PaperlayerView exact revision on arXiv

Source-linked AI summary

AI-driven ionic liquid discovery with unified chemical intelligence

Yuqi Yin, Yibo Fu, Siyuan Wang, Lizhen Huo, Peng Sun, Hongyu Wang, Xiaohui Wang, Chengli Pan, Zhiyong Li, Zhirong Liu, Jianji Wang, Zhaoxi Sun

arXiv:2511.11257v2cs.AIcs.CEcs.LG

TL;DR

Modeling ionic-liquid properties is challenging because multicomponent behavior emerges from collective interactions, while ionic species are represented inconsistently. AIonopedia is presented as a unified modeling framework for these systems; literature-grounded evaluations report predictive capability on structurally distinct ILs, with a stated caveat about sampling bias.

Abstract

from arXiv · show

The rational discovery of functional solvents remains challenging because macroscopic solvent behaviors emerge from complex many-body interactions across vast chemical spaces. Ionic Liquids (ILs) constitute one of the most chemically diverse solvent families, yet their exceptional structural tunability remains only partially explored owing to the limitations of existing experimental and computational approaches. Here we introduce AIonopedia, a large-language-model-orchestrated agentic framework for IL research. Built around a domain-specific multimodal foundation model, it integrates information retrieval, chemical reasoning and autonomous tool execution within a unified architecture. Pretrained on large-scale unlabeled corpora and fine-tuned using over 80,000 curated measurements, AIonopedia learns expressive joint representations of multicomponent ionic systems, enabling accurate prediction across diverse physicochemical properties. Through a series of literature-grounded evaluations, AIonopedia exhibits robust out-of-distribution generalization, while many-body-expansion-based analysis provides physically grounded explanations of structure-property relationships. By coupling molecular screening and optimization with wet-lab validation in a closed-loop workflow, AIonopedia identifies high-performance ILs and reveals transferable design principles for the capture of various volatile organic compounds. This work establishes an end-to-end AI-driven paradigm for understanding and exploring complex ionic systems, demonstrating the promise of chemically specialized foundation models as a generalizable strategy for scientific discovery.

Supporting Information

The supporting information describes experimental datasets spanning solute–solvent thermodynamics and bulk properties, alongside larger synthetic corpora for multimodal pretraining. It also reports dataset curation and analyses of chemical-space coverage.

S1.1. Experimental datasets.

  • Records were extracted from a literature-mining workflow using automated retrieval, LLM-assisted extraction and OCR when needed, followed by manual verification.The workflow began with approximately 10,000 retrieved papers.Source: S1.1. Experimental datasets · p. 28
  • About 30,000 solvation and transfer free-energy measurements cover roughly 80 ionic solvents and 150 solutes; bulk-property data span about 6,000 ILs formed from approximately 1,200 cations and 300 anions.The cation–anion–solute dataset is described as the largest curated dataset currently available for these properties.Source: S1.1. Experimental datasets · pp. 28–29
  • Similarity analyses indicate broad chemical coverage, while descriptor-based t-SNE projections separate ionic species from neutral molecules more clearly than traditional similarity metrics.Structurally related [BMIM]+ and [HMIM]+ cations remain close in descriptor space.Source: S1.1. Experimental datasets · p. 29

S1.2. Construction of the multimodal pretraining corpora.

  • Because experimental data cover limited chemical space, the authors generated about 2.8 million synthetic systems across four formats for modality-alignment pretraining.The formats include temperature-conditioned and temperature-independent bulk IL systems, IL–solute systems, and organic-solvent–solute systems.Source: S1.2. Construction of the multimodal… · pp. 29–30
  • Synthetic systems received descriptor and system-format pseudo-labels, while the IL-centric corpus strategy was intended to avoid representation biases from databases dominated by non-IL molecules.The corpora were built by expanding known IL species through similarity-based retrieval and combinatorial sampling.Source: S1.2. Construction of the multimodal… · pp. 29–30

S2. AIonopedia employs a multimodal foundation model for ionic solvents.

AIonopedia represents complete physicochemical systems rather than isolated molecules and processes them through language and graph modalities. This design aims to capture component interactions, composition, and experimental context in multicomponent IL systems.

S2.1. Overall architecture.

