目的 通过结合人工神经网络和激光冲击有限元仿真，以减少激光冲击强化最佳参数设计的迭代次数，提高参数优化效率。方法 构建基于冲击波压力幅值的激光冲击强化Abaqus有限元模型。采用Vdload子程序与对应二次开发脚本形成光斑重叠区域残余应力的初始数据集。建立人工神经网络(ANN)算法模型，采用测试集对ANN模型进行测试，对超参数进行优化，对比分析不同机器学习算法的R²得分、MAE和RMSE。设计并优化镍铝青铜模型表面的残余应力大小与分布，对比分析经机器学习预测后的模型表面残余应力分布情况。结果 经ANN预测整个面LSCP处理的模型表面形成了高达?413 MPa的残余压应力，并且预测了均匀与非均匀的残余压应力分布;RMSE均方根误差仅为1.189 1，既显示出较好的预测精度，又避免了模型的过拟合，保证了一定的泛化能力，模型综合性能远优于其他经典的ML算法回归模型。所预测的残余压应力分布模型均达到了较深的影响层深，且在1 Hz的脉冲重复频率下，最大效率达到了1.87 mm²/s。结论 激光冲击强化与机器学习的结合实现了易产生残余应力孔洞的镍铝青铜光斑重叠区域的最大残余压应力分布，且该方法优化出了整个表面的均匀与相对非均匀的残余压应力分布，为非均匀塑性应变引起镍铝青铜材料异质结构的形成开辟了新的设计途径。

Laser shock composite peening (LSCP) is one of the advanced methods for material surface enhancement, and it has gained significant attention recently due to its ability to induce beneficial residual stress fields. Traditional LSCP design methods involve selecting processing parameters through trial and error, which can be imprecise and time-consuming. These methods suffer from the complexities of internal stress wave transmission and non-uniform plastic strain under high strain rate loads. Machine learning (ML) algorithms offer a promising alternative by automating the design of critical LSCP parameters, thus reducing the iterative design process and associated costs. The work aims to leverage an Artificial Neural Network (ANN) algorithm to predict and regulate the residual compressive stress distribution on nickel-aluminum bronze surfaces, thus reducing the iterative design process and associated costs. An initial dataset of residual stresses in the spot overlap area was generated by an Abaqus finite element model with the Vdload subroutine and custom scripts. Before the regression analysis, the interquartile range (IQR) method was used to remove the top and bottom 10% of outliers, and the input data were standardized to eliminate the effect of data scale on the prediction results. The ANN model was trained and tested with the generated residual stress dataset, optimizing its hyperparameters for enhanced performance. Based on the process parameters in the training set, the residual stress values in areas prone to stress hole were predicted for 110 different process parameter sets. The results showed that almost all predictions were in close agreement with the actual values, confirming the strong prediction capability of the artificial neural network. The ANN accurately predicted residual compressive stress distributions, achieving an RMSE of 1.189 1, significantly outperforming other classical ML algorithms. The residual stress distributions were predicted and optimized, with the ANN model indicating compressive stress up to ?413 MPa across the treated surface. These predictions were validated by the test set, confirming the high prediction accuracy and robustness of the model against overfitting. Further analysis revealed that the predicted residual compressive stress distributions reached substantial effect depths, critical for material property enhancement. The LSCP process achieved a maximum efficiency of 1.87 mm2/s at a 1 Hz pulse repetition frequency. This method presents a novel approach to designing and regulating complex residual stress fields in LSCP, effectively addressing the challenge of residual stress hole in nickel-aluminum bronze. The integration of ML with LSCP not only optimizes the residual stress distributions, but also provides insights into the development of heterogeneous structures in the material due to non-uniform plastic strain. Future research will aim to incorporate more experimental data into the ML models to enhance their applicability to various types of metals and further refine the prediction accuracy of residual stress fields in complex material systems. This will involve collecting data from a broader range of experimental conditions and materials, thereby improving the robustness and versatility of the ML models. The goal is to create a comprehensive framework that can be applied to a wide array of materials and LSCP scenarios, ensuring that the benefits of this approach can be realized across different industries. This study contributes to the field of material surface enhancement by demonstrating the potential of combining advanced computational models with machine learning for more precise and efficient material treatment outcomes.