Hybrid Monte Carlo-MLP prediction of tunneling percolation and conductivity in conductive polymer composites under compressive strain

초록

Conductive polymer composites (CPCs) are promising materials for compression-responsive sensing applications. However, their deformation-induced nonlinear electrical response complicates efficient material design. This study proposes a hybrid computational framework that integrates three-dimensional Monte Carlo particlenetwork simulations with a multilayer perceptron (MLP) surrogate model to rapidly predict spanning-cluster evolution (quantified by the spanning cluster fraction, SCF) and to evaluate tunneling-governed network conductivity under compressive strain. A representative volume element (RVE) is populated with randomly dispersed, non-overlapping spherical fillers using a hard-core constraint, and compressive kinematics are modeled with Poisson's ratio to capture transverse expansion. Electrical connectivity is determined by a soft-shell tunneling criterion, and tunneling resistance is evaluated using the Simmons tunneling model to compute network-level conductivity. Conductive network formation is quantified by the spanning cluster fraction (SCF) in a finite RVE, defined as the fraction of particles belonging to the cluster that bridges the two opposite electrode faces. The spanning cluster is identified efficiently using a depth-first search (DFS) algorithm. A simulation database was generated from 100 independent runs per condition and was used to train the MLP surrogate by varying five design parameters: filler volume fraction (Vf), particle volume (Vp), Poisson's ratio (u), compressive strain (e), and tunneling cutoff distance (6). Permutation-importance analysis further quantified the relative influence of key design variables on percolation evolution, indicating that Vf dominates network formation (76.0%), while e and 6 primarily regulate sensing sensitivity, with an optimal sensing window near the percolation threshold (Vf approximate to 20 - 28%). The trained MLP achieved strong predictive performance for SCF, with a testset coefficient of determination (R2) of 0.9744 and a mean 5-fold cross-validated R2 of 0.9927, providing an interpretable and computationally efficient tool for CPC design-space screening under compressive loading. A literature-based comparison with reported compression-responsive CPC conductivity trends was additionally included to assess the qualitative consistency of the simulated network evolution.

키워드