The children cannot recognize the Arabic script handwriting because the recognizer is intolerant of high inter-individual variations. The issue is further compounded by other problems such as irregular strokes, interrupted forms and inconsistent aspect ratios. The study explores which CNN architectures are most suitable for robustly recognizing black-and-white Arabic letters and digits drawn from kids, aged 5-10. The dataset comprises 570 000 characters images. The six well known CNN architectures are LeNet 5, AlexNet VGG16, GoogLeNet, DenseNet, and ResNet50. In order for the experiments to be reproducible and easily verifiable, we used a supercomputer for all models training and tests. The ResNet50 model was shown to perform best of all models with a validation accuracy of 99.86%, a global F1 score of 99.89%, validation loss of 6% and 0.96 GFLOPS. Along with benchmarking, the proposed work provides optimized lightweight CNNs for 64×64 grayscale images of children’s handwritten Arabic characters. The suggested model achieves a recognition accuracy of 98.3% at a cost 41% lower than VGG16, while drastically reducing the number of parameters. According to the study, modeling various handwritten text types can benefit from residual learning. LeNet-5 and other lightweight models have demonstrated good performance with less processing power and can be applied to embedded systems. The results indicate that children's handwriting can be automatically analyzed using the improved CNNs. Additionally, they demonstrate the applicability of these CNNs in digital assessment systems, educational technologies, and multilingual writing processing. The study offers an AI method for interpretability and robustness and lays the foundation for hybrid CNN-Transformer models.