Background Microstructural features such as grain boundaries play a significant role in the macroscopic plastic response of polycrystalline metals. However, a quantitative link between plastic strain accumulation at grain boundaries and material response in plasticity dominated phenomena is still lacking. Objective Here we seek to develop predictive relations between a material’s granular microstructure and the accumulation of plastic strains at the microstructural level during plastic deformation. Methods A single-input neural network approach was applied to predict the residual plastic strain fields at regions surrounding grain boundaries of an austenitic stainless steel. The neural network was trained on data obtained by applying a very-high resolution digital image correlation (DIC) experimental technique that allows the measurement of grain-scale strains aligned to the underlying microstructure obtained from electron backscatter diffraction (EBSD) scans. Results The neural-network-predicted and the DIC-measured strain fields showed good correlation for most of the tested cases. Best individual agreement was found when each microstructure was used to predict fields in its own case. However, best overall average predictions were seen when multiple samples were used for the network training. Conclusions The results showed that the local geometrical angle between a grain boundary and the loading axes is in many cases a good predictor for the accumulation of strains at the given boundary. The expected limitations of this single parameter approach (grain boundary angle alone cannot be a good predictor for varying strains along a straight grain boundary, for example) were seen as the reason for the situations where predictions were not as good.

中文翻译：

---

多晶金属晶界应变累积的机器学习神经网络预测

背景 晶界等微观结构特征在多晶金属的宏观塑性响应中起着重要作用。然而，仍然缺乏在晶界处的塑性应变积累与塑性主导现象中的材料响应之间的定量联系。目的 在这里，我们寻求在塑性变形过程中开发材料的粒状微观结构与微观结构水平上塑性应变积累之间的预测关系。方法应用单输入神经网络方法预测奥氏体不锈钢晶界周围区域的残余塑性应变场。神经网络接受了通过应用超高分辨率数字图像相关 (DIC) 实验技术获得的数据进行训练，该技术允许测量与从电子背散射衍射 (EBSD) 扫描获得的底层微观结构对齐的晶粒尺度应变。结果 神经网络预测的和 DIC 测量的应变场在大多数测试案例中显示出良好的相关性。当每个微观结构用于预测其自身情况下的场时，发现了最佳的个体一致性。然而，当多个样本用于网络训练时，可以看到最佳的总体平均预测。结论 结果表明，在许多情况下，晶界和加载轴之间的局部几何角度是给定边界处应变积累的良好预测指标。