Expert-Guided LLM Reasoning for Battery Discovery: From AI-Driven Hypothesis to Synthesis and Characterization

Hannan Xu; Harry Guo; Huanxin Li; Shengchao Liu; Yan Ai; Yoshua Bengio

REVIEW 4 major objections 6 minor 52 references

Expert-Guided LLM Reasoning for Battery Discovery: From AI-Driven Hypothesis to Synthesis and Characterization

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read An expert-guided LLM pipeline, called ChatBattery, designed three cathode materials that outperformed NMC811 in synthesized coin cells.

desk verdict Real wet-lab cathode results, but the pipeline's selectivity is unproven and the core claim about LLM reasoning outruns the evidence. read the letter →

arxiv 2507.16110 v1 pith:IF6BOYMN submitted 2025-07-21 cs.AI cs.LG

Shengchao Liu , Hannan Xu , Yan Ai , Huanxin Li , Yoshua Bengio , Harry Guo This is my paper · ORCID

classification cs.AIcs.LG

keywords largelanguagemodelsbatterymaterialsdiscoveryNMC811cathodeexpert-guidedreasoningsynthesistheoreticalcapacityagenticpipelineelectrochemicalcharacterization

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Large language models have shown reasoning skill on math and coding, but this paper asks whether that reasoning can be steered by domain knowledge to invent battery materials. It introduces ChatBattery, a two-phase agentic pipeline that explores chemical space and then filters, ranks, and validates candidates, and it uses the pipeline to optimize the commercial cathode NMC811. The paper's central claim is that this pipeline identified three new cathode compositions, NMC-SiMg, NMC-SiCa, and NMC-MgB, that were actually synthesized, characterized, and tested in coin cells, reaching reversible capacities of 174, 169, and 160 mAh/g by the third cycle against roughly 135 mAh/g for NMC811. Those are gains of 28.8%, 25.2%, and 18.5%. The larger import is the demonstration that a complete design-to-synthesis-to-characterization loop can be driven by LLM reasoning within a few months.

What carries the argument

The load-bearing object is the ChatBattery pipeline itself: a two-phase loop with seven specialized agents. The mechanism that carries the argument is the substitution of slow experiments by cheap proxy functions at each filtering step: theoretical capacity computed from lithium count and molecular weight ($\mathrm{Capacity} = nF/(3.6M)$), an oxidation-state heuristic for total charge, element-count preparation complexity, and a large language model used as a qualitative voltage comparator whose pairwise answers are merged into a ranking. A second mechanism is the computational surrogate for structural validation: partially occupied candidate structures are expanded into fully occupied supercells and evaluated by a machine-learned interatomic potential, since direct density-functional calculations on large disordered cells are too costly. Together these proxies shrink a generated list of 100 candidates to 89 unique, then 29, then 20, then the final 3 that enter the wet lab.

What would settle it

Synthesize, under the identical protocol, several NMC811-derived control compositions matched for lower molecular weight or reduced nickel content but chosen at random rather than by the LLM ranking; if any control matches or exceeds 160-174 mAh/g by the third cycle, the ranking step is not doing the work. Alternatively, compute formation energies relative to the most stable competing phases for the three candidates; if any is positive at relevant temperatures, the thermodynamic-stability claim fails.

Watch

Extended reading notes

Core claim

ChatBattery organizes material discovery into eight stages and seven communicating agents. In the exploration phase, an LLM agent proposes modified formulas, a search agent checks databases for prior reports, a decision agent accepts only candidates whose formula-based theoretical capacity exceeds the input's, and a retrieval agent feeds back similar valid compounds when proposals fail. In the exploitation phase, a rank agent filters 100 generated candidates down to three using approximate total charge, preparation complexity, and an LLM-based qualitative voltage ranking; the survivors are then checked by supercell-based energy calculations and finally synthesized in the wet lab. The paper claims that the three surviving formulas retain the layered NMC structure and deliver 174, 169, and 160 mAh/g reversible capacity, respectively, compared with roughly 135 mAh/g for NMC811, under identical coin-cell testing. The authors state that this is the first full design-to-synthesis-to-characterization battery materials cycle driven by LLM reasoning.

Load-bearing premise

The pipeline assumes that its cheap formula-level proxies, namely theoretical capacity from lithium count and molecular weight, oxidation-state charge balance, element-count complexity, and an LLM's qualitative voltage comparisons, are informative enough to pick the best three candidates before any experiment, and it runs no control tests against simpler or random substitutions.

