AI-Powered Flow Field Prediction for Permeability Calculation in Liquid Composite Molding

In the fast evolving field of composite manufacturing, scientists are constantly striving to improve and create more robust processes. In Liquid Composite Molding (LCM), a key component of this advancement is understanding how resin moves through textile materials, which involves determining the material's permeability at various scales. Traditionally, this has required either expensive experiments or, more recently, intensive numerical methods and computational power. However, a new approach introduced here presents a more efficient method using artificial intelligence.

The Challenge: Measuring Textile Permeability Virtually

Textile permeability is a crucial parameter for the design of effective LCM manufacturing processes. To accurately measure it on a numerical basis, researchers have relied on complex numerical methods such as Finite Volume, Finite Element, and Lattice Boltzmann Methods. These approaches solve intricate mathematical equations (Stokes' equation) to simulate how liquid flows through fibrous structures at various spatial scales. However, these methods are computationally demanding and time consuming, especially when dealing with large 3D models necessary for accurate results.

Enter Deep Learning: A Faster, Smarter Solution

This study introduces a novel approach using deep learning to predict the flow of liquid through these fibrous materials. The network architecture used is called MS-Net and implements coupled 3D convolutional neural networks (3D-CNN) specifically tailored for this task. MS-Net learns to predict the velocity of liquid flow through porous media by analyzing different resolutions of the same computational domain, making the training more efficient and less computationally intensive.

Promising Initial Results with Minimal Data

 Leibniz-Institut für Verbundwerkstoffe`s research team has achieved promising initial results using a small training dataset of only eight samples. After 24 hours of training, the model was able to predict the flow velocity field on samples of the training set with high accuracy in just five seconds on a desktop computer. For flow parallel to the fiber direction, the model's error in the prediction of the average flow velocity was a mere 0.3%, demonstrating high precision while predicting flow in the transverse fiber direction proved to be more challenging.

Scaling Up with Advanced Techniques

Building upon this success, the researchers are refining the MS-Net architecture using state-of-the-art deep learning techniques. The dataset at hand is based on 4,284 synthetic 3D geometries and flow velocity fields with a size of 320³ voxels (approx. 33 million voxels), each with varying fiber volume content, diameter, and direction. By downscaling the geometry resolution and recalculating the velocity fields, they aim to accelerate training times and hyperparameter tuning.

Looking Ahead

This work in AI-driven permeability prediction holds significant promise for the future of composite manufacturing. The general approach has the potential to not only make permeability faster and more interpretable but is also transferable to other phenomena, e.g. temperature filed prediction for heat conduction problems. The addition of field information, coupled with fast prediction times, can help manufacturers and scientists to design better processes more efficiently, reducing costs and accelerating innovation. While it is still early days for this research, rapid advancement in deep learning technologies will drive the development of AI as an additional tool for engineers and scientists in the field of composite manufacturing.