论文部分内容阅读
Wire pull tests are generally conducted to assess the wire bonding quality. Using Cu/low-k technology, two failure modes are usually observed: Failure in the neck of the wire (neck break) and metal peel off (MPO). The objective of our study is to investigate the root cause of metal peel off by using a combined experimental and numerical approach. First, dedicated failure analyses are conducted to identify the failure locations. Scanning Electron Microscopy analysis for a large number of completely failed samples shows that the delaminated interface, after MPO has occurred, is always the interface on top of the third metal layer, near the centre of the stack. However, these inspections do not indicate where and when MPO initiated. To understand the initiation, incremental (non-destructive) wire pull tests are used. These samples have not failed completely, but may already show the initiating crack in either of the two possible regions. Combined with use of scanning acoustic tomography (SCAT) and focused ion beam (FIB) show that MPO initiates by delamination in the back-end structure at the interface on top of the third metal layer. Secondly, a 3D FEM model for a half bond pad with partial wire bond is used to simulate the wire pull test, in order to understand the failure mode. Analyses using stress as a failure index indicates, however, that the top interface is the most critical one. This does not match with the experimental observation. Therefore, an alternative, energy-based failure index is used, the so-called area release energy method (ARE). The ARE method approximately identifies the same critical interface as found from experiments. It is assumed that the presence of a stiff layer nearby a potential crack location restricts the elastic deformation upon release. This indicates that the initiation of a crack in the upper and lower interfaces results in the release of a lower amount of energy when compared to an interface in the centre, where the surrounding material constrains the release less. A 2D test model confirms this assumption.