Editorial extensions

If this is right

If the ranking proxies are predictive, the same eight-stage loop can be pointed at other cathode families or other materials classes, compressing candidate screening that traditionally takes years into months.
The three validated formulas become concrete starting points for further expert refinement, as the paper demonstrates by deriving a Li-rich variant from NMC-SiMg that reaches 181 mAh/g and a 34% improvement over NMC811.
Wet-lab results that agree with the LLM voltage ranking suggest that qualitative surrogate comparisons can replace expensive molecular dynamics in early-stage screening.
The completed cycle provides a template for human-AI collaboration in which the AI generates hypotheses and the human agent intervenes at problem framing and final verification.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Beyond the paper: because all three winners lower nickel content and add lighter elements, the capacity gain may partly follow from the theoretical-capacity proxy's dependence on molecular weight; a control with arbitrary light-element substitutions would separate that effect from the LLM's chemical reasoning.
Beyond the paper: the computational stage rests on negative total energies of sampled supercells rather than formation energies against competing phases, so thermodynamic stability is not yet established; a phase-competition calculation would settle it.
Beyond the paper: the pipeline's value would be sharpened by pre-registering all 100 generated candidates and releasing wet-lab results for more than the top three, since that would show whether the ranking actually orders performance.
Beyond the paper: the same surrogate-ranking recipe could apply to any field where a fast formula-level proxy exists, suggesting a general pattern for expert-guided LLM discovery in other multivariate sciences.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Short version: three new NMC variants were actually synthesized and tested, with third-cycle capacities of 174, 169, and 160 mAh/g versus 135 for NMC811. That is a real, reproducible-in-principle empirical result, and it is the first full loop I know of from LLM generation to coin cells. The paper deserves a referee. But the evidence does not support the claim that ChatBattery's expert-guided reasoning is what did the selecting.

What is genuinely new: the specific compositions NMC-SiMg, NMC-SiCa, and NMC-MgB, and their measured capacities. The synthesis and characterization look plausible: XRD shows the layered structure, SEM images are there, and coin-cell data are shown for three cycles. The authors also admit that negative MACE-MP energies are not enough for stability. That is honest.

The load-bearing weakness is in Stage 6 and the surrounding pipeline. The theoretical capacity filter (Eq. 5) is just inverse molecular weight, because every candidate has one Li and two O. All three winners have lower formula weight than NMC811, so the "expert-guided" selection is collinear with a trivial proxy. The charge and complexity heuristics cut candidates but do not predict capacity. The LLM voltage ranking picks by voltage, not capacity. And there are no wet-lab controls: no simple Mg-only or Si-only substitution, no random selection from the 89, and no replicate cells. The baseline itself, 135 mAh/g for NMC811, is lower than typical literature values, so the percentage gains are not pinned down. Also, the retrieval weights, tau, and ranking cutoffs were tuned on this same task, which is selection bias even if the wet-lab results are independent. Finally, the code and data links are not actually provided in the text.

If the measured capacities reproduce, this is a useful demonstration that LLM-guided pipelines can propose synthesizable cathodes. But the claim that ChatBattery's reasoning, rather than low molecular weight or chance, is responsible for the gains is unsupported as stated. The authors need replicate statistics, control substitutions, and a formation-energy comparison (or at least MACE-MP energies for NMC811) before that claim is credible.

For peer review: send it out. A serious referee can push for controls and error bars. I would not cite it yet, but I would follow the thread.

Referee Report

4 major / 6 minor

Summary. The paper introduces ChatBattery, an agentic LLM framework for battery materials discovery, and reports its application to the NMC811 cathode. The pipeline consists of an exploration phase (four stages with LLM generation, database screening, and retrieval feedback) and an exploitation phase (deduplication, ranking by total charge, preparation complexity, and LLM-based voltage comparison, followed by computational and wet-lab validation). From 100 generated candidates, 89 unique formulas remain, and three are selected and synthesized: NMC-SiMg, NMC-SiCa, and NMC-MgB. The authors report third-cycle reversible capacities of 174, 169, and 160 mAh/g, compared with about 135 mAh/g for NMC811, corresponding to claimed improvements of 28.8%, 25.2%, and 18.5%. They also claim this is the first successful full design-to-synthesis-to-characterization cycle for LLM-driven battery materials discovery.

Significance. If the attribution to the pipeline is valid, this is a significant end-to-end demonstration: real synthesis, XRD/SEM characterization, and electrochemical testing of LLM-proposed cathode compositions, with code and data links and detailed prompts. The wet-lab results are independent evidence that the three materials can be synthesized and exhibit reversible electrochemistry, which is a genuine strength. However, the central claim that ChatBattery's expert-guided ranking, rather than lower formula weight or chance, drives the measured improvements is not currently supported: the ranking proxies are unvalidated, the computational stability check is not a thermodynamic test, and the electrochemical data lack error bars, replicate counts, and control compositions. The contribution is therefore promising as a case study, but the validation falls short of establishing ChatBattery as a predictive discovery platform.

major comments (4)

