1. The short version / before you read / where this sits in the series

• There are three ways of searching for battery materials: (1) narrow down by calculation (DFT, or AI that imitates it), (2) learn from past experimental data, (3) run experiments automatically. The weight of the basis for an answer differs by type Sourced [Ref. 1, 4, 12].
• Read "predicted" and "confirmed by experiment" separately: a study that narrowed 12,831 known crystals to 21 solid-electrolyte candidates delivered a "prediction". Another, which narrowed 32.6 million candidates, synthesized four compositions, confirmed the predicted structure for one of them and measured its conductivity Sourced [Ref. 1, 4].
• What finally confirms is synthesis, measurement and cell testing: calculated and measured conductivities differ in conditions, and in one case differ by thousands of times. We could not confirm the cause of the gap from the paper Our calculation.

2. What we search for in battery materials — three ways of searching

Searches for battery materials fall into three types according to "what we want to decide". The table is this article's own arrangement.

The table is this article's own arrangement (commentary).

View original SVG, values and sources

Original SVG reference: Concept diagram (vector drawing). Three ways of searching for battery materials with MI. The basis of the answer differs: "calculation", "past experimental data" or "the experiment itself". This leads into the "ladder of confirmation" in Section 9. The grouping is this article's own; the examples are the studies listed in the references.

3. A vast map — feeling the number of candidates with our own calculation

Why do we need AI and calculation to narrow things down? Look at the number of candidates. Chen et al. started from 54 elements for battery materials [Ref. 4]. Just choosing sets of 3, 4 or 5 elements from them gives the following numbers of element choices (Our calculation).

• 3 elements: C(54, 3) = 24,804 ways
• 4 elements: C(54, 4) = 316,251 ways
• 5 elements: C(54, 5) = 3,162,510 ways

That counts only sets of elements. Even for the same set, differences in ratio (composition) and crystal structure multiply the candidates by orders of magnitude. In fact, Chen et al. made 32,598,079 initial candidates by replacing atoms in known crystal structures (the 2022 version of ICSD) with other elements of the same valence Sourced [Ref. 4]. Cubuk et al. reported a way, using physical insight, of screening billions of compositions for conductivity Sourced [Ref. 3].

The bottom row of the figure is the number of compositions Chen et al. actually tried to synthesize: 4. Looking at 32.6 million by calculation and four in the lab. This drop is why it matters to read "prediction" and "experiment" separately (our commentary). By analogy, it is like picking a few books from a whole library's collection and checking what is inside. The analogy has a limit: unlike a book, what a material "contains" can change with how it is synthesized.

4. Solid-electrolyte search 1 — a funnel from thousands of entries (Sendek et al.)

(1) What they did

In 2017, Sendek et al. reported a method that applies conditions one after another to narrow down 12,831 lithium-containing crystals in the Materials Project (a database of calculated materials [Ref. 11]). The flow is shown in Fig. 4 Sourced [Ref. 1].

• First narrowing: with conditions such as thermodynamic stability (energy above the convex hull, Ehull = 0 eV/atom), no transition metals, a band gap of 1 eV or more and an oxidation-side voltage of 4 V or more, 12,831 entries became 317 (a 92.2% reduction). The strictest condition was stability: only 11.5% (1,472) passed [Ref. 1].
• Conductivity classification: from literature values of measured lithium-ion conductivity (40 materials, of which 11 are superionic), logistic regression learned "superionic likeness". Superionic is defined there as 10⁻⁴ S/cm or more at room temperature. This classification took 317 entries to 21 [Ref. 1].
• Where the result stands: the 21 are "promising structures", and the authors say that few of them have been examined experimentally [Ref. 1]. So the outcome of this study is a prediction.

A follow-up from the same group reports a search in which machine learning chooses candidates and DFT molecular dynamics (DFT-MD, which follows the motion of atoms by calculation) computes conductivity. Compared with random search, it was 2.7 times more likely to find fast lithium-ion conductors, with at least a 45-fold improvement in the log-average room-temperature conductivity, and the classification model had an F1 score of 0.50, 3.5 times that of random guessing Sourced [Ref. 2 (preprint)]. The "conductivity" here is a value calculated by DFT-MD (single crystal), not a measured value. The candidates were also chosen from about 12,000 known materials that had been "synthesized and characterized experimentally" [Ref. 2].

