4.1.1 Implementation Details

The Monte-Carlo Ion Implantation simulator developed at the Institute for Microelectronics [56,57,58,59,60] is an implementation of this simulation methodology. It is fully integrated with the WAFER-STATE-SERVER and was re-implemented in C++ based on a previously developed FORTRAN implementation which was not WAFER-STATE conforming. The new implementation differs from the previous version in the following major points.

• The FORTRAN implementation takes as input a geometry that is stored on a PIF file. The output of the simulator is a dopant profile that is also stored on a PIF file. The simulator, however, completely ignores other attributes that are already existing in the input file. The histogram (damage information) that is used as input in consecutive invocations of the Monte-Carlo simulator, is stored on a separate file (not on a PIF file). This approach is not WAFER-STATE conforming since a topography simulation step can not be applied after a Monte-Carlo simulations without taking into account the extra (simulator specific) damage file.

  The new implementation manages all input and output data likewise. The damage information is stored as a Wafer attribute (i. e. on a grid) which is queried by the simulator at startup and written upon completion of the simulation. This is fully WAFER-STATE conforming in the sense that a topography altering process step does not need to take into account additional (simulator specific) data. Attributes that are existing on the input Wafer and ignored by the simulator are also present on the resulting output Wafer.

• The input PIF file for the FORTRAN simulator does not contain all informations necessary for a Monte-Carlo simulation. Therefore additional information has to be passed on the command line (e.g. the material type of a simulation domain) or is hard-coded (e.g. material density), respectively.

  The new implementation takes these informations from the input Wafer, where they are stored as constant attributes (as opposed to distributed attributes).

• In the FORTRAN implementation the simulation domain is discretized into leafs of a fixed size. Thus, the simulator has a fixed resolution that is determined at compile time^4.1. As a consequence the memory requirements are also fixed for every simulation problem. The resolution must be selected according to the most demanding geometry that is to be simulated. This is especially crucial if regions of largely differing aspect ratios have to be treated at the very same time (e.g. gate and bulk regions). A point bucket oct-tree, that associates leafs with a certain material type is used for the point location operation.

  The new implementation uses dynamic memory allocation and a variable resolution. The memory usage thereby is adapted to the simulation domain and is thus determined by the simulation problem. Geometry handling is completely delegated to the WAFER-STATE-SERVER. The geometry data are stored in a finite oct-tree which is also used to perform the point location operations (i. e. material locations). Therefore no point bucket oct-tree is necessary in the simulator.

• The most time consuming part of both implementations is the point location operation. Therefore, an optimization to minimize the number of point locations is implemented in both versions. The optimization estimates the number of collisions until an ion reaches a material interface and refrains from point locations during the determined number of collisions. The optimization must not overestimate this number or an error is introduced (one or more collisions with a wrong material type). Due to the lack of a geometry support in the FORTRAN implementation, the optimization is restricted to determining the distance to the surface within a leaf of the point bucket oct-tree.

  The C++ implementation uses the distance method of the WAFER-STATE-SERVER to compute the distance from a given point to the nearest interface. The implementation of this method uses a finite oct-tree that contains just the surface elements of each Wafer segment. Thereby, on the one hand side the number of searched elements is minimized. On the other hand side, the estimation of the number of collisions until an interface is reached is also better, thus a better reduction of point location operations is achieved.

• To achieve WAFER-STATE conformity the simulator has to store the resulting profile on an unstructured grid. This is realized by delegating meshing operations to the WAFER-STATE-SERVER (via the Gridder interface). Thereby, the WAFER-STATE-SERVER acts as a grid manager that is responsible to hold the grid data structures and to provide interpolation mechanisms to interpolate from the histogram information onto the tetrahedral grid elements. It is worth mentioning, that the main advantage of this new approach is that the simulator is capable of storing and reading polygonal information without having a built-in gridding mechanism.

The long simulation times of ion-implantation simulations and the comparably huge number of point location operations that are necessary during the simulation run, justify the usage of the oct-tree data structures for this application. Although the pre-processing time for building the oct-tree is much larger as with the alternative available jump-and-walk algorithm the total execution time is reduced.

2003-03-27