[Section 3.3, Table 1] The computational validation does not establish thermodynamic stability or discriminate among the 89 deduplicated candidates. Table 1 reports average 'total energies' predicted by the MACE-MP surrogate, not formation energies relative to competing phases; negative total energies are expected for any bound structure and cannot indicate stability. No NMC811 comparator or random-substitution baseline is included, and the table is labeled 'DFT total energy' even though the text states the values come from MACE-MP. The manuscript itself concedes at the end of §3.3 that 'negative total energy alone does not guarantee synthesizability or long-term phase stability,' so Stage 7 cannot rescue the ranking or support the claim that the selected materials are computationally validated.
[Section 3.2, Stage 6, and Eq. (5)] The Stage 6 ranking criteria are not validated as predictors of wet-lab capacity. The theoretical-capacity metric in Eq. (5), with a fixed Li count and the same two oxygen atoms per formula, is essentially inverse molecular weight; all three selected materials have lower formula weight than NMC811 and would pass this criterion automatically. The total-charge heuristic in Eq. (8) and Table 3 only filters for near charge neutrality and does not rank by capacity or stability. The third criterion is a qualitative LLM voltage comparison, but the reported outcome is capacity, not voltage, and no evidence is given that voltage ranking selects for high capacity. No control experiments are reported, such as simple Mg-, Al-, or Si-only substitutions, random selections from the 89 candidates, or the NMC811 baseline synthesized by the same protocol in the same batches. Without such controls, the measured capacity gains cannot be attributed specifically to ChatBattery's reasoning.
[Section 3.4, Table 2] The central quantitative claims rest on a single third-cycle capacity value per material, with no error bars, no replicate coin cells, and no indication of measurement variability. The NMC811 baseline of approximately 135 mAh/g is also reported without statistics. Given typical cell-to-cell variability in coin-cell testing, the differences among 160, 169, and 174 mAh/g versus 135 mAh/g may be meaningful, but the current presentation is insufficient to support the precise percentage improvements claimed in the abstract and Table 2. Replicate measurements and cell counts are needed for these numbers to be quantitative.
[Section 5.6, Eq. (6)] Several key hyperparameters—the retrieval weights (3, 7, 5, 10, 5, 1, 10), the deduplication threshold tau=0.1, and the ranking cutoffs 29/20/3—are reported as having been selected after testing on this same NMC811 task. This makes the evidence for pipeline efficacy partially circular: the pipeline's choices were tuned on the very task whose outcome they are used to explain. The wet-lab results are independent, so this is not fully self-referential, but an ablation showing robustness to these choices, or validation on an independent task, is needed to support the general claim that ChatBattery is an effective expert-guided reasoning platform rather than a fixed set of choices fitted to one example.

minor comments (6)

[Eq. (7)] In the deduplication range-match formula, the denominator reads 'max(CE,xi, CE,xi)' in the text; it should be 'max(CE,xi, CE,xj)'.
[Table 1] The table heading 'DFT total energy' is misleading because the values are predictions from the MACE-MP machine-learning surrogate, not from DFT; rename the heading to 'MACE-MP-predicted total energy'.
[Figure 3 caption] The caption calls the materials 'NMC-Si/Mg/Ca/B Composites,' but the paper describes substituted/doped layered oxides, not composites; use 'substituted NMC derivatives' or similar.
[Section 3.2, Stage 6] The text refers to 'total voltage' as a ranking metric, but the actual procedure is a qualitative LLM pairwise comparison of voltages; please clarify the wording to avoid suggesting a quantitative total-voltage calculation.
[Supplementary Table S3] Several total-charge values are floating-point artifacts (e.g., 4.44e-16 and 6.66e-16); these should be rounded to zero or presented with a tolerance notation.
[Section 5.4, Eq. (2)] The retrieval formula 'xretrieved = min_{xretrieved} ...' is notationally awkward; define the set over which the minimum is taken more explicitly, for example as the candidate database entries satisfying the decision condition.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the wet-lab capacities are external measurements and are not derived from any fitted parameter or self-citation.

full rationale

The pipeline's theoretical-capacity surrogate (Eq. 5) is used as a validity filter: the Decision Agent admits only candidates with Capacity_xoutput > Capacity_xinput, and because the generated formulas all keep one Li per formula unit, this is equivalent to requiring a lower formula weight than NMC811. That makes the sign of the theoretical-capacity gain an artifact of the filter, not of LLM reasoning. However, this is not circularity in the paper's central claim. The headline outcome - reversible capacities of 174, 169, and 160 mAh/g versus about 135 mAh/g for NMC811 - is measured in wet-lab coin cells (Section 3.4) and is external to the pipeline. The paper does not predict these numbers from Eq. 5, and the ranking stages (total charge, preparation complexity, LLM voltage comparisons) do not use the measured capacity. In fact, the theoretical-capacity ordering (SiMg 294.66 > MgB 293.08 > SiCa 287.29) differs from the measured ordering (SiMg 174 > SiCa 169 > MgB 160), showing the measured outcome is not forced by the surrogate. The self-citations (e.g., ref. 13 for retrieval feedback, refs 35-38 for GFlowNets) are methodological or forward-looking and are not load-bearing for the three-material result. There is no imported uniqueness theorem, no ansatz smuggled by citation, and no fitted parameter renamed as a prediction. The hyperparameter choices in Section 5.6 are described as tuned for stable retrieval, not for the wet-lab capacities. Labelling MACE-MP energies as 'DFT total energy' in Table 1 is a correctness/terminology problem, not circularity. The paper's stated limitation that most predicted materials remain close to known classes is a scope limitation, not a circular step. Overall, the core derivation chain is not circular: the final evidence is independent experimental characterization.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

