The stochastic nature of the model was used to characterize the uncertainty of soil parameters based on the random transformation of an artificial neural network architecture incorporating Monte Carlo Dropout (ANN_MCD) in order to address the issue that using deterministic analysis method is risky to control the post-construction settlement of ground improved by deep mixed columns (DMCs). The predictions were performed efficiently for post-construction settlement of ground improved by DMCs considering the uncertainty of soil parameters by combining the finite element simulations with a surrogate model. The probability distribution of post-construction settlement with different combinations of pile length, pile diameter, pile spacing and cushion thickness was obtained. The limit value of the post-construction settlement was determined from the target reliability index for serviceability limit state, and linked to the cost of ground improvement based on the nonlinear mapping. The structural design optimization was finally conducted by considering the benefit–cost ratio. The uncertainty of modified Cam-Clay model parameters can be derived from the plasticity index I_p of soft clay with ANN_MCD-based model, with the 95% prediction interval matching closely with the experimental data. The ANN-based model with a separated input layer can individually extract features from soil and structural parameters, avoiding a redundant architecture of ANN and achieving efficient and precise predictions for post-construction settlement of DMC composite foundation. The relationship between post-construction settlement and the lowest construction cost can be fitted by a Logistic curve. The threshold of benefit-cost ratio corresponds to the maximum curvature point of the Logistic curve, and the optimized design should be on the side of cost lower than C_v (highly cost-effective).