Hot Stamping Intelligence — Tensile Strength Prediction via ElasticNet Regression

"You don't need to know which regularisation to use — you need a framework that finds out for you."

---

🎯 Business Problem

Hot stamping — press hardening a high-strength steel blank above 900°C and quenching it inside the die — is the dominant process for automotive structural safety parts: B-pillars, door reinforcements, bumper beams. The tensile strength of the finished part is determined by a complex interaction of thermal, mechanical, and material variables.

The process engineer faces a choice no textbook resolves cleanly: the data has some correlated variables (furnace entry and exit temperatures track each other at r = 0.95) and some genuinely irrelevant ones. The right regularisation depends on which problem dominates — and that is not known in advance.

Lasso zeroes redundant features. Ridge stabilises correlated ones. ElasticNet does both — and lets the data decide the mixture. By searching a joint grid of penalty strength (α) and L1/L2 mix (l1_ratio) via 5-fold cross-validation, ElasticNetCV selects the optimal combination without requiring the engineer to bet on one regularisation strategy upfront.

In this case, the CV grid confirms the data rewards sparsity: l1_ratio = 1.0 (pure L1) achieves the best score. furnace_entry_temp_c is zeroed — its information is fully captured by furnace_exit_temp_c. ElasticNet found the correct answer. A naive Ridge model would have kept both sensors active and mislead the process team into monitoring a redundant thermocouple.

---

📊 Dataset

• 1,500 hot stamping cycles captured from the press PLC, furnace SCADA, and inline tensile testing records
• Target: tensile_strength_mpa — part tensile strength after press hardening (MPa, continuous)
• Range: 602.9 – 1,399.2 MPa | Mean: 989.6 MPa | Spec: ≥ 900 MPa | In-spec: 74.9%
• Steel grades: Usibor 1500 MPa class (mean 1,052.6 MPa) · Ductibor 500 MPa class (mean 986.2 MPa) · MBorian 700 MPa class (mean 931.4 MPa)

Column	Type	Description
furnace_entry_temp_c	float	Blank temperature entering the furnace (°C) — zeroed by ElasticNet
furnace_exit_temp_c	float	Blank temperature exiting the furnace (°C)
die_temp_c	float	Active die surface temperature (°C)
transfer_time_s	float	Time between furnace exit and die close (s)
press_force_kn	float	Maximum press force during forming (kN)
press_speed_mm_s	float	Ram closing speed (mm/s)
dwell_time_s	float	Time blank remains in closed die (s)
blank_thickness_mm	float	Sheet blank thickness (mm)
steel_grade	int	1 = Usibor · 2 = Ductibor · 3 = MBorian
active_cooling	int	1 = water-cooled die active · 0 = passive
cooling_pressure_bar	float	Cooling water pressure (bar)
cooling_flow_l_min	float	Cooling water flow rate (L/min)
tensile_strength_mpa	float	Target — part tensile strength after quench (MPa)

Data Origin (Real-World Perspective)

Variable(s)	Source System	Notes
furnace_entry_temp_c, furnace_exit_temp_c	Furnace SCADA / Thermocouple Array	Logged at blank entry and exit portals — two sensors on the same furnace zone
die_temp_c	Press PLC / Die Thermocouple	Die surface temperature from embedded thermocouple — logged at die close
transfer_time_s	Press PLC Timer	Timestamp delta between furnace exit signal and die-close signal
press_force_kn	Press PLC / Load Cell	Peak force recorded during the forming stroke
press_speed_mm_s	Press PLC / Encoder	Ram velocity at die approach — from the motion controller
dwell_time_s	Press PLC Timer	Closed die dwell time — programmable parameter logged per stroke
blank_thickness_mm	ERP / Material Cert	Nominal blank thickness from the coil certificate or incoming inspection
steel_grade	ERP / Material Master	Grade code from the production order — set at job setup
active_cooling, cooling_pressure_bar, cooling_flow_l_min	Cooling Circuit PLC / Flow Meter	Cooling system parameters logged per cycle from the die cooling controller
tensile_strength_mpa	Inline Tensile Tester / QMS	TARGET — destructive or ultrasonic tensile estimate per production lot

In real-world hot stamping, this dataset requires joining at least four systems: the furnace SCADA (temperatures), the press PLC (force, speed, dwell, transfer time), the ERP (material spec and grade), and the quality management system (tensile test results). The join key is typically the stroke counter and shift timestamp — and the tensile result may arrive hours after the production event.