No new physical entities are introduced. The central result is an empirical demonstration, but it rests on multiple hand-set thresholds, surrogate models, and database completeness assumptions. The wet-lab capacity data is independent evidence, while the pipeline's internal validity is entangled with free parameters.

free parameters (5)

retrieval distance weights (alpha, beta, gamma, delta, epsilon, zeta, eta) = (3, 7, 5, 10, 5, 1, 10)
Set after testing various combinations to produce 'stable and reasonable retrieval results' (Section 5.6); no held-out objective or external benchmark.
range-match threshold tau = 0.1
Used in Search Agent and deduplication (Eqs. 4 and 7); chosen by hand with no sensitivity analysis.
ranking cutoffs = 29 (charge), 20 (complexity), 3 (voltage)
Arbitrary cutoffs in the ranking tree (Section 5.5, Section 3.2); they determine which candidates reach the wet lab.
valence values for non-NMC elements = highest common oxidation states (Table 3)
Heuristic assignment used in total-charge ranking; not derived from formal charge analysis or DFT.
LLM generation hyperparameters = GPT-3.5 temperature 1, frequency penalty 0.2; GPT-o4 for voltage ranking
Chosen for generation and ranking; no ablation or reproducibility controls.

assumptions (5)

domain assumption The ICSD and Materials Project databases are sufficiently complete that absence from them implies the proposed formulas are novel.
Used in Stage 3 Search Agent to reject existing compounds; the novelty claim depends on database coverage.
domain assumption Theoretical capacity Eq. (5), assuming all Li ions participate and denominator 3.6*M, is a valid screening criterion.
Used by Decision and Retrieval agents; it is a standard formula but ignores voltage, kinetics, and structural factors.
domain assumption MACE-MP machine-learned potential energies are accurate enough for these doped disordered supercells to judge thermodynamic feasibility.
Table 1 reports MACE-MP average energies; no comparison to DFT for these compositions, and negative total energy is not a stability criterion without a hull reference.
domain assumption The substituted atoms occupy the transition-metal (0,0,0) site with full occupancy in R-3m, preserving the layered structure.
Initial CIF construction in Section 3.3 assumes this substitution pattern; XRD supports similarity but not site occupancy.
domain assumption A single third-cycle coin-cell capacity is a representative measure of practical capacity improvement.
Headline improvements in Table 2 rest on one reported cycle number per material with no replicate statistics.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Expert-Guided LLM Reasoning for Battery Discovery: From AI-Driven Hypothesis to Synthesis and Characterization." pith.science (2026). https://pith.science/paper/IF6BOYMN

@misc{pith2026250716110,
  author       = {Pith},
  title        = {Pith review of: Expert-Guided LLM Reasoning for Battery Discovery: From AI-Driven Hypothesis to Synthesis and Characterization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IF6BOYMN}},
  note         = {Machine review of arXiv:2507.16110}
}

read the original abstract

Large language models (LLMs) leverage chain-of-thought (CoT) techniques to tackle complex problems, representing a transformative breakthrough in artificial intelligence (AI). However, their reasoning capabilities have primarily been demonstrated in solving math and coding problems, leaving their potential for domain-specific applications-such as battery discovery-largely unexplored. Inspired by the idea that reasoning mirrors a form of guided search, we introduce ChatBattery, a novel agentic framework that integrates domain knowledge to steer LLMs toward more effective reasoning in materials design. Using ChatBattery, we successfully identify, synthesize, and characterize three novel lithium-ion battery cathode materials, which achieve practical capacity improvements of 28.8%, 25.2%, and 18.5%, respectively, over the widely used cathode material, LiNi0.8Mn0.1Co0.1O2 (NMC811). Beyond this discovery, ChatBattery paves a new path by showing a successful LLM-driven and reasoning-based platform for battery materials invention. This complete AI-driven cycle-from design to synthesis to characterization-demonstrates the transformative potential of AI-driven reasoning in revolutionizing materials discovery.

Figures

Figures reproduced from arXiv: 2507.16110 by the authors.

