Science Historical founder research

The scientific foundation predates the AI boom.

The pieces LILM is designed to integrate were developed and published separately, starting in 2011. This is the founder’s historical research, not current Avestiq validation.

A historical design case Historical founder research Published research

185 candidates to one, computationally.

A 2013 computer-aided design for toluene/n-heptane separation, benchmarked against sulfolane (AIChE J. 2013, 59, 4627–4640). Select a stage to see the constraint applied.

Designed IL vs incumbent solvent

Predicted properties of the designed ionic liquid compared with sulfolane reference values

β∞: UNIFAC prediction at infinite dilution. Measured finite-concentration selectivities of related ionic liquids (toluene/heptane, 40 °C) were 16.7–32.8, so the comparison is not like-for-like. The designed IL was not synthesized or tested in that study. Sulfolane values are the paper’s reference thresholds.

What it demonstrated

  • Systematic search can eliminate most candidates before synthesis
  • Structural trends can emerge computationally
  • Multiple constraints can be imposed simultaneously

What it left open

  • Uncertainty intervals
  • Synthesis feasibility
  • Process penalties (viscosity 5.5× the incumbent’s)
  • Cost
  • Toxicity
  • Life-cycle impact
  • Experimental confirmation

LILM is designed around the questions the early CAILD work left open.

Uncertainty Historical founder research

A model should know when it does not know.

In a published in-domain test, a COSMOtherm-based QSPR reproduced measured heat capacity and density of imidazolium ionic liquids closely. For a novel candidate outside the data, the error is unknown unless the model reports it.

Heat capacity, 8 ILs

Predicted vs experimental, J mol⁻¹ K⁻¹

View data

Density, 9 ILs

Predicted vs experimental, kg m⁻³

View data
Historical in-domain average relative error ≈1.5% (1.49%; maximum 3.52%). Shaded band: ±5%. Mehrkesh & Karunanithi, Comput. Chem. Eng. 2016, Table 6. [emim][PF6] heat capacity had no experimental value.

LILM requirements Planned

  1. Prediction intervals with stated coverage
  2. Applicability-domain score on every candidate
  3. Calibration checked prospectively
  4. Out-of-domain flagging
  5. No silent ranking outside the validated domain

These plots show historical model performance from published founder research. They do not represent LILM performance, which has not yet been measured.

Empirical-correlation lab Historical founder research Computed from published data

Measure a model against experiment before trusting it.

Three published comparisons of predicted and measured properties, plus one computed here: Peng–Robinson against the NIST reference equation of state for CO2. Choose a dataset, apply a correction, and read the error statistics. Recomputed from the tabulated values, the averages agree with the papers’ to within rounding.

Predicted vs experimental

Historical founder research

    Residuals

    Historical founder research
    Residual = (predicted − experimental) ÷ experimental. The dashed line is a least-squares trend through the uncorrected residuals; a slope means the error grows with the property value, which a linear correction can remove.

    Error statistics

    Error statistics for the selected dataset

    Residual map

    Historical founder research
    Heat capacity and density are arranged by cation and anion, which shows whether error follows a family. The Henry’s-constant test points have no compound identities in the source, and the saturation-pressure points are arranged by temperature.
    DefinitionsRMSE = √(Σ(ŷ − y)²/n) · MAE = Σ|ŷ − y|/n · MAPE = (100/n)·Σ|ŷ − y|/y · R² = 1 − Σ(ŷ − y)²/Σ(y − ȳ)² · bias = Σ(ŷ − y)/n · within ±b = share of points with |ŷ − y|/y ≤ b

    These are historical in-domain tests of published models. They are not measurements of LILM, which has not been benchmarked.

    Property correlations Historical founder research Computed from published data

    Which predicted properties move together, and which only seem to?

    Seven predicted properties of the ten candidates from the 2016 CO2 screen. Select a cell to see the scatter behind the coefficient.

    Correlation matrix, n = 10

    Computed from published data
    * p < 0.05 · ** below the Bonferroni threshold. Spearman ρ uses ranks with ties averaged; its p-value uses the same t approximation.

    Selected pair

    Computed from published data
    Points are labeled with the candidate’s rank in the paper. Dashed line: least-squares fit (Pearson view).

    Source: Farahipour, Mehrkesh & Karunanithi, Chem. Eng. Sci. 2016, Tables 4 and 6. All values are predictions. Correlation across ten selected candidates says nothing about the full design space, which the selection has already narrowed.

    Multi-objective design Historical founder research

    A single-objective optimum is not an industrial decision.

    Published thermal-energy-storage designs, replotted as a trade-off between storage density and melting point.

    Storage density vs melting point, published IL structures

    Pareto trade-off for thermal-energy-storage ionic liquidsScatter of predicted storage density against melting point for published candidate structures, marking Pareto-optimal feasible candidates, other feasible candidates, candidates failing the decomposition-temperature constraint, candidates failing the melting-point constraint, and two incumbent fluids.
      View data
      Group-contribution predictions from Mehrkesh & Karunanithi, Comput. Chem. Eng. 2016, 93, 402–412 (Tables 2–5; η = ρ·Cp at 300 °C). Constraints from the paper: Tm < 140 °C, Td > 400 °C. The paper maximized η; the Pareto replot is Avestiq’s. Incumbent values as reported in the paper.

