Sargassum SSP Synth Data

What is this project?

This project is part of the BIP Summer Institute 2026. It was mentored by Dr. Rita Appiah from Purdue University.

The goal is to create synthetic data for Sargassum seaweed energy conversion. This data supports the "TP2" research pipeline.

We use two methods:

1. Biological Block: Simulates methane production over 46 days using Fortran.
2. Thermochemical Block: Simulates pyrolysis and gasification yields using Fortran.

The generated CSV files are used to train Machine Learning models (LSTM and XGBoost) for biogas and syngas prediction.

Project Structure

• src/: Fortran source code for data generation.
• data/: Generated CSV datasets and analysis plots.
• scripts/: Gnuplot scripts for visualization.
• Makefile: Instructions to compile the Fortran code.

How to Run

You need gfortran and make installed.

1. Compile the code:

   make all

2. Run the biological simulation:

   ./sim_bip_dinamico.out

3. Run the thermochemical simulation:

   ./sim_bip_termo.out

Data Description

• sargassum_dynamic_data.csv: Time-series data (t=1..46 days) for LSTM training.
• sargassum_thermochemical_data.csv: Static mass/energy balance data for XGBoost/RF training.

Data Source

The synthetic data is based on real research:

• Biological Block: Uses separation coefficients and methane yields from Salgado-Hernández et al. (2023). The Liquid Fraction (LF) produces 159.7 L CH4/kg VS, while the Solid Fraction (SF) produces 83.45 L CH4/kg VS.
• Thermochemical Block: Uses thermodynamic equations for syngas Lower Heating Value (LHV) based on Velázquez-Hernández et al.
• Kinetics: Time-series data (46 days) follows the Modified Gompertz model parameters reported in the same study.

ML Pipeline Context

This dataset supports two machine learning tasks:

1. LSTM Networks: Trained on the 46-day dynamic series to predict cumulative methane production. LSTM is chosen because it handles time-series process dynamics effectively in biogas monitoring (Tisocco et al., 2026).
2. XGBoost & Random Forest: Trained on static thermochemical data to predict gas yields and LHV. XGBoost often leads in gasification gas prediction, while Random Forest remains competitive for pyrolysis product yields (Cihan, 2025; Dong et al., 2022). Both models are benchmarked here to compare performance per target.

References

1. Salgado-Hernández, E., et al. (2023). Methane Production of Sargassum spp. Biomass from the Mexican Caribbean: Solid–Liquid Separation and Component Distribution. International Journal of Environmental Research and Public Health, 20(1), 219. https://doi.org/10.3390/ijerph20010219

   Key source for biological block parameters, separation coefficients, and BMP yields.

2. Tisocco, S., et al. (2026). Machine learning vs. ADM1: Reliable biogas prediction with minimal data requirements in full-scale plants. Environmental Science and Ecotechnology, 29, 100662. https://doi.org/10.1016/j.ese.2026.100662

   Supports the use of LSTM networks for time-series methane prediction in waste-to-energy systems.

3. Cihan, P. (2025). Bayesian Hyperparameter Optimization of Machine Learning Models for Predicting Biomass Gasification Gases. Applied Sciences, 15(3), 1018. https://doi.org/10.3390/app15031018

   Benchmarks XGBoost performance for gasification gas composition and LHV prediction.

4. Dong, Z., Bai, X., Xu, D., & Li, W. (2022). Machine learning prediction of pyrolytic products of lignocellulosic biomass based on physicochemical characteristics and pyrolysis conditions. Bioresource Technology, 128182. https://doi.org/10.2139/ssrn.4191315

   Demonstrates Random Forest competitiveness for pyrolysis product yield prediction.

License

This project is for academic research purposes.