5. Solid-electrolyte search 2 — from tens of millions of candidates to synthesis (GNoME, Chen et al.)

(1) GNoME: the number of "candidates" found by calculation

A Google DeepMind group trained graph neural networks at scale to predict the stability of crystals, and reported more than 2 million structures "below the current convex hull" (more stable by calculation than known materials). As one example of downstream use, they screened solid-electrolyte candidates Sourced [Ref. 5]. Specifically, with a holistic screen that excludes transition metals and filters by lithium fraction, they found 528 "promising lithium-ion conductors", 25 times the number in the earlier study (Sendek et al. [Ref. 1]). According to the methods section, this number comes from filters based on the band gap and on stability against a lithium-metal anode [Ref. 5]. So 528 is the result of a computational screen, not a count of measured conductivities.

The abstract also says that 736 of the stable structures had already been independently realized in experiments [Ref. 5]. That count refers to the stable structures as a whole, not to the 528 conductors. Meanwhile, Cheetham and Seshadri examined the claims of this work and commented that they found scant evidence for compounds that satisfy novelty, credibility and utility together, and that expertise in synthesis and crystallography needs to be built in Sourced [Ref. 6]. We set out the claims of the study and the commentary side by side and do not judge which is right (the evaluation of crystal-structure prediction is covered in our explainer on crystal structure prediction and stability).

(2) Chen et al.: from 32.6 million candidates to four compositions synthesized

A group at Microsoft (Azure Quantum) and the U.S. Pacific Northwest National Laboratory (PNNL) combined machine learning with large-scale cloud computing to search for solid electrolytes, and reported the work through to experimental confirmation Sourced [Ref. 4]. The flow was as follows.

1. Generating candidates: choose 54 elements and, in known crystal structures, replace atoms with elements of the same valence, giving 32,598,079 initial candidates.
2. Predicting stability: with machine-learning potentials (M3GNet, CHGNet and others) and first-principles calculations, about 500,000 were predicted to be potentially stable.
3. Property filters: using the electrochemical stability window (ESW), band gap, conductivity from ab initio molecular dynamics (AIMD) and so on, they narrowed to 18 promising candidates (8,795 compounds entered the ESW filter).
4. Synthesis and measurement: four rare-earth halides at the top (the Pna2₁ phase of Li₃YCl₆, Li₅YCl₈, Li₇Y₂Cl₁₃ and Na₂LiYCl₆) were synthesized and their structure and conductivity examined.

Here are the experimental results alongside the calculated values (the same content as the upper table of Fig. 5).

Source: Table 1 and the text of Chen et al. (arXiv version). Sourced. The paper notes that the P-3m1 phase of Li₃YCl₆ is a known material in which Li and Y share sites.

Three things can be read from this table.

• The predicted phase is not always formed: in three of the four, the target structure was not obtained and the known Li₃YCl₆ (P-3m1) formed instead. In one (Na₂LiYCl₆) the predicted structure was confirmed. The authors say confirmation by single-crystal XRD is in progress Sourced [Ref. 4].
• Calculated and measured values cannot be compared directly: the calculated value for Na₂LiYCl₆ (300 K) is 0.12 mS/cm (1.2×10⁻⁴ S/cm), and the measurement is 2.2×10⁻⁶ S/cm at 100 °C. Converting the measurement to 300 K (assuming the activation energy of 0.6 eV is constant) gives about 2×10⁻⁸ S/cm, so the calculated value is about 5,000 to 6,000 times larger Our calculation. However, the calculation is for the Li conductivity of an ideal ordered structure, while the measurement is on a pressed powder in which Na and Li are mixed and grain boundaries exist. We could not confirm the cause of the gap from the paper (Inference: several factors probably overlap).
• Experiments are laborious: the paper states that the crystallinity of the samples was strongly affected by humidity during handling and measurement, that the background from Kapton tape and broadened peaks made it hard to refine lattice parameters, and that small amounts of NaCl and LiCl remained in some compositions [Ref. 4].

