Pith. sign in

REVIEW 3 major objections 4 minor 97 references

This paper shows that a hybrid model embedding matrix completion inside the Bromley activity equation can predict mean ionic activity and osmotic coefficients for unseen electrolytes, yielding a completed parameter set for 9,296 aqueous sal

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 13:23 UTC pith:6Y2FBF42

load-bearing objection Solid hybrid ML paper with a genuinely useful idea, but the 9,296-electrolyte headline outruns the validation. the 3 major comments →

arxiv 2607.19114 v1 pith:6Y2FBF42 submitted 2026-07-21 cs.LG physics.chem-ph

Predicting Activities in Aqueous Electrolyte Solutions with Hybrid Machine Learning

Zeno Romero , Maximilian Kohns , Fabian Jirasek This is my paper
classification cs.LG physics.chem-ph
keywords Bromley modelmatrix completionhybrid machine learningactivity coefficientsosmotic coefficientselectrolyte thermodynamicsaqueous solutionslow-rank factorization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to establish that a hybrid model, Bromley-MCM, can predict concentration-dependent mean ionic activity and osmotic coefficients for electrolytes never seen during training, across a far larger chemical space than earlier predictive models. It treats the single adjustable parameter of the physics-based Bromley activity model as a cation-by-anion matrix and completes that matrix using a low-rank factorization learned end-to-end from experimental activity data. The completed matrix covers 83 cations and 112 anions, yielding predictions for 9,296 aqueous electrolytes at 298 K, of which 8,818 have no literature data. In leave-one-electrolyte-out tests, the model reports substantially lower prediction errors than the ion-specific Bromley model and the Simoes model.

Core claim

The central claim is that the electrolyte-specific Bromley parameter B_CA — which classically must be fitted to experimental data for each salt — can be predicted as a dot product of a cation feature vector and an anion feature vector of length three, with both vectors inferred directly from measured activity and osmotic coefficients. Training on 478 electrolytes produces a complete 83-by-112 parameter matrix, so the Bromley model becomes predictive for 9,296 salts at 298 K. The paper demonstrates that for held-out electrolytes, the predicted activity and osmotic coefficients are markedly more accurate than those of two established ion-descriptor benchmarks, including in a controlled compari

What carries the argument

The Bromley-MCM model. It embeds a matrix completion method inside the Bromley activity equations: the Bromley parameter B_CA = θ_C · β_A is a rank-3 factorization of the cation-by-anion parameter matrix, learned in a Bayesian framework with Cauchy priors and a Cauchy likelihood on each measured γ± and φ value. Because training is end-to-end, the ion feature vectors are determined by thermodynamic data rather than by predefined ion descriptors. Once a parameter is completed, the Bromley equations convert it into smooth, Gibbs–Duhem-consistent activity and osmotic coefficients as functions of ionic strength.

Load-bearing premise

That every cation–anion interaction strength (the Bromley parameter) is well captured by a rank-3 dot product of ion-specific vectors, so that a pair of ions never jointly observed can be predicted from their behavior with other partners.

What would settle it

Measure γ± and φ experimentally at 298 K for a salt whose cation and anion each appear in only one measured electrolyte in the training set (so the pair is unmeasured and both ions are sparse), then compare the measurements with the Bromley-MCM prediction. If the errors greatly exceed the leave-one-electrolyte-out errors reported for well-connected ions, the low-rank completion is failing for sparse pairs.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • For any of the 9,296 cation–anion combinations spanned by the 83 cations and 112 anions, the mean ionic activity and osmotic coefficients at 298 K become computable without new experimental fitting.
  • The published complete parameter matrix can be dropped directly into existing engineering calculations that already use the Bromley model, such as solubility or process design.
  • Because the model is trained simultaneously on activity and osmotic data, it can give consistent predictions for salts for which only one of the two properties has ever been measured.
  • Accurate predictions are reported for electrolytes that challenge the benchmark models, including organic anions, multiply charged ions, and ionic coordination complexes.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A natural next test would be to let the ion feature vectors depend on temperature, since the current model is confined to 298 K; the same hybrid architecture could then cover the temperature-dependent form of the Bromley model.
  • The same matrix-completion-in-a-physical-model idea could be carried to other solvents or mixed-solvent systems, where the data are even sparser and prediction would be more valuable.
  • The largest blind spot is electrolytes whose cation or anion appears in only one measured salt: the leave-one-electrolyte-out evaluation cannot validate those, so a targeted measurement campaign on such rare-ion pairs would reveal whether the rank-3 assumption extrapolates.
  • The current paper reports point predictions; reporting the posterior spread of each completed B_CA would make the model more informative in engineering use.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper introduces Bromley-MCM, a hybrid model that replaces the electrolyte-specific Bromley parameter B_CA with a low-rank matrix completion factorization B_CA = θ_C · β_A (Eq. 15, K=3). The model is trained end-to-end on mean ionic activity coefficients and osmotic coefficients from a consolidated DDB dataset, and is evaluated by leave-one-electrolyte-out on a reduced database of 401 electrolytes whose ions appear in at least two measured pairs. The authors compare Bromley-MCM against a re-fitted ion-specific Bromley model (IBM) and the Simoes Pitzer-parameter model, reporting that Bromley-MCM outperforms both. They then train on the full database (478 electrolytes, 83 cations, 112 anions) and release a completed B_CA matrix for 9,296 electrolytes, of which 8,818 have no data at 298 K.

