This paper constructs a surrogate model for load inversion and response prediction of tubular joints based on sparse measurement data by integrating two-dimensional Chebyshev orthogonal polynomials and a radial basis function (RBF) neural network, to address the challenge of dynamic monitoring and structural response prediction for the service status of tubular joints in offshore jacket platforms and achieve accurate load identification and rapid response prediction. First, given the complexity of the surfaces at key tubular joints of the jacket platform, two-dimensional Chebyshev orthogonal polynomials are used to mathematically represent the distributed loads acting on these surfaces. Combined with truncated singular value decomposition (TSVD) and generalized cross-validation (GCV), high-precision and stable load identification is achieved from pre-set sparse measurement response data. Then, based on the inversion results, a high-quality training sample set consisting of 100 working conditions, covering spatial coordinates and load characteristic coefficients, is generated. Each sample uses spatial node coordinates and load coefficients as input features, with the corresponding equivalent stress as output. Finally, the RBF surrogate model is constructed and trained, with the coefficient of determination ( R²) used to evaluate model performance. The results show that the model achieves an R² of 0.913 78 on the validation set, enabling high-precision and high-generalization prediction of the stress field at key tubular joints of the jacket platform, providing effective technical support for real-time condition assessment and fatigue damage accumulation prediction of the jacket platform.