Comment on Predicting Aqueous Adsorption of Organic Compounds onto Biochars, Carbon Nanotubes, Granular Activated Carbons, And Resins with Machine Learning.

Zhang et al.[1] published a paper on machine learning based predictions of organic contaminant sorption onto carbonaceous materials and resins. The authors provide a novel approach to predict concentration-dependent sorption distribution coefficients (KD) to these materials, without the need to link it to any specific isotherm model. This study is a valuable contribution to the field that can stimulate the scientific discussion in the adsorption-modeling community regarding (i) mechanistic assumptions prior to model building, (ii) the parametrization of the model based on these assumptions, (iii) the grouping of data to train the algorithm, and (iv) data filtering strategies. We recently published a paper on a similar topic[2] and are confident that this discussion is valuable to improve the future applicability of machine learning techniques to sorption phenomena.

Zhang et al. used the BET specific surface area and total pore volume to describe the sorbent materials and state that “these two parameters are critical for deciding the adsorption of organic compounds through hydrophobic interactions and pore-filling, two key mechanisms for organic compounds to be adsorbed by various adsorbents.” These processes are of general importance. However, it is well accepted for carbonaceous sorbents that π–π electron donor–acceptor interactions are a key mechanism for the sorption of organic compounds.[3,4] These interactions can be related to the polarizability of the compound[5] as well as the aromaticity of the sorbent materials, which can be approximated by the broadly available molar H/C ratio of the materials elemental composition.[6]

The correct parametrization of the dominant sorption processes is crucial. Zhang et al. built their model on two highly correlated sorbent parameters, i.e., the BET specific surface area and total pore volume, both determined from N2 physisorption.[7,8] Hydrophobic interactions cannot be assigned directly to the BET specific surface area; instead, if hydrophobicity is a key driver for sorption of organic compounds, it would be important to include a sorbate hydrophobicity parameter such as log Dow or the hexadecane water partitioning coefficient (“L”), which is widely used in ppLFER models including models for carbonaceous materials.[5,9] Zhang et al. suggest the use of the McGowan-Volume as a hydrophobicity proxy. While it is true that hydrophobicity tends to increase with molecular size, other aspects, such as the polarity of a compound are not accounted for with the McGowan-Volume.

Zhang et al. subdivided their data sets into four categories (i.e., biochar, carbon nanotubes, granular activated carbon, and resins). Based on our recent study[2] sorption mechanism to the various carbonaceous sorbents is not fundamentally different and, when sorbent material properties are well parametrized, these data can be combined. In the case of Zhang et al. this would result in only two categories, i.e., carbonaceous materials and resins. Thereby the machine learning algorithm could be trained on a larger data set for carbonaceous materials, which might improve its generalization and forecasting capabilities.

With deterministic data filtering techniques such as cosine similarity used by Zhang et al. the data may be systematically filtered to a level that is not necessarily representative of the original trend. Random-based statistical techniques such as low entropy data removal or significance test measures may be better choices for data filtering. Since random allocation of data into different sets for training, validation, and testing would result in various goodness of fit, we suggest multitraining as a technique to increase model generalization. Thereby, even a poor set can contribute to the model predictability and performance.

Consideration of the above aspects will further improve the applicability of machine learning algorithms for studying contaminant dynamics.