Deep metric learning (DML) is designed to learn effective sample similarity metrics, which can improve the performance of various tasks, such as image retrieval and face recognition. In prior research, traditional DML methods predominantly rely on isotropic metrics, assuming uniform data distribution in all directions. However, in real-world data, this assumption often fails to hold, as data typically exhibit complex directional dependencies and manifold structures. To overcome these limitations, we propose NPFML: a Non-isotropic Potential Field based deep Metric Learning framework with Hierarchical Decay. The core of the NPFML lies in constructing a direction-sensitive model of attractive and repulsive forces. Unlike traditional isotropic metrics, this potential field model can differentiate sample relationships in different directions based on directional information of the data, thereby guiding the neural network to learn an embedding space with directional selectivity. We have conducted comprehensive evaluations on multiple widely used benchmark datasets. The experimental results clearly demonstrate that the proposed NPFML effectively maintains intra-class compactness while significantly enhancing inter-class separability.