Experimental design, machine learning, and genetic algorithm to support the production of conductive polymer nanocomposites

The development of conductive polymer nanocomposites presents a complex challenge due to the nonlinear interactions between polymer matrices, carbon-based fillers, and processing conditions. In this study, we combine experimental design, machine learning (ML), and genetic algorithms (GA) to optimize the formulation of conductive filaments for additive manufacturing applications. A full factorial experimental design was employed to evaluate different polymer types, carbonaceous fillers, and processing parameters. The electrical conductivity of the resulting composites was measured and used to train predictive models-namely, a multilayer perceptron (MLP) and a random forest (RF)-whose hyperparameters were optimized using Optuna. Both models achieved satisfactory performance metrics, with Nash-Sutcliffe Efficiency (NSE) values of 0.47 for MLP and 0.43 for RF on the test set. The trained models were then integrated into a genetic algorithm to explore optimal formulations, which identified PA66 as the most promising matrix and single-walled carbon nanotubes as the most effective conductive filler. Model interpretability was assessed via SHapley Additive exPlanations (SHAP), which corroborated GA results, particularly highlighting the models' preference for longer processing times and lower/medium rotation speeds-conditions that align with known percolation and dispersion requirements in conductive nanocomposites. This integrated approach illustrates the potential of materials informatics to support the design and discovery of advanced functional materials.