Evaluation of AI models for accuracy in the prediction of film thickness in polydisperse polymer films

Spin-coated polystyrene (PS) thin-films have many industrial applications, which vary greatly based on film thickness. A previous study by Wang et al. used a curve-fit machine learning model to create a 3D manifold relating molecular weight (M), polymer concentration in solution, and film thickness of spin-coated monodisperse PS samples [1]. Hence, this study aims to evaluate its effectiveness in predicting polydisperse PS film thickness. We simulated polydispersity by creating bidisperse solutions of PS with ratios of high and low M, M₁ and M₂, respectively, at varying concentrations. After spin-coating these samples onto silicon wafers [1,0,0], we determined film thickness through ellipsometry. For a given ratio and concentration, we found that the model used by Wang et al. consistently predicted lower thicknesses than actual experimental values. We were able to relate thickness, concentration, and ratio M₁:M₂ through a series of graphs, enabling the adjustment of their model. Our data agrees with the theory of constraint release (CR) induced dynamic tube dilation (DTD) for bidisperse mixtures with a wide separation of M₁ and M₂, which is further corroborated by the viscoelastic data on bidisperse PS solutions from Watanabe et al [2]. Our adjustment enables the application of the Wang et al. model to polydisperse PS mixtures, dramatically facilitating future research.

[1] Wang, A.C., et al. MRS Communications 14, 230–236 (2024).

[2] Watanabe, H. (1999). Progress in Polymer Science, 24(7), 1253-1403.