Data-driven modelling is gaining interest in geotechnical engineering as it offers an alternative to the often expensive, high fidelity, high-dimensional models based on standard computational methods for solids. Geotechnics is particularly demanding as it involves material and sometimes geometric non-linearity, mixtures of materials (e.g. soil structure interaction) and the need for a detailed 3D spatial representation. Here we implement a data-driven approach for predicting spatio-temporal geotechnical behaviour in reduced dimensions. Based on the precomputed high-fidelity simulations, that need be run once, and using the notion that behaviour can be represented by a limited number of modes, the reduced-order modes, which can be obtained by data-driven techniques, e.g., proper orthogonal decomposition and auto-encoders amongst others. Here we implement neural operators [1] to approximate the mappings between infinite-dimensional Banach spaces, e.g., mapping the parameters that can take an infinite number of values (geometry, soil properties, etc.) to the reduced-order modes. A data-driven technique is then used to reconstruct the high-fidelity responses (wall displacement, prop loads etc.) from the reduced-order modes. We demonstrate the whole process for consolidation analysis, footing problems, and braced excavations, highlighting the superior predictive accuracy and computational efficiency compared with existing methods. The proposed approach has wide application to other geotechnical problems which need complex high-fidelity modelling. REFERENCES [1] Lu, L., Jin, P., Pang, G. et al. Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators. Nature Machine Intelligence, Vol. 3, pp. 218-229, 2021.