6. Electrode materials — predicting voltage and capacity

(1) Capacity follows from the molar mass (our calculation)

The "capacity" of an electrode material is how much (charge of) lithium can be moved in and out per gram of material. When one Li carries one electron, the theoretical capacity is 26,801 mAh/mol divided by the molar mass (g/mol); that is the Faraday constant, 96,485 C/mol, divided by 3.6 to convert to mAh. Our calculation

• LiFePO₄ (molar mass 157.76): 26,801 / 157.76 ≈ 170 mAh/g
• LiCoO₂ (molar mass 97.87): ≈ 274 mAh/g if all the Li comes out; ≈ 137 mAh/g if only half is used (the "half" is an assumption for illustration)
• Graphite (LiC₆, per C₆): ≈ 372 mAh/g

This calculation helps with searching. It shows that, for capacity, materials built from light elements that can move as much Li as possible are advantageous. However, the amount of Li actually usable (not all can be removed), the voltage and the volume change need separate evaluation (our commentary).

(2) Voltage is computed from an energy difference

The idea of computing voltage is to take the energy (calculated by DFT) of the "lithium-inserted" and "lithium-removed" states, find the energy needed to remove lithium from the difference, and convert that to a voltage (our explanation). The 1997 paper by Aydinol et al., as its title says, calculated the intercalation voltage of lithium-transition-metal oxide electrodes from first principles [Ref. 15] (we checked it only at the title level). Hautier et al. evaluated phosphate cathode materials by high-throughput first-principles calculations and organized the voltage ranges by redox couple and the limits on capacity Sourced [Ref. 8].

(3) What machine learning predicts is a "calculated value"

Joshi et al. used voltage data from the Materials Project to predict electrode voltages with a deep neural network, a support vector machine and kernel ridge regression, and proposed nearly 5,000 candidates for Na- and K-ion batteries Sourced [Ref. 9]. Li et al. compiled calculated cathode voltages for 5,577 and 2,423 structure pairs (Li-ion and Li-free cathodes) and reported that a machine-learning model predicting voltage from chemical formulas showed state-of-the-art performance against two established models (Roost and CrabNet) Sourced [Ref. 10].

What matters is that the correct answers (labels) of these models are voltages calculated by DFT, not measured voltages. However accurate the AI, that accuracy is "how fast it can imitate DFT" (our commentary). How far DFT voltages differ from measured ones is a separate question, and we could not confirm it from the abstracts of these papers within the range we checked.

The right side of Fig. 6 is an example of how a voltage error acts on an energy estimate. For a material with a 3.5 V mean voltage, if the calculated voltage is off by 0.3 V, the specific-energy estimate is off by about 8.6% (Our calculation; the 0.3 V error is an assumption for illustration). Voltage multiplies directly into energy density, so the size of the calculation error bears on judging practicality.

7. Electrolytes and charging conditions — searching by running experiments automatically

So far we have looked at searching for materials by calculation. The other way is to run the experiment itself automatically and let AI choose the next condition. In batteries, the following have been reported (Fig. 7).

(1) Electrolyte formulation: Dave et al. (2022)

Dave et al. combined a robot (Clio) with an experiment planner based on Bayesian optimization (Dragonfly) and optimized electrolyte conductivity over one salt and a mixture of three solvents. In two working days and 42 experiments they identified six electrolytes suited to fast charging, and report that this was six times faster than random search by the same automated experiment. As a check in real batteries, in 220 mAh pouch cells (graphite and LiNi₀.₅Mn₀.₃Co₀.₂O₂), cells with the robot-developed electrolytes showed better fast-charging performance than a baseline electrolyte chosen in advance Sourced [Ref. 12].

The feature of this study is that it searched with measured values (conductivity) and then confirmed in cells. On the other hand, what was optimized was one metric, conductivity; lifetime, safety and cost are outside the study's evaluation (our commentary).

