On the CIFAR-10 (Canadian Institute for Advanced Research-10), ImageNet, and Penn Treebank datasets, Neural Architecture Search (NAS) algorithms obtained better results by computerizing the process of architectural design on the CIFAR-10, ImageNet, and Penn Treebank datasets. Even though the search time has been simplified, search algorithms count on performance prediction or controllers. When used on a new task with a fresh dataset, this may necessitate optimal structuring. The problem of architecture search is not solved because this is done by hand. Using continuous relaxation and gradient descent methods, Differentiable Architecture Search (DARTS) [1] avoids this issue. There are, however, plenty of intriguing methods to make DARTS better. In this paper, we first split the DARTS’s supernet into three (03) sub-supernets and applied neural message passing so that each node in the graph has information from other nodes. The three (03) sub-supernets by using gradient descent to find the best graph represent the whole search space. By adding parameter sharing and transfer learning, our method enhances the final accuracy of one-shot-based DARTS systems consistently. On CIFAR-10, it achieves 98.25% test set accuracy, according to the results.