This study combines multi-frequency dynamic mechanical analysis (DMA) with machine learning (ML) to characterize and predict the viscoelastic properties of a commercial polycarbonate/acrylonitrile–butadiene–styrene (PC/ABS) blend. DMA temperature sweeps at four frequencies (1–10 HzHz) in single cantilever mode yielded a glass transition range of 115.8–123.2∘°C (E′′ peak), frequency sensitivity of 7.18msupC7.18 °C/decade, and an apparent activation energy of 335 ± 85 kJ mol−1. Time–temperature superposition master curves were parameterized with a six-term Prony series (R2=0.998). Four data-driven models (RF, XGB, SVR, MLP) and a physics-informed NeuralWLF model were evaluated through a hierarchical validation framework. Temperature-blocked CV ranked MLP (R2¯=0.989) above RF (0.950) for interpolation; LOFO validation revealed that NeuralWLF achieved the best cross-frequency generalization (R2>0.92 for all targets) with interpretable WLF parameters (C1≈12.2, C2≈51.7∘C). A systematic block size sweep (5– 30∘C) revealed a validation inflation effect in which MLP tanδR2 dropped from 0.986 to 0.592 as the gap-to-FWHM ratio increased from 0.5 to 3.1, establishing the gap/FWHM ratio as a quantitative validation stringency criterion. A physics–data crossover was identified at gap/FWHM ≈2: beyond this threshold, NeuralWLF outperformed all data-driven models in tanδ prediction by up to +0.300 in R2, while curriculum learning (freezing the WLF layer for 300 epochs) further improved the most stringent 30∘C validation from R2=0.660 to 0.731. The integrated framework demonstrates that honest evaluation of DMA–ML models requires validation gaps exceeding the characteristic feature width and introduces a quantifiable physics-data crossover criterion for selecting between data-driven and physics-informed architectures.
Sun et al. (Sat,) studied this question.