1.1.3 Ion Implantation

  For the fabrication of integrated devices it is essential to introduce electrically charged particles into the target materials. The most common method is to scan a high energy ionized particle beam over the target material. The ion must have enough energy to penetrate the surface regions and loses then energy by scattering processes with the target lattice atoms or their interatomic potential, until it comes to rest in the lattice. Because of the high forces involved the target lattice is damaged by the implantation process. For the simulation of the subsequent annealing and isolation processes a reliable description of the as-implanted dopant profile as well as the damage generated defects is required. Therefore, the simulation of the ion implantation process is the first non-structural simulation task in the process simulation queue and fundamental for any shallow junction engineering. The rapidly diminishing thermal budget available to maintain shallow junction depths, however, places severe limitations on the junction profile prior to annealing. Generally speaking, the formation of shallow p -n junctions is more difficult than the formation of n -p junctions because the most commonly used p-type dopant, boron, has a high diffusivity and exhibits strong channeling tendencies. To avoid this channeling effect, preamorphization can be applied by implanting heavy ions such as silicon, or a thin oxide layer is deposited to randomize the ions. For both cases the radiation damage associated with a common implantation sequence must be removed to ensure good junction quality[Tan89]. Unfortunately, dopant diffusion during damage annealing limits the extent of this anneal. The tradeoff between dopant diffusion and damage removal is the most severe restriction for ion implantation in forming shallow junctions.

In recent years there were only two modeling approaches successful in the sense of computational efficiency and predictability of the doping profiles. First, the analytical description of the doping profiles based on statistical moments and distribution functions [Kri78] [Wil80] [Lor90], and second a Monte Carlo method based on statistical mechanics and atomic target-ion interactions [Bie84] [Gil86a] [Hob86] [Bar87] [Moo88]. Early efforts in the development of physically based implantation models which include the LSS theory [Lin61] and a full solution of the Boltzmann transport equation [Gil86b] did not gain any significance. The analytical method usually has the advantage of less CPU-time consumption in two-dimensions and even in three-dimensions. Thereby one-dimensional analytical distribution functions are combined to multi-dimensional profiles by a convolution method taking into account a dose matching rule and special numerical scaling algorithms to capture multilayered simulation structures (see Chapter 2). The Monte Carlo method is mainly used for the simulation of critical ion implantation steps. Due to the enormous CPU-time consumption the applicability is limited to TCAD systems and optimization loops, but it is not necessary to simulate every implantation step with a Monte Carlo simulator. There is a tradeoff between accuracy and simulation time. The extension of Monte Carlo simulators onto three-dimension simulation domains gives the opportunity to investigate several three-dimensional effects, like the dopant distribution near mask or oxide corners [Boh95]. The basic Monte Carlo algorithm is a follow each ion algorithm, where each single ion is traced on its trajectory through the target. The energy loss is calculated by nuclear and electronic stopping laws. If the ion has not enough energy to travel on, it comes to rest at a certain lattice position. One can imagine that the computational effort is extremely high to calculate one million of ion trajectories to get reasonable results even in three dimensions. To improve the speed of three-dimensional Monte Carlo simulators several simulation techniques were developed. The superposition method decreases the number of collision events by copying trajectories. This is only valid for amorphous targets, using crystalline targets each ion must be calculated, individually. As the basic Monte Carlo algorithm is always three-dimensional an appropriate geometry handling must be provided to detect material intersection with the ions trajectories. Octrees are commonly used [Sti93b]. To account for damage induced point defects within the crystalline target is also important for subsequent annealing steps. The most popular approach is the Kinchin-Pease model [Kin55] which calculates the number of generated lattice defects by probability methods. Unfortunately, presently there is no experimental way to measure the exact generated point defect distributions at the lattice sites. With the Rutherford Back-Scattering method (RBS) it is possible to determine the degree of damage after implantation, but the local defect concentration cannot be traced. These defect distributions are affecting the setup of the subsequent annealing process, however, the predictability of these models is rather poor and the amount of generated point defects overestimated. Realistic defect generation models should also include damage accumulation and defect recombination due to wafer self heating effects.