**Figure 1.** Overview of the ChatBattery pipeline. It includes two main phases, i.e., Exploration (left) and Exploitation (right), eight sequential stages (rounded rectangles), and seven specialized agents (legend at bottom). higher energy density. Battery materials discovery is a complex and multidisciplinary endeavor that requires the integration of computational modeling, materials synthesis, and experimental characterizatio… view at source ↗

**Figure 2.** Instantiation of ChatBattery. (a) Illustrates two cycles of Phase 1 in ChatBattery, from stage 1 to stage 4; Cycle 2 builds upon Cycle 1’s output and demonstrates the use of the Retrieval Agent. (b) Depicts stages 5 and 6 of Phase 2, where top candidate materials are progressively identified, evaluated, and selected (from left to right). thesize, and characterize three novel lithium-ion battery cathode materials 1 :… view at source ↗

**Figure 3.** Visualization of Structural and Electrochemical Characterization for NMC-Si/Mg/Ca/B Composites. (a) The side view and (b) the top view of crystal structure for NMC-SiMg, NMC-SiCa, and NMC-MgB; (c) XRD pattern for NMC-SiMg; (d) SEM image of NMC-SiMg; (e) SEM image of NMC-SiCa; (f) SEM image of NMC-MgB; The first 3 cycle charge discharge profiles for (g) NMC-SiMg, (h) NMC-SiCa, and (i) NMC-MgB. NMC-SiCa shows a smooth… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

52 extracted references · 41 canonical work pages

[1]

Highly accurate protein structure prediction with AlphaFold

Jumper, J., Evans, R., Pritzel, A., et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589. https://www.nature.com/articles/s41586-021-03819-2 (2021)

work page 2021
[2]

Probabilistic weather forecasting with machine learning

Lam, R., Ravuri, S., Brock, A., et al. Probabilistic weather forecasting with machine learning. Nature 620, 480–486. https://www.nature.com/articles/s41586-024-08252-9 (2024)

work page 2024
[3]

T., Lee, S

Chen, C., Nguyen, D. T., Lee, S. J., Baker, N. A., Karakoti, A. S., Lauw, L., Owen, C., Mueller, K. T., Bilodeau, B. A., Murugesan, V . & Troyer, M. Accelerating Computational Materials Discovery with Machine Learning and Cloud High-Performance Computing: from Large-Scale Screening to Experimental Validation. Journal of the American Chemical Society 146, ...

doi:10.1021/jacs.4c03849
[4]

& Viswanathan, V

Dave, A., Mitchell, J., Burke, S., Lin, H., Whitacre, J. & Viswanathan, V . Autonomous optimization of non-aqueous Li-ion battery electrolytes via robotic experimentation and machine learning coupling. Nature Communication 13. ISSN : 2041-1723. http://dx.doi.org/10.1038/s41467-022-32938-1 (Sept. 2022)

doi:10.1038/s41467-022-32938-1
[5]

Symmetry-Informed Geometric Representation for Molecules, Proteins, and Crystalline Materials

Liu, S., Du, W., Li, Y., Li, Z., Zheng, Z., Duan, C., Ma, Z., Yaghi, O., Anandkumar, A., Borgs, C., et al. Symmetry- Informed Geometric Representation for Molecules, Proteins, and Crystalline Materials.arXiv preprint arXiv:2306.09375 (2023)

work page Pith review arXiv 2023
[6]

S., Aykol, M., Cheon, G

Merchant, A., Batzner, S., Schoenholz, S. S., Aykol, M., Cheon, G. & Cubuk, E. D. Scaling deep learning for mate- rials discovery. Nature 624, 80–85 (2023)

work page 2023
[7]

A., MacKnight, R., Kline, B

Boiko, D. A., MacKnight, R., Kline, B. & Gomes, G. Autonomous chemical research with large language models. Nature 624, 570–578 (2023)

2023
[8]

Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D., Dhariwal, P ., Neelakantan, A., Shyam, P ., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D., Wu, J., Win- ter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford,...

work page 2020

Show all 52 references

[9]

Chain of Thought Prompting Elicits Reasoning in Large Language Models.arXiv preprint arXiv:2201.11903 (2022)

Wei, J., Bosma, M., Huang, T., Wang, X., Wainwright, M., Luan, Y., Liang, P ., Hovy, E., Xu, W.-t., Narang, S., et al. Chain of Thought Prompting Elicits Reasoning in Large Language Models.arXiv preprint arXiv:2201.11903 (2022)

2022 arXiv
[10]

OpenAI o3-mini System Card 2025

OpenAI. OpenAI o3-mini System Card 2025. https://cdn.openai.com/o3-mini-system-card-feb10.pdf

2025
[11]

DeepSeek-AI & etc., D. G. DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning 2025. arXiv: 2501.12948 [cs.CL]. https://arxiv.org/abs/2501.12948

2025 arXiv
[12]

Gemini 2.0 Flash Thinking https : / / deepmind

Google DeepMind. Gemini 2.0 Flash Thinking https : / / deepmind . google / technologies / gemini / flash - thinking/. Accessed: 2025-02-19. 2025

2025
[13]

