Live demonstration

Engine Cylinder Block · KE-CYL-V4-220

FG260 grey iron, IS 210 spec · primary customer OEM-TATA · slide the parameters to see predictions change.

— ms

Predicted Defect Risk

Probability this heat results in any defect class

0%
Low risk

Most likely defect

prediction loading…
Predicted disposition
OK
Yield: —%  ·  Severity:

Business impact (expected per casting) ? How each number is computed
  • Scrap cost: (P(Scrap) + 0.4 × P(Major_Rework)) × ₹6,500
    i.e. probability this casting ends up scrapped × scrap-cost-per-casting. The 0.4× factor accounts for ~40% of Major Rework cases that get scrapped instead of reworked when the cost-benefit doesn't justify it.
  • Rework cost: (P(Minor_Rework) + 0.6 × P(Major_Rework)) × ₹1,800
    Minor + the remaining 60% of Major Rework that goes through the rework loop. Cost includes inspection, machining time, requalification.
  • Delay cost: expected_delay_min × ₹450/min
    Expected delay = 8 min × P(Minor) + 18 min × P(Major) + 35 min × P(Scrap). Captures downstream line stoppage at machining/assembly.
  • OEM complaint: P(complaint_escape) × ₹85,000
    P(complaint_escape) ≈ 5% × P(any defect) + 5% × (warranty_risk/10). Cost includes logistics, investigation, customer concession per incident.
  • Warranty reserve: (warranty_risk / 10) × 0.04 × ₹4,50,000
    4% of complaints escalate to warranty claims. Each claim costs ₹4.5 L (part replacement, field service, brand exposure).
  • Total expected loss: sum of all five rows above.
All assumptions live in data/cost_assumptions.json — editable per plant. See full breakdown at /assumptions.

Scrap cost₹0
Rework cost₹0
Delay cost₹0
OEM complaint₹0
Warranty reserve₹0
Total expected loss₹0

Root cause (SHAP attribution)

For predicted defect:
Adjust a slider to see attributions.

Recommended corrective actions

Predictions will surface targeted actions here.

Warranty risk ? What this score predicts
The probability that this specific casting, once shipped, will trigger a warranty claim in the field within 24 months — i.e., a customer-side failure that costs ₹4.5 L per claim in part replacement, field service, and brand exposure.

What drives the score
  • Chemistry: Cr > 0.10% and P > 0.06% promote brittleness & hot cracking that show up months later. Mn/S violations leave free FeS at grain boundaries.
  • Dimensional capability: bore diameter drift from the 89.500 mm target — even within spec — accelerates ring-pack wear and head-gasket failures.
  • Cooling rate & superheat: off-optimum thermal history → residual stresses → field cold-cracks.
  • Predicted severity: heats classified as Major_Rework or Scrap carry baseline warranty risk even if reworked.
How to read the scale
0–2 safe · 3–5 monitor (inspect twice) · 6–8 high — hold for QC review · 9–10 field failure likely — scrap rather than ship.

Model: LightGBM regressor trained on 5,000 heats with R² 0.73 against synthetic warranty outcomes.

0–10 scale · probability this casting becomes a field warranty claim

0 · safe5 · monitor10 · field-failure likely
— / 10

If your plant ran 14,000 castings at these conditions →

This setpoint extrapolated
per-casting × 14,000 / yr
AI-optimal extrapolated
≈ ₹2.5 Cr
~₹1,800 × 14,000 / yr
Gap closed by AI
improvement if all heats matched AI-optimal
Note: these numbers are worst-case extrapolations — they assume every heat for the year looked like this one. Your real annual loss is a mix of good and bad heats. The point is the per-casting delta: every heat you steer toward AI-optimal recovers ~₹(this − 1,800) of expected loss.