Prioritizing Data Quality in Machine Learning for Thermophysical Property Prediction: A Case Study on Normal Boiling...

Accurate prediction of the normal boiling point (NBP) of organic compounds is essential for the design and optimization of chemical processes. This is especially true for modern practice since experimental measurements of many compounds of contemporary importance are difficult to perform due to cost, purity, and stability issues. Traditional methods, such as group contribution (GC) and quantitative structure-property relationship (QSPR) approaches, are limited by inflexibility and diminishing accuracy in these situations. Machine learning (ML) approaches have shown promise but are plagued by two weaknesses for the nonspecialist: ease of use (especially compared to GC and QSPR) and proven accuracy for compounds outside the training set. The large effort required to use an ML model of others, let alone develop a new model, is the source of the first problem. The second often arises due to the data-intensive nature of ML. In particular, the origin and reliability of the data used to train a model has to date been of secondary importance compared to finding the optimal ML techniques to use for a given task. This study addresses both these issues. First, we explore the effectiveness of ML models trained on two data sets: a smaller, rigorously curated experimental set from the DIPPR 801 database and a larger, publicly available but uncurated data set. Four distinct molecular featurization methods were evaluated, including RDKit descriptors, Joback groups, and two graph-based representations using graph neural networks (GNNs). Models were implemented using LightGBM and PyTorch, and performance was assessed through a 10-fold cross-validation across four evaluation frameworks. Despite conventional expectations that larger data sets yield superior models, results consistently showed that the curated data set of fewer but certifiably accurate values outperformed the uncurated one in terms of accuracy, bias reduction, and generalization. These findings emphasize the importance of data quality over quantity and advocate for a data-centric approach in ML applications for chemical property prediction. A user-friendly web application is also provided to facilitate access to the best-trained model for the nonspecialist, where all that is needed is a SMILES formula (a single and simple text string that is widely available) for the molecule.