PaperlayerView exact revision on arXiv

Source-linked AI summary

Prototype-guided transfer of sparse literature knowledge for electrolyte additive discovery

Weixiang Hong, Hongting Du, Jiayue Tang, Ruifeng Tan, Yangjian Quan, Jia Li, Jiaqiang Huang

arXiv:2609.02209v1physics.chem-phcs.LG

TL;DR

Electrolyte-additive discovery is constrained by sparsely reported successful additives and source-target distribution mismatch in boron-additive chemical space. ProtoMI transfers sparse literature-derived knowledge through molecular prototypes to guide candidate exploration in data-scarce chemical space. Representative candidate TNDB improved high-temperature cycling stability of LiFePO4||graphite cells relative to the baseline electrolyte.

  • Problem

    Electrolyte-additive discovery is constrained by sparsely reported successful additives and source-target distribution mismatch in boron-additive chemical space.

    Source: Dataset construction reveals… · pp. 5–6

  • Method

    ProtoMI transfers sparse literature-derived knowledge through molecular prototypes to guide candidate exploration in data-scarce chemical space.

    Source: Discussion and conclusion · pp. 22–23

  • Results

    Representative candidate TNDB improved high-temperature cycling stability of LiFePO4||graphite cells relative to the baseline electrolyte.

    Source: Discussion and conclusion · p. 23

  • Takeaways & Limitations

    ProtoMI identifies electrolyte-additive candidates whose performance can be evaluated in battery-relevant high-temperature cycling.

    Source: Discussion and conclusion · p. 23

    The current implementation focuses on boron-containing additives and does not explicitly incorporate electrochemical stability or reaction representations.

    Source: Discussion and conclusion · p. 24

Abstract

from arXiv · show

Electrolyte additive discovery remains challenging because experimentally validated molecules are sparse, whereas accessible chemical spaces are vast and largely unlabeled. This challenge is amplified in lithium-ion batteries, where additive performance arises from coupled interfacial reactions rather than a single molecular property. Here, we develop a prototype-guided molecular intelligence, ProtoMI, a literature-driven framework that learns transferable structural priors from reported electrolyte additives and uses them to prioritize candidates in unlabeled chemical space. For boron-containing additives, ProtoMI combines 126 literature-reported molecules with 179,977 unlabeled candidates. Graph contrastive learning identifies seven chemically interpretable prototypes from the reported additives, and prototype guided semi-supervised contrastive learning adapts these prototypes to the candidate space under source-target distribution mismatch. In retrospective temporal validation, ProtoMI achieves enrichment factors of 9.2-45.6 while screening less than 2% of the candidate space. A subsequent translation step identifies four commercially accessible candidates. One representative candidate, 4,4,5,5-Tetramethyl-2-[10-(1naphthyl)anthracen-9-yl]-1,3,2-dioxaborolane (TNDB), improves high-temperature LiFePO4||graphite cycling at 55 °C by 34.93% relative to the baseline electrolyte. An arsenal of characterizations and operando optical fiber Fourier transform infrared spectroscopy suggest that TNDB forms B-containing, F/P/O-modified inorganic interphases, suppresses solvent decomposition and reduces Fe deposition on graphite. This case study shows how sparse literature knowledge can guide experimentally efficient molecular discovery in data-scarce battery-additive spaces.

Results

The Results section presents ProtoMI, a prototype-guided framework that transfers sparse literature knowledge into unlabeled boron-containing chemical space to prioritize candidates efficiently. It covers prototype extraction and adaptation, validation and representation assessment, translation to commercially accessible molecules, and experimental evaluation of representative TNDB in high-temperature lithium-ion battery cells.

