SEMICONDUCTOR
Every step a deviation
goes unexplained
adds to its cost.
Defects and drifts are often understood after more wafers and tool time have been spent. Xpdeep localizes the variables and temporal segments that drove the model output. Engineers can then evaluate whether that evidence supports an earlier process decision.
ILLUSTRATIVE PROCESS · THE FAB'S CONTROL SYSTEM REMAINS IN CONTROL
01 PROCESS STEP
The defect may be visible later.
The relevant process evidence may belong to an earlier step.
Semiconductor manufacturing chains combine process history, equipment state, recipes, temporal signals and inspection evidence.
Xpdeep can expose which variables and temporal windows carry the learned output.
ILLUSTRATIVE VARIABLES · NOT ALL PROCESSES
- PROCESS STEP N-2
- PROCESS STEP N-1
- PROCESS STEP N
- INSPECTION / OUTCOME
Xpdeep localizes the variables and temporal segments that drove the model output. Engineers can then evaluate whether that evidence supports an earlier process decision. Predictive evidence directs the investigation toward relevant process steps; it does not by itself establish the physical cause of a defect. Customer engineering investigates the physical cause.
02 LOCALIZATION
The model should show
which signal regions
support its output.
- RAW TEMPORAL SIGNAL
- LOCALIZED WINDOW
- STRUCTURAL DEPENDENCY
- MODEL OUTPUT
On temporal data, Xpdeep can localize not only which variable mattered, but the segment and morphology that carried the model output.
The model localizes the learned driver.
The engineer names the physical cause.
POTENTIAL USES
- Artefact localization
- Process-deviation localization
- Measurement correction
- Fault attribution support
INSPECTION
Post-hoc maps estimate spatial input influence.
Xpdeep reads it from the model’s own structure.
KNOWN CONDITION
STRUCTURAL EXPLANATION · DOMAIN STATE: KNOWN
CHANGED CONDITION
ERROR BREAKDOWN · DOMAIN STATE: OUTSIDE
Xpdeep can expose which learned structures drove the output and whether the image remains inside the model’s learned domain.
Structural explanation · attributed domain state · error breakdown where relevant.
Post-hoc saliency methods approximate how input regions influenced an output from outside the model, and remain useful for exploration. Xpdeep’s architectural difference is that spatial input influence is returned by the model itself, alongside its domain state. With authorized access to the PyTorch architecture and weights, Xpdeep can structurally analyse the customer’s or supplier’s learned inspection model.
Structural integration requires read access to the PyTorch architecture and weights. Supplier-owned models require the supplier’s cooperation.
TECHNICAL BOUNDARY: Image-level prescription: not available; under study.
03 DOMAIN
A process or inspection model
should know when the current state
has moved beyond what it learned.
TEMPORAL DOMAIN
Graded distance by variable and time window
INSPECTION DOMAIN
Attributed domain departure / novel-class detection where appropriate
The objective is not a generic confidence score.
It is an attributed domain state that identifies what drove the departure.
Novel-class detection applies to classification tasks. It is distinct from generic anomaly detection and from prediction confidence.
EXAMPLES
- Unseen process condition
- Changed equipment behavior
- New defect morphology
- New visual class
- Process drift outside the trained domain
04 INLINE DECISION
Move the decision closer to the step
associated with the learned evidence.
TEMPORAL / TABULAR STAGE
Supported temporal or tabular models can add a separate constrained prescription where applicable.
- PROCESS STATE
- PREDICT
- EXPLAIN
- CHECK DOMAIN
- PRESCRIBE WHERE APPLICABLE
- JUSTIFY
- ACT / HOLD / ESCALATE
VISION / INSPECTION STAGE
Structural explanation + domain state + downstream decision. Image-level prescription is not available.
- INSPECTION MODEL
- STRUCTURAL EXPLANATION
- DOMAIN STATE
- DOWNSTREAM DECISION
The surrounding manufacturing and control system remains in control.
The fab’s process-control, inspection and quality systems remain responsible for applying operational rules. A model-level explanation or domain signal does not automatically justify changing a recipe, sampling plan or production release decision.
QUALITY
Not every deviation
should trigger the same response.
ROUTINE STATES
→ remain inside the system
TRUE EXCEPTIONS
→ HOLD / ESCALATE → EXPERT
When the model can state both its decision context and its domain, human review can focus on the cases that genuinely require intervention.
POTENTIAL OPERATIONAL VALUE
- Fewer unnecessary reviews
- Faster exception handling
- More equipment / process steps per expert
- Earlier intervention
- Less decision latency
DESIGN
Catch model and data problems
before they reach the line.
- INITIAL MODEL
- READ LEARNED STRUCTURE
- DIAGNOSE PROBLEM REGION
- TARGETED LOCAL EDIT (NATIVE)
- VERIFY EFFECT
REST OF MODEL FROZEN
Xpdeep can expose variables, temporal structures or learned regions that are not carrying the intended decision, allowing targeted engineering intervention before deployment.
