next up previous contents
Next: 4.3 Package Modeling Interface Up: 4. Integration between Semiconductor Previous: 4.1 Introduction

Subsections


4.2 Interfaces

The different interfaces identified can be seen in Figure 4.1. The shaded areas in Figure 4.1 indicate the different interfaces between ``reality'' and ``simulation''.

Figure 4.1: Scheme of identified interfaces between TCAD and semiconductor fabrication

\includegraphics[origin=c,height=2.1\textwidth,clip=true]{figures/overall_interfaces.2.ps}

4.2.1 Between Design/Layout as Process Simulation Input

According to the work flow outlined in Chapter 3, the layout of the masks is one of the two main inputs for process simulation.
Normally this layout is available in GDSII-binary format which can be read by any of the above mentioned commercially available TCAD tools. To mirror the activities carried out in the mask shop this data must undergo the same transformations as in reality listed in Subject 4 of Section 2.1. The ``Mask Generation Instructions'' define special boolean operations to modify the mask data in a way that certain effects of wafer processing are cancelled out. Examples for possible corrections like simple mask biasing or proximity corrections are given in Section 3.3.2. An example for typical ``Mask Generation Instructions'' is given in Table 4.1. The mask levels are defined by the GDSII mask numbers given in the layout data file. The mask description identifies every generated reticle by name and is the reference for the lithography step in fabrication. The working plate field gives the data type for each mask. ``DK'' means dark field and ``LT'' means light field. The type gives an indication if the layout data has to be inverted for mask generation. Light field masks are masks which are generated as defined by the GDSII data. Dark field masks are generated with the inverted GDSII data. The contact mask is an example for a dark field mask. Since the contact mask has to be used for etching a hole (which is later filled with metal) into the oxide layer covering the silicon wafer, the photoresist covers all areas which are NOT contact. Therefore the GDSII data showing the contacts in the layout has to be inverted before mask generation. Figure 4.2 shows this situation.
Figure 4.2: Example for a contact layer GDSII mask layer information (which is of dark field type) (a), the corresponding chrome mask recticle and resist shape (neglecting proximity correction) (b) and the resulting etched contact holes in the oxide layer (c)
\begin{figure}\begin{center}
\par
\input{figures/contact_layer.pstex_t}
\par\end{center}\end{figure}
The biasing colummn give the bias to be applied to the GDSII data to get the final chrome mask dimensions. The strongest bias has to be applied to the mask level 10 listed in Table 4.1 which compensates for the effect shown in Figure 3.4 in Section 3.3.2. Since modern lithography steppers use 5x projection, the structures on the recticle are five times bigger than they are projected onto the photoresist. For mask quality control the critical dimensions of the chrome masks are defined in the ``reticle CD'' column in Table 4.1.

Table 4.1: Example of mask generation instruction table including biasing and CD (critical dimension) information
RESIST TYPE: POSITIVE
Mask Mask Description Working Biasing 5x
Level   Plate $ \mu m$/side Reticle
    Field (1x) CD ($ \mu m$)
5 n-Well Mask DK 0 3
8 p-Well Mask LT 0  
10 Active Area Mask LT +0.08 2.4
8 p-Well Mask LT 0  
24 VTP-Implant Mask DK 0  
48 Anti Punch Through Mask DK 0  
14 Gate Oxide LT 0  
20 Poly 1 Mask LT +0.02 2
29 High Resistive Mask LT 0  
30 Poly 2 Mask LT -0.04 3.0
21 n-LDD Mask DK 0  
24 p-LDD Mask DK 0  
23 N+ S/D Mask DK 0  
24 P+ S/D Mask DK 0  
34 Contact Mask DK +0.02 2.2
35 Metal1 Mask LT 0 2.5
36 Via 1 Mask DK 0 2.5
37 Metal 2 Mask LT 0 3
38 Via 2 Mask DK 0 2.5
39 Metal 3 Mask LT 0 3
41 Via 3 Mask DK 0 2.5
42 Top Metal Mask LT 0 3
40 Pad Mask DK 0  


Figure 4.3 shows the detailed structure of the identified interface as motivated by the real mask generation and lithography process. The GDSII data is converted into the ASCII formatted CIF format (for easier processing) This data is then subject to boolean and biasing operations as defined by the mask generation instructions. To emulate the real shape of the photoresist a proximity correction is applied and the resulting contours are written back to a CIF format serving as mask information input during the process simulation.