Dataset construction reveals source-target mismatch in boron-additive chemical space

  • The extraction pipeline achieved 94% accuracy for additive identification, required approximately 11 s per article, and cost less than 60 CNY in total processing.These results indicate an efficient route to constructing a curated positive molecular set for downstream prototype.Source: Dataset construction reveals… · p. 6
  • Positive additives occupied restricted, low-to-moderate molecular-weight regions with moderate synthetic accessibility, whereas unlabeled candidates spanned broader and more heterogeneous chemical space.UMAP, clustering, molecular-weight–SA maps, graph descriptors, and node features consistently showed population-level and local-environment differences.Source: Dataset construction reveals… · p. 6
  • 179,977 PubChem candidates were contrasted with 126 literature-reported additives, establishing a sparse positive dataset against a largely unlabeled boron-containing chemical space.The positive set was extracted from Web of Science articles using an LLM-assisted heuristic pipeline.Source: Dataset construction reveals… · p. 5, Figure 2 · p. 7

Reported additives encode transferable molecular prototypes

  • GINE+ achieved silhouette scores of 0.853 with K-Means and 0.881 with hierarchical clustering, outperforming fingerprint-, SMILES-, and other graph-based representations.Augmentation-enhanced graph representations captured hidden structural organization in the reported additive set.Source: Reported additives encode transferable… · p. 8
  • Hierarchical clustering identified seven prototypes organized by recurring boron-related motifs, including oxalate, fluorinated aromatic, catechol-derived, fluorinated alkoxy, and tri-alkoxy frameworks.Functional annotations linked prototypes to interphase formation, solvation regulation, Lewis acidity, anion reception, and stabilization-oriented chemistry.Source: Reported additives encode transferable… · pp. 8–9
  • Seven chemically distinct prototypes emerged from unsupervised graph contrastive learning and captured recurring structural motifs and literature-supported functional tendencies without performance labels.The prototypes organized reported boron-containing additives into chemically meaningful molecular families and functional biases.Source: Reported additives encode transferable… · p. 9, SMILES · p. 11

Prototype-guided adaptation transfers additive knowledge to unlabeled chemical space

  • During training, adapted prototype centroids followed initial exploration, progressive refinement, and eventual stabilization; decreasing prototype drift indicated adaptation was not one-to-one copying and supported chemically meaningful transfer.Source: Prototype-guided adaptation transfers… · pp. 11–12
  • Removing top-k assignment, prototype decorrelation or EMA updating degraded recommendation quality by reducing confidence, diversity or cross-domain transfer stability.The ablations respectively impaired high-confidence selection, inter-prototype separation and training stability.Source: Prototype-guided adaptation transfers… · p. 12
  • 8.8–20.0% recall with 9.2–45.6 enrichment factors was achieved while screening only 0.19–1.27% of candidates in temporal validation.ProtoMI also reduced the effective search space by 79–517-fold while preserving substantial recall.Source: Prototype-guided adaptation transfers… · pp. 12–13, Figure 4 · p. 13

Knowledge-guided translation converts recommendations into testable candidates

  • Four low-cost, commercially available molecules were selected after filtering 30,799 recommendations to 68 supplier-accessible candidates.Filters incorporated identifier validity, molecular weight, synthetic accessibility, active-hydrogen exclusion, CAS availability and cost.Source: Knowledge-guided translation converts… · pp. 14–15
  • Four candidates remained after sequential filtering of 30,800 ProtoMI recommendations using validity, molecular-weight, synthetic-accessibility, functional-group, availability and cost constraints.The final candidates spanned multiple transferred prototype regions; NP4 and NP7 contributed no commercially accessible molecules because of practical accessibility constraints.Source: Knowledge-guided translation converts… · p. 15, Figure 5 · p. 16

Experimental validation in high-temperature LiFePO4||graphite cells

  • 34.93% improvement in capacity retention at the 100th cycle was achieved by 1 wt% TNDB versus the baseline electrolyte at 55 °C.TNDB was tested in LiFePO4||graphite cells using 1M LiPF6 in EC:DMC = 3:7 v/v.Source: Experimental validation in… · p. 17
  • TNDB formed a B/F- and P/O-containing inorganic graphite interphase, with increased LiₓPOFy relative to baseline and suppressed oxygen-rich, carbonate-derived decomposition products.XPS and ToF-SIMS showed TNDB-derived boron incorporation, reorganized LiF2⁻ and PO2⁻ distributions, and less organic carbonate decomposition through the interphase.Source: Experimental validation in… · pp. 17–18
  • TNDB-derived interphase chemistry progressively incorporated boron, attenuated solvent-derived organic fragments and reduced FeO⁻ signals, consistent with mitigated Fe contamination.These changes may suppress Fe-catalyzed electrolyte decomposition and interphase deterioration during high-temperature cycling.Source: Experimental validation in… · pp. 18–19
  • 89% and 71% lower Fe contents after formation and 100 cycles, respectively, were measured with TNDB relative to baseline graphite electrodes.Baseline Fe contents increased from 2.06 wt% after formation to 3.61 wt% after 100 cycles, whereas TNDB values were 0.23 wt% and 1.04 wt%.Source: Experimental validation in… · p. 19
  • At the end of charge, operando FTIR measured an integrated response of 0.25 a.u. cm⁻¹ for the baseline electrolyte versus 0.17 a.u. cm⁻¹ with TNDB, consistent with suppressed formation of carbonate-derived organic SEI species.Source: Experimental validation in… · p. 21