& Xiao, C

Liu, S., Wang, J., Yang, Y., Wang, C., Liu, L., Guo, H. & Xiao, C. Conversational Drug Editing Using Retrieval and Domain Feedback. https://openreview.net/forum?id=yRrPfKyJQ2 (2024)

2024
[14]

Towards an AI co-scientist

Gottweis, J., Weng, W.-H., Daryin, A., Tu, T., Palepu, A., Sirkovic, P ., Myaskovsky, A., Weissenberger, F., Rong, K., Tanno, R., et al. Towards an AI co-scientist. arXiv preprint arXiv:2502.18864 (2025)

2025 arXiv
[15]

& Anandkumar, A

Liu, S., Nie, W., Wang, C., Lu, J., Qiao, Z., Liu, L., Tang, J., Xiao, C. & Anandkumar, A. Multi-modal molecule structure–text model for text-based retrieval and editing. Nature Machine Intelligence 5, 1447–1457 (2023)

2023
[16]

& Anandkumar, A

Liu, S., Li, Y., Li, Z., Gitter, A., Zhu, Y., Lu, J., Xu, Z., Nie, W., Ramanathan, A., Xiao*, C., Tang*, J., Guo*, H. & Anandkumar, A. A Text-guided Protein Design Framework.Nature Machine Intelligence, 1–12 (2025)

2025
[17]

J., Shen, S., Lee, R., Wallace, E., Wei, J., Wainwright, M

Hu, E. J., Shen, S., Lee, R., Wallace, E., Wei, J., Wainwright, M. & Narang, S. LoRA: Low-Rank Adaptation of Large Language Models. arXiv preprint arXiv:2106.09685 (2021)

2021 arXiv
[18]

Materials and methods are available as supplementary material
[19]

I., Kasim, M

Saaid, F. I., Kasim, M. F., Winie, T., Elong, K. A., Azahidi, A., Basri, N. D., Yaakob, M. K., Mastuli, M. S., Shaffee, S. N. A., Zolkiffly, M. Z., et al. Ni-rich lithium nickel manganese cobalt oxide cathode materials: A review on the synthesis methods and their electrochemic...

2024
[20]

& Bloom, I

Son, S.-B., Robertson, D., Tsai, Y., Trask, S., Dunlop, A. & Bloom, I. Systematic study of the cathode compositional dependency of cross-talk behavior in Li-ion battery. Journal of The Electrochemical Society 167, 160508 (2020)

2020
[21]

& Rehme, S

Zagorac, D., Müller, H., Ruehl, S., Zagorac, J. & Rehme, S. Recent developments in the Inorganic Crystal Structure Database: theoretical crystal structure data and related features.Journal of applied crystallography52, 918–925 (2019). 18 REFERENCES ChatBattery

2019
[22]

P ., Hautier, G., Chen, W., Richards, W

Jain, A., Ong, S. P ., Hautier, G., Chen, W., Richards, W. D., Dacek, S., Cholia, S., Gunter, D., Skinner, D., Ceder, G., et al. Commentary: The Materials Project: A materials genome approach to accelerating materials innovation.APL materials 1 (2013)

2013
[23]

R., Zanin, H., Monteiro, R

Oliveira Filho, H. R., Zanin, H., Monteiro, R. S., Barbosa, M. H. & Teófilo, R. F. High-nickel cathodes for lithium- ion batteries: From synthesis to electricity. Journal of Energy Storage 82, 110536 (2024)

2024
[24]

& Cadars, S

Okhotnikov, K., Charpentier, T. & Cadars, S. Supercell program: A combinatorial structure-generation approach for the local-level modeling of atomic substitutions and partial occupancies in crystals. Journal of Cheminformatics 8 (Mar. 2016)

2016
[25]

& DeCost, B

Choudhary, K. & DeCost, B. Atomistic Line Graph Neural Network for improved materials property predictions. npj Computational Materials 7, 1–8 (2021)

2021
[26]

Machine-learning interatomic potentials for materials science

Mishin, Y. Machine-learning interatomic potentials for materials science. Acta Materialia 214, 116980. ISSN : 1359-
[27]

Zhang, L., Han, J., Wang, H., Car, R. & E, W. Deep Potential Molecular Dynamics: A Scalable Model with the Accuracy of Quantum Mechanics. Physical Review Letters 120. ISSN : 1079-7114. http://dx.doi.org/10.1103/ PhysRevLett.120.143001 (Apr. 2018)
[28]

V ., Thompson, A

Zuo, Y., Chen, C., Li, X., Deng, Z., Chen, Y., Behler, J., Csányi, G., Shapeev, A. V ., Thompson, A. P ., Wood, M. A. & Ong, S. P . Performance and Cost Assessment of Machine Learning Interatomic Potentials. The Journal of Physical Chemistry A 124, 731–745. ISSN : 1520-5215. h...

doi
[29]

