BINARY DISTILLATION SURROGATE MODEL

Project Overview

This project develops ML surrogate models to replace computationally expensive process simulations with fast, accurate machine learning predictions for binary distillation columns. Using data generated from DWSIM (an open-source chemical process simulator), by systematically varying key operating parameters to predict:

• Distillate Purity (xD): Mole fraction of ethanol in the distillate product stream
• Reboiler Duty (QR): Energy consumption required for separation (in kW)

The surrogate models enable rapid design exploration, optimization studies, and sensitivity analysis without running full process simulations.

Context: This project was completed as a screening task for the FOSSEE IIT Bombay Internship.

Dataset Information

• Source: Self-generated using DWSIM Process Simulator
• System: Ethanol-Water binary mixture at 1 atm
• Property Package: NRTL (Non-Random Two-Liquid)
• Raw Data: 373 records with 8 features
• Clean Data: 307 records (after removing 66 duplicates)

Input Features:

Feature	Description	Range/Values
Reflux_Ratio	Ratio of liquid returned to column vs. distillate product	0.8 to 5.0
Boilup_Ratio	Ratio of vapor generated in reboiler vs. distillate flow	Variable
Feed_MoleFraction	Ethanol mole fraction in feed stream	0.2 to 0.95
Feed_FlowRate	Total molar feed rate (kmol/h)	±30% variation
Number_of_Stages	Theoretical trays in the column	15, 20, 25 (discrete)
Feed_ThermalCondition	Feed thermal state	0 (subcooled) or 1 (saturated)

Output Variables (Targets):

Variable	Description	Unit
Distillate_MoleFraction (xD)	Purity of ethanol in distillate	Mole fraction (0-1)
Reboiler_Duty (QR)	Heat energy supplied to reboiler	kW

DWSIM Data Generation Protocol

To generate your own dataset:

1. Download DWSIM from https://dwsim.org/

2. Set up flowsheet:

   • Create new steady-state simulation
   • Add compounds: Ethanol and Water
   • Select NRTL property package
   • Add material streams (Feed, Distillate, Bottoms)
   • Add distillation column and energy streams

3. Configure parameters:

   • Set feed conditions (temperature, pressure, composition, flowrate)
   • Specify column configuration (reflux ratio, boilup ratio, stages)
   • Define feed thermal condition

4. Run simulations:

   • Use spreadsheet feature to define parameter ranges
   • Run batch simulations and verify convergence
   • Export results to CSV

For detailed instructions, refer to Report.docx in the repository.

Model Workflow

This project follows a structured machine learning pipeline from simulation to model deployment.

1. Data Generation (DWSIM)

   • Set up binary distillation flowsheet for Ethanol-Water system.
   • Configure material streams, distillation column, and energy streams.
   • Systematically vary input parameters across specified ranges.
   • Export converged simulation results to CSV.

2. Data Preprocessing & EDA

   • Load dataset and remove duplicates (373 → 307 samples).
   • Verify unit consistency for flowrates (kmol/h) and duties (kW).
   • Perform univariate analysis (histograms, boxplots) for distribution and outlier detection.
   • Conduct bivariate analysis (scatterplots) to identify input-output correlations.
   • Generate correlation heatmaps and pairplots for multivariate relationships.

3. Feature Engineering

   • Apply StandardScaler to numerical features (Reflux Ratio, Boilup Ratio, Feed Mole Fraction, Feed Flowrate, Number of Stages).
   • Apply OneHotEncoder to categorical feature (Feed Thermal Condition).
   • Split data into train/validation/test sets using operating space blocks (Reflux Ratio 3.5-4.5 as validation).

4. Model Training & Selection

   • Train multiple regression models: Linear Regression, Polynomial Regression, Random Forest, AdaBoost, SVR, XGBoost.
   • Perform hyperparameter tuning using RandomizedSearchCV.
   • Evaluate models using R², RMSE, and MAE on validation and test sets.
   • Select best-performing model based on generalization performance.

5. Model Diagnostics

   • Generate parity plots (predicted vs. actual) to assess prediction accuracy.
   • Plot residuals to check for systematic errors and heteroscedasticity.
   • Analyze error distributions to confirm model robustness.

Performance Metrics

The project evaluates multiple regression models to predict Distillate Mole Fraction (xD) and Reboiler Duty (QR). The table below summarizes performance metrics for both outputs.

