Experimental Dataset and Predictive Model for Multi-Product Yield Prediction from Cocoa Residues

Description

This repository contains the complete experimental dataset (84 laboratory trials), the trained Artificial Neural Network model, and an interactive Google Colab notebook developed for the valorization of cocoa residues (Theobroma cacao L.) in Santander and Norte de Santander, Colombia. The resources support the research article "A hybrid empirical-AI model for multi-product yield prediction and stochastic optimization from cocoa residues" published in Sustainability (2026). The dataset includes operational variables (temperature, time, pH, solvent concentration, extraction method, cocoa variety) and yields for five target compounds: bioethanol, essential oils, antioxidants, pectins, and paraffins. The predictive model integrates a hybrid empirical-AI architecture with dimensional consistency, trained on 100% empirical data, combined with Monte Carlo sensitivity analysis (10,000 simulations) for uncertainty quantification.

Steps to reproduce

1. Access the repository files: Download all files from this Mendeley Data repository to a local directory, or open the Google Colab notebook directly in your browser by clicking the "Open in Colab" badge (if available) or by uploading the .ipynb file to Google Colab. 2. Software requirements: Python 3.8 or higher TensorFlow 2.10+ (automatically available in Google Colab) Required Python packages: numpy, pandas, matplotlib, openpyxl, tensorflow (see requirements.txt for complete list) 3. Execute the inference notebook: Open the file Cocoa_Biorefinery_Inference_Model.ipynb in Google Colab or Jupyter Notebook and execute the cells sequentially: Cell 1: Import dependencies and configure the environment Cell 2: Load model parameters (reference values, calibration table, categorical variables) Cell 3: Define preprocessing functions (non-dimensionalization, one-hot encoding, dimensional reconstruction) Cell 4: Upload the experimental dataset (dataset_cocoa.xlsx) Cell 5: Train the Artificial Neural Network model (1 hidden layer, 16 neurons, linear output activation with post-prediction clipping to [0.5, 1.5]) or load the pre-trained model (cocoa_biorefinery_model.h5) Cell 6: Validate the model against experimental data (10 random samples) Cell 7: Run interactive demonstrations with predefined scenarios (CCN-51 and Criollo varieties) Cell 8: Perform custom predictions by modifying input parameters Cell 9: Execute dimensional consistency verification tests 4. Interpret the outputs: The model predicts a dimensionless relative efficiency factor (η)* for each input sample, which is then reconstructed into physical yields (mL/kg, mg GAE/g, %, g/kg) using the deterministic empirical scaling factors (reference yields and variety correction factors) as detailed in the associated manuscript. The notebook outputs include: Model training metrics (R², RMSE, MAE) Validation results comparing predicted vs. experimental yields Interactive predictions for custom operational conditions Dimensional consistency verification tests 5. Reproduce Monte Carlo sensitivity analysis: The Monte Carlo simulation protocol (10,000 iterations with Latin Hypercube Sampling) is described in Section 2.11.2 of the manuscript. The probability distributions for operational variables (temperature, pH, time, solvent concentration) are parameterized based on instrumentation precision specifications reported in the study.

Institutions

• Universidad Nacional Abierta y a Distancia

  Bogota D.C., Bogotá

Categories

Licence