Significance. If the central claim were fully supported, this would be a practically valuable extension of the Bromley model to a much broader electrolyte space, and the hybrid matrix-completion architecture is a sensible design. The paper has genuine strengths: the leave-one-electrolyte-out protocol provides true holdout predictions for electrolytes in the reduced database; the K=2 control is a reasonable attempt to isolate the effect of architectural expressiveness from parameter count; the IBM baseline is honestly re-fitted to the same data; and the authors flag weak/protonatable electrolytes that the Bromley model cannot describe. However, the validation design is narrower than the headline claim: the completed 9,296-electrolyte matrix is extrapolated far beyond the evaluated regime, both for ions that appear in only one measured pair and for weak electrolytes explicitly excluded from the accuracy assessment. The significance of the work depends on closing or honestly re-scoping that gap.

major comments (3)
  1. [Database / Table 1; Conclusions] The leave-one-electrolyte-out evaluation is performed on the reduced database (401 electrolytes, 57 cations, 62 anions), which retains only ions with at least two measured cation–anion pairs. The full database adds 26 cations and 50 anions that appear only in the 77 electrolytes dropped from the reduced set. For any electrolyte containing one of these sparse ions, the model is never tested in a held-out pair with a different partner, so the reported accuracy provides no evidence for the completed entries involving those ions. The 9,296- and 8,818-electrolyte statements in the Abstract and Conclusions are therefore extrapolations beyond the evaluated regime. The only support for them is the low-rank ansatz of Eq. (15) with K=3, whose hyperparameter was selected on the reduced database (Fig. S3). Please provide a direct validation for sparse-ion pairs — e.g., a leave-one-out analysis on th
  2. [Database, 'Electrolyte Activity Models' / 'Hybrid Bromley-MCM'] The consolidated database excludes electrolytes for which the Bromley fit residual exceeds 5%, so the reported prediction errors in Figures 4 and 5 are conditional on electrolytes that the Bromley model can describe. Yet the completed matrix includes weak and protonatable electrolytes (acetic acid, phosphoric acid, etc.), which the authors themselves flag as outside the Bromley model's validity. These flagged electrolytes are counted in the 9,296 completed parameters and in the 'no data available' subset of 8,818. The abstract's accuracy claim therefore does not transfer to a portion of the delivered parameter matrix. Please quantify how many of the 9,296 entries are flagged as weak/protonatable, and either report their expected accuracy separately or exclude them from the headline counts.
  3. [Results, 'Comparison of Bromley-MCM to the Simoes Model'; Fig. 5 and Fig. S7] In Figure 5, the Simoes model is used with its published fitted parameters, so for many electrolytes its results are correlations rather than true predictions, while Bromley-MCM is evaluated as a true prediction. The authors acknowledge this and report a separate seen/unseen split in Fig. S7, which is helpful. However, the main-text statement that 'Bromley-MCM substantially outperforms the Simoes model' is based on a mixed comparison; the paper would be clearer if the main box plot reported the unseen-electrolyte subset (Fig. S7 right) and presented the seen subset only as a sanity check.
minor comments (4)
  1. [Throughout] There are numerous typos and inconsistent spellings, e.g., 'thermodnyamics' in the Introduction, 'electrolyts' in Results, 'Cadmiun(II)' and 'Timethylsulfonium' in SI Tables S1/S2, and 'plantinum(IV)' in Table S1. A careful proofread is needed.
  2. [Computational Details and Evaluation, Fig. S3] The claim that the K=2 comparison 'does not depend on the number of trainable parameters' is a bit strong: the IBM and the K=2 MCM differ in functional form (BIBCA = B_C + B_A + δ_C δ_A vs. an unrestricted inner product), not only in parameter count. The comparison is still informative, but the wording should be softened.
  3. [Results, Ref. to Figure S3] Figure S3 is difficult to read: the axis tick labels and legends are dense and partially garbled in the provided rendering. The hyperparameter values actually chosen (K=3, σ0=0.2, λ=0.1) should be listed in a short table or caption text.
  4. [Abstract / Conclusions] The phrase 'predicting activities for unstudied electrolytes' is used as if the 9,296 entries are equivalently validated. Given the validation-gap issue above, the abstract should state that the leave-one-out results apply to the reduced database, and that the remaining entries are model extrapolations.

Circularity Check

0 steps flagged

No significant circularity: the Bromley-MCM derivation is an empirically evaluated low-rank completion, not a restatement of its inputs.

full rationale

The claimed chain is that the electrolyte-specific Bromley parameter is modeled as B_CA = θ_C · β_A (Eq. 15) inside the Bromley equations (9)-(13), with θ and β inferred end-to-end from experimental γ± and φ data by Bayesian inference and evaluated by leave-one-electrolyte-out. No step defines a predicted quantity in terms of the same quantity used as its target: for a held-out electrolyte, B_CA is determined by feature vectors learned from other electrolytes, not by refitting that electrolyte's own data. Eq. 15 is an explicit low-rank matrix-completion assumption, not a derived result, and it is tested against held-out electrolytes rather than asserted from a uniqueness theorem. The heavy citation of the authors' own matrix-completion work describes the inference framework and prior applications, but the central equations and evaluation protocol are stated in this paper, so the self-citations are not load-bearing. The comparison with the ion-specific Bromley model is also not a renaming: the IBM form B_C + B_A + δ_C·δ_A is a special case of the more flexible rank-3 bilinear form, and the paper demonstrates predictive differences. The manuscript itself flags two validation concerns: hyperparameters (K, σ0, λ) were selected by leave-one-out on the same reduced database later used to report final errors (Fig. S3 vs. Fig. 4), and the full 9,296-electrolyte claim includes weak electrolytes and sparse ions that the leave-one-out evaluation does not test and that the authors explicitly caution are outside Bromley model validity. These are correctness and extrapolation risks, not definitional circularity: the predictions are not equivalent to the training inputs by construction. Therefore the appropriate circularity finding is a non-finding with score 0.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

The central claim rests on the low-rank factorization and on a data-selection protocol that retains only electrolytes the Bromley model describes within 5%; the learned latent features are fitted to the training data, and hyperparameters (K, σ0, λ) are tuned on the same reduced dataset. No new physical entities are introduced.

