pith. sign in

arxiv: 2606.24983 · v1 · pith:GGJR4JQDnew · submitted 2026-06-23 · ⚛️ physics.chem-ph · cs.LG

ConSolv: Solvent-Conditional Machine Learning Implicit Solvent Potential

Linying Zhang , Julija Zavadlav This is my paper

Pith reviewed 2026-06-25 22:27 UTC · model grok-4.3

classification ⚛️ physics.chem-ph cs.LG
keywords implicit solventmachine learning potentialsolvation free energyorganic solventsattention mechanismtransferable modelmolecular simulation
0
0 comments X

The pith

A single attention-conditioned MLP predicts solvation free energies across 66 organic solvents using combined experimental and ab initio data.

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

The paper introduces ConSolv, an implicit solvent machine learning potential that conditions on solvent identity through an attention mechanism. It trains one set of weights on a mix of experimental solvation free energies and ab initio calculations so the same model works for dozens of non-aqueous solvents. A sympathetic reader cares because most prior ML implicit solvents were limited to water, yet organic synthesis, battery electrolytes, and many other applications require accurate treatment of other solvents. The model is shown to outperform both classical explicit-solvent calculations and selected ab initio implicit methods on benchmark solvation energies while also reproducing experimental NMR shifts for a test case in chloroform. The architecture is presented as readily extensible to new solvents and larger chemical spaces.

Core claim

ConSolv is a solvent-conditional MLP that inserts an attention-based solvent-embedding block to modulate solute interactions according to solvent identity. Trained on a combination of experimental solvation free energy data and ab initio calculations, a single set of network weights becomes transferable across 66 common organic solvents, outperforming classical explicit solvent methods and selected ab initio implicit solvent approaches on multiple solvation free energy benchmarks while generalizing to solvents absent from training. The same model also produces NMR chemical shifts for γ-fluorohydrin molecules in chloroform that agree closely with experiment.

What carries the argument

attention-based solvent-embedding block that explicitly incorporates solvent effects on solute interactions

If this is right

  • Solvation free energy predictions improve over both classical explicit-solvent simulations and selected ab initio implicit models on standard benchmarks.
  • The model produces solvation free energies for solvents not encountered during training without additional fitting.
  • NMR chemical shifts computed with the implicit solvent model match experimental values for γ-fluorohydrin in chloroform.
  • The architecture supports extension to larger chemical spaces or different training data mixtures while retaining a single weight set.

Where Pith is reading between the lines

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

  • The attention weights inside the solvent-embedding block could be inspected to rank which solvent properties most influence particular solute interactions.
  • Replacing the current training mixture with data from a narrower solvent class might improve accuracy inside that class at the cost of broader transferability.
  • Coupling the model to explicit solvent molecules only in the first solvation shell could test whether the implicit treatment already captures the dominant long-range effects.

Load-bearing premise

The attention-based solvent-embedding block captures the dominant solvent-dependent effects on solute interactions well enough that one set of weights stays accurate for both seen and unseen solvents without per-solvent retraining.

What would settle it

Measure solvation free energies for a solvent outside the original 66 and check whether the model's predictions deviate systematically from the new experimental values.

Figures

Figures reproduced from arXiv: 2606.24983 by Julija Zavadlav, Linying Zhang.

Figure 1
Figure 1. Figure 1: ConSolv architecture (a) and training (b). In a conventional message passing graph neural network [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Root mean squared error (RMSE) with respect to solvent using Solv@TUM test set (a) and solute [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: (a) Parity plot of solvation free energy predictions for 1-hexanol solute molecule across six solvents [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Generalization to solvents unseen during ConSolv training (o-xylene, acetonitrile, dimethyl sulfox [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Root mean squared error (RMSE, kcal/mol), Pearson correlation coefficient [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Per-atom attention weights (head3) for 1-hexanol solute molecule and solvent descriptors of methanol (top, ε = 32.61, polar) and 2,2,4-trimethylpentane (bottom, ε = 1.93, non-polar). On the other hand, when TMP is used as a solvent, the attention pattern shifts markedly. The descriptors that receive the highest attention weights are still concentrated on hydroxyl oxygen, but they are now MaxPartialCharge (… view at source ↗
read the original abstract