      The candidate may win one objective and lose on another.

      Industrial decisions must balance properties, processability, cost, EHS and uncertainty simultaneously.

      The 2016 optimum matched the molten salt on storage density and exceeded it on decomposition margin in the model. It was not synthesized or tested in that study, and cost was outside the model.

      Design space Historical founder research

      Change the anion and the answer changes.

      For the best cation in each of five families, the 2016 study varied the anion. Predicted storage density moved by up to 0.34 MJ m⁻³ K⁻¹ within one family, and the melting-point and stability limits removed 12 of the 25 combinations.

      Predicted storage density η, 5 cation families × 5 anions

      η = ρ·Cp at 300 °C, MJ m⁻³ K⁻¹ · hover a cell for Tm and Td

      Heatmap of predicted storage density by cation family and anionTwenty-five cells, one per cation family and anion. Color encodes predicted storage density from 2.30 to 2.88 megajoules per cubic meter per kelvin. Hatched cells fail the melting-point limit; cells marked Td fail the decomposition limit. Imidazolium with tetrafluoroborate has the highest value, 2.88.
      View data
      Group-contribution predictions, Mehrkesh & Karunanithi, Comput. Chem. Eng. 2016, 93, 402–412, Table 3 (best cation per family, anion varied). Limits from the paper: Tm < 413.15 K (140 °C), Td > 673.15 K (400 °C); Td is the apparent decomposition temperature. Predictions, not measurements.

      The anion with the best storage density is not always the safest choice.

      [BF4]− gave the highest η in every family, while [Tf2N]− gave the lowest melting point in every family. Chloride failed the melting-point limit in every family, and both halides failed the stability limit for pyridinium and pyrrolidinium.

      A design tool has to carry every constraint at once. Ranking on one property and checking the others afterward discards the trade-off this grid makes visible.

      Life cycle Historical founder research

      Nonvolatile is not the same as low-impact.

      The published founder work found that production impacts could dominate modeled use-phase release under the study assumptions. On average, 83% of production-phase ecotoxicity came from chemicals released in upstream synthesis and 17% from life-cycle energy.

      Production-phase freshwater ecotoxicity by emission source

      CTUe per kg of ionic liquid produced (USEtox-based characterization)

      Stacked bars of production-phase ecotoxicityFor three ionic liquids, stacked horizontal bars split production-phase ecotoxicity into upstream emission sources: halobenzene from the butyl-halide route, chloramine from the pyridine route, metals from the 1-methylimidazole route, and other or not itemized.
        View data
        Mehrkesh & Karunanithi, Environ. Sci. Technol. 2016, 50, 6814–6821 (Table 5). Modeled values with stated uncertainty; the study assumed 0.4–2% of the ionic liquid reaches freshwater during use. “Other” combines benzene, dihalobenzene and the remainder not itemized in the table.

        The same study reported freshwater ecotoxicity characterization factors for direct release of five common ionic liquids, from 624 to 1,768 CTUe/kg.

        Toxicity and life-cycle results are screening estimates. They do not establish that any chemical is safe or environmentally preferable.

        Direct-release characterization factors

        CF = EF × FF × XF · bubble area ∝ CF (CTUe/kg)

        Freshwater ecotoxicity characterization factors of five ionic liquidsBubble chart. The horizontal axis is the effect factor, 4 to 13 PAF cubic meters per kilogram; the vertical axis is the fate factor, 90 to 150 days. Bubble area is proportional to the characterization factor: [Bmim][Br] 624, [Bmim][Cl] 748, [Bmim][BF4] 823, [Bmim][PF6] 927 and [BPy][Cl] 1,768 CTUe per kilogram. Dashed curves mark constant characterization factor.
        View data
        Mehrkesh & Karunanithi, Environ. Sci. Technol. 2016, Table 3 (USEtox-based; exposure factor XF ≈ 0.9996–0.9998). Dashed curves: constant CF = EF × FF × XF, with XF = 0.9996.

        Production-phase ecotoxicity: energy vs chemical release

        Share of cradle-to-gate freshwater ecotoxicity, %

        Energy-related versus chemical-related ecotoxicity of ionic-liquid productionOne hundred percent stacked bars for five ionic liquids. Energy-related shares: [Bmim][Br] 8%, [Bmim][Cl] 20%, [Bmim][BF4] 27%, [Bmim][PF6] 21%, [BPy][Cl] 8%. The remainder is chemical-related. The average energy share is about 17%.
        View data
        Mehrkesh & Karunanithi, Environ. Sci. Technol. 2016, Table 4. Chemical-related: direct release of materials and chemicals in upstream production of reagents and precursors.

        Life-cycle impact belongs inside candidate ranking—not after molecular selection.

        Privacy

        Terms