For microplastic (MP) spectral classification, models built using convolutional neural networks (CNNs) have been reported to be more accurate and faster than library searches due to CNNs' superior feature extraction capabilities and the elimination of the need to compare each spectrum with all reference data. However, the CNN models reported in the currently available literature were trained on limited spectral databases that might lack representativeness, i.e., insufficient diversity in plastic types, plastic aging stages, secondary materials (e.g., additives, pigments) in the MPs, and environmental contexts where the MPs are extracted, and etc. Consequently, the robustness of these models may diminish when used by other MP researchers. To benefit a wider MP research community, we constructed a comprehensive dataset that includes self-collected and validated infrared (IR) spectra as well as IR spectra gathered from various other representative sources (e.g., Open Specy), and based on this extensive dataset we trained eight CNN models: one-dimensional and two-dimensional LeNet, AlexNet, VGG16, and ResNet34. The dataset we established covers both virgin (micro)plastics and (micro)plastics from the environment. These (micro)plastics were placed on various substrates for spectrum collection and have different thicknesses, sizes, degrees of aging, and cleanliness, and contain different secondary materials of different amounts. Before training, the dataset was balanced using data augmentation, resulting in a final dataset of 19,000+ spectra (covering 19 plastic types), with 80 Also see: https://micro2024.sciencesconf.org/559538/document

In MICRO 2024: Plastic Pollution from MACRO to nano