  • Unlike conventional molecular foundation models that encode molecules independently, AIonopedia treats the entire physicochemical system as its representation unit.A system can include an ionic solvent alone for bulk-property prediction or solvents together with a neutral solute for thermodynamic tasks.Source: S2.1. Overall architecture · p. 33, S2.2. System representation · p. 33
  • The system-level formulation is designed to capture intermolecular interactions, compositional dependencies, and thermodynamic context that determine measured properties.The framework combines language and graph modalities through alignment, bidirectional cross-modal fusion, and multi-task supervised learning.Source: S2.1. Overall architecture · p. 33

S2.2. System representation.

  • For each modality, molecular components and experimental conditions are organized together into a unified representation of the complete system.The language input serializes component SMILES and conditions; the graph input combines molecular graphs with system-level tokens.Source: S2.2. System representation · pp. 33–34

S2.3. Multimodal encoding.

The model encodes each system in language and graph branches, then aligns their global representations with system-level contrastive learning. Momentum encoders and queued candidates support this alignment before later fusion.

S2.3. Multimodal encoding.

  • A pretrained decoder-only LLM processes serialized system prompts, while a shared Graph Transformer encodes molecular topology and a graph-fusion encoder combines components and conditions.The two branches are intended to capture complementary chemical semantics and structural information.Source: S2.3. Multimodal encoding · p. 34

S2.4. Cross-modal alignment.

  • The language and graph representations are projected into a common space, then aligned using momentum graph-text contrastive learning with positive pairs defined over complete physicochemical systems.The method uses queued momentum features as additional candidates and optimizes text-to-graph and graph-to-text losses.Source: S2.3. Multimodal encoding · p. 35, S2.4. Cross-modal alignment · pp. 35–36

S2.5. Bidirectional cross-modal fusion.

After alignment, bidirectional cross-attention lets language and graph representations exchange information. Multi-task pretraining adds matching, graph-conditioned text generation, and pseudo-label prediction objectives to support downstream property learning.

S2.5. Bidirectional cross-modal fusion.

  • Bidirectional Transformer cross-attention lets each modality attend to the other, producing a shared representation for downstream learning.The language-query branch uses a causal attention mask to preserve the language model's autoregressive property.Source: S2.5. Bidirectional cross-modal fusion · pp. 36–37

S2.6. Multi-task learning objectives.

  • Pretraining combines graph-text matching, graph-grounded text generation, and supervised pseudo-label prediction with the cross-modal contrastive objective.The pseudo-label objective predicts molecular descriptors and system-format classes from the fused representation.Source: S2.6. Multi-task learning objectives · pp. 37–38
  • Graph-text matching is trained as binary classification using matched pairs and two mismatched negatives per positive, with negatives sampled by cross-modal similarity.The generation objective trains autoregressive next-token prediction conditioned on graph-derived representations.Source: S2.6. Multi-task learning objectives · pp. 37–38
  • The combined pretraining objective weights graph-text matching by 2, contrastive alignment and graph-grounded generation by 0.2 each, and adds the supervised objective.After modality alignment, auxiliary prediction heads are replaced by task-specific regression heads for downstream property prediction.Source: S2.6. Multi-task learning objectives · p. 39

S3. Training procedures.

AIonopedia was implemented with multiple language-model backbones and a shared molecular-graph representation, then pretrained through self-supervised modality alignment. The procedures also specify parameter-efficient backbone adaptation and the training setup.

S3.1. Backbone implementations.

  • Galactica, Qwen3, and Gemma3 backbones were paired with a shared molecular encoder and system-level graph sequence for downstream tasks.The lightweight Galactica-125m was also included as a baseline, and missing molecular components were represented by empty graph placeholders for a unified input format.Source: S3.1. Backbone implementations · p. 40
  • Language backbones were adapted with LoRA rather than full-parameter optimization, using rank 16, lora_alpha 32, and dropout 0.1.LoRA layers were inserted into attention projections.Source: S3.1. Backbone implementations · p. 40

S3.2. Self-supervised modality alignment.

  • The multimodal model underwent five epochs of self-supervised modality alignment on eight A100 80GB GPUs, with an effective global batch size of 512.The Qwen3-0.6b model required approximately 14 hours for pretraining.Source: S3.2. Self-supervised modality alignment · pp. 40–41

S3.3. Property-specific adaptation.

Property-specific adaptation uses chemically meaningful cross-validation and task-specific regression fine-tuning, with dedicated protocols for out-of-distribution evaluation and hydration prediction. The adjacent framework sections describe an iterative tool-using planner and retrieval procedures that ground scientific and chemical inputs.