Discussion and conclusion

  • ProtoMI transfers knowledge at the prototype level, adapting higher-order organization from reported additives across source-target distribution mismatch rather than memorizing individual molecules.This enables exploration of structurally broader candidate regions and concentrates limited experiments on chemically promising areas.Source: Discussion and conclusion · p. 23

Experimental validation in high-temperature LiFePO4||graphite cells

  • TNDB improved high-temperature cycling stability of LiFePO4||graphite cells relative to the baseline electrolyte, without a supported quantitative improvement percentage.TNDB incorporated B-containing species into the graphite interphase, modified F/P/O-containing inorganic SEI components, suppressed carbonate-solvent decomposition, and reduced Fe deposition on graphite.Source: Discussion and conclusion · p. 23

Discussion and conclusion

  • Retrospective temporal validation showed that literature-derived prototype knowledge can enrich future-reported additives while screening only a small fraction of the candidate space.The experimental validation should be interpreted as a proof of concept for one representative ProtoMI-recommended candidate rather than exhaustive validation of all recommendations.Source: Discussion and conclusion · p. 23

Methods

  • The study used 179,977 boron-containing molecules as the search space and 126 reported boron electrolyte additives as positive samples.The current implementation focuses on boron-containing additives and primarily structural molecular representations, without explicitly incorporating electrochemical stability, reaction pathways, solvation structures, synthesis routes, or cell-level performance labels.Source: Methods · p. 24, Discussion and conclusion · p. 24
  • Unsupervised graph contrastive learning encodes reported additive molecules and clusters their embeddings into molecular prototypes using cosine distance and silhouette-based selection.Two graph views are contrasted, and hierarchical clustering with outlier filtering defines the prototype groups.Source: Methods · pp. 25–26
  • ProtoMI transfers prototype knowledge from experimentally reported boron additives to unlabeled molecules using prototype semi-supervised contrastive learning.The model adapts prototypes toward the broader unlabeled space while preserving knowledge extracted from reported additives.Source: Methods · pp. 26–27
  • Prototype-guided training selects top-k molecules most similar to each centroid, reducing low-confidence unlabeled influence and encouraging distinct prototype representations through decorrelation.Prototype centroids are updated with an exponential moving average during adaptation.Source: Methods · pp. 27–28
  • ProtoMI evaluates embedding separation with cosine-distance silhouette scores, prototype similarity and coverage with EWPS, and future-additive recovery through temporal validation and recall.Higher EWPS indicates recommendations that are both representative of learned prototypes and broadly distributed across them.Source: Methods · pp. 28–29
  • Enrichment factor quantified how many times recommendation outperformed random selection, with higher values indicating more efficient prioritization.The metric uses the number of recommended molecules and the candidate-pool size to assess screening efficiency.Source: Methods · p. 29
  • Four molecules were manually selected from 68 commercially available candidates after six sequential chemical and practical filters.Filters included PubChem CID, molecular weight, synthesizability, active-hydrogen exclusion, CAS availability, and cost considerations.Source: Methods · pp. 29–30
  • The study used fingerprints, SMILES, and molecular graphs, with augmented molecular graphs identified as the best-performing representation for model training.Graph features encoded atom and bond properties, including elemental identity, aromaticity, ring membership, hybridization, bond type, and connectivity attributes.Source: Methods · p. 30
  • Multi-scale graph augmentation perturbed molecular features, local structure, and global topology to improve representation robustness under limited labeled data.Augmentations included feature shuffling, Gaussian noise masking, node dropping, weighted edge perturbation, and Personalized PageRank-based connectivity changes.Source: Methods · pp. 30–31

M9. Hierarchical clustering algorithm with outlier filtering. To identify

The methods establish prototype discovery through hierarchical clustering and outlier filtering, then define contrastive objectives for molecular and prototype-guided representation learning. Electrochemical testing and complementary surface, chemical, and morphological characterizations assess additive effects in high-temperature LiFePO4||graphite cells.

