Group Contribution Revisited: The Enthalpy of Formation of Organic Compounds with “Chemical Accuracy” Part II

Abstract

Group contribution (GC) methods to predict thermochemical properties are eminently important to process design. We present a group contribution parametrization for the heat of formation of organic molecules exhibiting chemical accuracy, maximum 1 kcal/mol (4.2 kJ/mol) difference between experiment and model values while minimizing the number of parameters avoiding overfitting and therewith avoiding reduced predictability. Compared to the contemporary literature, this was successfully achieved by employing available literature high-quality and consistent experimental data, optimizing parameters group by group, and introducing additional parameters when chemical understanding was obtained supporting these. A further important result is the observation that the applicability of the group contribution approach breaks down with increasing substitution levels, i.e., more heavily alkyl-substituted molecules, the reason being a serious influence of substitution on the conformation of the flexible part of the entire molecule within particular valence angles and torsional angles affected, which cannot be accounted for by additional GC parameters with fixed numerical values.

Combining Group-Contribution Concept and Graph Neural Networks Toward Interpretable Molecular Property Models

“…An example of such a descriptor is the functional group segmentation approach known as group-contribution methods (GC models) used in chemical engineering, where the molecule is represented through a feature vector of the occurrence of a set of predefined subfragments of the molecule and where each group has a fixed contribution to the property . GC models are especially attractive since some properties such as the enthalpy of formation abide by the additivity rules for e.f. homologous series, a fact well known for chemists. , Various rules have been developed to segment the molecules into their constituent groups, and some are more in line with how chemists define different entities that constitute a group (functional groups). , However, in their aim of developing more generalizable and accurate models, various new definitions of groups have been adopted with no general consensus on the rules and are largely based on heuristics and which led to the development of many GC models. , Some of the GC models include those developed by Lydersen (42 groups), Joback and Reid (41 groups), Tu, Benson and Buss, and Klincewicz and Reid, all differing in how the groups are defined. Among the most exhaustive group definitions are those developed by Hukkerikar et al, − which are extensions and revisions of the methods developed by Joback and Reid and Constantinou and Gani .…”

“…5,6 Various rules have been developed to segment the molecules into their constituent groups, and some are more in line with how chemists define different entities that constitute a group (functional groups). 5,6 However, in their aim of developing more generalizable and accurate models, various new definitions of groups have been adopted with no general consensus on the rules and are largely based on heuristics and which led to the development of many GC models. 5,6 Some of the GC models include those developed by Lydersen (42 groups), 7 Joback and Reid (41 groups), 8 Tu, 9 Benson and Buss, 10 and Klincewicz and Reid, 11 all differing in how the groups are defined.…”

“…5,6 However, in their aim of developing more generalizable and accurate models, various new definitions of groups have been adopted with no general consensus on the rules and are largely based on heuristics and which led to the development of many GC models. 5,6 Some of the GC models include those developed by Lydersen (42 groups), 7 Joback and Reid (41 groups), 8 Tu, 9 Benson and Buss, 10 and Klincewicz and Reid, 11 all differing in how the groups are defined. Among the most exhaustive group definitions are those developed by Hukkerikar et al, 12−14 which are extensions and revisions of the methods developed by Joback and Reid 8 and Constantinou and Gani.…”

“…4 GC models are especially attractive since some properties such as the enthalpy of formation abide by the additivity rules for e.f. homologous series, a fact well known for chemists. 5,6 Various rules have been developed to segment the molecules into their constituent groups, and some are more in line with how chemists define different entities that constitute a group (functional groups). 5,6 However, in their aim of developing more generalizable and accurate models, various new definitions of groups have been adopted with no general consensus on the rules and are largely based on heuristics and which led to the development of many GC models.…”

“…(5) The descriptors developed are intended for all-purpose applications which might result in redundant groups for some properties. (6) For some properties, the groups defined do not provide all the necessary and influential structural information, which is highlighted by low accuracy in the prediction of some properties such as the enthalpy of fusion and solubility parameters. 12,13 The recent surge in deep-learning algorithms has pushed the state-of-the-art models across many applications and especially in the field of molecular property prediction.…”