Defect classification API

Defect patterns, classified.
Yield decisions, accelerated.

Wafertune identifies defect signatures in your wafer maps using a purpose-trained ML model — results in your data pipeline in under 3 seconds.

Request Pilot Access Explore the Pattern Library

Annotated wafer map showing classified defect patterns including ring exclusion, linear scratch, and random cluster signatures

180+ defect pattern classes in model taxonomy

< 3s average classification latency per wafer map

4 node families — analog, power, RF, MEMS — trained on specialty data

How it works

Three steps. One API call.

Wafertune plugs into your existing data pipeline without additional infrastructure.

Send your wafer map

STDF, CSV grid, or image upload via REST API. No preprocessing required — Wafertune handles format normalization.

Model classifies patterns

Convolutional spatial analysis runs against Wafertune's 180+ pattern taxonomy. Confidence scores assigned to each detected signature.

Structured response, your pipeline

JSON labels, confidence scores, and spatial coordinates — ready for your yield management system or custom analysis workflow.

See the full technical walkthrough

Pattern taxonomy

From the model's training data

Each pattern class is documented with its spatial signature, process origin, and confidence characteristics.

RING_EDGE_EXCL

Ring Edge Exclusion

Edge

Ring of flagged dies at wafer periphery. Edge bevel contamination or edge-focused etch profile.

SCRATCH_LINEAR

Linear Scratch

Linear

Diagonal line of flagged dies. Wafer handling artifact or CMP pad scratch mechanism.

CLUSTER_RANDOM

Random Cluster

Clustered

Localized cluster of failed dies without spatial regularity. Particle fallout during deposition or etch.

LITHO_REPEAT

Litho Repeater Pattern

Periodic

Periodic repeating defect pattern aligned to reticle field boundaries. Reticle contamination signature.

See all 180+ patterns in the library

POST /v1/classify

# Request
POST https://api.wafertune.com/v1/classify
Authorization: Bearer wft_sk_live_xxxx
Content-Type: application/json

{
  "wafer_id": "LOT42_W03_EWS",
  "map_data": "<base64_stdf_payload>",
  "format": "stdf"
}

# Response (avg 2.4s)
{
  "wafer_id": "LOT42_W03_EWS",
  "classification_time_ms": 2381,
  "pattern_classes": [
    {
      "class_id": "RING_EDGE_EXCL",
      "confidence": 0.94,
      "bbox": [0, 0, 300, 300],
      "process_origin_hint": "edge_bevel"
    }
  ],
  "yield_risk_score": 0.72
}

See full API reference

REST API — no SDK required. Standard HTTP POST with JSON body. Works with any language or HTTP client.
API key auth with RBAC scopes. Separate classify, batch, and manage permissions. Per-team key management.
3-second SLA on classification. P99 under 4 seconds. Batch endpoint for multi-wafer lots.
Works with STDF, CSV, and PNG wafer maps. Wafertune normalizes format differences so your pipeline doesn't have to.

Why specialty fabs

Built for the fabs most analytics tools ignore.

Most yield analytics platforms were built around leading-edge logic and memory. The 200mm analog fab running BCD process, the MEMS foundry, the power-device fab — these have different defect physics, different inspection workflows, and different data formats.

Wafertune's model was trained specifically on specialty-node defect signatures. Ring exclusion patterns, edge bevel contamination, LDMOS gate oxide scratches — patterns that generic models misclassify or miss entirely.

See supported node types

Stylized map of Arizona showing semiconductor fab cluster locations in the Phoenix metro area

"Defect classification was a manual step everyone in the fab accepted as slow. We thought: if the model is trained on the right patterns, it should be faster and more consistent than a human review."

Jonas Falk, Founder — Jonas built Wafertune after spending five years on wafer defect analysis tools in computational imaging research.

From the Research