Abstract: During foundation pit excavation,the deformation of influenced-zone soil generally remains within the small-strain range.Traditional approaches for determining the parameters of the Hardening Soil Small-Strain (HSS) model have difficulty reflecting the effects of sampling disturbance and testing errors on soil stiffness characteristics.Hence,developing an inversion method for the deformation parameters of the HSS constitutive model based on measured lateral displacements of retaining piles is an urgent need for numerical analysis of deep excavations.Based on a deep excavation project in Fuzhou as a case study,a dataset of deformation parameters and lateral displacements was established by conducting finite element analyses of excavation under various combinations of deformation parameters.A convolutional neural network (CNN) was introduced to extract local features of the lateral displacement,while a bidirectional long short-term memory (BiLSTM) network was used to capture the short-term and long-term dependency characteristics of the lateral displacement response.An attention mechanism was adopted to adaptively weight critical spatial-temporal features associated with lateral displacement.These components were integrated into a sequential CNN-BiLSTM-Attention machine learning model for deformation parameter inversion.The study demonstrates that finite element analysis with optimal deformation parameters inversely derived from measured lateral displacements of retaining piles during the initial two excavation phases generated results varying between 0.82 and 1.18 times the field measurement values.The calculation accuracy is significantly better than the results based on single-stage data.After further intergrating the third-stage data,the prediction accuracy only increases by about 3%.Compared with other neural network models,the CNN-BiLSTM-Attention model performs better than other dual neural network combinations in terms of average absolute percentage error MAPE,coefficient of determination R²,and root mean square error RMSE.The proposed CNN-BiLSTM-Attention accurately captures the long-term and short-term dependencies of key local features of the horizontal lateral displacement of the retaining piles,and optimizes the deformation parameters of the HSS constitutive model based on the real-time horizontal lateral displacement data of the retaining piles,effectively improving the prediction accuracy of subsequent foundation pit excavation deformations.