Figure 4.3: Interface of layout-data to simulation mask-data

\includegraphics[angle=0,origin=c,height=0.5\textheight,clip=true]{figures/layout_interface.ps}

A detailed description of the algorithm is given in Chapter 5.


4.2.2 Between Process Flow Description and Process Simulation Command File

The overall process flow information is typically documented in the database of a manufacturing execution system like PROMIS [149] from Brooks-PRI, SiView [150] from IBM, or WorkStream [151] from AppliedMaterials. The information relevant for process simulation inside this database is numerous, depending on the detailed level of the simulation models (as outlined in Table 3.1). Normally the following datasets are needed for process simulation:
  1. Sequence of process steps representing the semiconductor process flow (oxidation $ \Rightarrow$ layer thickness measurement $ \Rightarrow$ implantation $ \Rightarrow$ diffusion $ \Rightarrow$ material deposition $ \Rightarrow$ lithography etc.)

  2. Blocks of process sequences which are carried out on the same semiconductor fabrication equipment and are normally organized as program sequences like oxidation/diffusion programs which can consist of up to dozens of single process steps with different temperatures (temperature ramps) and gas ambients (gas steps or ramps).

  3. The detailed process parameter set of one single process step (e.g. ion species and composition, ion dose and energy, angle of incidence and rotational orientation of ion beam with respect to wafer, ion beam divergence, etc., for ion implantation)

  4. Positions in the full semiconductor process flow where selected physical characteristics like layer thickness or sheet resistances are measured by using metrology tools on wafer level.

Subject 1 through Subject 3 represent the hierarchy levels from highest to lowest. It is advisable to use an abstract representation of data as an intermediate format between the process step information and the simulation command file. Since commercial simulators and simulators from university are still under heavy development, syntax changes of the simulator command files are happening frequently. Therefore a direct translator between process flow information and simulation is inflexible and difficult to extend. With an intermediate abstract format different types of simulators can be supported by a single source of data (see Figure 4.4).

Figure 4.4: Interface of process flow information to process simulation command files

\includegraphics[angle=0,origin=c,width=1.4\textwidth,clip=true]{figures/process_interface.ps}

Typically the Manufacturing Execution System (MES) is not storing the full details of the diffusion and oxidation recipes, the plasma etch programs or the etch and deposition rates of the wet chemistry used during semiconductor processing. This information is stored in separate databases as outlined in Figure 4.4. Normally there is only a reference to a machine recipe or a etch sink given in the MES flow. Again the strategy of exactly mirroring the real situation was chosen to set up the interfaces. The recipes are transferred by a converter into the meta-syntax and are then converted into the simulator syntax of the process simulator chosen. This procedure is shown on the left hand side of Figure 4.4. The converted recipes are then transferred into a subroutine format, provided by every commercial or university simulator available. Thus this approach can be used for every TCAD environment available. Furthermore, through the use of a meta-syntax the interpretation of the data formats inside the manufacturing system (MES and recipe databases) and the conversion into the different simulator syntaxes can be treated in a more systematic way, since both tasks cannot interfere in a single program but are performed in a modular way. Last but not least the meta-syntax enables a very compact and concise overview about the details of the semiconductor process flow. Thus, this syntax can be used as a source for a very sophisticated process flow description for documentation and training (see Section 6.2).
The details of the implementation and examples for the meta-syntax used are given in Chapter 5.

