Machine learning-driven optimization of high-velocity impact resistance for three-dimensional biphasic composites

Under high-velocity loading conditions, composite materials inevitably exhibit wave propagation and localized temperature rise, yet their complex microstructural arrangement contributes to the characteristic stress localization and attenuation behaviors. To mitigate this challenge and optimize the impact resistance of composite materials, we developed a three-dimensional convolutional neural network (3D-CNN) with experimental and simulation methods to quantitatively correlate microstructural configurations with the dynamic response of biphasic composites. This methodology synergizes experimental data from split Hopkinson pressure bar (SHPB) dynamic tests with finite element (FE) simulations, enabling a comprehensive characterization of the dynamic behavior of composites. The optimized microstructures improve dynamic stress uniformity from 21.54% to 97.45% and increase dynamic load capacity by approximately 22% to 57% compared with random designs at the same stiff-phase fractions. This machine learning-driven approach can improve dynamic stress equilibrium properties for soft and stiff composites. The thermal transport module enables integrated evaluation of impact resistance and heat-spreading performance, revealing multi-material architectures that sustain high dynamic mechanical performance while allowing effective thermal conductivity through phase connectivity and percolation pathways. This proposed 3D CNN framework facilitates rapid exploration of the design space and accurate prediction of thermomechanical properties, thereby enabling the inverse design of next-generation impact-resistant composites for aerospace and automotive applications.