M9. Hierarchical clustering algorithm with outlier filtering

  • Seven clusters were selected by cosine-distance silhouette score, with cluster-wise outlier filtering followed by recomputed hierarchical clustering and retention of prototype-construction data.Clustering used average linkage on learned molecular graph embeddings; outliers exceeded cluster-specific mean-distance thresholds by 1.5 standard deviations, while clusters smaller than two samples were exempt.Source: M9. Hierarchical clustering algorithm… · p. 31

M10. Loss functions

  • InfoNCE contrastive learning maximized similarity between two augmented views of each molecular graph relative to other batch samples treated as negatives.Embeddings were L2-normalized and cosine similarity was used, with temperature τ controlling distribution concentration.Source: M9. Hierarchical clustering algorithm… · p. 32
  • Prototype semi-supervised contrastive learning treated each molecule’s assigned prototype as positive and other prototypes as competing logits, while decorrelation promoted nonoverlapping patterns.Positive samples were selected by top-k similarity within each batch, and the total objective combined prototype contrastive and decorrelation losses.Source: M9. Hierarchical clustering algorithm… · p. 32

M11. Electrochemical measurements

  • Electrochemical testing used an electrolyte containing 1 wt% additive; cells were cycled at 55 °C between 2.5 and 3.65 V, with three formation cycles at 0.33C charge/discharge, followed by long-term cycling at 0.33C charge and 1C discharge.Source: M9. Hierarchical clustering algorithm… · p. 33

M12. Post-mortem sample preparation

  • Post-mortem graphite electrodes were disassembled and rinsed inside an Ar-filled glovebox before air-free transfer to characterization chambers.Electrodes were rinsed three times with anhydrous DMC to remove residual electrolyte and loosely attached salts.Source: M9. Hierarchical clustering algorithm… · p. 33

M13. X-ray photoelectron spectroscopy

  • XPS analyzed SEI surface chemical composition and bonding environments across the C 1s, O 1s, F 1s, P 2p, B 1s, and Fe 2p regions when applicable, using C-C calibration and background subtraction to determine relative component contents.Source: M9. Hierarchical clustering algorithm… · p. 33

M14. Time-of-flight secondary ion mass spectrometry

  • ToF-SIMS mapped SEI-related fragments and depth profiles to compare solvent-derived, carbonate, LiF, phosphate/fluorophosphate, and Fe-containing species.Negative-ion analysis included PO2⁻, CH3CO⁻, LiF2⁻, CO3⁻, CH3COO⁻, and FeO⁻ fragments.Source: M9. Hierarchical clustering algorithm… · pp. 33–34

M15–M16. SEM and EDS mapping

  • SEM examined cycled graphite morphology, while EDS mapping compared spatial Fe accumulation after formation and long-term cycling in baseline and additive-containing electrolytes.EDS evaluated transition-metal deposition on representative graphite-electrode regions.Source: M9. Hierarchical clustering algorithm… · p. 34

M17. Operando IR spectroscopy collection.

Operando IR spectra were collected in a customized electrochemical cell with an embedded fiber during cycling. Measurements used a liquid-nitrogen-cooled MCT detector across 5000–600 cm^-1 at 1-minute intervals and 4 cm^-1 resolution.

M17. Operando IR spectroscopy collection.

  • Operando IR measurements used a Bruker Invenio R spectrometer and customized electrochemical cell containing an embedded IR fiber.The manually polished fiber was inserted through symmetric through-holes and sealed with epoxy resin for 12 h.Source: M17. Operando IR spectroscopy collection · p. 34
  • The cell stacked graphite, the IR fiber, a 120 μL-electrolyte-wetted GF/D separator, and a 12 mm LFP electrode within a PTFE stabilizing ring.Transmitted IR signals were collected after a 1 h rest.Source: M17. Operando IR spectroscopy collection · p. 34
  • Spectra were acquired every 1 min during cycling from 5000–600 cm^-1 at 4 cm^-1 resolution with 32 scans and approximately 10,000 signal intensity.A liquid-nitrogen-cooled MCT detector was used to maintain a high signal-to-noise ratio.Source: M17. Operando IR spectroscopy collection · p. 34

Competing interests

The authors report a pending Chinese patent application related to the work and provide public source code, while datasets are available from the corresponding author upon reasonable request.