free parameters (5)
  • Matrix rank K = 3
    Chosen from preliminary leave-one-out experiments (Fig. S3); sets the number of latent features per ion and controls model capacity.
  • Latent ion feature vectors θ_C and β_A = 357 numbers (83×3 + 112×3) learned by ADVI; also a K=2 variant with 238 numbers
    These are the actual model parameters fit end-to-end to activity/osmotic data; they encode all predictions and are not fixed from prior literature.
  • Prior scale σ0 = 0.2 for K=3; 0.02 for K=2
    Hyperparameter selected by preliminary leave-one-out experiments on the reduced dataset; controls shrinkage of latent features.
  • Likelihood scale λ = 0.1
    Cauchy likelihood scale, selected via the same hyperparameter search on the reduced dataset.
  • Electrolyte inclusion threshold / Bromley fit residual = 5% relative mean residual
    Electrolytes for which the Bromley model cannot reproduce φ or γ± within 5% are excluded from the database; directly controls which systems enter training and evaluation.
axioms (6)
  • domain assumption Complete dissociation of all considered electrolytes (Eq. 1)
    The Bromley extension for associating ions is not considered; weak electrolytes are flagged in SI but predictions are still generated for them.
  • domain assumption Bromley model equations (9)–(13) with fixed constants 0.06, 0.6, and c=1.5/|zC zA|
    Adopted from ref. 10; not derived in this paper. This is the physical backbone that maps B_CA to γ± and φ.
  • domain assumption Bromley model valid up to ionic strength I=6 in water at 298 K
    Data points at I>6 were excluded based on the literature upper bound; predictions outside this range rely on extrapolation.
  • ad hoc to paper Low-rank factorization B_CA = θ_C · β_A with K=3
    Central inductive bias of the matrix completion; not derived from physics. Validated only indirectly via leave-one-out on observed electrolyte pairs.
  • domain assumption Gibbs-Duhem consistency of the consolidated data
    Used to verify simultaneous fits of φ and ln γ± within 5%; assumes the Bromley model is thermodynamically consistent and the underlying data mutually consistent.
  • domain assumption ADVI posterior approximation is adequate
    Automatic differentiation variational inference is used; no convergence diagnostics or posterior predictive checks are reported.

pith-pipeline@v1.3.0-alltime-deepseek · 34047 in / 10444 out tokens · 99009 ms · 2026-08-01T13:23:16.336463+00:00 · methodology

0 comments
read the original abstract

Activities in aqueous electrolyte solutions, usually described by ionic activity and osmotic coefficients, are important properties for modeling many processes in industry and nature. Established activity models, such as those of Pitzer or Bromley, require fitting to experimental data for each electrolyte of interest and thus cannot predict properties for unstudied systems. While some predictive approaches exist, they are typically limited in scope and rely on additional ion-specific descriptors. In this work, we introduce a new hybrid model that combines the physics-based Bromley model with a matrix completion method (MCM) from machine learning. The MCM is employed to predict the electrolyte-specific parameters of the Bromley model, exploiting the fact that these parameters can be arranged in a matrix with cations and anions as rows and columns, respectively. Due to the lack of experimental data for many electrolytes, the initial parameter matrix is sparsely populated, making the prediction of the Bromley parameters for unstudied electrolytes a matrix completion problem. The hybrid model, Bromley-MCM, was trained end-to-end on experimental data for mean ionic activity coefficients and osmotic coefficients of aqueous solutions of 478 electrolytes at 298 K from the Dortmund Data Bank. As output, we obtain a completed matrix of Bromley parameters for 83 cations and 112 anions, enabling consistent prediction of concentration-dependent activities in aqueous solutions of 9,296 electrolytes at 298~K. This substantially extends the applicability of the Bromley model while maintaining high predictive accuracy, as demonstrated through evaluations on electrolytes excluded from model training.

Figures

Figures reproduced from arXiv: 2607.19114 by Fabian Jirasek, Maximilian Kohns, Zeno Romero.

Figure 1
Figure 1. Figure 1: Experimental data availability (blue cells) for [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Experimental data availability for γ˜± and/or ϕ for cation-anion pairs at T = 298±1 K in aqueous systems in the Simoes database. For the numbers identifying the cations and anions, see Table S3 in the Supporting Information. White cells indicate electrolytes for which no experimental data are available [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Schematic representation of the hybrid Bromley-MCM for predicting [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Box plot of the ln ˜γ± and ϕ prediction errors of Bromley-MCM and IBM on the reduced database evaluated using leave-one-electrolyte-out analysis. Boxes represent interquartile range (IQR), and whiskers denote 1.5 IQR. For both models, we observe higher prediction errors for ln ˜γ± than for ϕ. Given that the original Bromley model – the physical model underlying both approaches – is Gibbs-Duhem￾consistent, … view at source ↗
Figure 5
Figure 5. Figure 5: shows the box plot of the prediction errors of ln ˜γ± and ϕ for Bromley-MCM and Simoes on the Simoes horizon (cf [PITH_FULL_IMAGE:figures/full_fig_p018_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Prediction of ln ˜γ± for exemplary electrolytes in aqueous solutions at 298 K as a function of ionic strength. Lines: Predictions with different models. Circles: experimental data from DDB.39 20 [PITH_FULL_IMAGE:figures/full_fig_p020_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Prediction of ϕ−1 in exemplary aqueous electrolyte solutions at 298 K as a function of ionic strength. Lines: Predictions with different models. Circles: experimental data from DDB.39 The agreement between the experimental data and the predictions with Bromley-MCM is 21 [PITH_FULL_IMAGE:figures/full_fig_p021_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

