Manual prediction
Enter the heat process parameters to obtain a nitrogen content prediction before tapping.
Current prediction (ppm)
—
Sievert N eq (ppm)
—
η attenuation
—
Δ vs. actual (ppm)
—
NitroPINN prediction result
Predicted N (ppm)
—
— wt%
Sievert N eq (ppm)
—
— wt%
η attenuation
—
—
Batch CSV upload
Upload a CSV file with columns matching the 15 NitroPINN input variables. The results will be added to history.
CSV file format
The CSV must contain a header with exact column names (comma-separated):
heat_id,C,Mn,P,S,Fe_slag,MnO_slag,SiO2_slag,CaO_slag,slag_basicity,Ttap,Oactivity,tblow,treblow,scrap_ratio,Crate,actual_N
Columns heat_id and actual_N are optional.
Drag and drop a CSV here or click
Supported formats: .csv · Max. 5 MB
Prediction history
All manual and batch predictions from the current session.
Predicted vs. actual
Visualization of model accuracy. Requires records with actual measured values.
Not enough data
Add actual measured values (Actual N) to prediction history.
About the model NitroPINN
Physics-Informed Neural Network for predicting nitrogen in liquid steel before tapping from a BOF converter.
Architecture
- 🔷 Input: 16 normalized features (Z-score)
- 🔷 Backbone: FC 64→32→16, Tanh, Dropout 0.15
- 🔷 η-head: Linear(16,1) + Sigmoid → [0,1]
- 🔷 δ-head: Linear(16,1) → residual correction
- 🔷 Output: Npred = η·Neq + δ
Physical constraints
- ⚛️ Sieverts law: log Neq = −3750/T − 1.154 + 0.5·log pN₂ − log fN
- ⚛️ Wagner formalism: log fN = Σ eⱼ·[j] (C, Mn, P, S, O)
- ⚛️ η penalty: [0.05, 0.95] bounds
Performance (5-fold CV)
7.98
RMSE (ppm)
5.60
MAE (ppm)
0.325
R²
16 input features
C, Mn, P, S (steel)Fe, MnO, SiO₂, CaO (slag)
Slag basicityTapping temperature Ttap
Oxygen activityBlowing time tblow
Reblowing time treblowScrap ratio
Decarburization rate ĊNeq (Sievert)