Bitumen optical microscopy as a quality assurance tool through deployment of trained Siamese neural network model

Complexity of bitumen chemistry creates challenges in ensuring consistent bitumen quality and integrity during the selection, design, and construction phases of a project. Traditional methods for bitumen fingerprinting are often time-consuming, expensive, and require specialised equipment or proficiency, making them less suitable for frequent use in field setups. The present study focuses on developing a machine learning-based tool that utilises optical microscopy for imaging bitumen samples to verify their source. In this research, microscopic images of 31 bitumen samples from 12 producers were captured and used to train a deep learning model. The Siamese neural network model, trained on images taken on unaged bitumen samples at 100× magnification, achieved good performance metrics regarding same source identification and distinguishing among disparate sources. Specifically, the model demonstrated an accuracy of 93.17% with F1 scores of 73.08% for unaged samples, indicating that the bitumen source(s) can be viably identified through the model. On the other hand, when short-term aged 100× images were used for testing the model, the results did not show good correspondence with any (unaged) sources. The lack of similarity shows that major physicochemical change and corresponding change in microstructure can be identified through the machine learning model. It should be noted that the actual modification of bitumen and intrinsic chemical changes are not known beforehand and there is possibility of multiple refineries using the same crude source, while variability might be imparted due to differing processes. It can be concluded that the machine learning model, based on bitumen optical microscopy, has the potential for being used as a rapid source verification utility.