All models available on MLPROP are developed by the Laboratory of Engineering Thermodynamics (LTD) at RPTU Kaiserslautern and require the molecular structure in the form of a SMILES as input. Below is a list of the currently available models on MLPROP with a brief description and a link to the corresponding publication.

HANNA
Hard-constraint Neural Network for Consistent Activity Coefficient Prediction

  • Model for the excess Gibbs energy \(g^\mathrm{E}\)
  • Strictly thermodynamically consistent
  • New version trained to VLE and LLE data
  • New version applicable to multi-component mixtures
  • Published in Nature Communications
image

GRAPPA
Graph Neural Network for Predicting the Parameters of the Antoine Equation

  • Vapor pressure and boiling point prediction
  • Based on the Antoine equation:
    • \(\ln(p^{\mathrm{s}} / \mathrm{kPa}) = A - \frac{B}{T / \mathrm{K} + C}\)
  • Paper in Chemical Engineering Journal Advances
image

UNIFAC 2.0 and mod. UNIFAC 2.0

  • Enhanced classical group-contribution methods
  • Missing parameters predicted with matrix completion
  • Higher prediction accuracy than the original methods
  • UNIFAC 2.0 paper, mod. UNIFAC 2.0 paper
image

COGNAC

  • Graph neural network that writes the complete mod. UNIFAC (Dortmund) parameter table for the queried molecules: \(R_k\), \(Q_k\) and the interaction polynomial
    • \(\Psi_{nm} = \exp\left(-\frac{a_{nm} + b_{nm} T + c_{nm} T^2}{T}\right)\)
  • Every atom is its own group, so the parameters are specific to the molecules asked for instead of being drawn from a fixed group alphabet
  • COGNAC evaluates the closed-form mod. UNIFAC group sum and can therefore be exported to a process simulator; the variant COGNAC (implicit) differs only in the residual closure, which solves the COSMOSPACE self-consistency condition instead
  • Also available combined with GRAPPA for VLE predictions and as a source for fitted NRTL parameters
  • Pre-print coming soon

ESE - Enhanced Stokes-Einstein

  • Prediction of binary diffusion coefficients at infinite dilution
  • Hybrid model incorporating the Stokes–Einstein equation
  • Physically consistent temperature dependence
  • Pre-print available on arXiv
image

HADES - Hybrid Architecture for Consistent Diffusion Coefficient Estimation Across Solvents

  • Composition-dependent self-diffusion coefficients in mixtures of any number of components
  • Enforces physically consistent temperature and composition dependence by construction
  • Requires only pure-component viscosities by assuming ideal mixing in the application of the Stokes–Einstein equation
  • Pre-print coming soon
image

EVE - Enhanced Vignes Equation

  • Composition-dependent mutual diffusion coefficients in binary liquid mixtures
  • Physically consistent composition dependence by construction
  • Preprint coming soon
image

Prediction of Binary and Ternary Phase Equilibria using the CEM

  • Prediction of binary and ternary phase diagrams
  • Simultaneous calculation of VLE and LLE
  • Activity coefficients by \(g^\mathrm{E}\) model of choice
  • Vapor pressure calculated by GRAPPA
  • Applicable to multi-component mixtures
  • Based on the convex envelope method (CEM)
image

Prediction of Vapor-Liquid Equilibria

  • Prediction of isothermal and isobaric binary VLE
  • Vapor pressure calculated by GRAPPA
  • Activity coefficients by \(g^\mathrm{E}\) model of choice
  • Based on extended Raoult's law:
    • \(p_i^{\mathrm{s}}\,x_i\,\gamma_i= p\, y_i\)
image