S3.3. Property-specific adaptation.

  • Five-fold chemically meaningful cross-validation replaces random splitting, including cation-based and ion-pair splits for bulk-property evaluation.Ternary thermodynamic tasks additionally use a split that keeps each solute-cation-anion combination exclusively in training or validation, preventing memorization of combinations.Source: S3.3. Property-specific adaptation · p. 41
  • Hydration free energy is estimated without an additional property-specific model by combining gas-to-IL solvation and IL-to-water transfer predictions through virtual IL intermediates.The protocol averages estimates from ten deduplicated virtual ILs per solute; it slightly underestimates performance variance while reducing small-sample bias.Source: S3.3. Property-specific adaptation · pp. 41–42

S4.1. Agent architecture.

  • The GPT-5 planner follows a Thought-Action-Observation cycle, decomposing requests into subtasks, selecting tools, and incorporating their outputs into subsequent reasoning.The cycle supports multi-step workflows combining symbolic reasoning and numerical computation.Source: S4.1. Agent architecture · p. 44

S4.2. Knowledge retrieval and chemical grounding.

  • The web retrieval pipeline uses scholarly search and metadata services, with general web search as a fallback when scholarly coverage is insufficient.Responses are grounded in externally verifiable sources.Source: S4.2. Knowledge retrieval and chemical… · p. 44
  • Chemical grounding maps names and synonyms to SMILES, then uses reasoning-based correction to reconstruct valid ionic forms and assign charge states when needed.The authors state that this improves robustness of IL structure identification.Source: S4.2. Knowledge retrieval and chemical… · p. 45

S4.3. Property prediction interface.

The prediction interface maps system composition, molecular identities, temperature, and target property into a unified multimodal representation for diverse bulk and thermodynamic predictions. The discovery workflow uses predictions to guide constrained candidate screening, refinement, and experimental follow-up.

S4.3. Property prediction interface.

  • The unified predictor supports bulk properties including density, viscosity, surface tension, and melting point, alongside solvation, transfer, and hydration free energies.The interface extracts system-level inputs including composition, solute and solvent identities, temperature, and target property.Source: S4.3. Property prediction interface · p. 45
  • Predicted properties serve as intermediate signals in the agentic loop for property-guided ion screening and refinement, not only as final outputs.The planner transforms the extracted system inputs into the model's unified multimodal representation.Source: S4.3. Property prediction interface · pp. 45–46

S4.4. IL discovery with agentic AI.

  • Discovery is framed as constrained exploration using similarity-guided screening and targeted structure modification around experimentally validated hits.Candidates are evaluated iteratively, with experimental feedback used to select promising hits for further local chemical modification.Source: S4.4. IL discovery with agentic AI · p. 46

S5.1. Hydration free energy calculation.

The hydration benchmark combines extensive nonequilibrium simulations with a comparison of RESP charge-generation choices. The traditional HF/6-31G* charges performed strongly across several reported metrics, while implicit-solvent DFT charges were close to that baseline.

S5.1. Hydration free energy calculation.

  • The benchmark simulations used one net-neutral solute with TIP3P water, equilibrated the system, and sampled 100 equilibrium snapshots before fast-switching simulations.Each fast-switching run lasted 1.2 ns, with van der Waals coupling followed by charging.Source: S5.1. Hydration free energy calculation · p. 48
  • Nonequilibrium work values were processed with Jarzynski's scheme, equivalent here to exponential averaging, to estimate solvation free-energy differences.Uncertainty was estimated by numerical bootstrap resampling.Source: S5.1. Hydration free energy calculation · p. 48
  • 28 RESP charge sets across 84 solutes at approximately 220 ns per solute-charge-set combination amount to approximately 0.5 ms of all-atom simulation.The authors describe this as the largest reported benchmark of solvation free energy calculations to their knowledge.Source: S5.1. Hydration free energy calculation · p. 48
  • The authors retained HF/6-31G*-derived RESP charges for subsequent modeling involving experimental hydration free energies.They note that implicit-solvent models are important when elevating the theory level to DFT, and attribute the traditional option's performance to fortuitous error cancellation.Source: S5.1. Hydration free energy calculation · p. 49
  • For hydration-free-energy benchmarks, HF/6-31G* RESP charges rank highest by MAE, Kendall τ, and PI; no gas-phase DFT selection or HF with def2-TZVPP/def2-QZVPP outperforms gas-phase HF/6-31G*.By RMSE, several DFT selections perform similarly to traditional HF/6-31G*.Source: S5.1. Hydration free energy calculation · pp. 49–50

S5.2. Molecular simulations of ionic solvents.