& Ong, S

Chen, C. & Ong, S. P . A universal graph deep learning interatomic potential for the periodic table. Nature Compu- tational Science 2, 718–728. ISSN : 2662-8457. http://dx.doi.org/10.1038/s43588-022-00349-3 (Nov. 2022)

2022 doi
[30]

Deng, B., Zhong, P ., Jun, K., Riebesell, J., Han, K., Bartel, C. J. & Ceder, G. CHGNet as a pretrained universal neural network potential for charge-informed atomistic modelling. Nature Machine Intelligence, 1–11 (2023)

2023
[31]

M., Kovács, D

Batatia, I., Benner, P ., Chiang, Y., Elena, A. M., Kovács, D. P ., Riebesell, J., Advincula, X. R., Asta, M., Avaylon, M., Baldwin, W. J., et al. A foundation model for atomistic materials chemistry. arXiv preprint arXiv:2401.00096 (2023)

2023 arXiv
[32]

P ., Jha, G., Youn, D

Pender, J. P ., Jha, G., Youn, D. H., Ziegler, J. M., Andoni, I., Choi, E. J., Heller, A., Dunn, B. S., Weiss, P . S., Penner, R. M., et al. Electrode degradation in lithium-ion batteries. ACS nano 14, 1243–1295 (2020)

2020
[33]

C., Wiseman, P

Mizushima, K., Jones, P . C., Wiseman, P . J. & Goodenough, J. B. LixCoO2 (0< x<-1): A new cathode material for batteries of high energy density. Materials Research Bulletin 15, 783–789 (1980)

1980
[34]

& Goodenough, J

Yuan, L.-X., Wang, Z.-H., Zhang, W.-X., Hu, X.-L., Chen, J.-T., Huang, Y.-H. & Goodenough, J. B. Development and challenges of LiFePO 4 cathode material for lithium-ion batteries. Energy & Environmental Science 4, 269–284 (2011)

2011
[35]

& Bengio, Y

Bengio, E., Jain, M., Korablyov, M., Precup, D. & Bengio, Y. Flow network based generative models for non- iterative diverse candidate generation. Advances in Neural Information Processing Systems 34, 27381–27394 (2021)

2021
[36]

F., Ekbote, C

Jain, M., Bengio, E., Hernandez-Garcia, A., Rector-Brooks, J., Dossou, B. F., Ekbote, C. A., Fu, J., Zhang, T., Kilgour, M., Zhang, D., et al. Biological sequence design with gflownets in International Conference on Machine Learning (2022), 9786–9801

2022
[37]

& Bengio, Y

Jain, M., Deleu, T., Hartford, J., Liu, C.-H., Hernandez-Garcia, A. & Bengio, Y. Gflownets for ai-driven scientific discovery. Digital Discovery 2, 557–577 (2023)

2023
[38]

J., Jain, M., Elmoznino, E., Kaddar, Y., Lajoie, G., Bengio, Y

Hu, E. J., Jain, M., Elmoznino, E., Kaddar, Y., Lajoie, G., Bengio, Y. & Malkin, N. Amortizing intractable inference in large language models. ICLR (2024)

2024
[39]

Characterization of Li-rich-NMC-SiMg is available as supplementary material
[40]

Synthesis of NMC-SiMg, NMC-SiCa, NMC-MgB, and Li-rich-NMC-SiMg is available as supplementary material
[41]

& Hou, J

Li, L. & Hou, J. Capacity detection of electric vehicle lithium-ion batteries based on X-ray computed tomography. RSC advances 8, 25325–25333 (2018)

2018
[42]

A., Chen, J., Ding, Z., Marie, J.-J., Juelsholt, M., House, R

Song, W., Pérez-Osorio, M. A., Chen, J., Ding, Z., Marie, J.-J., Juelsholt, M., House, R. A., Bruce, P . G. & Nellist, P . D. Visualization of Tetrahedral Li in the Alkali Layers of Li-Rich Layered Metal Oxides.Journal of the American Chemical Society 146, 23814–23824 (2024)

2024
[43]

S., Widiyandari, H., Purwanto, A., Arif, A

Wijareni, A. S., Widiyandari, H., Purwanto, A., Arif, A. F. & Mubarok, M. Z. Morphology and particle size of a synthesized NMC 811 cathode precursor with mixed hydroxide precipitate and nickel sulfate as nickel sources and comparison of their electrochemical performances in an...