Implicit solvent machine learning potentials (MLPs) offer a powerful route to bridging the gap between accuracy and efficiency in molecular simulations. However, existing models have largely focused on aqueous environments, overlooking the diverse and important roles of non-aqueous solvents in areas such as organic synthesis and battery technology. Here, we present ConSolv, a solvent-conditional MLP architecture that explicitly incorporates solvent effects on solute interactions through an attention-based solvent-embedding block. By combining experimental solvation free energy data with ab initio data, we train a single implicit solvent MLP that is transferable across 66 common organic solvents. ConSolv outperforms classical explicit solvent methods and selected ab initio implicit solvent approaches across multiple solvation free energy benchmarks, and demonstrates generalization to unseen solvents. Beyond solvation free energies, the model shows close agreement with experimental nuclear magnetic resonance (NMR) data for $\gamma$-fluorohydrin molecules in chloroform. ConSolv's architecture is readily extensible to broader chemical spaces and alternative training strategies, while its attention-based design supports explainable artificial intelligence (AI) analysis that can help elucidate complex, solvent-dependent molecular interactions.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The manuscript introduces ConSolv, a solvent-conditional MLP architecture featuring an attention-based solvent-embedding block. Trained on a mixture of experimental solvation free energy data and ab initio calculations, the model is presented as transferable across 66 common organic solvents, outperforming classical explicit-solvent methods and selected ab initio implicit-solvent approaches on multiple solvation free energy benchmarks while also generalizing to unseen solvents and reproducing experimental NMR shifts for γ-fluorohydrin in chloroform.

Significance. If the reported transferability and benchmark outperformance are substantiated, the work would meaningfully extend machine-learned implicit solvent models beyond water to a chemically diverse set of organic solvents relevant to synthesis and energy storage, while the attention mechanism provides a route to interpretable solvent-dependent interactions.

major comments (1)
  1. The central claims of outperformance and generalization to unseen solvents rest on benchmark comparisons whose construction, data splits, error bars, and statistical controls are not visible in the provided text; without these, it is impossible to determine whether the reported gains survive proper controls or reduce to differences in training data composition.
minor comments (2)
  1. [Abstract] The abstract states that the model 'outperforms classical explicit solvent methods' but does not name the specific methods, force fields, or quantitative metrics (e.g., MAE or RMSE values) used for comparison.
  2. [Abstract] No information is given on the dimensionality of the solvent embedding, the attention mechanism implementation, or how solvent identity is encoded as input to the MLP.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for highlighting the need for greater transparency in our benchmark methodology. We will revise the manuscript to include the requested details on data construction, splits, error bars, and controls, ensuring the claims can be properly evaluated.

read point-by-point responses

  1. Referee: The central claims of outperformance and generalization to unseen solvents rest on benchmark comparisons whose construction, data splits, error bars, and statistical controls are not visible in the provided text; without these, it is impossible to determine whether the reported gains survive proper controls or reduce to differences in training data composition.

    Authors: We agree that the current text lacks sufficient methodological detail on the benchmarks. In revision we will add an expanded Methods section (and supplementary tables) that explicitly describes: (i) the full composition and sources of the training set (experimental solvation free energies vs. ab initio calculations, with counts per solvent), (ii) the train/validation/test splits, including the precise protocol used to designate 'unseen' solvents, (iii) how error bars were computed (standard deviation across random seeds or bootstrap resampling), and (iv) the statistical tests applied to compare ConSolv against baselines. We will also clarify that all reported test-set molecules and solvents were strictly held out from training. These additions will allow readers to assess whether performance differences arise from the architecture or from data composition. revision: yes

Circularity Check

0 steps flagged

No significant circularity; model training and generalization claims are independent of test benchmarks

full rationale

The paper trains a solvent-conditional MLP on a combination of experimental solvation free energy data and ab initio calculations, then evaluates transferability on held-out solvents and external NMR benchmarks. No equations, self-citations, or uniqueness theorems are invoked that would make reported predictions equivalent to the training inputs by construction. The attention-based embedding and mixed-data training strategy constitute an architectural choice whose performance is assessed against independent observables rather than being tautological.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review provides no equations or methods section, so free parameters, axioms, and invented entities cannot be enumerated from the supplied text.