The ionic-solvent simulations adapt fixed-charge force fields and simulation protocols to account for ionic charge transfer and the convergence challenges of ionic liquids. They evaluate bulk densities and solvation free energies across multiple compositions and temperatures.

S5.2. Molecular simulations of ionic solvents.

  • RESP charges are combined with GAFF2, and ionic charge scaling approximates charge-transfer effects that may be significant in these species.The scaling factor is 0.8 for IL-forming ions; net-neutral solutes remain unscaled.Source: S5.2. Molecular simulations of ionic… · p. 51
  • Only molecules forming the ionic solvent receive uniform charge scaling; net-neutral solutes remain untouched.A scaling factor of 0.8 is a common selection.Source: S5.2. Molecular simulations of ionic… · p. 51
  • IL boxes contain at least 120 ion pairs, a benchmark-supported minimum for avoiding finite-size artifacts in mass densities and solvation thermodynamics.The box is built with about 13,000 atoms using PACKMOL, and simulations include temperatures such as 298, 328, and 348 K.Source: S5.2. Molecular simulations of ionic… · p. 51
  • Bulk simulations use 10 ns NPT equilibration followed by 100 ns NPT production, with density averaged from trajectory samples every 20 ps.PubChem coordinates are used for solutes, while structure-prediction outputs are used for molecules absent from existing databases.Source: S5.2. Molecular simulations of ionic… · p. 51
  • Solute-in-IL free-energy switching lasts 2 ns, with longer van der Waals and electrostatic coupling times than in hydration simulations to address poor IL convergence.The coupling times are 1.5 ns for van der Waals and 0.5 ns for electrostatics; the external datasets cover multiple IL compositions and temperatures.Source: S5.2. Molecular simulations of ionic… · pp. 51–52

S6.1. Chemical grounding benchmark.

The chemical-grounding benchmark tests whether models correctly interpret abbreviations for ions that form ionic liquids. GPT-5 was selected as AIonopedia’s planner, while domain-specific modules further improved performance under the strictest criterion.

S6.1. Chemical grounding benchmark.

  • The benchmark assessed 50 commonly used IL-forming-ion abbreviations using exact canonical-SMILES, structural-equivalence, and full-name criteria.Human experts manually assessed outputs under the three increasingly lenient criteria.Source: S6.1. Chemical grounding benchmark · p. 55
  • Domain-specific modules further improved performance under the strictest chemical-grounding criterion, supporting their role in reliable downstream reasoning.GPT-5 was selected as the planner after achieving the highest overall accuracy among the evaluated web-augmented LLMs.Source: S6.1. Chemical grounding benchmark · p. 55
  • The evaluated models included GPT-5, Gemini 2.5 Pro, Claude Sonnet 4.5, and DeepSeek-R1-0528, with the selections reflecting model availability at evaluation time.DeepSeek-R1-0528 was chosen as a fixed, reasoning-focused baseline; Sonnet 4.5 was selected as Anthropic’s then-positioned strongest model.Source: S6.1. Chemical grounding benchmark · p. 55

S6.2. Cross-validation benchmarks for property prediction.

This section presents cross-validation benchmarks for property prediction, comparing AIonopedia and baselines across datasets and split schemes and listing viscosity, surface tension, and mass density among the properties considered.

S6.2. Cross-validation benchmarks for property prediction.

