Pioneering Steps in Modeling Liquid Composite Molding based on Artificial Intelligence


In modern manufacturing processes for fiber reinforced polymers (FRP), Liquid Composite Molding (LCM) has emerged as a key technique for creating complex structural parts in high volumes. LCM parts are found in various industries, including aerospace and automotive. However, predicting the behavior of the textile reinforcement during the manufacturing process has been challenging scientists and engineers for years, particularly when it comes to permeability predictions of such fiber structures.

Bridging the Gap between Physics Based Numerical Modeling and Artificial Intelligence (AI)

Researchers at Weierstrass Institute for Applied Analysis and Stochastics (WIAS) and Leibniz-Institut für Verbundwerkstoffe (IVW) are exploring ways to make these processes more predictable. Conventionally, they are leveraging physics-based numerical modeling to predict permeabilities of fiber structures. Recently, an innovative AI approach that couples traditional physics with neural networks, known as Physics-Informed Neural Networks (PINNs), has been investigated for the aforementioned purposes.

Why Is Permeability Prediction So Important?

In LCM, understanding how the resin flows through the fibrous materials is crucial. Wrong predictions can lead to defects such as voids and compromised structural integrity. Accurate permeability prediction is essential for a robust process design, that provides high quality of the final product and cost savings in the long run.

The Role of Physics-Informed Neural Networks

Unlike traditional neural networks, which rely solely on data, PINNs incorporate the laws of physics into the learning process. This means that the network not only learns from the data but also respects the physical laws governing the behavior of the material. In the context of Liquid Composite Molding, this approach holds the promise of fast and accurate predictions of resin flow characteristics in complex fiber structures while being very data efficient. Another advantage comes from the meshless nature of PINNs, which makes them especially well-suited to complex geometries of porous media, e.g. fibrous structures, compared to standard mesh-dependent numerical methods for permeability prediction.

A Cautious Step Forward

In this work, PINNs were trained on simplified geometries of fibrous microstructures in order to assess their capabilities of predicting resin flow fields. The permeability of a structure can subsequently be derived from such predictions. After finding the correct neural network architecture and training parameters, the initial findings are promising and the results are in good agreement with those provided by conventional numerical methods (Figure 1).

Looking Ahead

While it is still early days for this research, rapid advancements in hardware and deep learning technologies will drive the development of PINNs as a complementary or even enhancing method to conventional numerical simulations. As a result, PINNs promise to become an increasingly effective and reliable tool for engineers in the field of Liquid Composite Molding simulations.

The results of this collaboration between WIAS and IVW will be presented at the 1st Conference on Artificial Intelligence in Material Science and Engineering (AIMSE) on November 22-23 in Saarbrücken, Germany.

Horizontal (u) and vertical (v) velocity components approximated by PINNs (left), calculated by conventional methods (center), and the corresponding point wise error (right).


Stefano Cassola

Wiss. Mitarbeiter Prozesssimulation

↰ News