The mechanical properties of glass fiber reinforced polymers (GFRP) are significantly governed by the orientation of the fibers in the composite. Micro X-ray computed tomography (CT) imaging offers a way of determining the fiber’s orientation in a non-destructive fashion. Various approaches have been presented to compute the direction of the fibers based on fiber tracking or weighted volume algorithms. In this work we present two novel approaches, one employing convolutional neural networks (CNNs) and the other directional analysis by inertia tensor evaluation (ITE). We establish a workflow based on molecular dynamics simulations to efficiently create synthetic training data for the CNN. The two methods are applied to two experimental CT scans of the GFRP polyamide 66 of two different components, featuring different CT resolutions, fiber lengths and volume fractions.

The CNN model trained by synthetic data predicts fiber orientations consistently and with similar accuracy as best-in-class commercially available products. We observe an increase in computational speeds of at least a factor of 4 on CPUs or about a factor of 50 on GPUs, respectively. A striking feature of this approach is that the ground truth of our training data is perfectly known and no time-consuming manual labeling of fibers for training is needed. The proposed ITE method is very robust and particularly suited to lower resolution CT scans, as the evaluation of gradients is not necessary.

Both methods extend the toolbox of weighted volume approaches and are well-suited to predict orientations of densely-packed fibers that are often encountered in industrial practice. In addition, predictions by the trained CNN model can be run on standard office hardware which makes them particularly interesting for industrial environments.

玻璃纤维增​​强聚合物 (GFRP) 的机械性能在很大程度上受复合材料中纤维取向的控制。显微 X 射线计算机断层扫描 (CT) 成像提供了一种以非破坏性方式确定纤维方向的方法。已经提出了各种方法来基于纤维跟踪或加权体积算法来计算纤维的方向。在这项工作中，我们提出了两种新方法，一种采用卷积神经网络 (CNN)，另一种采用惯性张量评估 (ITE) 进行方向分析。我们建立了一个基于分子动力学模拟的工作流程，以有效地为 CNN 创建合成训练数据。这两种方法应用于两种不同组分的 GFRP 聚酰胺 66 的两次实验 CT 扫描，具有不同的 CT 分辨率，

由合成数据训练的 CNN 模型可以一致地预测纤维方向，并具有与同类最佳市售产品相似的准确度。我们观察到 CPU 的计算速度至少提高了 4 倍，GPU 的计算速度提高了约 50 倍。这种方法的一个显着特点是我们训练数据的基本事实是完全已知的，并且不需要耗时的人工标记训练纤维。所提出的 ITE 方法非常稳健，特别适用于较低分辨率的 CT 扫描，因为不需要对梯度进行评估。

这两种方法都扩展了加权体积方法的工具箱，非常适合预测工业实践中经常遇到的密集纤维的方向。此外，经过训练的 CNN 模型的预测可以在标准办公硬件上运行，这使得它们在工业环境中特别有趣。