Model	xD - R²	xD - RMSE	QR - R²	QR - RMSE	Notes
Linear Regression	0.65	High	0.40	High	Weak baseline, high bias
Polynomial Regression	0.72	Medium	0.55	Medium	Improved fit but overfits
Random Forest	0.92	Low	0.91	Low	Strong, robust performance
AdaBoost	0.88	Low	0.85	Medium	Decent accuracy
SVR	0.90	Low	0.88	Low	Accurate but computationally expensive
XGBoost	0.95	Very Low	0.93	Very Low	Best overall performer

Best Model: XGBoost Regressor

• Highest R² scores (>0.93) for both outputs
• Lowest RMSE across all metrics
• Excellent generalization on test set
• 99%+ reduction in computation time compared to full DWSIM simulations

Key Insights

For Distillate Purity (xD):

• Strongest predictors: Feed Mole Fraction (0.67 correlation) and Reflux Ratio (0.61 correlation)
• Feed Flowrate shows moderate positive correlation (0.47)
• Boilup Ratio shows slight negative correlation (-0.22)

For Reboiler Duty (QR):

• Strongest predictor: Feed Flowrate (0.66 correlation) - biggest energy driver
• Distillate purity and reboiler duty show positive correlation (0.57) - energy-purity tradeoff
• Moderate correlations with Feed Thermal Condition (0.38), Reflux Ratio (0.29)

Instructions to Use

Explore the Notebooks

• EDA Notebook: View data distributions, correlations, and feature relationships
• Model Training Notebook: See model training, hyperparameter tuning, and evaluation

Generate Your Own Data

• Download DWSIM from https://dwsim.org/
• Follow the data generation protocol in the detailed report
• Set up Ethanol-Water distillation flowsheet
• Vary input parameters and export results

Train Models

• Load preprocessed data from notebook/data/distill_data.csv
• Apply feature scaling and encoding
• Train models using provided pipelines
• Evaluate performance on test set

Project Structure

Binary-Distillation-Surrogate-Model/
├── README.md                           # Project documentation
├── requirements.txt                    # Python dependencies
├── setup.py                            # Package setup
│
├── notebook/                           # Jupyter notebooks
│   ├── EDA.ipynb                       # Exploratory Data Analysis
│   ├── ModelTraining.ipynb             # Model development & evaluation
│   └── data/
│       └── distill_data.csv            # Generated dataset from DWSIM
│
├── src/                                # Source code modules
│   ├── __init__.py
│   ├── components/                     # Pipeline components
│   │   ├── data_ingestion.py
│   │   ├── data_transformation.py
│   │   └── model_trainer.py
│   └── utils.py                        # Utility functions
│
├── Data Visualisation/                 # EDA plots and visualizations
│   ├── histograms/
│   ├── boxplots/
│   ├── scatterplots/
│   ├── pairplots/
│   └── correlation_heatmaps/
│
├── Evaluation/                         # Model diagnostic plots
│   ├── parity_plots/
│   ├── residual_plots/
│   └── error_distributions/
│
└── Report.docx                         # Detailed technical report

Key Technologies & Libraries

• Process Simulation: DWSIM - Open-source chemical process simulator
• Machine Learning & Modeling: Scikit-learn, XGBoost, Random Forest, SVR, AdaBoost, Polynomial Regression
• Data Processing: Pandas, NumPy, StandardScaler, OneHotEncoder
• Visualization: Matplotlib, Seaborn, Plotly
• Model Evaluation: R² Score, RMSE, MAE, RandomizedSearchCV

Use Cases

• Process Engineers: Rapid estimation of distillation column performance without running full simulations
• Design Optimization: Integrate surrogate models into optimization loops for faster design iterations
• Sensitivity Analysis: Quickly evaluate impact of operating condition changes
• Energy Studies: Visualize energy-purity tradeoffs across operating ranges
• Education: Demonstrate ML applications in chemical engineering

Installation

Prerequisites

• Python 3.8 or higher
• pip package manager
• DWSIM (optional - for data generation only)

Local Setup

• Clone the repository

git clone https://github.com/aditi-gupta-git/Binary-Distillation-Surrogate-Model.git
cd Binary-Distillation-Surrogate-Model

• Create virtual environment

python -m venv venv
source venv/bin/activate   # Windows: venv\Scripts\activate

• Install dependencies

pip install -r requirements.txt

• Launch Jupyter Notebook

jupyter notebook

• Navigate to notebooks
  • Open notebook/EDA.ipynb for exploratory analysis
  • Open notebook/ModelTraining.ipynb for model training

Acknowledgments

• FOSSEE IIT Bombay - For providing this internship screening task opportunity

⭐ If you found this project helpful, please consider giving it a star!