Domain-continual learning for expanding the design space of deep generative modelling in nonlinear analysis of masonry

Application of conditional generative adversarial network (cGAN) offers a promising approach for predicting the nonlinear behaviour of masonry. However, the large variability in masonry’s mechanical properties makes developing comprehensive models highly time- and resource-intensive. This paper presents a continual learning (CL) approach to expand the predictive capabilities of a pre-trained cGAN model, designed to predict full mechanical response fields of masonry panels, to new domains of unseen material property combinations. Elastic weight consolidation (EWC) regularisation is adopted to mitigate catastrophic forgetting in the initial training domain. The effects of fine-tuning hyperparameters, trainable blocks, and fine-tuning subset configurations, are investigated to optimise fine-tuning performance. The fine-tuned model demonstrates excellent capability in predicting the strain maps and reaction forces and capturing extreme strain values within the expanded domain, while avoiding catastrophic forgetting. This approach outperforms costly full re-training from scratch, demonstrating a viable and computationally efficient solution for extending the generalisation capabilities of data-driven models. • Expanding a pre-trained cGAN’s design domain to unseen material properties in masonry analysis using continual learning. • Mitigating catastrophic forgetting of previously learnt knowledge through elastic weight consolidation during fine-tuning. • Optimising fine-tuning process through investigating fine-tuning hyperparameters and trainable blocks. • Identifying a practical number of fine-tuning cases necessary for achieving robust generalisation among the expanded domain. • Demonstrating the efficiency of fine-tuning over full re-training in terms of computational cost and predictive accuracy.