(2) Charging conditions: Attia et al. (2020) and Severson et al. (2019)

Attia et al. combined an early-prediction model (predicting lifetime from the first few cycles) with Bayesian optimization and identified a long-cycle-life protocol among 224 candidates of 6-step, 10-minute fast-charging protocols in 16 days. They state that an exhaustive search without early prediction would take more than 500 days (about 31 times, by Our calculation) Sourced [Ref. 13]. The study of Severson et al., on which that early prediction was based, tested 124 commercial lithium iron phosphate/graphite cells under fast-charging conditions to produce data with widely varying lifetimes of 150 to 2,300 cycles, and predicted lifetime from the first 100 cycles with a test error of 9.1% (classifying lifetime into two groups from the first 5 cycles had an error of 4.9%) Sourced [Ref. 14].

8. The data problem — collecting experimental conductivity data

MI starts from data. There is an example of organizing experimental conductivity data in the solid-electrolyte field. Hargreaves et al. released an expert-curated dataset of lithium solid-electrolyte conductivities. It holds 820 records measured by AC impedance spectroscopy, collected from 214 sources, each with a composition, an expert-assigned structural label and the conductivity at a particular temperature (5 to 873 °C). There are 403 unique compositions with a conductivity near room temperature (15 to 35 °C). With this data they trained a CrabNet-based classifier of high versus low ionic conductivity, as "a practical tool to aid experimentalists in prioritizing candidates" Sourced [Ref. 7].

Sendek et al. trained on 40 materials [Ref. 1]. Compared with that, the number of compositions is about ten times larger (403). But more data do not always mean higher accuracy. Even for the same composition, conductivity can change a great deal with the synthesis method, density, grain boundaries, and the temperature and method of measurement (our commentary). The data in papers are collections of values that contain such differences. The quality of the data and the record of measurement conditions set the ceiling of prediction.

9. Telling prediction from experiment — the ladder of confirmation

Here we line up the reports so far by "how far they were confirmed". For battery materials, it helps to divide the stages of confirmation into five rungs (this article's own arrangement).

1. Calculated prediction: prediction of stability, voltage or conductivity by calculation or AI
2. Synthesized, structure confirmed: actually make it, and confirm by powder XRD and so on that it is the target phase
3. Material property measured: measure conductivity and so on
4. Performance confirmed in a cell: make a battery and measure charge-discharge, lifetime and fast charging
5. Independent check, long-term, mass production: replication by other groups, long-term reliability, reproduction at scale

As Fig. 9 shows, the rung reached differs by report. The 528 of GNoME and the 21 of Sendek et al. are at rung 1 (prediction). Chen et al. went through rungs 1 to 3 (Na₂LiYCl₆), and confirmation in a cell is described as in progress Not yet confirmed. Dave et al. searched electrolyte formulations by experiment and reached rungs 3 and 4 (conductivity measurement and pouch cells). Among the reports covered in this article there was none for which rung 5 (independent verification, mass production) could be confirmed.

10. A materials engineer's view

What can be left to AI, and where expert judgment is needed

The table is this article's own arrangement (commentary).

11. Limits, what could not be confirmed, and summary

(1) Papers whose published text we could not check

We checked Chen et al. in the text of the preprint (arXiv) and only the abstract of the published version (JACS). The 2019 follow-up by Sendek et al. was checked from the abstract of the preprint (the final wording of the published version is unchecked). The 2017 paper of Sendek et al. was checked in the authors' public PDF. Hautier et al., Severson et al. and Aydinol et al. are cited from the abstract or the title.

(2) The cause of the gap between calculated and measured values

For Na₂LiYCl₆ of Chen et al., we showed by our conversion that calculated (AIMD) and measured conductivities differ by about 5,000 to 6,000 times. We could not confirm the cause of the gap (the difference between an ideal structure and a real sample, grain boundaries, mixed Na and Li, the assumption in the temperature extrapolation and so on) from the paper.

(3) From laboratory to mass production

Among the reports covered, there were none for which independent verification, long-term reliability or mass production could be confirmed. We do not cover applications of MI in companies' battery development, patents or commercialization.

(4) Materials and methods not covered

We do not cover negative-electrode materials, binders, electrolyte additives, interfaces (SEI), individual discussions of sulfide, oxide and polymer solid electrolytes, materials for sodium and multivalent-ion batteries, or extraction of battery data from the literature by large language models. All-solid-state batteries are covered in another field of this series.

(5) Developments in Japan

We did not survey efforts on MI for battery materials in Japan (national research projects, data infrastructure at research institutes).