Competing interests

Data and code availability

  • The ProtoMI source code is publicly available, and the generated and analyzed datasets can be requested from the corresponding author.Source: Data and code availability · p. 35

Supplementary information

The supplementary information details the structural analyses, prototype-guided adaptation, candidate translation, and molecular-feature and literature-mining pipelines underlying the study. It also documents the four final candidates and supporting validation and interphase analyses for TNDB.

Supplementary information

  • Reported boron-containing additives cluster into recurring structural families, including lithium borates, fluorinated borates, and aryl borate derivatives.The dendrogram indicates that the reported additives are enriched in recurrent scaffold-level motifs rather than randomly scattered isolated examples.Source: Supplementary information · p. 40

N71 CHI_TETRAHEDRAL_CC

  • Seven molecular prototypes were selected by silhouette-score analysis, and learned embeddings form seven well-separated additive groups with structurally and functionally distinguishable regions.Hierarchical clustering reached its highest silhouette score at k = 7, while K-means showed a consistent optimum in the same region.Source: Figure S2 · p. 43, Figure S3 · p. 43, N71 CHI_TETRAHEDRAL_CC · pp. 43–44
  • Prototype-guided adaptation preserves chemically meaningful structural information, with distinct transferred prototypes enriched for different boron-containing motifs rather than arbitrary candidate-space clusters.Analyzed motifs include tetra-alkoxy borate, tri-alkoxy borate, boron oxalate, BF3-like motifs, fluorinated alkoxy groups, C-B(OR)2 motifs, pentafluorophenyl groups, and catechol boronate.Source: Figure S6 · p. 57, N71 CHI_TETRAHEDRAL_CC · p. 58
  • ProtoMI distributes recommended molecules across multiple prototype regions while retaining high prototype similarity, avoiding collapse into a single prototype.Its recommendations show the highest overall prototype similarity among compared strategies, while similarity to reported additives spans both close and lower-similarity candidates.Source: Figure S7 · p. 58, Figure S8 · p. 59, N71 CHI_TETRAHEDRAL_CC · p. 59
  • 30,800 PCL-recommended molecules were reduced to 68 translation candidates and four experimentally actionable candidates, all assigned to NP3.Filtering used molecular validity, molecular weight, synthetic accessibility, active hydrogen exclusion, CAS availability, cost, commercial availability, and experimental feasibility.Source: Figure S10 · p. 61, N71 CHI_TETRAHEDRAL_CC · p. 61
  • Four final molecules were identified, with one candidate selected for validation at a listed price of 557.9￥/g.The remaining candidates were marked as under preparation at 240.9￥/g, 122.9￥/g, and 68.9￥/g.Source: Table S9 · p. 66, N71 CHI_TETRAHEDRAL_CC · p. 66
  • Supplementary XPS and ToF-SIMS analyses characterize graphite interphases after formation and 100 cycles at 55 °C with and without TNDB.The analyses assign carbon, boron, LiF-related, phosphate, carbonate, oxygenated-organic, and Fe-containing fragments within the interphase.Source: Figure S11 · p. 67, Figure S12 · p. 68, N71 CHI_TETRAHEDRAL_CC · p. 68
  • Supplementary SEM images compare graphite-electrode surface morphology after formation and 100 cycles at 55 °C in baseline and TNDB-containing electrolytes.The images cover electrodes recovered from LiFePO4||graphite cells.Source: Figure S16 · p. 70, N71 CHI_TETRAHEDRAL_CC · p. 70
  • The supplementary data describe a large-language-model-assisted pipeline for content identification, guided extraction, and literature-based auto-correction of electrolyte and interphase information.The dataset collection deadline is before July 2025, with prompts provided for extraction and correction.Source: Figure S17 · p. 71, Figure S18 · p. 72, Figure S19 · p. 73
  • A supplementary table maps molecular graph edge features to their corresponding physical meanings.The mapping is provided in Table S10.Source: Table S10 · p. 74
Loading 2609.02209v1…