Manufacturing · SteelData Scientist · Noodle.ai PQAINov 2019 – Apr 2023
Steel Coil Mechanical Property Precognition
Machine learning and deep learning models predicting Yield Strength, Tensile Strength, and Elongation from alloy composition and process recipes — enabling recipe simulation to slash quality variability.
20–40%
Quality Variability
Mechanical property stabilization
Predictive
Recipe Simulation
Prototype before physical run
// THE PROBLEM
Problem formulation
Steel coil mechanical properties drift batch-to-batch, causing customer rejections and rework. Physical testing is slow and expensive, making trial-based recipe optimization impractical.
// DATA FOUNDATIONS
Data engineering & ingestion
- Aggregated ladle chemistry, rolling schedules, and mechanical test results
- Aligned discrete and continuous process variables into unified prediction windows
- Engineered features for alloy composition, thermal history, and deformation parameters
// MODELING STRATEGY
Mathematical & algorithmic approach
- Random Forest regression as interpretable baseline for Yield, Tensile, Elongation
- Deep Neural Networks with Transfer Learning to adapt across alloy families
- Hyperparameter search with cross-validation on plant-level holdouts
// EXPLAINABILITY
Interpretability & governance
- SHAP summary plots identifying dominant metallurgical drivers
- PDP curves quantifying how carbon, manganese, and thermal variables shift properties
- Uncertainty surfaced for out-of-distribution alloy compositions
// PRODUCTION
Deployment & integration
- Batch inference enabling 'what-if' recipe simulation before production
- Integration with planning systems to flag high-risk product mixes
- Monitoring dashboards tracking prediction drift against physical test feedback
// TECH STACK
PythonPyTorchTensorFlowTransfer LearningSHAPPDPPandasNumPy