4.2.3 Between Electrical Test and Device Simulation

After finished processing of the silicon wafers the first electrical test is the measurement of simple test structures and devices organized in Process Control Monitors (PCM) in the scribe-lines of the wafer. These measurements are carried out on automated tester systems on wafer level. The measurement procedures are again hierarchically oriented in the following way:

  1. Measurement program set up for actual technology node.

  2. Subprogram defined for actual PCM (normally several PCMs are inserted in the scribe-lines).

  3. Module for Device Under Test (DUT) consisting of single program statements measuring relevant electrical parameters.

  4. Single measurement algorithms for e.g. CMOS threshold voltage, or diffusion sheet-resistance.

  5. Single steps of carrying out the measurement algorithm for e.g. CMOS threshold voltage in saturation. These steps define how the device terminals are connected to the voltage and current sources of the automated tester and how the currents and voltages of the DUT are measured.

Subject 3 to Subject 5 are mirrored on the device simulation side to provide comparable electrical data of measurements and simulation.
Since the algorithms under Subject 4 and Subject 5 are not changing frequently (the algorithms under Subject 5 are fixed4.1 with the hardware of the automated tester used and measurement algorithms defined under Subject 4 are only changing, if a completely new device type is introduced) these algorithms are not converted on a daily basis. The structure of the interface is shown in Figure 4.5. The subprogram conversion of the DUT modules is carried out much more frequently on a daily basis.

Figure 4.5: Interface of electrical wafer acceptance test information to device simulation command files

\includegraphics[origin=c,width=1.40\textwidth,clip=true]{figures/wat_interface.rot.ps}


There are two main application areas existing for converting electrical test programs. First, this conversion is used for the automated generation of big device test-chips during process development including new device architectures. Second, the standard PCM structure measurement algorithms have to be converted to match the simulation results with the PCM measurements. Both methods are described in Chapter 5 in detail.


4.2.4 Between Device Characterization and Device Modeling (SPICE)

This interface deals with the generation of reliable device models for circuit modeling (e.g. SPICE). The main devices (NMOS/PMOS transistors for standard CMOS processes and, in addition, bipolar transistors for BiCMOS processes) of any new process fabrication must be characterized completely in terms of output characteristics, transfer characteristics, amplification, etc. This task results in scalable electrical models (BSIM3 for CMOS, VBIC for Bipolar transistors) or compact models for circuit simulation. In the TCAD fabrication integration scheme the source for this fitting procedure can be twofold: First, the usual way of measuring the characteristics on semiconductor wafer material and second, by simulating these characteristics with device simulation. The second approach has the enormous advantage of getting worst case predictions [152], [153], which are directly related to process parameter changes by applying statistical variations on selected semiconductor process step parameters (e.g. selected implantation doses). Furthermore, combined process and device simulations without the existence of any semiconductor material can generate preliminary models very early in the process development stage. Unfortunately the generation of SPICE models (e.g. BSIM3.3 or BSIM4 with hundreds of free parameters) is not an automated straightforward task. The process characterization engineers have to set up many initial values for starting the optimization of actual SPICE models and have to follow a complicated iterative strategy to get a good model with reasonable accuracy. Therefore an automated global optimizer for generating a good SPICE model is not available. Currently the only way to get so called ``TCAD based models'' is to generate characteristics with device simulation as they were measured on a real device and submit this information to process characterization engineers for the generation of SPICE models. Nevertheless, this approach enables the generation of a design environment of a new technology in a very early stage of a process development. The time to market for new process technologies is thus significantly reduced.



Footnotes

... fixed4.1
these algorithms are typically provided by the tester vendor in the form of test libraries

next up previous contents
Next: 4.3 Package Modeling Interface Up: 4. Integration between Semiconductor Previous: 4.1 Introduction

R. Minixhofer: Integrating Technology Simulation into the Semiconductor Manufacturing Environment