Automated Tariff Classification of Organic Chemicals Using Molecular Fingerprints and Interpretable Structural Features

Tariff classification of chemical substances has traditionally relied on manual expert review, due to the need for in-depth chemistry knowledge and interpretation of complex rules. The volume of classification work is large and growing, and expert time is the limiting resource. This work investigates whether a machine learning classifier based on molecular structure can predict eight-digit tariff codes for organic chemicals, i.e., Chapter 29 of the Combined Nomenclature (CN). Labelled data for over 70,000 chemicals was downloaded from the European Customs Inventory of Chemical Substances (ECICS). Morgan fingerprints combined with other structural features were used to train three different machine learning algorithms. The best performance was achieved with the XGBoost algorithm: top-1 accuracy on a held-out 20% test set was 97.0% for the four-digit heading and 94.3% for the eight-digit CN code. For top-3 accuracy, the figures increased to 99.6% and 99.0%, respectively. For the eight-digit classifier, these figures do not include codes with fewer than 50 substances, for which a separate approach was used. The combined end-to-end performance was a top-1 accuracy of 91.2% and a top-3 accuracy of 97.9%. End-to-end performance on a temporal test set (not available at any time during model development) was similar: a top-1 accuracy of 90.3% and a top-3 accuracy of 97.3%.