Steel Production Optimization Model

Project Report: Utilization of Intelligent Systems for Developing a Steel Optimization Model Date: December 15, 2025

Project Overview

This project enables the selection of production process parameter settings based on input data to support ArcelorMittal engineers in steel production optimization. The system implements two primary modeling approaches:

1. Forward Model: Predicts mechanical parameters (Yield Strength, Tensile Strength, Elongation) based on chemical composition.
2. Inverse Problem: Takes existing mechanical parameters as input to output the required chemical composition ("recipe") and process parameters.

Data Sources

The models utilize two datasets to simulate complex and non-linear characteristics:

• Proprietary Database: 123 unique records sourced from TotalMateria containing chemical and mechanical parameters.
• Extended Dataset: Approximately 312 records derived from scientific publications for training purposes.

Methodology

Machine Learning Models

The project evaluates several regression algorithms to avoid overfitting on limited data:

• Support Vector Regression (linear, rbf, poly kernels)
• Gradient Boosting Regression (XGBoost)
• Random Forest Regression
• K-Nearest Neighbors Regression

Deep Learning

• TabNet architecture implemented using PyTorch.

Optimization & Validation

• Hyperparameter Tuning: Optuna tool used to minimize Mean Squared Error (MSE).
• Validation: Bootstrapping method used to determine confidence intervals for standard deviation. Root Mean Squared Error (RMSE) used for error analysis.
• Explainability: SHAP (SHapley Additive exPlanations) analysis included for model transparency.

Project Structure

• std-model/: Standard machine learning models (XGBoost, SVR, RF, KNN) and tuning scripts.
• inv-model/: Prototype for the inverse prediction problem.
• tabnet-models/: Deep learning implementations (TabNet).
• shap/: Model explainability analysis and plots.

Key Results

• Best Performance: Group 3's machine learning model achieved the highest reliability with an $R^2$ of 0.92 on the kaggle dataset.
• Inverse Model Accuracy:
  • Carbon (%C): $R^2 \approx 0.75$
  • Titanium (%Ti): $R^2 \approx 0.68$
  • Nickel (%Ni): $R^2 \approx 0.58$

Requirements

• Python 3.x
• scikit-learn
• xgboost
• optuna
• torch (PyTorch)
• pandas
• numpy
• tkinter (for GUI prototype)

Source:

https://www.kaggle.com/datasets/fuarresvij/steel-test-data?resource=download [ Downloaded on 2025-11-19 ]