Introduction
Semiconductor fabrication is an immensely complex process, often involving hundreds of sequential steps such as etching, doping, and oxidation. While industry-standard numerical simulators like SUPREM provide high-precision data, they often function as "black boxes" regarding causality. For instance, if a finished wafer exhibits abnormally high resistance, a numerical simulator cannot easily pinpoint which specific step in the 200-step sequence was responsible. This paper introduces a qualitative simulation framework that moves beyond traditional numerical modeling by representing manufacturing steps as symbolic actions. This system generates a "wafer history" that explicitly links final device attributes to specific processing steps, enabling automated diagnosis of manufacturing failures and more intuitive recipe design.
Methodology
The authors developed a symbolic simulator rooted in techniques from AI planning and qualitative physics. Instead of a complex 3D mesh, the wafer is modeled as a 2D cross-section divided into vertical strips. Each strip consists of a sequence of 1D layers with attributes like material, thickness, and dopant concentration.
The simulator utilizes an extension of discrete action languages (similar to STRIPS) to handle:
- Conditional Effects: e.g., "If the material is silicon, then the thickness decreases."
- Universal Quantification: Applying effects to all layers simultaneously.
- Relative Changes: Describing effects qualitatively as "increases" or "decreases" rather than using precise numerical values.
As the simulation runs, it records a "wafer history"—a dependency graph that connects every structural attribute of the final chip back to the specific parameters of the operations that created or modified it.
Validation and Diagnostic Pruning
The model was validated using a real-world 48-step recipe for an oxide-isolated bipolar process. A key feature of the system is its ability to perform "diagnostic pruning" using specific geometric configurations called CUS (Collector-Under-Silicon) and CUO (Collector-Under-Oxide) test structures.
By comparing measurements from different test structures on the same wafer, the system uses its causal knowledge to eliminate suspected faulty parameters. If a normal measurement in one area depends on a specific parameter, that parameter is likely not the cause of a failure in another area, allowing the system to "exonerate" certain process factors and narrow down potential defects.
Key Results
The system successfully simulated the complex bipolar process, generating a graphical representation of the resulting NPN transistor. Beyond simple simulation, the tool can generate symbolic expressions for attributes like layer thickness. By computing partial derivatives of these expressions, the system performs automated sensitivity analysis, helping engineers prioritize diagnostic hypotheses based on which parameters a measurement is most sensitive to.
Limitations and Future Work
While promising, the current model is primarily 1D/2D and ignores lateral effects like "encroachment," which are critical in sub-micron processes. Additionally, dopant concentrations are modeled as constant step functions rather than the continuous gradients found in real-world profiles. Future work aims to automate knowledge acquisition and expand the model to handle "gross errors," such as skipped or repeated process steps.
Summary updated