Rational Design Of High-Energy-Density Polymer Composites By Machine Learning Approach

Rationally designing polymer composite structures, including physical parameters of nanofillers, nanofiller-matrix interface characteristics, and geometric distribution of nanofillers, is thought to be an effective approach to achieve the desired dielectric properties such as breakdown strength (E-b), permittivity (epsilon(r)), and energy density (U-e) in wide applications. However, the work is difficult to complete through merely high-cost and time-consuming trial-and-error experiments. A machine learning (ML) driven approach, trained on hundreds of experimentally measured data, is presented to rationally design polymer composites with desired properties. The doping scheme of nanofillers is finger-printed with a string of characters considering the physical parameters, shape, distribution of fillers, and shell properties in core-shell structures, and then the Gaussian process regression algorithm is trained to establish the linkage between the filler doping scheme and the dielectric properties. The dielectric properties of the randomly generated tens of millions of candidate composites are calculated with the resulting ML model, and representative composites with high E-b, epsilon(r), and U-e are presented. The results indicate that the effects of nanofiller permittivity and bandgap on E-b and epsilon(r) follow exactly the opposite trend, hence it is difficult to simultaneously improve E-b and epsilon(r) by choosing the type of nanofiller. Fortunately, the trade-off between E-b and epsilon(r) can be improved by tailoring the shape, orientation, and distribution of the nanofillers, for instance, by using horizontally orientated nanosheets and orthotropic nanowires with high permittivity.