Abstract: To achieve soft sensor modeling in a chemical process with strong nonlinearity and complexity, a recurrent neural network method based on square-root cubature Kalman filter (SCKF) training algorithm is proposed. The state-space model of the Elman recurrent neural network is first established, and then the network is trained by the SCKF algorithm; all its weights are considered the state of the system to be updated. Cubature Kalman filter (CKF) generates cubature points by third-order spherical-radial standards and then uses cubature points to approximate the posterior distribution of the state, which makes computing the numerical solution of the multivariate integral of high-dimensional nonlinear filtering possible. On the basis of CKF, SCKF propagates the square root factor of prediction and posteriori error covariance matrix to further improve the numerical stability of the algorithm. The employed method is applied to instances of soft sensor modeling, which include the estimation of the butane concentration in the bottom flow of a debutanizer column and the estimation of the concentrations of SO₂ and H₂S in sulfur recovery unit tail gas composition. Compared with feedforward neural network method based on extended Kalman filter (EKF) and the SCKF algorithm and recurrent neural network method based on the EKF algorithm, the proposed method can obtain better modeling estimation accuracy under the same condition. Experimental results demonstrate the effectiveness of the proposed method.