# Fileset [rm5c00040_si_001.pdf](https://mdr.nims.go.jp/filesets/e55df56c-ec9b-4f66-b717-f05f3594259b/download) ## Creator [Yusuke Hibi](https://orcid.org/0000-0003-4006-1070), Shiho Uesaka, Kiyotaka Hitomi, Ken-ichi Niihara, Asami Imai, [Sadaki Samitsu](https://orcid.org/0000-0002-4139-1656), [Masanobu Naito](https://orcid.org/0000-0001-7198-819X) ## Rights [Creative Commons BY Attribution 4.0 International](https://creativecommons.org/licenses/by/4.0/) ## Other metadata [A Nearly Zero-Cost Lot-by-Lot Inspection of Recycled Plastics: Prediction of Mechanical Properties from Viscosity Evolution during Melt Kneading](https://mdr.nims.go.jp/datasets/543df61c-d036-45b7-91ee-4c771da74eca) ## Fulltext 1 Support Information for A Nearly Zero-cost Lot-by-lot Inspection of Recycled Plastics: Prediction of Mechanical Properties from Viscosity Evolution during Melt Kneading Yusuke Hibi*1, Shiho Uesaka1, Kiyotaka Hitomi1, Ken-ichi Niihara2, Asami Imai2, Sadaki Samitsu1, Masanobu Naito*1 1Data-driven Polymer Design Group, Research Center for Macromolecules and Biomaterials, National Institute for Materials Science (NIMS); 1-2-1, Sengen, Tsukuba, Ibaraki 305-0047, Japan. 2Toyama Kankyo Seibi; 3-3, Fuchu-machi Yoshitani, Toyama, 939-2638, Japan. *Corresponding authors. Emails: hibi.yusuke@nims.go.jp; naito.masanobu@nims.go.jp This PDF file includes: Supplementary Figures 1-4 Supplementary Table 1-2 Captions for Supplementary Data 1-2 Other Supplementary Materials for this manuscript include the following: Data 1-2 are attached in csv format. mailto:hibi.yusuke@nims.go.jp2 Supplementary Figures and Tables Fig. S1. Data augmentation by adding Gauss noise with standard deviation of 0.15 to the original curve (black bold curve). 3 Fig. S2. Confusion matrices representing the results of tensile predictions based on (A) raw data, (B) raw and mean-zero offset data, (C) raw and standardized data, (D) mean-zero offset and standardized data, and (E) all three data. 4 Fig. S3. Distribution of averaged Charpy impact energies from five trials for each sample. See Table S1 for the numerical data. With a standard deviation of 0.07–0.1 kJ/m² across five trials, splitting classes at 3.95 kJ/m² appears arbitrary. Initially, samples with impact energies of 3.90–3.95 kJ/m² were labeled as Class 1, but this led to an imbalanced class distribution: Class 0: 3, Class 1: 14, Class 2: 6. This class imbalance caused prediction errors, as the small size of Class 0 (only 3 members) made feature extraction unreliable. Therefeore, we reclassified the seven samples with impact energies of 3.90–3.95 kJ/m² as Class 0. This adjustment resulted in a more balanced dataset: Class 0: 10, Class 1: 7, Class 2: 6. Although this reclassification may introduce mislabeling due to the relatively large standard deviation, soft labeling mitigated this uncertainty. The soft label for Class 0, defined as [0.75, 0.2, 0.05], reflects the possibility that Class 0 samples could actually belong to Class 1 with a 0.2 probability or to Class 2 with a 0.05 probability. 5 Fig. S4. Flowchart illustrating the relationship between viscosity data, tensile classes, and impact classes. During LOOCV, test data for both tensile and impact classes remained unlabeled and were never used for training. Despite differences in conditions between viscosity acquisition and the preparation of Charpy test specimens, transfer learning from tensile prediction to impact prediction enabled sufficiently accurate predictions, as shown in Fig. 5D. In practical applications, the top orange process represents pelletization, where viscosity data is acquired, while the bottom teal process corresponds to the customer’s final product manufacturing. The target property—herein corresponding to Charpy impact energy but not limited to it—can be predicted from viscosity data obtained during pelletization, allowing customers to select pellets optimized for their specific applications. 6 Table S1. Numerical data of the 23 recycled PP samples used in this study, including tensile properties (averaged values over two trials), impact energy per unit area (averaged values over five trials), and assigned classes. Y % Tensile classes were determined through hierarchical clustering based on the three tensile properties (Fig. 2), while impact classes were assigned by thresholding at 3.95 and 4.0 kJ/m² 7 (Fig. S3). The standard deviations of impact energy across five trials ranged from 0.07 to 0.1 kJ/m², introducing uncertainty in impact class labeling. This uncertainty was addressed through the application of soft labeling and label distribution learning (see main text). 8 Caption for Data S1. The dataset includes melt viscosity measurements and tensile test results from two trials for each sample. The tensile data, comprising absolute strain (mm) and force (N), can be converted to relative strain (%) and stress (MPa) using the specimen dimensions (width: 5.08 mm, depth: 2.04 mm, length: 52.3 mm). Caption for Data S2. Python source code for model implementation using the PyTorch package.