AI-Driven inverse design of 3D-printed continuous fiber composites using multiscale homogenization and dual neural networks
Authors:
Preprint
- 6 October 2025
- preprint
- Published by Springer Nature in Research Square
Abstract
Accurately predicting the mechanical behavior of 3D-printed continuous fiber-reinforced composites remains a complex task due to their heterogeneous and anisotropic microstructures, which arise from fiber orientation, overlap, and spatial distribution. This study proposes a novel computational framework that combines multiscale homogenization and artificial intelligence to enable both predictive modeling and inverse design of composite materials. A finite element method (FEM) was employed to simulate 195 representative volume element (RVE) configurations, capturing various fiber overlaps, horizontal and vertical spacing parameters. The generated dataset was used to train a dual neural network architecture comprising a forward neural network (FNN) for property prediction and an inverse neural network (INN) for microstructural design. Advanced regularization, early stopping, batch normalization, and dropout techniques were applied to optimize model performance. The FNN model accurately predicted key mechanical properties including Young's modulus (E), shear modulus (G), and Poisson's ratio (v) with an average R² score above 92%. The INN enabled rapid inverse design of RVE configurations that satisfy targeted mechanical properties, achieving a mean prediction accuracy of 96.06% with response time under one second. Sensitivity analysis revealed that interlayer spacing parameters significantly affect material anisotropy and stiffness. The proposed hybrid FEM–AI framework demonstrates high accuracy and computational efficiency in both forward prediction and inverse microstructure design. It offers a scalable solution for the rapid development of custom-tailored composite materials and holds strong potential for high-performance applications in aerospace, automotive, and biomedical engineering.Keywords
All Related Versions
This publication has 44 references indexed in Scilit:
- A machine learning based prediction model for the impact mechanical response of composite laminates considering microstructure sensitive transverse propertiesPolymer Composites, 2024
- An efficient deep learning-based topology optimization method for continuous fiber composite structureActa Mechanica Sinica, 2024
- Model performance evaluation of build time using geometric shape complexity and process parameters in material extrusionAdditive Manufacturing, 2024
- Deep learning, deconvolutional neural network inverse design of strut-based lattice metamaterialsComputational Materials Science, 2024
- Deep Learning in Mechanical Metamaterials: From Prediction and Generation to Inverse DesignAdvanced Materials, 2023
- Calculation of dislocation binding to helium-vacancy defects in tungsten using hybrid ab initio-machine learning methodsActa Materialia, 2023
- Data-driven models for structure-property prediction in additively manufactured steelsComputational Materials Science, 2022
- Additive manufacturing: expanding 3D printing horizon in industry 4.0International Journal on Interactive Design and Manufacturing (IJIDeM), 2022
- A Convolutional Neural Network (CNN) classification to identify the presence of pores in powder bed fusion imagesThe International Journal of Advanced Manufacturing Technology, 2022
- Effects of Surface Energy Reducing Agents on Adhesion Force in Liquid Bridge MicrostereolithographyAdditive Manufacturing, 2020