  • Among similarly sized AIonopedia variants, Gemma3-1b was weakest overall, consistent with a possible benefit from the science-oriented pretraining reported for Galactica and Qwen3.The comparisons instantiated AIonopedia with Galactica, Qwen3, and Gemma3 backbones.Source: S6.2. Cross-validation benchmarks for… · p. 57
  • AIonopedia performed best on nearly all datasets and split schemes, while Qwen3-0.6b ranked first in 18 comparisons and achieved strict win rates above 85% against every baseline.Qwen3-0.6b was selected as the default model; Tables S9–S11 report the numerical results underlying the rankings and win rates.Source: Table S9 · p. 56, S6.2. Cross-validation benchmarks for… · pp. 57–58
  • ILBERT was the strongest overall baseline and performed particularly well on melting-point prediction, but its relative performance declined substantially on viscosity prediction.SPMM showed a similar, more pronounced pattern, with viscosity RMSE nearly twice that of most other baselines.Source: S6.2. Cross-validation benchmarks for… · p. 57
  • The authors suggest that viscosity’s sensitivity to molecular size and chain length may be harder for encoder-style models using 1-D CNN aggregation or a [CLS] token to capture, while noting architectural and pretraining differences as alternatives.The proposed explanation is presented as a possibility, not a definitive cause.Source: S6.2. Cross-validation benchmarks for… · p. 57
  • For solute–solvent interaction datasets, sparse experimental coverage of individual ions can produce uneven cation allocation across folds and relatively large variance, so overlapping error bars alone do not establish equivalent performance.The error bars in Fig. 2C represent standard deviations across five cross-validation folds.Source: S6.2. Cross-validation benchmarks for… · pp. 57–58
  • The listed property labels are viscosity (log scale), surface tension, and mass density.Source: S6.2. Cross-validation benchmarks for… · p. 62
  • Table S12 is titled “AARD of viscosity.”Source: Table S12 · p. 62

S6.3. System-specific OOD and additional benchmarks.

AIonopedia retained accurate predictions on system-specific out-of-distribution ionic-liquid tests, including cases where ions were absent from training. It also performed best or comparably on additional density and hydration-free-energy benchmarks.

S6.3. System-specific OOD and additional benchmarks.

  • AIonopedia achieved the lowest errors in most system-specific solvation-free-energy comparisons, with RMSE = 0.060~0.464 kcal mol-1.Molecular-dynamics simulations had RMSE = 0.524~0.897 kcal mol-1, while learning-based methods varied more with system composition.Source: S6.3. System-specific OOD and… · p. 63, Table S13 · p. 64
  • The excluded [QUIN8]⁺[TF2N]- ion pair was an easier OOD case because both constituent ions were well represented individually, enabling near-perfect predictions.Each complete ion pair was excluded from training in this system-specific evaluation.Source: S6.3. System-specific OOD and… · p. 63
  • AIonopedia retained accurate predictions on phosphonium systems with one or both ions absent from training, where several baselines degraded substantially.For [P66614]⁺[L-Lact]-, ILBERT’s RMSE exceeded 1 kcal mol-1, the commonly accepted chemical-accuracy threshold.Source: S6.3. System-specific OOD and… · pp. 63–64
  • Across additional mass-density and hydration-free-energy benchmarks, AIonopedia achieved the best or comparable performance among evaluated approaches.These results support applicability across diverse IL-related thermodynamic properties.Source: S6.3. System-specific OOD and… · p. 64, Table S14 · p. 66

S6.4. Ablation analysis of the multimodal property predictor.

Ablations show that the full multimodal predictor performs best, while modality alignment, both input modalities, and physicochemical supervision each contribute to predictive performance.

S6.4. Ablation analysis of the multimodal property predictor.

S7. Selection criteria and detailed discussion for literature case studies used for mechanistic interpretation of biphasic transitions.

The literature case studies were selected for experimental relevance, chemical representability, and independence from training data. Complex phases and polymer inputs were simplified for model inference, limiting the interpretation of some predictions to trends.

S7. Selection criteria and detailed discussion for literature case studies used for mechanistic interpretation of biphasic transitions.

  • Studies were selected for substantial experimental content and properties predicted by AIonopedia, or observables mechanistically related to predicted quantities.Source: S7. Selection criteria and detailed… · p. 69
  • The selection excluded systems that could not be represented reliably and redundant studies, while reported measurements were essentially absent from training data.Redundancy was judged by overlap in case-defining experimental variables.Source: S7. Selection criteria and detailed… · p. 69
  • The selected independent studies spanned multiple application scenarios to assess physically meaningful interpretation beyond predictive performance.Source: S7. Selection criteria and detailed… · p. 69

S7.1. Representation of complex multiphase systems.

  • Because inputs do not explicitly represent multicomponent solvent mixtures, phases were approximated by dominant components to preserve principal chemical contrast rather than full composition.Source: S7.1. Representation of complex… · p. 70
  • For iodine extraction, salt solutions were represented as virtual IL phases to evaluate ionic contributions, identifying [EMIM]+ as the primary extraction driver over its anion.Source: S7.1. Representation of complex… · p. 70

S7.2. Oligomeric representation of polymer phases.