97 extracted references · 36 canonical work pages · 1 internal anchor

  1. [1]

    Database for liquid phase diffusion coefficients at infinite dilution at 298 K and matrix completion methods for their prediction , journal =

    Oliver Gro. Database for liquid phase diffusion coefficients at infinite dilution at 298 K and matrix completion methods for their prediction , journal =. 2022 , publisher =. doi:10.1039/d2dd00073c , url =

    work page doi:10.1039/d2dd00073c 2022

  2. [2]

    Stan Development Team, Stan Modeling Language Users Guide and Reference Manual, Version 2.35 , date =

  3. [3]

    Prediction of Henry

    Nicolas Hayer and Fabian Jirasek and Hans Hasse , journal =. Prediction of Henry. 2022 , month =. doi:10.1002/aic.17753 , eid =

    work page doi:10.1002/aic.17753 2022

  4. [4]

    Jirasek, Fabian and Alves, Rodrigo A. S. and Damay, Julie and Vandermeulen, Robert A. and Bamler, Robert and Bortz, Michael and Mandt, Stephan and Kloft, Marius and Hasse, Hans , year =. Machine learning in thermodynamics: Prediction of activity coefficients by matrix completion , volume =. The Journal of Physical Chemistry Letters , publisher =. doi:10.1...

    work page doi:10.1021/acs.jpclett.9b03657

  5. [5]

    Hybridizing physical and data-driven prediction methods for physicochemical properties , volume =

    Jirasek, Fabian and Bamler, Robert and Mandt, Stephan , year =. Hybridizing physical and data-driven prediction methods for physicochemical properties , volume =. Chemical Communications , publisher =. doi:10.1039/d0cc05258b , number =

    work page doi:10.1039/d0cc05258b

  6. [6]

    Perspective: Machine learning of thermophysical properties , volume =

    Jirasek, Fabian and Hasse, Hans , year =. Perspective: Machine learning of thermophysical properties , volume =. doi:10.1016/j.fluid.2021.113206 , journal =

    arXiv 2021

  7. [7]

    Blei and Alp Kucukelbir and Jon D

    David M. Blei and Alp Kucukelbir and Jon D. McAuliffe , title =. 2017 , month = apr, publisher =. doi:10.1080/01621459.2017.1285773 , url =

    arXiv 2017

  8. [8]

    Predicting activity coefficients at infinite dilution for varying temperatures by matrix completion , volume =

    Damay, Julie and Jirasek, Fabian and Kloft, Marius and Bortz, Michael and Hasse, Hans , year =. Predicting activity coefficients at infinite dilution for varying temperatures by matrix completion , volume =. Industrial and Engineering Chemistry Research , publisher =. doi:10.1021/acs.iecr.1c02039 , number =

    work page doi:10.1021/acs.iecr.1c02039

  9. [9]

    Prediction of temperature-dependent Henry’s law constants by matrix completion , volume =

    Hayer, Nicolas and Hasse, Hans and Jirasek, Fabian , year =. Prediction of temperature-dependent Henry’s law constants by matrix completion , volume =. The Journal of Physical Chemistry B , publisher =. doi:10.1021/acs.jpcb.4c07196 , number =

    work page doi:10.1021/acs.jpcb.4c07196

  10. [10]

    Blei , title =

    Alp Kucukelbir and Dustin Tran and Rajesh Ranganath and Andrew Gelman and David M. Blei , title =. Journal of Machine Learning Research , year =

  11. [11]

    Hierarchical matrix completion for the prediction of properties of binary mixtures , ISSN =

    Gond, Dominik and Sohns, Jan-Tobias and Leitte, Heike and Hasse, Hans and Jirasek, Fabian , year =. Hierarchical matrix completion for the prediction of properties of binary mixtures , ISSN =. doi:10.1016/j.compchemeng.2025.109122 , journal =

    arXiv 2025

  12. [12]

    Greg Landrum and Paolo Tosco and Brian Kelley and Ricardo Rodriguez and David Cosgrove and Riccardo Vianello and sriniker and Peter Gedeck and Gareth Jones and Nadine Schneider and Eisuke Kawashima and Dan Nealschneider and Andrew Dalke and Matt Swain and Brian Cole and Samo Turk and Aleksandr Savelev and Alain Vaucher and Maciej Wojcikowski and Ichiru Ta...

    2024

  13. [13]

    Prediction of parameters of group contribution models of mixtures by matrix completion , volume =

    Jirasek, Fabian and Hayer, Nicolas and Abbas, Rima and Schmid, Bastian and Hasse, Hans , year =. Prediction of parameters of group contribution models of mixtures by matrix completion , volume =. Physical Chemistry Chemical Physics , publisher =. doi:10.1039/d2cp04478a , number =

    work page doi:10.1039/d2cp04478a

  14. [14]

    Advancing thermodynamic group-contribution methods by machine learning: UNIFAC 2.0 , volume =

    Hayer, Nicolas and Wendel, Thorsten and Mandt, Stephan and Hasse, Hans and Jirasek, Fabian , year =. Advancing thermodynamic group-contribution methods by machine learning: UNIFAC 2.0 , volume =. doi:10.1016/j.cej.2024.158667 , journal =

    arXiv 2024

  15. [15]

    Prediction of pair interactions in mixtures by matrix completion , volume =

    Hoffmann, Marco and Hayer, Nicolas and Kohns, Maximilian and Jirasek, Fabian and Hasse, Hans , year =. Prediction of pair interactions in mixtures by matrix completion , volume =. Physical Chemistry Chemical Physics , publisher =. doi:10.1039/d4cp01492h , number =

    work page doi:10.1039/d4cp01492h

  16. [16]

    MLPROP – An Interactive Web Interface for Thermophysical Property Prediction with Machine Learning , volume =

