Model parameter inference is crucial for all computational science and engineering disciplines. It establishes a link between computational models and real-world observations. Geomaterials can be modelled at the microscale using the discrete element method, at the mesoscale with constitutive laws, or at the macroscale using the finite element method. A unified approach is lacking not only in integrating data, whether computational or experimental, from various scales for model inference but also in achieving a cross-scale understanding of the mechanical behaviour embedded in the model parameters. The bottleneck in the inference of geomechanical models is twofold: (1) the computational cost when running these models and (2) the model complexity reflected by parameter correlations. Reduced-order modelling and clustering algorithms can be used to emulate what the computational models can predict and to learn parameter uncertainties in a reduced space. In this talk, we will introduce probabilistic and data-driven techniques that enable the efficient inference of geomechanical model parameters. We will illustrate how data-driven and physics-based models are combined for the inference of geomechanical model parameters at various scales. Furthermore, we will address the question of which neural networks are best suited as model surrogates for problems at a given scale. Examples of the implications of this hybrid approach for geotechnical practices will be provided in the context of dike safety.