S8. Selection criteria and detailed discussion for literature case studies used for mechanistic interpretation related to bulk properties.

This section presents literature-grounded bulk-property case studies for evaluating AIonopedia and interpreting ionic-liquid properties through many-body expansion and symbolic density correction. It also documents ionic-liquid preparation and testing for volatile-compound absorption, including supporting experimental procedures and measurements.

S8. Selection criteria and detailed discussion for literature case studies used for mechanistic interpretation related to bulk properties.

  • The external-validation set used quantitative bulk-property measurements from original studies and included specially designed ions with structures unlike conventional ILs.The selection excluded reviews and secondary data compilations; bulk properties served as auxiliary descriptors rather than standalone explanations of wet-lab observations.Source: S8. Selection criteria and detailed… · p. 72
  • Maximum absolute errors were 0.603 (mPa·s, Log scale) for viscosity, 0.0547 g/cm3 for density, and 43.0 K for melting point.These results were reported for 13 structurally distinct ILs drawn from four independent studies and assessed against viscosity, density, and melting-point measurements.Source: S8. Selection criteria and detailed… · p. 72
  • With no eligible primary study for surface tension, the authors used a review example showing close agreement but excluded it from external validation.At 293 K, reported versus predicted surface tensions were 46.7 versus 46.6 mN/m for [(mPEG2)2IM]+[I]- and 45.6 versus 45.6 mN/m for [(mPEG2)2IM]+[PF6]-.Source: S8. Selection criteria and detailed… · p. 72
  • RMSEs were 0.271 (mPa·s, Log scale) for viscosity, 0.0329 g/cm3 for density, and 24.1 K for melting point, comparable to dataset-level benchmarks.The melting-point error was substantially lower than in the corresponding benchmark, but the authors cautioned that limited literature validation data may have caused sampling bias.Source: S8. Selection criteria and detailed… · p. 72, Figure S3 · p. 73

S9. Finite-difference formulation of perturbative cooperativity analysis.

  • The XAI framework integrates many-body expansion with functional-group attribution to interpret IL properties and provide mechanistic insights and design guidance.It treats first-order terms as individual substituent effects and second-order terms as pairwise nonadditivity; higher-order contributions can also be defined.Source: S9. Finite-difference formulation of… · pp. 74–75

S10. Interpretable symbolic correction for IL density prediction.

  • For density, the authors explored symbolic correction as a complementary route to express information learned by AIonopedia in a global, compact, analytically usable form.Mass density was selected as a representative property because the model achieved its best average learning performance for it.Source: S10. Interpretable symbolic correction… · p. 77
  • The density correction augments the analytical formula with structure-derived descriptors and an RDKit-computed geometric volume for accessible, reproducible estimation.The descriptor set combines cation and anion values for 21 modality-alignment descriptors with molar volume as a 22nd descriptor.Source: S10. Interpretable symbolic correction… · pp. 77–78
  • Combining the LogP and constant terms achieved an RMSE of 0.0530 g/cm3, a 15.7% reduction from the baseline.The LogP term alone reduced RMSE by 8.1%, while the constant term alone improved it by 4.5%.Source: S10. Interpretable symbolic correction… · p. 78, Table S16 · p. 79
  • The density formula’s symbolic correction remained associated with experimental residuals after controlling for AIonopedia’s correction, supporting a role for descriptor supervision in encoding domain knowledge.The authors interpret this as evidence that the supervised pseudo-label objective is an important route for incorporating domain prior knowledge into the learned representation.Source: S10. Interpretable symbolic correction… · p. 79

S11.1. Materials.

  • The VOC-absorption experiments used ILs purchased commercially or synthesized when literature approaches were reported, alongside commercial acetonitrile, toluene, and ethanol.The chemicals were used without further purification.Source: S11.1. Materials · p. 81

S11.2. Synthesis of ILs.

  • ILs were prepared by either neutralization, using a hydroxide-form IL and the target-anion acid, or metathesis, mixing salts containing the target cation and anion before extraction and drying.The hydroxide-form IL was titrated, and neutralization used a 1:1 molar ratio.Source: S11.2. Synthesis of ILs · p. 81

S11.2.1 Synthesis of Intermediates.

