If the process‐induced distortions (PIDs) of asymmetrical laminates can be predicted accurately and tailored at the early design stage, the production of curved panels from flat molds could be an attractive technique in a cost‐driven production environment. A data‐driven computational methodology which integrates the finite element method (FEM) and artificial neural network (ANN) is presented to rapidly predict the maximum PID and to perform high‐throughput screening of thermosetting‐matrix composites of an asymmetrical laminate for a targeted maximum PID. We performed a grid search on ANN architectures and hyper‐parameters using cross‐validation and obtained a well‐trained ANN model with high generalization performance. For the forward problem, the ANN model was adopted to predict the maximum PIDs of CYCOM X850 and CYCOM 977‐2 prepregs, which were subsequently verified experimentally. For inverse design, a large‐scale screening method based on the ANN model was utilized to determine the candidates for a targeted maximum PID, with an experimental demonstration using one of these candidates. The well‐trained ANN model provides an alternative approach to faster computation with high accuracy for the maximum PID prediction and further guides the discovery of materials with desired distortion behaviors.

中文翻译：

---

基于人工神经网络的复合材料变形行为的快速预测与逆设计

如果可以在设计的早期阶段准确地预测并定制不对称层压板的加工引起的变形（PID），则在成本驱动的生产环境中，用平板模具生产弯曲面板可能是一种有吸引力的技术。提出了一种将有限元方法（FEM）和人工神经网络（ANN）集成在一起的数据驱动计算方法，以快速预测最大PID，并对不对称层压板的热固性基质复合材料进行高通量筛选，以达到目标最大PID。我们使用交叉验证对ANN架构和超参数进行了网格搜索，并获得了训练有素且具有较高泛化性能的ANN模型。对于前向问题，采用了ANN模型来预测CYCOM X850和CYCOM 977-2预浸料的最大PID，随后进行了实验验证。对于逆向设计，利用基于ANN模型的大规模筛选方法来确定目标最大PID的候选对象，并使用其中一个候选对象进行了实验演示。训练有素的人工神经网络模型提供了另一种方法，可实现更快的计算速度和高精度，以实现最大的PID预测，并进一步指导发现具有所需变形行为的材料。