    Hoffmann, Marco and Specht, Thomas and Hayer, Nicolas and Hasse, Hans and Jirasek, Fabian , year =. MLPROP – An Interactive Web Interface for Thermophysical Property Prediction with Machine Learning , volume =. Chemie Ingenieur Technik , publisher =. doi:10.1002/cite.70004 , number =

    work page doi:10.1002/cite.70004

  17. [17]

    Combining machine learning with physical knowledge in thermodynamic modeling of fluid mixtures , volume =

    Jirasek, Fabian and Hasse, Hans , year =. Combining machine learning with physical knowledge in thermodynamic modeling of fluid mixtures , volume =. Annual Review of Chemical and Biomolecular Engineering , publisher =. doi:10.1146/annurev-chembioeng-092220-025342 , number =

    doi:10.1146/annurev-chembioeng-092220-025342

  18. [18]

    Making thermodynamic models of mixtures predictive by machine learning: matrix completion of pair interactions , volume =

    Jirasek, Fabian and Bamler, Robert and Fellenz, Sophie and Bortz, Michael and Kloft, Marius and Mandt, Stephan and Hasse, Hans , year =. Making thermodynamic models of mixtures predictive by machine learning: matrix completion of pair interactions , volume =. Chemical Science , publisher =. doi:10.1039/d1sc07210b , number =

    doi:10.1039/d1sc07210b

  19. [19]

    Modified UNIFAC 2.0-A Group-Contribution Method Completed with Machine Learning , volume =

    Hayer, Nicolas and Hasse, Hans and Jirasek, Fabian , year =. Modified UNIFAC 2.0-A Group-Contribution Method Completed with Machine Learning , volume =. Industrial & Engineering Chemistry Research , publisher =. doi:10.1021/acs.iecr.5c00077 , number =

    work page doi:10.1021/acs.iecr.5c00077

  20. [20]

    Prediction of activity coefficients by similarity-based imputation using quantum-chemical descriptors , volume =

    Hayer, Nicolas and Specht, Thomas and Arweiler, Justus and Gond, Dominik and Hasse, Hans and Jirasek, Fabian , year =. Prediction of activity coefficients by similarity-based imputation using quantum-chemical descriptors , volume =. Physical Chemistry Chemical Physics , publisher =. doi:10.1039/d4cp04341c , number =

    work page doi:10.1039/d4cp04341c

  21. [21]

    , year =

    Argersinger, William J. , year =. Activity Coefficients of Electrolytes in Mixed Aqueous Solution from Electromotive Force Data , volume =. The Journal of Physical Chemistry , publisher =. doi:10.1021/j150519a021 , number =

    work page doi:10.1021/j150519a021

  22. [22]

    Bousfield, W. R. , year =. Iso-piestic solutions , volume =. Transactions of the Faraday Society , publisher =. doi:10.1039/tf9181300401 , number =

    work page doi:10.1039/tf9181300401

  23. [23]

    and Sinclair, Donald A

    Robinson, Robert A. and Sinclair, Donald A. , year =. The Activity Coefficients of the Alkali Chlorides and of Lithium Iodide in Aqueous Solution from Vapor Pressure Measurements , volume =. Journal of the American Chemical Society , publisher =. doi:10.1021/ja01324a003 , number =

    work page doi:10.1021/ja01324a003

  24. [24]

    Young, T. F. , year =. The Complete Calculation of Activity Coefficients from Freezing Point Data. , volume =. Chemical Reviews , publisher =. doi:10.1021/cr60044a008 , number =

    work page doi:10.1021/cr60044a008

  25. [25]

    Thermodynamik der Mischungen , ISBN =

    Heintz, Andreas , year =. Thermodynamik der Mischungen , ISBN =. doi:10.1007/978-3-662-49924-5 , publisher =

    work page doi:10.1007/978-3-662-49924-5

  26. [26]

    Predicting temperature‐dependent activity coefficients at infinite dilution using tensor completion , volume =

    Damay, Julie and Ryzhakov, Gleb and Jirasek, Fabian and Hasse, Hans and Oseledets, Ivan and Bortz, Michael , year =. Predicting temperature‐dependent activity coefficients at infinite dilution using tensor completion , volume =. Chemie Ingenieur Technik , publisher =. doi:10.1002/cite.202200230 , number =

    work page doi:10.1002/cite.202200230

  27. [27]

    Improvement of Diffusion Coefficient Prediction by Active Learning , volume =

    Romero, Zeno and M\". Improvement of Diffusion Coefficient Prediction by Active Learning , volume =. The Journal of Physical Chemistry B , publisher =. 2025 , month = aug, pages =. doi:10.1021/acs.jpcb.5c03825 , number =

    work page doi:10.1021/acs.jpcb.5c03825 2025

  28. [28]

    Prediction of Diffusion Coefficients in Mixtures with Tensor Completion

    Romero, Zeno and M\". Prediction of Diffusion Coefficients in Mixtures with Tensor Completion , publisher =. 2026 , copyright =. doi:10.48550/ARXIV.2602.23142 , url =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2602.23142 2026

  29. [29]

    Balancing molecular information and empirical data in the prediction of physico-chemical properties , volume =

    Zenn, Johannes and Gond, Dominik and Jirasek, Fabian and Bamler, Robert , year =. Balancing molecular information and empirical data in the prediction of physico-chemical properties , volume =. Digital Discovery , publisher =. doi:10.1039/d4dd00154k , number =

    work page doi:10.1039/d4dd00154k

  30. [30]

    and Singh, Daljit and Ray, Parthasarathi and Sridhar, Srinivasan and Read, Stanley M

    Bromley, Leroy A. and Singh, Daljit and Ray, Parthasarathi and Sridhar, Srinivasan and Read, Stanley M. , year =. Thermodynamic properties of sea salt solutions , volume =. AIChE Journal , publisher =. doi:10.1002/aic.690200218 , number =

    work page doi:10.1002/aic.690200218

  31. [31]