  • Intermediate examples included phosphonium, quinuclidinium, and ammonium bromide salts prepared by heating their amine or phosphine precursors with alkyl bromides.Reported conditions included heating at 120 °C overnight or for 24 h under N2, and refluxing in MeCN for 12 h under N2.Source: S11.2.1 Synthesis of Intermediates · pp. 81–82

S11.2.2 Ionic Liquids for Ethyl acetate and Acetone Absorption.

  • For ethyl acetate and acetone absorption, the passage describes ion exchange preparations of [EMIM]+[TCB]- and [HMIM]+[TCB]- and begins a [BMpy]+[TDI]- preparation.The TCB salts were made from the corresponding bromide and potassium tetracyanoborate in dichloromethane, followed by extraction, filtration, concentration, and drying.Source: S11.2.2 Ionic Liquids for Ethyl… · p. 82

S11.2.2 Ionic Liquids for Ethyl acetate and Acetone Absorption

  • Ethyl acetate and acetone case-study ionic liquids included TDI salts made by metathesis and TF2N salts prepared through anion exchange followed by neutralization.The passages identify [BMIM]+[TDI]-, [HMIM]+[TF2N]-, and [EMIM]+[TF2N]-.Source: S11.2.2 Ionic Liquids for Ethyl… · p. 82

S11.2.3 Ionic Liquids for Toluene and o-Xylene Absorption

  • Toluene and o-xylene case-study ionic liquids were prepared by converting bromide salts to hydroxides and neutralizing them with target anions.Examples include phosphonium and ammonium salts paired with pyrrole-2-carbonitrile, bis(trifluoromethane)sulfonimide, or 4,5-dicyanoimidazole.Source: S11.2.3 Ionic Liquids for Toluene and… · p. 83

S11.3. Volatile Vapor Absorption Test.

  • Volatile-vapor absorption capacity and curves were measured gravimetrically after drying ionic liquids at 70 ℃ for 24 hours and purging samples at 100 ℃ with argon for about 8 hours.Approximately 100 mg of dried IL was tested at 50 sccm argon flow; the sample was weighed until constant mass.Source: S11.3. Volatile Vapor Absorption Test · pp. 83–84

S11.4. Results of wet-lab experiments.

2-CNPyr

This section examines selective VOC absorption through a toluene absorption measurement and atomistic simulations across 90 IL–VOC systems. Cluster-based DFT and energy decomposition analyze VOC–ionic-cage interactions, distinguishing local electronic stabilization from absorption free energies that also depend on entropy and the surrounding bulk solvent.

2-CNPyr

  • Under literature-reported experimental conditions, [P8883]+[BF4]- reached a measured absorption capacity of 4137 mg g-1, exceeding the literature-reported capacity of 2837 mg g-1, with no clear indication of saturation.Because of time constraints, the ultimate saturated absorption capacity could not be determined.Source: 2-CNPyr · p. 89

S12. Physical origin of selective VOC absorption.

  • The study assembled 90 IL–VOC systems for simulations and quantum-chemical calculations: 40 carbonyl-associated and 50 aromatic-associated systems.Four IL candidates were paired with ten carbonyl analogues, and five IL candidates with ten aromatic analogues; the aromatic candidates included [P88810]+[2-CNPyr]-.Source: S12. Physical origin of selective VOC… · p. 92, Table S25 · p. 92

S12.1. Atomistic Modelling.

  • Nonequilibrium solute growth generated five independent equilibrated configurations per system, accelerating first-shell relaxation while avoiding unfavorable initial contacts.The protocol used fixed-charge all-atom models and switched on solute–solvent van der Waals interactions and electrostatics in stages.Source: S12.1. Atomistic Modelling · p. 93

S12.2. Construction of representative IL-solute clusters

S12.3. DFT-based energy decomposition calculations

  • Five clusters per IL–VOC system were evaluated by single-point DFT at their unoptimized MD geometries, with CounterPoise correction applied at the fixed geometry.Calculations used B3LYP-D3(BJ)/6-31+G(d,p); diffuse functions were included for the ionic cage’s anionic constituents.Source: S12.3. DFT-based energy decomposition… · pp. 95–96
  • Each cluster partitioned a neutral central VOC from a complete eight-ion cage, whose net charge was −2, 0, or +2, with ion–ion interactions internal to the cage.Source: S12.3. DFT-based energy decomposition… · pp. 95–96
  • sobEDAw [143] gives a four-component decomposition of VOC–cage interactions by redistributing DFT correlation between exchange-repulsion and dispersion while leaving electrostatic and orbital terms unchanged.Its dispersion-to-electrostatics ratio closely agrees with SAPT, without additional electronic-structure calculations.Source: S12.3. DFT-based energy decomposition… · pp. 96–97
  • These fixed-geometry interaction energies describe local electronic stabilization, not absorption free energies, which also depend on entropy and the surrounding bulk solvent.Source: S12.3. DFT-based energy decomposition… · p. 97

