfoxides with additional functional groups of different types.) Derivates from carboxylic acids and urea may also be interesting for further investigation, but are not included in this study.
A comparison with the literature supports the validity of our screening results: several compound classes have been identified to fail as Li–air battery electrolyte solvents, for example organic carbonates, sulfonates, pure esters, lactones and ethers. Only one compound class that was previously excluded due to experimental or theoretical findings is found in our list of best performing ones. The compounds incorrectly listed are phosphinic acid esters, which were shown to be susceptible to nucleophilic attacks by Bryantsev et al.19 Though we do not cover this type of reactivity, pKas were lower than those for DMSO and MeIm and thus lower chemical stabilities are predicted. Our study is further supported by the fact that many compound classes previously identified as promising are among our hits. Examples are lactams, amides, phosphine oxides and N-heterocycles.4,16,19 We do not list nitriles, as we find comparably poor Li+ solubilities on average, but otherwise reasonably good properties. Our findings also indicate that when screening existing databases, further reactivity estimates beyond the pKa need to be included. Additional estimates may for example take the reactivity of double or ester bonds or highly strained ring structures into account. These findings can on the other hand easily be incorporated as rules for structure generation when generating new databases, as we will show in the following.
Based on our analysis of the average performance of compound classes, phosphine-oxides and heterocycles mainly containing nitrogen, but also oxygen and sulfur were chosen for further investigations. Turning away from the known chemical space of the PubChem database, we screened the full chemical sub-spaces of these compound classes within certain structural constraints. The Molgen algorithm41 was used to construct all possible structures for the relevant sub-spaces. To keep the number of structures manageable, structure generation was broken down into parts. For phosphine-oxides for instance we first looked at aliphatic structures and had to keep the overall number of atoms low, while in the second step we looked at cyclic phosphine-oxides, where a much larger overall number of atoms was possible, because many atoms were bound to end up in the enforced ring motif. Turning to heterocycles we first looked at mono- and bi-cycles, but constrained to aromatic systems and only considering nitrogen heteroatoms. As bi-cycles did not give promising results, we turned to 5–6 membered mono-cycles, still only considering nitrogen but now also non-aromatic systems. To investigate also oxygen and sulfur systems and N/O/S mixed ones we had to turn to constructing simple N/O/S 5- and 6-ring heterocycles first and add aliphatic rests to these core rings later on.
Overall, five different investigations were carried out: first, all phosphine-oxides P1O1C3–6 and their sulfur analogs P1S1C3–6 were constructed with no rings other than 5- to 7-membered ones allowed. We then applied the screening protocol outlined above for the PubChem database, starting at the DFT level (step 5). Out of 362 compounds we identified 10 Pareto-optimal structures with a promising balance of the chemical potential of Li+ and O2− (cf. ESI,† Section S8). Unfortunately all structures with a low melting point are most likely reactive, because they incorporate double/triple bonds or allene structures. The most promising structures, an aliphatic phosphinan-oxide and an aromatic phosphol-oxide (ID po138, po922), have an outstanding balance of the chemical potential of Li+ and O2− and good safety features, but are barely liquid. In a second run, all cyclic phosphine-oxides P1O1C5–10 were constructed with one ring enforced, only 5–6 membered rings allowed, and double or triple bonds except aromatic ones forbidden. Screening was again started at the DFT level with 926 compounds, but all were showing high μ(O2−) values. As a third step, all aromatic N-heterocycles N1–3C2–12 were constructed with the same constraints as for the cyclic phosphine-oxides, i.e. all aromatic N-based mono- and bi-cycles. It should be noted that the structure generator does only count ring-wise fully conjugated double bonds as aromatic, so that structures like MeIm are not included in this set. Screening was started at the DFT level (step 5) with 28356 structures, but no compound turned out to be competitive. As a forth step all (including non-aromatic) N-heterocycles Ni=1–3C3–(10−i) were constructed with one ring enforced, only 5–6 membered rings allowed, but now also double bonds other than aromatic ones allowed. 113
140 structures were evaluated at the SQM level (step 4), and 1290 at the DFT level (step 5). 102 candidates are competitive to MeIm, of which 39 are Pareto-optimal, and 18 of the latter are ‘simple’ in the sense that they contain no reactive binding motifs (like double or triple bonds outside the ring). As a fifth step, all simple, unsubstituted 5- and 6-ring heterocycles containing up to 3 nitrogen, oxygen or sulfur atoms were constructed (865 structures) and then all possibilities of attaching up to three carbon atoms to these core rings were evaluated (204
695 structures). 458 candidates are competitive to MeIm, of which 74 are Pareto-optimal, and 13 of the latter are ‘simple’ in the sense that they contain no reactive binding motifs and only one heteroatom species. (Compounds with only one heteroatom species should be more easily accessible to the experiment.) We list all 31 ‘simple’ heterocycle hits in ESI,† Section S9, but estimated melting points indicate that most of these compounds are again not very likely to be liquid at room temperature.
Further experiments are clearly necessary to find out if higher μ(O2−) can be tolerated to allow for lower μ(Li+) values at low viscosities. To make suggestions for the systematic experimental investigation into the optimal balance of Li+/O2− solubility and viscosity, we again turn to the PubChem database, thereby making sure that suggested compounds are (more or less) readily available. Section S10 (ESI†) gives a compilation of low-viscosity compounds with very different balances of the two relevant chemical potentials. From all compounds of the final set with a viscosity lower than 2 cP and a chemical potential for Li+ lower than 10 kcal mol−1, compounds with the lowest chemical potential for O2− are given for each 1 kcal mol−1 interval of μ(Li+).
The PubChem and heterocycle data are readily available for re-evaluation with adjusted filter thresholds if higher μ(O2−) can indeed be tolerated. Raw data for all screening runs will accordingly be made available on our project web page to encourage further investigations also by other researchers. As a first example we give a list of the most promising motifs for the case that the solubility of the negative species is actually of lesser importance (i.e. looking for compounds with a viscosity lower than 2 cP in combination with a lithium cation chemical potential below 10 kcal mol−1, as well as pKas higher than 25 and melting points lower than 10 °C) in ESI,† Section S11. Very different compounds are found in this analysis in comparison to the previous ones, now with an emphasis on amide and amine motifs, thus nicely illustrating how important the choice of selection criteria (and therefore an input from the experiment) is.
Footnote |
† Electronic supplementary information (ESI) available: See DOI: 10.1039/c5cp02937f. Additionally, raw data will be made available as a web-accessible database on http://qmcathome.org/clean_mobility_now.html. |
This journal is © the Owner Societies 2015 |