POTENTIAL VALUE
- Cleaner model structure
- Fewer unnecessary inputs
- Earlier detection of model / data design issues
- Better KPI alignment
05 YIELD ECONOMICS
The value is not another defect score.
It is an earlier, better-attributed decision.
ECONOMIC SURFACES TO MEASURE
YIELD
Less value lost before the issue is understood
QUALITY
Predictive evidence localized in process history
SCRAP / REWORK
Potentially fewer downstream losses
THROUGHPUT
Less delay from unnecessary review or stoppage
ENGINEERING TIME
Faster diagnosis of model / process issues
TEST / INSPECTION
Effort directed where the evidence and domain gaps point
No quantified improvement is claimed without application-specific evidence. Unit-specific performance grading, typical-versus-guaranteed specification recovery, targeted testing based on domain gaps and reduced margin accumulation across the supply chain remain hypotheses, not delivered semiconductor results.
DATA · VALIDATION · EVIDENCE
What the fab measures, and where the evidence stands.
RELEVANT DATA AND MODEL TYPES
- Process-step and equipment time series
- Recipe, metrology and electrical test data (tabular)
- Wafer and defect images for learned inspection models
WHAT THE CUSTOMER MEASURES OR VALIDATES
- Localization and domain signals against its own engineering investigations
- Detection performance and investigation time against agreed KPIs
- Any sampling-plan, recipe, grade or release change, through its qualification and quality agreements
- Integration with FDC, MES, APC and quality systems
PUBLISHABLE EVIDENCE AND MATURITY
- LIVE · TEMPORAL MODELS
- Signal-level localization, structural explanation, attributed domain awareness and prescription.
- ESTABLISHED · VISION MODELS
- Structural analysis, attributed domain state and design-loop diagnosis on learned vision models.
- PLATFORM
- Core engines, XpViz, XpAct and the platform layer are live.
- MODEL PERFORMANCE
- On time-series data, Xpdeep models achieve accuracy at least equivalent to standard deep learning and frequently superior.
- TO BE ESTABLISHED
- Semiconductor-specific yield, test reduction, grading and economic outcomes remain to be demonstrated against agreed KPIs, including the effect of process decisions on material loss. Generic Xpdeep benchmarks are not attributed to a semiconductor production programme.
INTEGRATION
Xpdeep changes the model.
Not your fab stack.
- EQUIPMENT / SENSORS / INSPECTION
- DIGITIZED PROCESS DATA
- XPDEEP MODEL
- FDC / MES / APC / CONTROLLER / QUALITY SYSTEM
Fault-detection and classification, inspection and process-control systems remain in place. Xpdeep provides structural model evidence and attributed domain state; temporal and tabular stages can also return explained constrained interventions.
No proprietary transport or logging framework imposed.
Semiconductor engineering boundaries.
EXPAND
- APC, FDC, MES, inspection systems and metrology remain in place.
- Structural integration requires read access to the PyTorch architecture and weights, or a Native model.
- Prescription applies to temporal and tabular stages; image-level prescription for inspection models is not available; under study.
- Predictive evidence does not by itself establish the physical origin of a defect.
- The customer’s surrounding system defines and authorizes actions.
- Attributed domain state covers the retained model inputs; alone, it does not authorize action or establish system safety.
- Structural justification does not automatically establish physical causality.
- Minimal means minimal feasible under the stated constraints and optimization procedure.
- Targeted editing and controlled correction require Native mode.
- System validation and certification remain with the customer and relevant authority.
- Reduced requalification scope requires application-specific evidence and acceptance.
START WITH ONE FUNDED SEMICONDUCTOR PROBLEM
Start where a deviation
is expensive to understand too late.
PROCESS DEVIATION
A temporal model tied to equipment or recipe behaviour where deviations are understood too late
DEFECT INVESTIGATION
A defect whose relevant process history is not yet localized
MODEL DIAGNOSIS
A process or inspection model engineering cannot yet explain
INSPECTION DOMAIN
A vision model that needs attributed domain awareness
YIELD
A yield-related decision with an agreed success measure
A first engagement may focus entirely on diagnosis and engineering value, without prescription or autonomous control.
Phase 1 is a funded engagement, typically three to six months, on a consequential model, decision, function or programme you have prioritized. Agree the economic, operational, quality, mission, product or assurance success criterion at scoping; work on your data within your engineering environment and constraints.
Xpdeep supplies model-level technology and repeatable software deployment, not a replacement engineering lifecycle. Your teams and implementation partners can integrate and validate the model using your own tools and test processes.
01 →
One funded problem
A model, engineering problem, product function, subsystem or decision the business already owns.
02 →
Agreed criterion
Data, duration, success criterion and validation method defined at scoping.
03 →
Evidence on your application
Results assessed through the customer’s validation process; a side-by-side model comparison is one possible method.
04
Define the extension
Identify the possible Phase 2 scope at the start; use Phase 1 evidence to determine its scope.