In this research project, an attempt is made to fuse the fields of structural mechanics and machine learning. The goal is to find out if models can be created that are capable of predicting the outcomes of (nonlinear) finite element analyses. These models are created by means of Artificial Neural Networks, which is a powerful method in the domain of machine learning. The focus will be on stiffened steel plated structures that are part of a sea lock gate. The power of a trained neural network is that it is able to compute the output for a given set of input parameters within a fraction of a second. Running a complete finite element analysis on the other hand can take a significant amount of time, especially in case of geometrically and/or physically nonlinear analyses. When relying on nonlinear finite element analyses for performing a structural design optimization, a trained network can therefore save a huge amount of time. It also allows to evaluate many more design options, possibly finding a more optimal design than what would be possible with a manual design optimization. An automated procedure has been created to generate datasets by running FE analyses in batch mode. Parametric models are set up in ANSYS FE software, for which random sets of input parameters are generated. After running the analyses, the output is collected and organised in datasets that can be used for training. The sizes of the datasets and the dimensionality of the design spaces are varied in order to study the influence of these quantities on the accuracies of the predictions produced by the neural networks. Genetic algorithms, which is another machine learning technique, are deployed for the optimization of the \textit{hyperparameters} of the neural networks, which are basically the \textit{settings} of the network which determine the learning behaviour. Three standard interpolation techniques (Kriging and polynomial interpolation) are also fitted to the same datasets in order to compare the performance of the neural networks to these interpolation techniques. The final result is an overview of the accuracies of the predictions made with the neural networks on validation datasets. It was found that the neural networks produced accurate predictions on the maximum deflection of a simply supported, stiffened steel plate loaded by a uniform pressure. Most of the relative errors were within a range of 5\% error for design problems with 4 dimensions. Predictions of the linear buckling load of stiffened steel plates were found to be mostly within the range of 10\% error for design problems with 4 dimensions. When increasing the number of free design variables from 4 to 8, the errors were found to be mostly within the range of 20\% error. The predictions of maximum equivalent stresses in stiffened steel plated sections obtained by geometrically nonlinear FE analyses were found to be in the range of $\pm$ 20\%. These models had 13 free design variables. Improvements were made on the available options of hyperparameters for neural networks. With these improvements, new predictions were made on the maximum equivalent stresses and the accuracies were found to be slightly better, with errors ranging between $\pm$ 15\%. For implementation of the predictive model in a design optimization algorithm, these errors are considered to be too high. It is expected that the dimensionality of this problem (13 design parameters) combined with very irregular results due to the presence of peak stresses and different buckling shapes, resulted in these deviations. Additional datasets are generated with results of FE analyses of simplified, unstiffened steel plates. Both geometrically and physically nonlinear analyses are performed, with uni-directional compression applied to the plate edge directly as a displacement. The number of free design variables was set equal to either 1 or 4 free variables. It was found that even with small datasets (with 32 training samples), the neural networks produced very accurate results on predicting maximum equivalent stresses, maximum equivalent mechanical strains (in case of physically nonlinear analyses) and total reaction force. A neural network was found to be capable of producing a nonlinear load-displacement curve of a compressed rectangular plate with elastic-plastic material model.