This paper presents a novel post-earthquake damage assessment model employing pre-trained convolutional neural networks (CNN) to automate local element structural failure mode prediction in low-to-medium rise reinforced concrete (RC) buildings using image-based data. To address the prevalent challenge of insufficient real-world datasets within earthquake and structural engineering domain to train pre-trained CNN models, this research innovatively integrates newly generated data, composed of tailored data augmentation and synthetic data generation techniques, to enhance the dataset size and diversity. Multiple pre-trained CNN architectures, such as MobileNet, DenseNet, GoogleNet, Xception, AlexNet, ResNet and VGGNet, were systematically evaluated, with MobileNet demonstrating optimal fittingness to the target dataset relevant to this research. The chosen CNN architecture’s hyper-parameters were meticulously optimised through leveraging model regularisation techniques informed by an in-depth analysis conducted, uniquely aligning the model with the dataset-specific characteristics. The resulting novel model, titled ‘FailureModeNet’, demonstrated a strong ability to predict local element structural failure modes of low-to-medium rise RC buildings with all four performance indicators (accuracy, precision, recall and F1 score) exceeds 0.89 (89%), with superior domain generalisability to new and unseen data without overfitting, confirming the applicability of this model in diverse post-disaster scenarios. Overall, the model presented in this paper serves as a testament to the potential of artificial intelligence in driving interdisciplinary research, paving the way for robust and actionable solutions to enhance post-earthquake disaster resilience and management.