    , year =

    Bromley, Leroy A. , year =. Thermodynamic properties of strong electrolytes in aqueous solutions , volume =. AIChE Journal , publisher =. doi:10.1002/aic.690190216 , number =

    work page doi:10.1002/aic.690190216

  32. [32]

    A simple empirical model describing the thermodynamics of hydration of ions of widely varying charges, sizes, and shapes , volume =

    Marcus, Yizhak , year =. A simple empirical model describing the thermodynamics of hydration of ions of widely varying charges, sizes, and shapes , volume =. Biophysical Chemistry , publisher =. doi:10.1016/0301-4622(94)00051-4 , number =

    work page doi:10.1016/0301-4622(94)00051-4

  33. [33]

    and Hughes, Kevin J

    Simoes, Marcus C. and Hughes, Kevin J. and Ingham, Derek B. and Ma, Lin and Pourkashanian, Mohamed , year =. Estimation of the Pitzer Parameters for 1–1, 2–1, 3–1, 4–1, and 2–2 Single Electrolytes at 25 °C , volume =. Journal of Chemical & Engineering Data , publisher =. doi:10.1021/acs.jced.6b00236 , number =

    work page doi:10.1021/acs.jced.6b00236

  34. [34]

    and Mayorga, Guillermo , year =

    Pitzer, Kenneth S. and Mayorga, Guillermo , year =. Thermodynamics of electrolytes. II. Activity and osmotic coefficients for strong electrolytes with one or both ions univalent , volume =. The Journal of Physical Chemistry , publisher =. doi:10.1021/j100638a009 , number =

    work page doi:10.1021/j100638a009

  35. [35]

    A revised LIQUAC and LIFAC model (LIQUAC*/LIFAC*) for the prediction of properties of electrolyte containing solutions , volume =

    Mohs, André and Gmehling, J\". A revised LIQUAC and LIFAC model (LIQUAC*/LIFAC*) for the prediction of properties of electrolyte containing solutions , volume =. 2013 , month = jan, pages =. doi:10.1016/j.fluid.2012.09.023 , journal =

    work page doi:10.1016/j.fluid.2012.09.023 2013

  36. [36]

    Modified LIQUAC and Modified LIFACA Further Development of Electrolyte Models for the Reliable Prediction of Phase Equilibria with Strong Electrolytes , volume =

    Kiepe, J\". Modified LIQUAC and Modified LIFACA Further Development of Electrolyte Models for the Reliable Prediction of Phase Equilibria with Strong Electrolytes , volume =. Industrial & Engineering Chemistry Research , publisher =. 2006 , month = mar, pages =. doi:10.1021/ie0510122 , number =

    work page doi:10.1021/ie0510122 2006

  37. [37]

    A gE model for single and mixed solvent electrolyte systems , volume =

    Li, Jiding and Polka, Hans-Martin and Gmehling, J\". A gE model for single and mixed solvent electrolyte systems , volume =. 1994 , month = mar, pages =. doi:10.1016/0378-3812(94)87052-7 , journal =

    work page doi:10.1016/0378-3812(94)87052-7 1994

  38. [38]

    and Talbot, Nicola L.C

    Cawley, Gavin C. and Talbot, Nicola L.C. , year =. Efficient leave-one-out cross-validation of kernel fisher discriminant classifiers , volume =. Pattern Recognition , publisher =. doi:10.1016/s0031-3203(03)00136-5 , number =

    work page doi:10.1016/s0031-3203(03)00136-5

  39. [39]

    and Rard, Joseph A

    Clegg, Simon L. and Rard, Joseph A. and Pitzer, Kenneth S. , year =. Thermodynamic properties of 0–6 mol kg–1aqueous sulfuric acid from 273.15 to 328.15 K , volume =. J. Chem. Soc., Faraday Trans. , publisher =. doi:10.1039/ft9949001875 , number =

    work page doi:10.1039/ft9949001875

  40. [40]

    Electrolyte Solutions: Thermodynamics, Crystallization, Separation methods

    Kaj Thomsen. Electrolyte Solutions: Thermodynamics, Crystallization, Separation methods. 2009. doi:10.11581/dtu:00000073

    work page doi:10.11581/dtu:00000073 2009

  41. [41]

    and Silvester, Leonard F

    Pitzer, Kenneth S. and Silvester, Leonard F. , year =. Thermodynamics of electrolytes. VI. Weak electrolytes including H3PO4 , volume =. Journal of Solution Chemistry , publisher =. doi:10.1007/bf00645465 , number =

    work page doi:10.1007/bf00645465

  42. [42]

    Stokes, R. H. , year =. Thermodynamics of Solutions , ISBN =. doi:10.1201/9781351069472-1 , booktitle =

    work page doi:10.1201/9781351069472-1

  43. [43]

    Prediction of the density of aqueous electrolyte solutions with matrix completion methods , journal =

    Maximilian Kohns and Pascal Zittlau and Fabian Jirasek , keywords =. Prediction of the density of aqueous electrolyte solutions with matrix completion methods , journal =. 2025 , issn =. doi:https://doi.org/10.1016/j.fluid.2025.114454 , url =

    arXiv 2025

  44. [44]

    and Lazarou, G

    Kohns, M. and Lazarou, G. and Kournopoulos, S. and Forte, E. and Perdomo, F.A. and Jackson, G. and Adjiman, C.S. and Galindo, A. , title =. Physical Chemistry Chemical Physics , year =. doi:10.1039/c9cp06795g , keywords =

    work page doi:10.1039/c9cp06795g

  45. [45]

