The purpose of Network Dismantling (ND) is to find an optimal set of nodes and removing these nodes can greatly decrease the network connectivity. However, current dismantling methods are mainly focus on traditional simple network which only consider the pairwise interaction between two nodes, while the hypernetwork, which can model higher order groupwise relation among arbitrary nodes, is more suitable for modeling real world. Due to the structural difference between simple network and hypernetwork, current dismantling methods cannot be directly applied to hypernetwork dismantling. Although some centrality measures in hypernetwork can be used for hypernetwork dismantling, they face the problem of banlancing effect and efficiency. Therefore, in this paper, we propose a novel HyperNetwork Dismantling methods based on hypergraph neural network, called HND. Specifically, our method first generates plenty of node ranking samples with the help of synthetic hypernetwork generator. Then, a node betweenness approximation model in hypernetwork is built based on hypergraph neural network. And this model is trained on those ranking samples generated in previous step until convergence. Finally, the well-trained model is utilized to approximate the nodes' betweenness in real world hypernetworks and further used for dismantling. To confirm the effectiveness of our method, we conduct extensive experiments on five real world hypernetworks. The experimental results demostrate that the HND outperforms various baselines.