pith-pipeline@v0.9.1-grok · 5725 in / 1095 out tokens · 18901 ms · 2026-06-25T22:27:07.114760+00:00 · methodology

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

300 extracted references · 2 canonical work pages

  1. [1]

    Uncertainty Quantification for Molecular Models via Stochastic Gradient MCMC , pages =

    Thaler, Stephan and Zavadlav, Julija , booktitle =. Uncertainty Quantification for Molecular Models via Stochastic Gradient MCMC , pages =

  2. [2]

    ACS Nano , year=

    Tuning the Dielectric Response of Water in Nanoconfinement through Surface Wettability , author=. ACS Nano , year=

  3. [3]

    Zavadlav, Julija and Arampatzis, Georgios and Koumoutsakos, Petros , title =. Sci. Rep. , volume =. 2019 , doi =

    2019

  4. [4]

    Open-boundary molecular dynamics of a DNA molecule in a hybrid explicit/implicit salt solution , author=. Biophys. J. , volume=. 2018 , publisher=

    2018

  5. [5]

    Multiscale simulation of protein hydration using the SWINGER dynamical clustering algorithm , author=. J. Chem. Theory Comput. , volume=. 2018 , publisher=

    2018

  6. [6]

    Molecular dynamics simulation of high density DNA arrays , author=. Comput. , volume=. 2018 , publisher=

    2018

  7. [7]

    Adaptive resolution simulations coupling atomistic water to dissipative particle dynamics , author=. J. Chem. Phys. , volume=. 2017 , publisher=

    2017

  8. [8]

    Order and interactions in DNA arrays: Multiscale molecular dynamics simulation , author=. Sci. Rep. , volume=. 2017 , publisher=

    2017

  9. [9]

    Adaptive resolution simulation of an atomistic DNA molecule in MARTINI salt solution , author=. Eur. Phys. J. Spec. Top. , volume=. 2016 , publisher=

    2016

  10. [10]

    Adaptive resolution simulation of a DNA molecule in salt solution , author=. J. Chem. Theory Comput. , volume=. 2015 , publisher=

    2015

  11. [11]

    Interface Focus , volume=

    SWINGER: a clustering algorithm for concurrent coupling of atomistic and supramolecular liquids , author=. Interface Focus , volume=. 2019 , publisher=

    2019

  12. [12]

    Adaptive resolution simulation of supramolecular water: the concurrent making, breaking, and remaking of water bundles , author=. J. Chem. Theory Comput. , volume=. 2016 , publisher=

    2016

  13. [13]

    Adaptive resolution simulation of polarizable supramolecular coarse-grained water models , author=. J. Chem. Phys. , volume=. 2015 , publisher=

    2015

  14. [14]

    Adaptive resolution simulation of an atomistic protein in MARTINI water , author=. J. Chem. Phys. , volume=. 2014 , publisher=

    2014

  15. [15]

    Multiscale simulation of soft matter: From scale bridging to adaptive resolution , author=. Annu. Rev. Phys. Chem. , volume=. 2008 , publisher=

    2008

  16. [16]

    Discovery of self-assembling -conjugated peptides by active learning-directed coarse-grained molecular simulation , author=. J. Phys. Chem. B , volume=. 2020 , publisher=

    2020

  17. [17]

    A guide to supramolecular polymerizations , author=. Polym. Chem. , volume=. 2020 , publisher=

    2020

  18. [18]

    Calculation of effective interaction potentials from radial distribution functions: A reverse Monte Carlo approach , author=. Phys. Rev. E , volume=. 1995 , publisher=

    1995

  19. [19]

    Deriving effective mesoscale potentials from atomistic simulations , author=. J. Comput. Chem. , volume=. 2003 , publisher=

    2003

  20. [20]

