🌍 Agri-business is responsible for 30% of the world's carbon emissions.
💡 Along with 8 other students, we created an SaaS B2B platform to assist product designers and purchasing managers in the eco-conception and eco-supply of agri-food products. You can find in this repository the notebooks that conducted the analysis, as well as the final functions and the final presentation.
- Platform description
- Outputs of the project
- Datasets
- Recommendation systems' criterias
- Insight of the final tool
The platform is composed of three modules :
🌾 an ingredient recommendation system that takes into account the carbon emissions, the seasonality and the similarity with the initial ingredient.
♻️ a packaging recommendation system based on the packaging material's carbon emission, if it is recycled and its recyclable rate (% of material that will be recycled in the future).
🚜 a supplier recommendation system that optimises the supplier's carbon emissions based on the distance to the factory, the transport carbon emissions average, the supplier's low carbon, PDO (Protected Designation of Origin) and bio labels, and the carbon impact of packaging used for transport.
⛏️ The tool was built on DataStudio and is available here :
https://datastudio.google.com/reporting/37b9ff78-ddf0-48a0-9c39-5edfd342b7c8
A .pdf version of it is also in this repository.
✅ The final presentation was held in front of 6 people having sustainability positions of responsibility in different sectors :
- Vincent Poncet, Cloud Sustainability specialist at Google
- Raphaël de Cormis, VP Thales Digital Factory at Thales
- Frédérique Rousseau, Holding and Other activites - Human Resources Director at LVMH
- Isaelle Ryl, Director of PRAIRIES (PaRis Artificial Intelligence Research InstitutE) at INRIA
- Thomas Nivard, VC investor at 360 Capital
- Guillaume Roques, Senior Director - Google Coud Marketing EMEA at Google
- Pierre Deron, Director at CleanTech Open France
- Damien Gromier, CEO at Startup Inside
🏆 For this project, our team was awarded with the 1st prize.
We used 3 open-source datasets :
- ADEME, Base carbone : carbon emissions of ingredients and composed products caused by production
- Agribalyze : carbon emissions of ingredients and composed products caused by production
- Open Food Facts : compositions of agri-food products
Then, we use information given by the suppliers and distributors on :
- production location
- means of transportation
- packaging and their recyclability rate
This module first relies on a similarity algorithm.
The similarity is obtained by cleaning 80 000 recipes collected from Open Food Fact and applying the Google NLP model word2vec on them. The model identifies the relationships and roles of the ingredients within the recipes.
This allows us to output the top 20 most similar items of any ingredient present in the recipes.
The second criteria of the module is the category of the ingredients, gathered from Agribalyze's dataset. Indeed, the final tool would first recommend ingredients of the same category as the initial one.
The next step is the carbon emissions, collected both from Agribalyze and from the Base Carbone dataset of ADEME. We only output ingredients with less carbon emissions than the initial one.
Finally, we integrate a seasonality dataset built ourselves. It allows us to recommend only the ingredients of the season the manufacturer wants to produce at.
The first criteria of our packaging recommendation system is its carbon impact, taken from the ADEME base carbon dataset.
The second criteria is the recyclability of the matter : is the packaging recycled and what proportion will be recycled in the future ?
Finally, we apply a filter on the type of product concerned, whether the packaging is a bottle or a label.
The packaging recommendation algorithm associates a score between 0 and 1.
-
$carbon.score$ : score between 0 and 1 representing the normalised carbon impact of transport and packaging measured as explained previously (1 is the suppliers highest carbon footprint and 0 the smallest), sourced from ADEME Base Carbone -
$recyclable.rate$ : rate between 0 and 1 representing the percentage of material recycled after usage - sourced from CITEO -
$recycled.factor$ : binary value specifying if the product is recycled (0 for non-recycled and 1 for recycled material)
The weights are personnalisable to fit the client's priorities.
Default weights :
The scoring used can be decomposed into the transportation footprint, the packaging footprint, and the labels' factors. The final tool computes a score measuring the carbon impact for each supplier.
-
$label_{low Carbon}, label_{bio}, label_{PDO}$ : binary values representing the presence of the labels specified below (0 for absent label and 1 for present label) -
$carbon.score$ : score between 0 and 1 representing the normalised carbon impact of transport and packaging measured (1 is the suppliers highest carbon footprint and 0 the smallest) :$carbon.score = transport.footprint+packaging.footprint$ $transport.footprint = transport.footprint_{CO2kg/km,tonnes(vehicule),kg(product)} * distance(supplier, manufacture)_{km}*vehicule.weight$ $packaging.footprint = packaging.footprint_{CO2kg/kg(packaging),kg(product)} * packaging.weight_{kg}$
The weights are personnalisable to fit the client's priorities.
Default weights :
The data are given by the following sources :
ADEME, Base Carbone
-
$transport.footprint_{CO2kg/km,tonnes(vehicule),kg(product)}$ : carbon emissions per km and kg of the transportation means used to deliver the supply -
$vehicule.weight$ : average weight of the transportation (in tonnes) estimated -
$packaging.carbon.footprint_{CO2kg/kg(packaging),kg(product)}$ : carbon emissions per kg of the packaging and per kg of the product
Supplier
-
$supplier.loc$ : location of the supplier -
$packaging.weight_{kg}$ : average weight per type of packaging per product
Datact user
-
$manufacture.loc$ : location of the manufacture that transforms the agri-food
- Page 1 : ingredient and packaging recommendation systems
- Page 2 : supplier recommendation system