    Open circuit voltage of an all-vanadium redox flow battery as a function of the state of charge obtained from UV-Vis spectroscopy , volume =

    Heiß, Jana and Kohns, Maximilian , year =. Open circuit voltage of an all-vanadium redox flow battery as a function of the state of charge obtained from UV-Vis spectroscopy , volume =. Energy Advances , publisher =. doi:10.1039/d4ya00360h , number =

    work page doi:10.1039/d4ya00360h

  46. [46]

    Application of the Pitzer model for describing the evaporation of seawater , volume =

    Keller, Alexander and Burger, Jakob and Hasse, Hans and Kohns, Maximilian , year =. Application of the Pitzer model for describing the evaporation of seawater , volume =. doi:10.1016/j.desal.2020.114866 , journal =

    arXiv 2020

  47. [47]

    GRAPPA—A hybrid graph neural network for predicting pure component vapor pressures , volume =

    Hoffmann, Marco and Hasse, Hans and Jirasek, Fabian , year =. GRAPPA—A hybrid graph neural network for predicting pure component vapor pressures , volume =. doi:10.1016/j.ceja.2025.100750 , journal =

    arXiv 2025

  48. [48]

    HANNA: hard-constraint neural network for consistent activity coefficient prediction , volume =

    Specht, Thomas and Nagda, Mayank and Fellenz, Sophie and Mandt, Stephan and Hasse, Hans and Jirasek, Fabian , year =. HANNA: hard-constraint neural network for consistent activity coefficient prediction , volume =. Chemical Science , publisher =. doi:10.1039/d4sc05115g , number =

    doi:10.1039/d4sc05115g

  49. [49]

    Thermodynamically consistent machine learning model for excess Gibbs energy , volume =

    Hoffmann, Marco and Specht, Thomas and G\". Thermodynamically consistent machine learning model for excess Gibbs energy , volume =. Nature Communications , publisher =. doi:10.1038/s41467-026-71430-y , number =

    work page doi:10.1038/s41467-026-71430-y

  50. [50]

    Artificial intelligence in thermodynamics: hybrid modeling of thermophysical properties of fluids , volume =

    Hasse, Hans and Schmitt, Sebastian and Jirasek, Fabian , year =. Artificial intelligence in thermodynamics: hybrid modeling of thermophysical properties of fluids , volume =. doi:10.1016/j.coche.2026.101236 , journal =

    arXiv 2026

  51. [51]

    Open circuit voltage of an all-vanadium redox flow battery as a function of the state of charge obtained from UV-Vis spectroscopy

    Heiß, J.; Kohns, M. Open circuit voltage of an all-vanadium redox flow battery as a function of the state of charge obtained from UV-Vis spectroscopy. Energy Advances 2024, 3, 2597–2603

    2024

  52. [52]

    Application of the Pitzer model for describing the evaporation of seawater

    Keller, A.; Burger, J.; Hasse, H.; Kohns, M. Application of the Pitzer model for describing the evaporation of seawater. Desalination 2021, 503, 114866

    2021

  53. [53]

    Bousfield, W. R. Iso-piestic solutions. Transactions of the Faraday Society 1918, 13, 401

    1918

  54. [54]

    A.; Sinclair, D

    Robinson, R. A.; Sinclair, D. A. The Activity Coefficients of the Alkali Chlorides and of Lithium Iodide in Aqueous Solution from Vapor Pressure Measurements. Journal of the American Chemical Society 1934, 56, 1830–1835

    1934

  55. [55]

    Young, T. F. The Complete Calculation of Activity Coefficients from Freezing Point Data. Chemical Reviews 1933, 13, 103–110

    1933

  56. [56]

    Thermodynamik der Mischungen; Springer Berlin Heidelberg, 2017

    Heintz, A. Thermodynamik der Mischungen; Springer Berlin Heidelberg, 2017

    2017

  57. [57]

    Argersinger, W. J. Activity Coefficients of Electrolytes in Mixed Aqueous Solution from Electromotive Force Data. The Journal of Physical Chemistry 1954, 58, 792–795

    1954

  58. [58]

    Thomsen, K. Electrolyte Solutions: Thermodynamics, Crystallization, Separation methods; Technical University of Denmark, Department of Chemical and Biochemical Engineering , CERE – Center for Energy Ressources Engineering, 2009

    2009

  59. [59]

    S.; Mayorga, G

    Pitzer, K. S.; Mayorga, G. Thermodynamics of electrolytes. II. Activity and osmotic coefficients for strong electrolytes with one or both ions univalent. The Journal of Physical Chemistry 1973, 77, 2300–2308

    1973

  60. [60]

    Bromley, L. A. Thermodynamic properties of strong electrolytes in aqueous solutions. AIChE Journal 1973, 19, 313–320

    1973

  61. [61]

    A gE model for single and mixed solvent electrolyte systems

    Li, J.; Polka, H.-M.; Gmehling, J. A gE model for single and mixed solvent electrolyte systems. Fluid Phase Equilibria 1994, 94, 89–114

    1994

  62. [62]

    Modified LIQUAC and Modified LIFACA Further Development of Electrolyte Models for the Reliable Prediction of Phase Equilibria with Strong Electrolytes

    Kiepe, J.; Noll, O.; Gmehling, J. Modified LIQUAC and Modified LIFACA Further Development of Electrolyte Models for the Reliable Prediction of Phase Equilibria with Strong Electrolytes. Industrial & Engineering Chemistry Research 2006, 45, 2361–2373

    2006

  63. [63]

    A revised LIQUAC and LIFAC model (LIQUAC*/LIFAC*) for the prediction of properties of electrolyte containing solutions

    Mohs, A.; Gmehling, J. A revised LIQUAC and LIFAC model (LIQUAC*/LIFAC*) for the prediction of properties of electrolyte containing solutions. Fluid Phase Equilibria 2013, 337, 311–322

