Combination of explainable machine learning and conceptual density functional theory: applications for the study of key solvation mechanisms

Abstract

We present explainable machine learning approaches for the accurate prediction of solvation free energies, enthalpies, and entropies for different ion pairs in various protic and aprotic solvents. As key input...

SolvBERT for solvation free energy and solubility prediction: a demonstration of an NLP model for predicting the properties of molecular complexes

“…31,35 Using machine learning and deep learning to predict the relevant properties of solvation can save the expensive time and cost of performing experiments or computations, and also help to nd important features that contribute to the solvation properties. 36 Recent studies by Vermeire et al and Zhang et al both pre-trained graph-based neural networks on computational solvation free energy datasets and then used smaller experimental solvation free energy datasets for transfer learning. A potential drawback of using supervised learning in the pre-training phase is that it may lead to negative transfer (i.e., result in undesirable degradation of model performance) when transferring knowledge from pre-training on one attribute (e.g., solvation free energy) to the prediction of another attribute (e.g., solubility).…”

Insights into Hildebrand Solubility Parameters – Contributions from Cohesive Energies or Electrophilicity Densities?**

“…In previous publications, it was shown that decision-tree based methods usually provide the most accurate results for small datasets. [75] The source code was written in Python 3.9.1 [76] in combination [20,83] in combination with HOMO (E HOMO ) and LUMO energies (E LUMO ) as well as molecular volumes V m from the GEPOL method [70] for different solvents with D < 5 Debye reflecting poor protic behavior (0) in accordance with Reference [83]. with the modules NumPy 1.19.5, [77] Scikit-earn 1.0.1 [78] and Pandas 1.2.1.…”

“…As most accurate method for predictions, we identified the Extra Trees Method [74] with a number of estimators of 100 and a minimum sample split of 2. In previous publications, it was shown that decision‐tree based methods usually provide the most accurate results for small datasets [75] . The source code was written in Python 3.9.1 [76] in combination with the modules NumPy 1.19.5, [77] Scikit‐earn 1.0.1 [78] and Pandas 1.2.1 [79]…”

Scientific Deep Machine Learning Concepts for the Prediction of Concentration Profiles and Chemical Reaction Kinetics: Consideration of Reaction Conditions

“…In addition to scaling approaches, novel concepts from machine learning (ML) also attracted emerging attention in molecular sciences and chemical reaction design . As few examples, ML approaches are used for the prediction of novel molecular compounds, − the design and study of chemical reactions, ,,− and the computation of the corresponding activation energies and thermodynamic properties. ,− However, it has to be noted that most of these approaches focus on the thermodynamic and molecular properties of reacting compounds, while the kinetic aspects of the reactions remain rather unexplored . A recent perspective article addresses the use of ML-based methods for the prediction of chemical reaction profiles and kinetics .…”