    Faraday Discuss

    Systematic coarse-graining of molecular models by the Newton inversion method , author=. Faraday Discuss. , volume=. 2010 , publisher=

    2010

  21. [21]

    EPL , volume=

    Interatomic potentials from first-principles calculations: the force-matching method , author=. EPL , volume=. 1994 , publisher=

    1994

  22. [22]

    A multiscale coarse-graining method for biomolecular systems , author=. J. Phys. Chem. B , volume=. 2005 , publisher=

    2005

  23. [23]

    Fitting coarse-grained distribution functions through an iterative force-matching method , author=. J. Chem. Phys. , volume=. 2013 , publisher=

    2013

  24. [24]

    The relative entropy is fundamental to multiscale and inverse thermodynamic problems , author=. J. Chem. Phys. , volume=. 2008 , publisher=

    2008

  25. [25]

    Relative entropy as a universal metric for multiscale errors , author=. Phys. Rev. E , volume=. 2010 , publisher=

    2010

  26. [26]

    Top-down Multiscale Approach To Simulate Peptide Self-Assembly from Monomers , author=. J. Chem. Theory Comput. , volume=. 2019 , publisher=

    2019

  27. [27]

    Polymers , volume=

    Mesoscale Simulations of Polymer Solution Self-Assembly: Selection of Model Parameters within an Implicit Solvent Approximation , author=. Polymers , volume=. 2021 , publisher=

    2021

  28. [28]

    Evaluating Classical Force Fields against Experimental Cross-Solvation Free Energies , author=. J. Chem. Theory Comput. , volume=. 2020 , publisher=

    2020

  29. [29]

    Combining molecular dynamics and machine learning to predict self-solvation free energies and limiting activity coefficients , author=. J. Chem. Inf. Model. , volume=. 2020 , publisher=

    2020

  30. [30]

    Validation of the generalized force fields GAFF, CGenFF, OPLS-AA, and PRODRGFF by testing against experimental osmotic coefficient data for small drug-like molecules , author=. J. Chem. Inf. Model. , volume=. 2019 , publisher=

    2019

  31. [31]

    A biomolecular force field based on the free enthalpy of hydration and solvation: the GROMOS force-field parameter sets 53A5 and 53A6 , author=. J. Comput. Chem. , volume=. 2004 , publisher=

    2004

  32. [32]

    Developing force fields when experimental data is sparse: AMBER/GAFF-compatible parameters for inorganic and alkyl oxoanions , author=. Phys. Chem. Chem. Phys. , volume=. 2017 , publisher=

    2017

  33. [33]

    Optimizing solute--water van der waals interactions to reproduce solvation free energies , author=. J. Phys. Chem. B , volume=. 2012 , publisher=

    2012

  34. [34]

    FreeSolv: a database of experimental and calculated hydration free energies, with input files , author=. J. Comput. Aided Mol. Des. , volume=. 2014 , publisher=

    2014

  35. [35]

    Absolute hydration free energies of blocked amino acids: implications for protein solvation and stability , author=. Biophys. J. , volume=. 2013 , publisher=

    2013

  36. [36]

    Fluid Phase Equilib

    The correlation and prediction of infinite dilution activity coefficients of compounds in water at 298.15 K , author=. Fluid Phase Equilib. , volume=. 2017 , publisher=

    2017

  37. [37]

    Estimation of solvation quantities from experimental thermodynamic data: Development of the comprehensive compSol databank for pure and mixed solutes , author=. J. Phys. Chem. Ref. Data , volume=. 2017 , publisher=

    2017

  38. [38]

    Solvation thermodynamics of organic molecules by the molecular integral equation theory: approaching chemical accuracy , author=. Chem. Rev. , volume=. 2015 , publisher=

    2015

  39. [39]

    Infinite dilution activity coefficients as constraints for force field parametrization and method development , author=. J. Chem. Theory Comput. , volume=. 2019 , publisher=

    2019

  40. [40]

    SAMPL4, a blind challenge for computational solvation free energies: the compounds considered , author=. J. Comput. Aided Mol. Des. , volume=. 2014 , publisher=

    2014

  41. [41]