    2013

  64. [64]

    C.; Hughes, K

    Simoes, M. C.; Hughes, K. J.; Ingham, D. B.; Ma, L.; Pourkashanian, M. Estimation of the Pitzer Parameters for 1–1, 2–1, 3–1, 4–1, and 2–2 Single Electrolytes at 25 °C. Journal of Chemical & Engineering Data 2016, 61, 2536–2554

    2016

  65. [65]

    Perspective: Machine learning of thermophysical properties

    Jirasek, F.; Hasse, H. Perspective: Machine learning of thermophysical properties. Fluid Phase Equilibria 2021, 549, 113206

    2021

  66. [66]

    MLPROP – An Interactive Web Interface for Thermophysical Property Prediction with Machine Learning

    Hoffmann, M.; Specht, T.; Hayer, N.; Hasse, H.; Jirasek, F. MLPROP – An Interactive Web Interface for Thermophysical Property Prediction with Machine Learning. Chemie Ingenieur Technik 2025, 97, 1052–1056

    2025

  67. [67]

    Combining machine learning with physical knowledge in thermodynamic modeling of fluid mixtures

    Jirasek, F.; Hasse, H. Combining machine learning with physical knowledge in thermodynamic modeling of fluid mixtures. Annual Review of Chemical and Biomolecular Engineering 2023, 14, 31–51

    2023

  68. [68]

    GRAPPA—A hybrid graph neural network for predicting pure component vapor pressures

    Hoffmann, M.; Hasse, H.; Jirasek, F. GRAPPA—A hybrid graph neural network for predicting pure component vapor pressures. Chemical Engineering Journal Advances 2025, 22, 100750

    2025

  69. [69]

    HANNA: hard-constraint neural network for consistent activity coefficient prediction

    Specht, T.; Nagda, M.; Fellenz, S.; Mandt, S.; Hasse, H.; Jirasek, F. HANNA: hard-constraint neural network for consistent activity coefficient prediction. Chemical Science 2024, 15, 19777–19786

    2024

  70. [70]

    Thermodynamically consistent machine learning model for excess Gibbs energy

    Hoffmann, M.; Specht, T.; G\" o ttl, Q.; Burger, J.; Mandt, S.; Hasse, H.; Jirasek, F. Thermodynamically consistent machine learning model for excess Gibbs energy. Nature Communications 2026, 17

    2026

  71. [71]

    Artificial intelligence in thermodynamics: hybrid modeling of thermophysical properties of fluids

    Hasse, H.; Schmitt, S.; Jirasek, F. Artificial intelligence in thermodynamics: hybrid modeling of thermophysical properties of fluids. Current Opinion in Chemical Engineering 2026, 51, 101236

    2026

  72. [72]

    Jirasek, F.; Alves, R. A. S.; Damay, J.; Vandermeulen, R. A.; Bamler, R.; Bortz, M.; Mandt, S.; Kloft, M.; Hasse, H. Machine learning in thermodynamics: Prediction of activity coefficients by matrix completion. The Journal of Physical Chemistry Letters 2020, 11, 981–985

    2020

  73. [73]

    Prediction of Diffusion Coefficients in Mixtures with Tensor Completion

    Romero, Z.; M\" u nnemann, K.; Hasse, H.; Jirasek, F. Prediction of Diffusion Coefficients in Mixtures with Tensor Completion. 2026; https://arxiv.org/abs/2602.23142

    Pith/arXiv arXiv 2026

  74. [74]

    Hybridizing physical and data-driven prediction methods for physicochemical properties

    Jirasek, F.; Bamler, R.; Mandt, S. Hybridizing physical and data-driven prediction methods for physicochemical properties. Chemical Communications 2020, 56, 12407–12410

    2020

  75. [75]

    Predicting activity coefficients at infinite dilution for varying temperatures by matrix completion

    Damay, J.; Jirasek, F.; Kloft, M.; Bortz, M.; Hasse, H. Predicting activity coefficients at infinite dilution for varying temperatures by matrix completion. Industrial and Engineering Chemistry Research 2021, 60, 14564–14578

    2021

  76. [76]

    Hierarchical matrix completion for the prediction of properties of binary mixtures

    Gond, D.; Sohns, J.-T.; Leitte, H.; Hasse, H.; Jirasek, F. Hierarchical matrix completion for the prediction of properties of binary mixtures. Computers & Chemical Engineering 2025, 109122

    2025

  77. [77]

    Prediction of activity coefficients by similarity-based imputation using quantum-chemical descriptors

    Hayer, N.; Specht, T.; Arweiler, J.; Gond, D.; Hasse, H.; Jirasek, F. Prediction of activity coefficients by similarity-based imputation using quantum-chemical descriptors. Physical Chemistry Chemical Physics 2025, 27, 4307–4315

    2025

  78. [78]

    Balancing molecular information and empirical data in the prediction of physico-chemical properties

    Zenn, J.; Gond, D.; Jirasek, F.; Bamler, R. Balancing molecular information and empirical data in the prediction of physico-chemical properties. Digital Discovery 2025, 4, 683–693

    2025

  79. [79]

    Predicting temperature‐dependent activity coefficients at infinite dilution using tensor completion

    Damay, J.; Ryzhakov, G.; Jirasek, F.; Hasse, H.; Oseledets, I.; Bortz, M. Predicting temperature‐dependent activity coefficients at infinite dilution using tensor completion. Chemie Ingenieur Technik 2023, 95, 1061–1069

    2023

  80. [80]

    Prediction of Henry s law constants by matrix completion

    Hayer, N.; Jirasek, F.; Hasse, H. Prediction of Henry s law constants by matrix completion. AIChE Journal 2022, 68, e17753

    2022

Showing first 80 references.