---

🤖 Model

Algorithm: ElasticNet Regression (L1 + L2 combined) — sklearn.linear_model.ElasticNet + ElasticNetCV

ElasticNet combines both regularisation penalties into a single framework:

$$\text{Loss} = \sum_{i=1}^{n}(y_i - \hat{y}_i)^2 + \alpha \left[ \rho \sum_{j}|\beta_j| + \frac{1-\rho}{2} \sum_{j}\beta_j^2 \right]$$

Two hyperparameters are selected jointly: α (overall penalty strength) and ρ / l1_ratio (L1/L2 mixture). Setting l1_ratio = 0 gives pure Ridge; l1_ratio = 1 gives pure Lasso; values in between blend both behaviours.

Why ElasticNet over Ridge or Lasso alone? Because the structure of the hot stamping data was not known upfront. The furnace temperature collinearity (r = 0.95) suggested Ridge; the presence of at least one redundant sensor suggested Lasso. ElasticNet searches the full mixture landscape and lets the data answer. In this case, the CV confirmed l1_ratio = 1.0 — the data is Lasso-like. ElasticNet found the right answer without requiring a prior guess.

Alpha and l1_ratio tuning: ElasticNetCV searched 60 × 8 = 480 hyperparameter combinations (60 alpha values on a log scale × 8 l1_ratio candidates) using 5-fold cross-validation. Selected: alpha = 0.1597 · l1_ratio = 1.00.

Preprocessing: StandardScaler on all 12 features — mandatory for both L1 and L2 penalties.
Split: 80/20 train/test, random_state=42.

---

📈 Key Results

Metric	ElasticNet	Ridge	Lasso	Operational Meaning
R²	0.968	0.968	0.968	96.8% of tensile variance explained — all three converge
RMSE	24.16 MPa	24.3 MPa	24.3 MPa	2.7% relative error on 900 MPa spec — tight enough for recipe guidance
MAE	19.00 MPa	19.1 MPa	19.0 MPa	Median absolute miss — adequate for process decision support
Features active	11/12	12/12	11/12	ElasticNet and Lasso zero entry temp; Ridge keeps it
Train / Test	1,200 / 300 cycles	—	—	80/20 split, random_state=42

All three regularisations converge on accuracy. The value of ElasticNet is not higher R² — it is that the framework selected the correct penalty automatically. Ridge would have kept furnace_entry_temp_c active with a small misleading coefficient. ElasticNet zeroed it and told the truth.

---

🔍 ElasticNet Coefficients — Standardised (MPa per σ)

Feature	Coefficient	Direction	Metallurgical Interpretation
press_force_kn	+114.46	↑ Increases strength	Dominant driver — forming pressure determines martensite transformation completeness
steel_grade	−49.60	↓ Reduces strength	Grade encoding: Grade 3 (MBorian) < Grade 2 (Ductibor) < Grade 1 (Usibor) by design
die_temp_c	−27.66	↓ Reduces strength	Hotter die slows quench — reduces martensite fraction — primary operator adjustment lever
blank_thickness_mm	−11.68	↓ Reduces strength	Thicker blanks cool more slowly — lower quench rate — lower martensite content
active_cooling	+6.69	↑ Increases strength	Water-cooled die enforces faster quench — direct strength gain
furnace_exit_temp_c	+6.38	↑ Increases strength	Higher austenising temperature → more complete transformation potential
transfer_time_s	−2.67	↓ Reduces strength	Longer transfer = more cooling before die close = less martensite
cooling_flow_l_min	+2.60	↑ Increases strength	Higher flow rate → faster heat extraction → stronger part
cooling_pressure_bar	+1.31	↑ Increases strength	Pressure ensures flow uniformity across die channels
press_speed_mm_s	+0.27	↑ Increases strength	Faster closure reduces cooling time in open air before quench
dwell_time_s	−0.08	↓ Reduces strength	Minimal — longer dwell can allow slight tempering if die temp is high
furnace_entry_temp_c	0.000	—	Zeroed by ElasticNet — fully redundant with furnace_exit_temp_c (r = 0.95)

The zeroed feature is not a bad sensor — it is a redundant one. Monitoring furnace_entry_temp_c in real-time adds nothing to the prediction once furnace_exit_temp_c is known.