    Model performances evaluated for infinite dilution activity coefficients prediction at 298.15 K , author=. Ind. Eng. Chem. Res. , volume=. 2019 , publisher=

    2019

  42. [42]

    Approaches for calculating solvation free energies and enthalpies demonstrated with an update of the FreeSolv database , author=. J. Chem. Eng. Data , volume=. 2017 , publisher=

    2017

  43. [43]

    Extremely precise free energy calculations of amino acid side chain analogs: Comparison of common molecular mechanics force fields for proteins , author=. J. Chem. Phys. , volume=. 2003 , publisher=

    2003

  44. [44]

    Structure and dynamics of the homologous series of alanine peptides: a joint molecular dynamics/NMR study , author=. J. Am. Chem. Soc. , volume=. 2007 , publisher=

    2007

  45. [45]

    PhysNet: a neural network for predicting energies, forces, dipole moments, and partial charges , author=. J. Chem. Theory Comput. , volume=. 2019 , publisher=

    2019

  46. [46]

    Predicting small-molecule solvation free energies: an informal blind test for computational chemistry , author=. J. Med. Chem. , volume=. 2008 , publisher=

    2008

  47. [47]

    Uncovering Differences in Hydration Free Energies and Structures for Model Compound Mimics of Charged Side Chains of Amino Acids , author=. J. Phys. Chem. B , volume=. 2021 , publisher=

    2021

  48. [48]

    The effect of increasing dipolar distance on the activities of aliphatic amino acids in aqueous solution at twenty-five degrees , author=

    Thermodynamic properties of solutions of amino acids and related substances IV. The effect of increasing dipolar distance on the activities of aliphatic amino acids in aqueous solution at twenty-five degrees , author=. J. Biol. Chem. , volume=. 1940 , publisher=

    1940

  49. [49]

    Measuring and modeling activity coefficients in aqueous amino-acid solutions , author=. Ind. Eng. Chem. Res. , volume=. 2011 , publisher=

    2011

  50. [50]

    Osmotic pressure simulations of amino acids and peptides highlight potential routes to protein force field parameterization , author=. J. Phys. Chem. B , volume=. 2016 , publisher=

    2016

  51. [51]

    PLoS Comput

    Are current atomistic force fields accurate enough to study proteins in crowded environments? , author=. PLoS Comput. Biol. , volume=. 2014 , publisher=

    2014

  52. [52]

    Molecular dynamics simulations of intrinsically disordered proteins: force field evaluation and comparison with experiment , author=. J. Chem. Theory Comput. , volume=. 2015 , publisher=

    2015

  53. [53]

    Improved parameterization of amine--carboxylate and amine--phosphate interactions for molecular dynamics simulations using the CHARMM and AMBER force fields , author=. J. Chem. Theory Comput. , volume=. 2016 , publisher=

    2016

  54. [54]

    Molecular dynamics simulations of 441 two-residue peptides in aqueous solution: Conformational preferences and neighboring residue effects with the Amber ff99SB-ildn-NMR force field , author=. J. Chem. Theory Comput. , volume=. 2015 , publisher=

    2015

  55. [55]

    Are current molecular dynamics force fields too helical? , author=. Biophys. J. , volume=. 2008 , publisher=

    2008

  56. [56]

    CHARMM36m: an improved force field for folded and intrinsically disordered proteins , author=. Nat. Methods , volume=. 2017 , publisher=

    2017

  57. [57]

    Optimization of the additive CHARMM all-atom protein force field targeting improved sampling of the backbone , and side-chain 1 and 2 dihedral angles , author=. J. Chem. Theory Comput. , volume=. 2012 , publisher=

    2012

  58. [58]

    ff19SB: Amino-acid-specific protein backbone parameters trained against quantum mechanics energy surfaces in solution , author=. J. Chem. Theory Comput. , volume=. 2019 , publisher=

    2019

  59. [59]

    Residue-specific force field improving the sample of intrinsically disordered proteins and folded proteins , author=. J. Chem. Inf. Model. , volume=. 2019 , publisher=

    2019

  60. [60]

