Abstract <p>This paper proposes a novel machine learning method of finite element stress recovery, which is based on the feature variables of coordinates and displacements, to improve stress recovery accuracy. Four distinct machine learning algorithms—Backpropagation (BP) Neural Network, Radial Basis Function (RBF) Neural Network, Random Forest (RF), and Support Vector Regression (SVR)—are applied to both linear and nonlinear stress fields using linear and quadratic finite elements. Results indicate that the proposed method achieves higher accuracy in recovering the three node stress components compared to traditional methods, especially in small sample scenarios where its advantages are more evident. For linear stress fields with limited samples and a higher prevalence of outliers, the SVR algorithm demonstrated superior recovery performance, outperforming traditional finite element software. In contrast, for nonlinear stress fields with larger sample sizes and complex data relationships, the RBF Neural Network demonstrated the best recovery outcomes, also surpassing standard finite element software. Furthermore, quadratic elements consistently showed higher stress recovery accuracy than linear elements. These findings offer valuable insights and practical implications for enhancing stress recovery methods in finite element analysis.</p>