5 Void evolution and Interconnect Resistance Change Models

The description of the time period up to void nucleation must be directly followed by a model which handles void evolution. Once a void is nucleated, material transport on the void metal interface dominates [54,64,60,61,63]. The driving forces are electromigration proportional to the local tangential component of the electric field and the chemical potential. Since the most critical voids are nucleated at the copper/barrier and copper/cap-layer interfaces, which also represent the fastest diffusion paths, available models [64,60] have to be extended in order to include these phenomena.
Void evolution was originally handled applying explicit void surface tracking methods [64]. A further development was introducing numerically efficient diffuse interface models [56,57,58,59]. However, not until introducing an adaptive mesh approach into the diffuse interface method [62], realistic interconnect structures could be handled.
Applying a diffuse interface model we are able to predict the intrinsic void behavior from the nucleation until failure. The simulation of the evolving void/metal interface is accompanied with electrical field calculation and resistance extraction.

Once the void is nucleated, the dominating factor for the void evolution is the self-diffusion on the void surface and in the vicinity of this surface. The driving forces are again the chemical potential and the tangential component of the electrical field gradient. The chemical potential in this case is comprised of components due to the free energy of the surface and the local elastic strain energy. The material is transported along the void surface and into the metal bulk. This effect gives rise to the movement of the whole void and it's shape change. In this context the crystal orientation must be considered because of the anisotropic characteristics of self-diffusion [54].

Since the problem is a moving boundary problem, the first attempts to handle it was the application of explicit tracking of the void/metal interface by means of finite element methods [64,60]. These methods are quite complicated to implement and also have rather poor numerical stability. To overcome the problems mentioned above, we apply a diffuse interface model. In a diffuse interface model void and metal area are presented through an order parameter $\phi$ which takes values $+1$ in the metal area, $-1$ in the void area and $-1 < \phi < +1$ in the void-metal interface area. Demanding explicit tracking of the void-metal interface is not needed and models do not require boundary conditions to be enforced at the moving boundary.

The evolving void/metal interface is described through the time dependent distribution of the order parameter $\phi$. The actual diffuse interface is a belt area around the iso-surface $\phi=0$. This belt area has some finite width. With shrinking width of this belt area the accuracy of the simulation increases. The diffuse interface governing equation (DIGE) is a central part of each diffuse interface based modeling. The equation itself comprises electromigration and stress driven material transport on the void/metal interface and on this basis regulates the distribution of the order parameter. This equation is an extension of the well-known Chan-Hilliard equation [56,57,58,85,86]. Concerning the solution of Chan-Hilliard type equations several approach are proposed in the literature [85,86]. The gap between fast and reliable DIGE solving approaches which are already known from the literature and a solution approach which is also applicable for arbitrary interconnect geometries was bridged by introducing an adaptive grid technique [62].

The main idea of the adaptive grid method was to use local grid refinement and a coarsening algorithm, a fine grid is set dynamically adjusted only in the vicinity of the void/metal interface. The geometry of the interconnect can be arbitrary and, due to the economical distribution of the grid density, computation can be performed efficiently.

Additional equations to be solved such as the Laplace equation for the electric field and mechanical stress-strain model equations are adapted to the diffuse interface model through an order parameter dependence of the electrical conductivity and the elastic compliance [59]. These equations are solved using common finite element schemes assuming a finite width of the diffuse interface area. Generally the accuracy of the calculation will increase by reducing this width. There is a rigorous asymptotic theory which proves that the diffuse interface model converges to the sharp void/interface physical situation both in the equations' solution and in the compliance with boundary conditions [56,57,58,59].

The initial void is placed on the spot predicted by the first stage module and the second stage module is invoked, which simulates evolution of the void under the impact of driving forces, until it reaches a critical size which changes the initial interconnect resistivity to such an extent that it can be considered as actual interconnect failure.
Thereby the resistance change up to the interconnect failure and, the time needed for this failure to occur are provided. We call this time period void evolution time and added to the void nucleation time calculated in the first stage module, it defines the interconnect time to failure.