    Machine learning of solvent effects on molecular spectra and reactions , author=. Chem. Sci. , volume=. 2021 , publisher=

    2021

  61. [61]

    Bioinformatics , volume=

    Generalized Born radii computation using linear models and neural networks , author=. Bioinformatics , volume=. 2020 , publisher=

    2020

  62. [62]

    Machine Learning Implicit Solvation for Molecular Dynamics , author=. J. Chem. Phys. , volume=

  63. [63]

    Machine learning-guided approach for studying solvation environments , author=. J. Chem. Theory Comput. , volume=. 2019 , publisher=

    2019

  64. [64]

    Variational optimization of an all-atom implicit solvent force field to match explicit solvent simulation data , author=. J. Chem. Theory Comput. , volume=. 2013 , publisher=

    2013

  65. [65]

    Machine learning antimicrobial peptide sequences: Some surprising variations on the theme of amphiphilic assembly , author=. Curr. Opin. Colloid Interface Sci. , volume=. 2018 , publisher=

    2018

  66. [66]

    Prediction of amphiphilic cell-penetrating peptide building blocks from protein-derived amino acid sequences for engineering of drug delivery nanoassemblies , author=. J. Phys. Chem. B , volume=. 2020 , publisher=

    2020

  67. [67]

    https://doi.org/10.21203/rs.3.rs-505801/v1 , year=

    Self-assembling Peptide Discovery: Overcoming Human Bias With Machine Learning , author=. https://doi.org/10.21203/rs.3.rs-505801/v1 , year=

    work page doi:10.21203/rs.3.rs-505801/v1

  68. [68]

    Beyond Tripeptides Two-Step Active Machine Learning for Very Large Data sets , author=. J. Chem. Theory Comput. , volume=. 2021 , publisher=

    2021

  69. [69]

    Biomacromolecules , volume=

    A comprehensive study on self-assembly and gelation of C13-dipeptides—from design strategies to functionalities , author=. Biomacromolecules , volume=. 2019 , publisher=

    2019

  70. [70]

    Molecular co-assembly as a strategy for synergistic improvement of the mechanical properties of hydrogels , author=. Chem. Commun. , volume=. 2017 , publisher=

    2017

  71. [71]

    Biomacromolecules , volume=

    Coassembly of C13-Dipeptides: Gelations from Solutions and Precipitations , author=. Biomacromolecules , volume=. 2020 , publisher=

    2020

  72. [72]

    Nanoscale , volume=

    The effect of self-sorting and co-assembly on the mechanical properties of low molecular weight hydrogels , author=. Nanoscale , volume=. 2014 , publisher=

    2014

  73. [73]

    Nanoscale , volume=

    Expanding the structural diversity of peptide assemblies by coassembling dipeptides with diphenylalanine , author=. Nanoscale , volume=. 2020 , publisher=

    2020

  74. [74]

    Virtual screening for dipeptide aggregation: Toward predictive tools for peptide self-assembly , author=. J. Phys. Chem. Lett. , volume=. 2011 , publisher=

    2011

  75. [75]

    Tunable morphology and functionality of multicomponent self-assembly: A review , author=. Mater. Des. , volume=. 2021 , publisher=

    2021

  76. [76]

    Biomacromolecules , volume=

    Self-assembling hydrogels based on a complementary host--guest peptide amphiphile pair , author=. Biomacromolecules , volume=. 2019 , publisher=

    2019

  77. [77]

    How should multicomponent supramolecular gels be characterised? , author=. Chem. Soc. Rev. , volume=. 2018 , publisher=

    2018

  78. [78]

    Multicomponent self-assembly as a tool to harness new properties from peptides and proteins in material design , author=. Chem. Soc. Rev. , volume=. 2018 , publisher=

    2018

  79. [79]

    Theranostics , volume=

    Peptide-modulated self-assembly as a versatile strategy for tumor supramolecular nanotheranostics , author=. Theranostics , volume=. 2019 , publisher=

    2019

  80. [80]

    Computational prediction of tripeptide-dipeptide co-assembly , author=. Mol. Phys. , volume=. 2019 , publisher=

    2019

Showing first 80 references.