2022
[44]

& Wen, C

Ahsan, Z., Cai, Z., Wang, S., Moin, M., Wang, H., Liu, D., Ma, Y., Song, G. & Wen, C. Recent development of phosphate based polyanion cathode materials for sodium-ion batteries. Advanced Energy Materials 14, 2400373 (2024)

2024
[45]

& Jiaqiang, E

Kang, S., Wang, C., Chen, J., Meng, T. & Jiaqiang, E. Progress on solvo/hydrothermal synthesis and optimization of the cathode materials of lithium-ion battery. Journal of Energy Storage 67, 107515 (2023)

2023
[46]

& Salavati-Niasari, M

Ghiyasiyan-Arani, M. & Salavati-Niasari, M. Effect of Li2CoMn3O8 nanostructures synthesized by a combustion method on montmorillonite K10 as a potential hydrogen storage material. The Journal of Physical Chemistry C 122, 16498–16509 (2018)

2018
[47]

& Sainio, T

Meshram, P ., Virolainen, S. & Sainio, T. Solvent extraction for separation of 99.9% pure cobalt and recovery of Li, Ni, Fe, Cu, Al from Spent LIBs. Metals 12, 1056 (2022)

2022
[48]

Ultra-thin Fe3C nanosheets promote the adsorption and conversion of polysulfides in lithium-sulfur batteries.Energy storage materials18, 338– 348 (2019)

Li, H., Ma, S., Cai, H., Zhou, H., Huang, Z., Hou, Z., Wu, J., Yang, W., Yi, H., Fu, C.,et al. Ultra-thin Fe3C nanosheets promote the adsorption and conversion of polysulfides in lithium-sulfur batteries.Energy storage materials18, 338– 348 (2019)

2019
[49]

& Kuang, Y

Li, H., Ma, S., Li, J., Liu, F., Zhou, H., Huang, Z., Jiao, S. & Kuang, Y. Altering the reaction mechanism to eliminate the shuttle effect in lithium-sulfur batteries. Energy Storage Materials 26, 203–212 (2020)

2020
[50]

Ampere-hour-scale soft-package potassium-ion hybrid capacitors enabling 6-minute fast-charging

Li, H., Gong, Y., Zhou, H., Li, J., Yang, K., Mao, B., Zhang, J., Shi, Y., Deng, J., Mao, M., et al. Ampere-hour-scale soft-package potassium-ion hybrid capacitors enabling 6-minute fast-charging. Nature communications 14, 6407 (2023). 20 ChatBattery A Experimental Details for...

2023
[1901]

https://proceedings.neurips.cc/paper_files/paper/2020/file/1457c0d6bfcb4967418bfb8ac142f64a- Paper.pdf

2020
[6454]

http://dx.doi.org/10.1016/j.actamat.2021.116980 (Aug. 2021)

2021

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Highly accurate protein structure prediction with AlphaFold

Jumper, J., Evans, R., Pritzel, A., et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589. https://www.nature.com/articles/s41586-021-03819-2 (2021)

work page 2021

[2] [2]

Probabilistic weather forecasting with machine learning

Lam, R., Ravuri, S., Brock, A., et al. Probabilistic weather forecasting with machine learning. Nature 620, 480–486. https://www.nature.com/articles/s41586-024-08252-9 (2024)

work page 2024

[3] [3]

T., Lee, S

Chen, C., Nguyen, D. T., Lee, S. J., Baker, N. A., Karakoti, A. S., Lauw, L., Owen, C., Mueller, K. T., Bilodeau, B. A., Murugesan, V . & Troyer, M. Accelerating Computational Materials Discovery with Machine Learning and Cloud High-Performance Computing: from Large-Scale Screening to Experimental Validation. Journal of the American Chemical Society 146, ...

doi:10.1021/jacs.4c03849

[4] [4]

& Viswanathan, V

Dave, A., Mitchell, J., Burke, S., Lin, H., Whitacre, J. & Viswanathan, V . Autonomous optimization of non-aqueous Li-ion battery electrolytes via robotic experimentation and machine learning coupling. Nature Communication 13. ISSN : 2041-1723. http://dx.doi.org/10.1038/s41467-022-32938-1 (Sept. 2022)

doi:10.1038/s41467-022-32938-1

[5] [5]

Symmetry-Informed Geometric Representation for Molecules, Proteins, and Crystalline Materials

Liu, S., Du, W., Li, Y., Li, Z., Zheng, Z., Duan, C., Ma, Z., Yaghi, O., Anandkumar, A., Borgs, C., et al. Symmetry- Informed Geometric Representation for Molecules, Proteins, and Crystalline Materials.arXiv preprint arXiv:2306.09375 (2023)

work page Pith review arXiv 2023

[6] [6]

S., Aykol, M., Cheon, G

Merchant, A., Batzner, S., Schoenholz, S. S., Aykol, M., Cheon, G. & Cubuk, E. D. Scaling deep learning for mate- rials discovery. Nature 624, 80–85 (2023)

work page 2023

[7] [7]

A., MacKnight, R., Kline, B

Boiko, D. A., MacKnight, R., Kline, B. & Gomes, G. Autonomous chemical research with large language models. Nature 624, 570–578 (2023)

2023

[8] [8]

Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D., Dhariwal, P ., Neelakantan, A., Shyam, P ., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D., Wu, J., Win- ter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford,...

work page 2020