Predicting finite non-linear deformations and associated microstructural changes is key to effective development and application of materials in various industrial sectors. Examples include severe microstructural changes of polymeric materials subject to extreme environments such as rapid gas decompression (RGD) and/or finite non-linear deformations. Recent advances in machine learning (ML) methods offer promising means for constructing surrogate predictive models due to their computational efficiency compared to standard physics-based approaches and numerical methods. Deep learning methods that exploit convolutional neural network (CNN) architecture are particularly suited to developing robust surrogates that make image-based predictions of the microstructural changes of materials. However, their ability to capture fields such as strain or stress has been less explored. Therefore, this study develops an ML approach for developing surrogate models that can capture both microstructural changes and reconstruct fields of strain/stress during non-linear material deformations. We use an adaptation of the simple video prediction (SimVP) model implemented using the Python library PyTorch. The model uses mean squared error (MSE) loss, Adam optimiser and OneCycleLR scheduler. It is trained for 100 epochs with a batch size of 16 and initial learning rate of 1e-3. The encoder extracts the spatial features; the translator learns the spatiotemporal evolution and the decoder reconstructs the ground truth. As the SimVP model uses only convolutional layers, it is computationally simple while maintaining state-of-the-art performance. The approach is used to develop a surrogate model for microstructural evolution and strain fields in a representative volume element (RVE) of a hyperelastic material with heterogeneities (cavities) subject to varying internal pressure, as associated with the RGD phenomenon. The surrogate predictions were compared against synthetic data, generated using finite element method (FEM), for a range of cavity volume fractions, material parameters values (shear modulus, bulk modulus) and internal cavity pressures. Comparing the predictions made by the ML model and those made by FEM-based simulations, the ML approach achieved sufficient accuracy (with structural similarity index measure around 1.0 and MSE around 0.7), while reducing prediction time to 2 seconds compared to 83 seconds for the equivalent accuracy numerical method.