Causal AI Risk Modeling for Faster Semiconductor Defect Diagnosis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Modern semiconductor manufacturing generates vast amounts of data with natural dependencies, making it time-consuming to detect and diagnose manufacturing defects and perturbations caused by malicious actors.
Innovation Solution
A computer-implemented method and system using Causal Relational Artificial Intelligence (CRAI) and a risk framework to predict causal effects by identifying interventions, collecting process dependency data, creating intervention models, and training CRAI models to determine intervention efficacy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional defect detection methods are used in semiconductor manufacturing, then comprehensive analysis can be performed on vast amounts of data, but the process becomes extremely time-consuming and arduous
Solution Approach 1:
The patent introduces a causal model as an intermediary between raw manufacturing data and defect diagnosis. This causal model encodes domain knowledge about process-defect relationships, acting as a mediator that guides the analysis process. The causal model includes nodes representing process parameters and defects, with edges representing causal relationships, enabling focused analysis rather than exhaustive search through all data.
Solution Approach 2:
The patent performs preliminary action by pre-defining causal models and dependency graphs before actual defect detection occurs. These models are constructed in advance based on manufacturing domain knowledge, specifying potential causal relationships between process parameters and defects. When a defect is detected, the pre-established causal model immediately provides a framework for analysis, eliminating the need to build analysis structures during the diagnostic process.
2Reliability
If comprehensive data analysis is performed to detect manufacturing defects, then detection thoroughness is improved, but the complexity of the detection system increases
Solution Approach 1:
The patent segments the complex detection system into distinct modular components: data collection modules for gathering manufacturing data, causal model construction modules for building dependency graphs, analysis modules for evaluating causal relationships, and diagnosis modules for identifying root causes. Each module handles a specific aspect of the detection process, making the overall system more manageable and maintainable while preserving comprehensive analysis capabilities.
Solution Approach 2:
The patent changes parameters by transforming raw manufacturing data into structured causal representations. Instead of analyzing raw data directly, the system transforms data into causal graphs where nodes and edges represent meaningful relationships. This parameter transformation simplifies the analysis by working with structured representations rather than raw data, reducing system complexity while maintaining detection reliability.
3Reliability
If traditional perturbation analysis methods are used, then security issues can be identified, but the effort required to diagnose malicious perturbations is extremely time-consuming
Solution Approach 1:
The patent implements feedback mechanisms where the causal model continuously learns from detected perturbations and updates its structure. When malicious perturbations are detected, the system uses feedback to refine the causal relationships and improve future detection. The feedback loop includes comparing predicted outcomes from the causal model with actual observed outcomes, identifying discrepancies, and updating the model accordingly, thereby improving both detection accuracy and speed over time.
Data Source
AI summary
In an approach to CRAI and risk framework for manufacturing applications, there is thus provided a computer-implemented method for causal effect prediction, the computer-implemented method including: identifying, by one or more computer processors, an intervention, wherein the intervention is selected from the group consisting of threats, failures, corrections, and relevant outputs; collecting, by the one or more computer processors, process dependency data; creating, by the one or more computer processors, an intervention model; combining, by the one or more computer processors, the process dependency data and the intervention model to create a combined process dependency graph; training, by the one or more computer processors, a causal relational artificial intelligence (CRAI) model; and determining, by the one or more computer processors, an estimate of an intervention efficacy.