12. How to read the yardsticks and standards

How good MI for battery materials is depends on what is taken as the correct answer (a DFT value or a measurement), how the evaluation data are split, what it is compared against (random search, exhaustive search, a baseline electrolyte) and how far it was confirmed (the ladder of Section 9). There are international standards for measuring the performance of batteries (cells) themselves, but we could not find an international standard specific to MI for battery materials (prediction models or searches). Here is what we could confirm (checked 22 October 2026).

The arrangement of the table is Our calculation (arrangement). The numbers, titles, editions and publication dates of IEC standards were checked on the official IEC pages (webstore.iec.ch) (22 October 2026). The texts of standards are paid, so only their scope is noted.

Cautions when reading (for materials engineers)

• A benchmark ranking is not the ability on your material: Matbench Discovery and the like evaluate "stability prediction" on calculated data. Ease of actual synthesis and performance in a battery must be confirmed separately.
• Tell prediction from experimental confirmation: always check which rung of the ladder in Section 9 a result is at. "Found by AI" is often a rung 1 story.
• Pitfalls of leakage and splitting: if materials of the same family straddle training and evaluation, accuracy is overstated (see our explainers on reading benchmarks and evaluation and on reading MI results and reproducibility). When training data are as few as the 40 of Sendek et al., the method of evaluation matters all the more.
• How strong a principle or standard is: the FAIR principles are recommendations, and IEC standards bind only when cited in law or a contract (the latter is our commentary). A project that has closed (BIG-MAP) and evaluations in progress are treated as Not yet confirmed.