S12.4. Physical interpretation of AIonopedia-derived solvent recommendations using complementary interaction analyses

This section examines AIonopedia-derived solvent recommendations through complementary interaction analyses and demonstrations of its workflow. It spans VOC-specific ion choices for ammonia, toluene, and CO2, alongside analyses that organize systems by VOC family.

S12.4. Physical interpretation of AIonopedia-derived solvent recommendations using complementary interaction analyses

  • Raw, z-score-standardized, and L2-normalized fingerprints retained broad carbonyl–aromatic separation, and analyses averaged across five conformations yielded consistent patterns.The separation was less sharply resolved without L2 normalization.Source: S12.4. Physical interpretation of… · pp. 97–98
  • AIonopedia’s final-token embeddings showed a cleaner block-diagonal family organization, with substantially less within-block noise than the interaction fingerprints.This suggests the model consolidates family-level organization captured by sobEDAw fingerprints.Source: S12.4. Physical interpretation of… · p. 98
  • The first two PCA components explained 93.7% and 97.1% of variance in standardized and raw fingerprints, respectively, preserving carbonyl–aromatic separation.Aromatic-associated systems formed more compact groupings than carbonyl-associated systems, though less distinctly than in the L2-normalized projection.Source: S12.4. Physical interpretation of… · p. 98
  • Across VOC analogues, total interaction-energy distributions were broadly similar, with ethyl acetate shifted less negative; dispersion favored aromatic systems, whereas electrostatics favored carbonyl systems.Exchange-repulsion and orbital-interaction energy distributions varied but remained comparatively similar across systems.Source: S12.4. Physical interpretation of… · pp. 97–98, Figure S8 · p. 99
  • sobEDAw fingerprints and IGMH maps provide complementary energetic and spatial views of solute–ion interactions, supporting physically meaningful, VOC-family-specific candidate prioritization.The family-level correspondence between fingerprints and AIonopedia’s final-token embeddings links data-driven recommendations to distinct solvation requirements.Source: S12.4. Physical interpretation of… · p. 100
  • Acetone in [EMIM]+[TCB]- showed continuous attractive regions supporting a C-H···O hydrogen-bond contact, with an approximately 2.26 Å H···O distance and 155.5° angle.The contact lies between acetone’s carbonyl oxygen and the imidazolium C2-H group.Source: S12.4. Physical interpretation of… · p. 100, Figure S9 · p. 101
  • Toluene in [N88810]+[2-CNPyr]- showed diffuse, disconnected weak van der Waals contacts around its aromatic ring and methyl group, consistent with greater dispersion.Source: S12.4. Physical interpretation of… · p. 100, Figure S9 · p. 101

S13. Example demonstrations of AIonopedia agent task execution and online deployment.

  • S13 demonstrates agentic information retrieval, prediction, screening, molecular refinement, and image–text reasoning, and provides a web interface for direct access to prediction models.Source: S13. Example demonstrations of… · p. 102

S13. Example demonstrations of AIonopedia agent task execution and online deployment

S12.4. Physical interpretation of AIonopedia-derived solvent recommendations using complementary interaction analyses

  • −1.7204 kcal/mol was the predicted CO2 solvation free energy for EMIM B(CN)4, versus −0.5964 kcal/mol for EMIM SCN.Calculations were at 298 K; more negative values indicate stronger CO2 absorption.Source: S13. Example demonstrations of… · p. 109
  • EMIM TFSI (−1.6346 kcal/mol) and EMIM FSI (−1.3857 kcal/mol) were strong alternatives, while EMIM TCM reached −1.3686 kcal/mol.The recommendation identifies TFSI or FSI as options when balancing uptake with viscosity or availability.Source: S13. Example demonstrations of… · p. 109

Related papers

Loading 2511.11257v2…