The specific volume of polymers is
a critical parameter that has
offered many highly promising applications in focus as a basic material
property. The specific volume (v) as a function of
commonly pressure (p) as well as temperature (T), however, does not satisfy accuracy requirements with
respect to experimental results. In addition, polymer forming is a
rapid cooling process, and the cooling rate (dT/dt) has a significant effect on the pvT behavior
of polymers. However, today’s commercial testing devices cannot
obtain the pvT data for high cooling rates. Hence,
the prediction of specific volume of polymers for rapid cooling applications
remains a challenge. This work deals with the potential application
of artificial neural networks (ANNs) to model the v in dependence of p, T, and dT/dt within their experimental uncertainty.
The experimental pvT data of the semicrystalline
polypropylene (PP) at the temperature range of 40–240 °C,
the pressure range of 200–2200 bar, and the set cooling rate
range of 2–20 °C/min were used for the ANN training. Besides
the data of p and T, time (t) and phase-transition temperature (T
t) were also used as input features of the training datasets
to improve the extrapolation ability for high cooling rates. A hyperparameter
search was conducted to identify the optimized hyperparameter sets.
The experimental values of T
t at atmospheric
pressure (1 bar) and cooling rates (2–300 000 °C/min)
obtained through differential scanning calorimetry (DSC) and flash
DSC were used to indirectly validate the prediction accuracy of the
ANNs. The newly developed continuous two-domain equations of state
(EoS) were also applied in comparison. The ANN approach showed the
most reliable predicting results both for interpolations and extrapolations
when trying to predict the specific volume beyond the pvT measurement range (the cooling rate is up to 6000 °C/min and
the pressure is the atmospheric pressure of 1 bar).