Platform · LILM Planned architecture

One model from the ion pair to the plant decision.

LILM, the Large Ionic Liquid Model, is Avestiq’s first engine. This page describes its planned architecture, the split between physics and machine learning, and how it will be tested. Modules are added only where held-out and prospective tests justify them.

Schematic of LILM’s seven planned layers stacked from representation to experiment, with results feeding back to the top layer

The seven layers Planned architecture

Seven layers from molecule to experiment.

Physics where physics is known. ML where data add value. Experiments where uncertainty matters. Select a layer to see what it is designed to do.

LILM (Large Ionic Liquid Model): “large” refers to breadth across ion families, properties, hazards, process conditions, life-cycle endpoints and tasks, not to parameter count.

Data flow Planned architecture

Every input keeps its provenance.

Public, Avestiq-owned and customer-private data are designed to travel on separate lines. Partner data stay in an isolated workspace and train shared models only where an agreement grants that right.

From data sources to a decision

Planned · schematic
Planned data flow through LILM, schematicFive data sources on the left: NIST ILThermo, published literature, quantum chemistry and COSMO-RS calculations, toxicity and life-cycle inventory databases, and Avestiq experiments. They feed the ion-pair representation, scientific models with learned residuals, process, life-cycle and cost models, and multi-objective ranking with uncertainty, which ends in a decision: make, do not make, or learn next. Make and learn-next decisions lead to experiments whose results return to Avestiq experiments. Partner process data follow a separate line into an isolated partner workspace that joins only at the decision.
    Line widths and particle counts are schematic and do not represent data volumes. Flows describe the planned architecture; they are not built.

    What the largest public IL database contains

    Public source
    NIST ILThermo data points by system typeDonut chart of 863,364 experimental data points: 163,123 for pure ionic liquids (18.9%), 480,069 for binary mixtures (55.6%) and 220,172 for ternary mixtures (25.5%).
    View data
    NIST ILThermo (SRD 147), database statistics, update of July 22, 2026. Shares computed from the published counts.

    81%

    of ILThermo data points describe mixtures, not pure ionic liquids.

    The measurements cover 2,749 ionic liquids, spread across many properties, temperatures, pressures and compositions. Sparse, heterogeneous data are the reason LILM is designed to lean on physical models and to report an applicability-domain score with each prediction.

    Purity and water content change measured properties. LILM’s representation layer is designed to carry them as inputs rather than discard them.

    Hybrid physics + ML Planned

    Physics carries the structure. ML learns the residual.

    ŷ(x,c) = fphys(x,c;θ) + gML(x,c)

    reported with σ(x,c) + applicability-domain score

    Physical models

    • Thermodynamics and phase equilibrium
    • COSMO-RS and group contribution
    • Mass and energy balances

    Machine learning

    • Residual correction gML
    • Predictive uncertainty σ
    • Applicability-domain score

    Experiments

    • Synthesis and independent measurement
    • Failed and negative results
    • Process and pilot data

    A decision with evidence

    • Make · Do not make · Learn next
    • Expected performance with an interval
    • The next most informative experiment

    Planned architecture. None of these modules is built yet.

    When a physical model already encodes conservation, thermodynamics or known structure, LILM is designed not to ask a neural network to rediscover it. Machine learning is intended to learn residual behavior, uncertain parameters and patterns the physical model does not capture.

    x: ion pair · c: conditions (temperature, pressure, composition, water, impurities) · θ: physical parameters · σ: predictive standard deviation · applicability-domain score: distance from the data the model was trained on.

    QuantityPhysical basisML role

    Decision explorer Historical founder research Computed from published data Planned criteria

    A ranked candidate answers part of the decision.

    A published CO2-capture screen (Farahipour, Mehrkesh & Karunanithi, Chem. Eng. Sci. 2016, 145, 126–132) narrowed 13,356 ionic liquids to ten on predicted thermodynamics, viscosity and melting point. Its values sit here in one candidate table with the seven benchmark ionic liquids the paper compared against (from Zhang et al., Appl. Energy 2016). Hatched cells are criteria that study did not assess and that LILM is designed to add. Select a candidate in any view; the others follow.

    Constraint survival: how many candidates each stage left

    Published screening stages, then the stages LILM is designed to add

    Historical founder researchComputedPlanned
    Counts from the paper’s text and Tables 2–4. The 310 is 10 cations × 31 anions as stated; the paper’s text describes 20 anions (200 ionic liquids) while its Table 2 tabulates 21 (210), and the melting-point count uses those 210 cells. Planned stages carry no counts because no candidate has been through them.

    Candidate × criterion feasibility map

    Color scales each value between the column’s worst and best (log scale for viscosity); brighter = better. Select a column heading to sort.

    Historical founder researchPlanned
      View data
      Values from Tables 4–7 of the paper; all are predictions (corrected COSMO-RS Henry’s constants; group-contribution viscosity, melting point and heat capacity). “—”: not reported for the benchmark set. Amber corner: above the Tm < 270 K limit stated in the paper’s text; Table 4 lists two selected candidates at 271.25 and 271.36 K. Benchmark viscosities are reproduced as tabulated in the paper; measured values for some of these ionic liquids differ (see NIST ILThermo).

      Parallel coordinates: drag along an axis to brush a range

      Every axis is oriented so the preferred direction is up, except heat capacity, which has no preferred direction here

      Historical founder research
      Brushes filter the table, this chart and the Pareto explorer together. Tap an axis without dragging, or press Escape on it, to clear its brush. A line stops where a value was not reported.

      Pareto explorer

      Choose any two criteria; the dashed line is the non-dominated front

      Computed from published data

      This explorer replays a 2016 study. It is not an output of LILM, and none of these candidates has been synthesized or tested by Avestiq.

      Validation Planned validation protocol

      Freeze the benchmark before seeing the answer.

      Under the planned protocol, the benchmark is written and shared with the industrial partner before any candidate is ranked, so a result cannot be reinterpreted after the fact.

      One prospective benchmark

      Planned validation protocol
      Planned prospective benchmark, schematicA protocol is frozen first. Two lanes then run in parallel: LILM’s top-k candidates and the best baseline’s top-k candidates. Both sets are synthesized and measured blind by an independent laboratory. Metrics are computed, the result is reported whether it passes or fails, and the model is updated only after the test closes.
      k, metrics and baselines are fixed with the industrial partner before ranking. No benchmark has been run yet.

      Splits

      • Random
      • Ion family held out
      • Time split
      • Prospective

      Metrics

      • Mean absolute error
      • Relative error
      • 95% interval coverage
      • Top-k enrichment
      • Screens per qualified candidate
      • Expert hours per decision

      Controls

      • Protocol frozen before ranking
      • Baseline synthesized alongside the candidate set
      • Independent measurement
      • All results reported
      • Model updated only after the test closes
      top-k enrichment = (qualified hits in LILM’s top k) ÷ (qualified hits in the best baseline’s top k) · same k · both sets synthesized

      A failed benchmark is a result.
      It is not reinterpreted after the fact.

      Privacy

      Terms