---

🔧 Simulation & Scenarios

Scenario	Configuration	Predicted	Status
A — Usibor, Optimised	Grade 1 · 920°C · 1,350 kN · die 140°C · active cooling · 4.0 s transfer	1,224 MPa	✅ Pass (+324 MPa margin)
B — MBorian, Passive Cooling	Grade 3 · 870°C · 1,050 kN · die 165°C · passive · 6.5 s transfer	734 MPa	⚠ Below 900 MPa spec
C — MBorian, Corrected Recipe	Grade 3 · 870°C · 1,050 kN · die 145°C · active cooling · 4.5 s transfer	812 MPa	⚠ Below 900 MPa spec

Important context on Scenarios B and C: MBorian is a 700 MPa target class designed for controlled ductility, not maximum strength. Its own product specification does not require 900 MPa. The simulator correctly shows that even an optimised recipe (Scenario C) reaches only 812 MPa for this grade — a physically accurate result, not a model failure. Activating cooling recovers +77.6 MPa vs Scenario B, confirming the cooling system is the primary lever for any grade.

The 2D response surface (Press Force × Die Temperature for Usibor with active cooling) extends this into a full process window map, showing every force/temperature combination that meets the 900 MPa spec.

---

🗂️ Repository Structure

ElasticNet_Hot_Stamping/
├── 11_ElasticNet_Hot_Stamping.ipynb   ← Notebook (no outputs)
├── hot_stamping_data.csv               ← 250-row sample dataset (GitHub public)
├── README.md
└── requirements.txt

📦 Full Project Pack — complete dataset (1,500 records), notebook with full outputs, presentation deck (PPTX), and app.py tensile strength simulator available on Gumroad.

---

🚀 How to Run

Option 1 — Google Colab: Click the badge above.

Option 2 — Local:

pip install -r requirements.txt
jupyter notebook 11_ElasticNet_Hot_Stamping.ipynb

---

💡 Key Learnings

1. ElasticNet is not a compromise — it is a search strategy. Choosing between Ridge and Lasso requires knowing the data structure in advance. ElasticNetCV runs both and finds the answer. When the data is clearly Lasso-like (l1_ratio = 1.0), ElasticNet confirms it. When it is clearly Ridge-like (l1_ratio = 0.0), it confirms that too. The framework earns its place before you know which regularisation is right.

2. A zeroed coefficient is a process insight. furnace_entry_temp_c eliminated means the entry thermocouple adds zero predictive value once exit temperature is known. This is not a data quality finding — it is an instrumentation decision. If the sensor is expensive to maintain, the model just justified decommissioning it.

3. All three regularisations converging on R² = 0.968 confirms strong signal. When OLS, Ridge, Lasso, and ElasticNet agree on accuracy, the dominant relationships are real and robust. The regularisation methods are not fighting over noise — they are all finding the same strong signal underneath. That is a process worth trusting.

4. Press force dominates because physics dominates. A coefficient of +114 MPa/σ on press_force_kn is not a modelling artefact — it reflects the metallurgical reality that forming pressure determines martensite transformation completeness. When a model's hierarchy matches first-principles engineering, the model is doing its job.

5. Steel grade is a variable, not a constant. The grade coefficient (−49.60 per σ) captures the metallurgical hierarchy across Usibor, Ductibor, and MBorian in a single parameter. A model that predicts tensile strength across three grade families simultaneously is a process library, not just a process predictor.

---

👤 Author

Luis Lozano | Operational Excellence Manager · Master Black Belt · Machine Learning
GitHub: LozanoLsa · Gumroad: lozanolsa.gumroad.com

Turning Operations into Predictive Systems — Clone it. Fork it. Improve it.