Inception Petroleum Analysis

This repository shows the code used in "Inception for Petroleum Analysis" for cetane number prediction. Study published in Fuel (Elsevier), 2025.

The IPA Deep CNN architecture

Convolutions with kernel size 3 are shown in blue, convolutions with kernel size 1 are shown in red. The numbers of neurons are shown under each computational blocks. For more details, verify on the paper under the section -> Methodology / Deep learning.

Getting started

Matlab

Install MATLAB R2022b and Eigenvector PLS_Toolbox version 9.2.

Launching environment

Matlab

Open the script within the eponymous software.

Python (Docker)

Build the docker image (using CPU):

docker build -t inception-petroleum-analysis .

Run and access to the output the of docker image:

docker run -it --rm inception-petroleum-analysis

Python (Conda)

This environment expect miniconda/anaconda already installed.

Follow the next commands:

conda env create -f environment.yml

Activate the environment and launch models calibration with:

conda activate inception-petroleum-analysis && python -m scripts.models-calibration

This script can be adapted to different needs.

Inside the main() function, three different approaches are possible:

• comparing the two models using the compare_models() function,
• using a very simple approach via the basic_usage() function,
• modifying a configuration via config_from_dict().

This code has been rewritten to be more modular and allow for testing different model configurations, as well as extension to other architectures. To do this, a practitioner can develop the following classes (found on eponym file under directory "src.model_utils"): ModelType, BaseModelConfig, BaseModelPipeline, and finally update class ModelFactory (especially methods: _registry and _config_registry.

Notes about our data

Information : all y are already normalized to follow entreprise policy, the scaler is not given. Therefore, calculated metrics will be roughly 10 times smaller than on the publication.

• PLS will be fed using X and y from 'data/raw-kennard-full_range.mat', y is already normalized (centered) but X are raw.
• IPA and DeepSpectra will be fed using y from 'y_train_scaled.csv' and 'y_test_scaled.csv' (mean 0 & std 1), the X comes raw from 'data/raw-kennard-full_range.mat' or feature selected from 'data/raw-kennard-reduced_range.mat'