13. Glossary

14. References

1. Sendek A. D., Yang Q., Cubuk E. D., Duerloo K.-A. N., Cui Y., Reed E. J. "Holistic computational structure screening of more than 12 000 candidates for solid lithium-ion conductor materials", Energy Environ. Sci. 10, 306–320 (2017), DOI 10.1039/c6ee02697d (text checked in the authors' public PDF) — doi.org/10.1039/c6ee02697d
2. Sendek A. D. et al. "Machine Learning-Assisted Discovery of Solid Li-Ion Conducting Materials", Chem. Mater. 31, 342–352 (2019), DOI 10.1021/acs.chemmater.8b03272 (the abstract of the preprint arXiv:1808.02470 was used for checking) — arxiv.org/abs/1808.02470
3. Cubuk E. D., Sendek A. D., Reed E. J. "Screening billions of candidates for solid lithium-ion conductors: A transfer learning approach for small data", J. Chem. Phys. 150, 214701 (2019), DOI 10.1063/1.5093220 — doi.org/10.1063/1.5093220
4. Chen C. et al. "Accelerating Computational Materials Discovery with Machine Learning and Cloud High-Performance Computing: from Large-Scale Screening to Experimental Validation", J. Am. Chem. Soc. 146, 20009–20018 (2024), DOI 10.1021/jacs.4c03849 (the text of the preprint arXiv:2401.04070 was used to check details) — arxiv.org/abs/2401.04070
5. Merchant A. et al. "Scaling deep learning for materials discovery", Nature 624, 80–85 (2023), DOI 10.1038/s41586-023-06735-9 — nature.com/articles/s41586-023-06735-9
6. Cheetham A. K., Seshadri R. "Artificial Intelligence Driving Materials Discovery? Perspective on the Article: Scaling Deep Learning for Materials Discovery", Chem. Mater. 36, 3490–3495 (2024), DOI 10.1021/acs.chemmater.4c00643 (open access) — doi.org/10.1021/acs.chemmater.4c00643
7. Hargreaves et al. "A database of experimentally measured lithium solid electrolyte conductivities evaluated with machine learning", npj Comput. Mater. 9, 9 (2023), DOI 10.1038/s41524-022-00951-z (open access) — doi.org/10.1038/s41524-022-00951-z
8. Hautier G. et al. "Phosphates as Lithium-Ion Battery Cathodes: An Evaluation Based on High-Throughput ab Initio Calculations", Chem. Mater. 23, 3495–3508 (2011), DOI 10.1021/cm200949v — doi.org/10.1021/cm200949v
9. Joshi R. P. et al. "Machine Learning the Voltage of Electrode Materials in Metal-Ion Batteries", ACS Appl. Mater. Interfaces 11, 18494–18503 (2019), DOI 10.1021/acsami.9b04933 — doi.org/10.1021/acsami.9b04933
10. Li H. H., Chen Q., Ceder G., Persson K. A. "Voltage Mining for (De)lithiation-Stabilized Cathodes and a Machine Learning Model for Li-Ion Cathode Voltage", ACS Appl. Mater. Interfaces 16, 69379–69387 (2024), DOI 10.1021/acsami.4c15742 — doi.org/10.1021/acsami.4c15742
11. Jain A. et al. "Commentary: The Materials Project: A materials genome approach to accelerating materials innovation", APL Mater. 1, 011002 (2013), DOI 10.1063/1.4812323 — doi.org/10.1063/1.4812323
12. Dave A. et al. "Autonomous optimization of non-aqueous Li-ion battery electrolytes via robotic experimentation and machine learning coupling", Nat. Commun. 13, 5454 (2022), DOI 10.1038/s41467-022-32938-1 (open access) — doi.org/10.1038/s41467-022-32938-1
13. Attia P. M. et al. "Closed-loop optimization of fast-charging protocols for batteries with machine learning", Nature 578, 397–402 (2020), DOI 10.1038/s41586-020-1994-5 — doi.org/10.1038/s41586-020-1994-5
14. Severson K. A. et al. "Data-driven prediction of battery cycle life before capacity degradation", Nat. Energy 4, 383–391 (2019), DOI 10.1038/s41560-019-0356-8 — doi.org/10.1038/s41560-019-0356-8
15. Aydinol M. K. et al. "Ab initio calculation of the intercalation voltage of lithium-transition-metal oxide electrodes for rechargeable batteries", J. Power Sources 68, 664–668 (1997), DOI 10.1016/s0378-7753(96)02638-9 — doi.org/10.1016/s0378-7753(96)02638-9
16. Dunn A. et al. "Benchmarking materials property prediction methods: the Matbench test set and Automatminer reference algorithm", npj Comput. Mater. 6, 138 (2020), DOI 10.1038/s41524-020-00406-3 (open access) — doi.org/10.1038/s41524-020-00406-3
17. Riebesell J. et al. "A framework to evaluate machine learning crystal stability predictions", Nat. Mach. Intell. 7, 836–847 (2025), DOI 10.1038/s42256-025-01055-1 (open access) — nature.com/articles/s42256-025-01055-1
18. IEC "IEC 62660-1:2018 Secondary lithium-ion cells for the propulsion of electric road vehicles — Part 1: Performance testing" (IEC Webstore page) — webstore.iec.ch/en/publication/28965
19. IEC "IEC 61960-3:2017 Secondary cells and batteries containing alkaline or other non-acid electrolytes — Secondary lithium cells and batteries for portable applications — Part 3: Prismatic and cylindrical lithium secondary cells and batteries made from them" (IEC Webstore page) — webstore.iec.ch/en/publication/29603
20. BIG-MAP "BIG-MAP (Battery Interface Genome – Materials Acceleration Platform)" (official site) — big-map.eu
21. Wilkinson M. D. et al. "The FAIR Guiding Principles for scientific data management and stewardship", Sci. Data 3, 160018 (2016), DOI 10.1038/sdata.2016.18 — doi.org/10.1038/sdata.2016.18
22. BIG-MAP "BIG-MAP" (project description page; used to check the name and the goal) — big-map.eu/big-map

15. Claim-to-source audit

Last updated 22 October 2026. Sources are limited to primary material (peer-reviewed papers, arXiv papers, and the official pages of IEC and BIG-MAP). The texts of IEC standards are paid and could not be checked. The calculation examples are computed by us from reported values and atomic masses, not measurement results on real batteries or materials. An international standard specific to MI for battery materials, a standard for measuring conductivity, the cause of the gap between calculated and measured conductivity, and reports in which independent verification or mass production could be confirmed could not be confirmed and are not stated here. All figures are explanatory concept graphics or drawings that include our calculations. Figs. 1 and 8 are ImageGen-generated conceptual illustrations, and none of them shows a real instrument or measurement result.