Machine Learning Prediction of Chemical Resistance of Resin Materials Based on Crystallinity

In semiconductor manufacturing equipment, resin materials in piping and other components are constantly in contact with solvents, necessitating long-term chemical resistance. However, testing the chemical resistance lifespan is time and cost-consuming. To reduce testing costs, we are developing a machine learning model to predict the chemical resistance lifespan of resin materials against solvents. As a preliminary step, we are working on a machine learning model to classify the chemical resistance of resin-solvent combinations.^[1] This presentation reports on the results of improving the accuracy of the chemical resistance prediction model by incorporating polymer crystallinity as a feature.
We created a chemical resistance dataset by assigning binary classification labels based on chemical resistance indices from the literature for resin-solvent combinations. To represent polymer crystallinity, we assigned a label of 1 to thermoplastic resins generally classified as crystalline polymers and 0 to those classified as amorphous polymers. We created a chemical resistance classification model based on a dataset of 3,237 entries, consisting of 30 resins and 218 organic solvents. As explanatory variables for machine learning, we calculated the Force Field Kernel Mean Descriptor (FFKM) from the SMILES of polymers and solvents.^[2] FFKM calculates the discrete distribution of 10 force field parameters included in Molecular Dynamics (MD) calculations from SMILES and generates continuous distributions through kernel mean embedding to create explanatory variable vectors. For polymers, we calculated FFKM for cyclic SMILES created by connecting 10 repeating units, and for solvents, we calculated FFKM for single molecule SMILES. We used the Gradient Boosting Decision Tree (GBDT) Classifier as the classification algorithm. For chemical resistance prediction accuracy, we calculated the F1 Score using Leave One Out Cross Validation (LOOCV) for 30 resins and Leave Cluster Out Cross Validation (LCOCV) based on 10 clusters of solvents grouped by solubility. The average F1 Score for the 30 resins showed improvement when crystallinity labels were added as explanatory variables. Additionally, the average F1 Score for LCOCV based on 10 clusters of solvents grouped by solubility also showed improved prediction accuracy when both FFKM and crystallinity labels were used. From the improved classification accuracy of the machine learning model, we confirmed that considering crystallinity is important for the chemical resistance of resins.

References:
[1] Qisong X., Jianwen J., ACS Appl. Polym. Mater. 2020, 2, 8, 3576–3586
[2] Kusaba, M., Hayashi, Y., Liu, C., Wakiuchi, A., Yoshida, R., Physical Review B, 2023, 108 134107