Cepio — Club Cepio, our method

How can we predict mushroom growth?

1Modelling mushroom growth

Unlike the tech giants and their “big data” approach, we chose to build on scientific expertise to predict when mushrooms appear (what is usually called an “expert” model).
Everyone believes you need millions of data points to build an AI. For mushroom foraging, there is no reliable record of past harvests. The reality on the ground is scattered, subjective and incomplete data.
What our model produces is a single indicator: the growth index. Here is how to read it:
- Scarce: low probability — better to wait
- Limited: a few active patches — worth a trip, but be selective
- Moderate: good potential — aim for the favourable areas
- Abundant: ideal conditions — grab your basket!

2There are two ways to model mushroom growth

Method 1

The data-driven approach

Principle

“Give me a million examples and I'll work out what to put together to predict the future”

Sources: Your personal data, and more besides
Advantages: Uncovers invisible correlations
Drawbacks: Needs huge amounts of quality data

Visualisation

Mountain of data → AI black box → Predictions

Method 2 · our choice

The knowledge-driven approach

Principle

“I know the mechanisms, so I model them”

Sources: Scientific publications, field expertise
Advantages: Works with little data, and can be explained
Drawbacks: Limited by what we currently know

Visualisation

Scientific knowledge → Explicit model → Predictions

3Why we chose the knowledge-driven approach

Mushrooms remain a challenge, because the way they grow is complex and uncertain. That complexity is exactly what interests our scientists: we have 50+ years of mycological research and scientific literature at our disposal.
The biological mechanisms behind how mushrooms develop and grow are well documented.
Controlled studies in the field and in the laboratory give a fine picture of the kinetics (the speed) of fruiting and growth in our mushrooms.
Missing data: there is no “Netflix of ceps” with millions of observations, photos and quality annotations
Human variability: every forager has their own secret spots
Observation bias: only the successes get reported, never the failures!

A closer look at what a “knowledge-driven” model is:

Analogy: you learn from a flight manual instead of learning by crash-testing aeroplanes
A concrete example: our weather models (the physics of the atmosphere) versus pure AI

4How the Cepio growth model was built

Step 1: we take mushroom biology apart (ceps/boletes)

The life cycles of mushrooms: Mycelium → Stimulation → Fruiting
The trigger factors identified in the literature:
- Thermal shock (cold stress)
- Rehydration (chemical signal)
- Water stress (reproductive urgency)
Our model looks for episodes of thermal shock. A certain number of cold hours is needed to trigger the fruiting of the sporophores. Ideally that combines with a heavy spell of rain and then a dry period, and there you have it.

Step 2: we turn the biological phenomena into simple mathematical equations

Temperature thresholds: based on physiological studies
Time windows: drawn from controlled observations
Interactions: modelling the synergies

A concrete example:

IF   (Temperature < 12°C for 72h)
AND  (Rain > 15mm over the next 5 days)
AND  (No rain for 5-8 days afterwards)
THEN Fruiting probability = 80%

Step 3: we calibrate with the data available today

Weather data: 30 years of records, available right across Europe
Field observations: cases documented in the scientific literature (chambers of agriculture, major European research bodies)
Regional adjustment: we adapt to local conditions

Step 4: we run our model

Major advantage: we can explain why it works and account for our result, the growth index. A predictive AI model could not.
Confidence: we can understand where the limits lie
Improvement: we can easily aim future research at sharpening our accuracy

Step 5: we validate our model by walking the woods with a test group

We validate without Big Data, through cross-validation tests (trying our model in different regions)
We put our model up against experienced mycologists and specialist groups

Step 6: we launch cepio.fr, the mushroom forecast

We put a first model online, tuned for ceps and boletes: it is free at the weekend, and 12,99 euros buys you forecasts a week ahead (to cover our costs).

Step 7: we look at the feedback and the needs, and we keep working

We analyse what comes back from the field: time spent searching, the forager's experience, photos of the mushrooms picked, weight of the harvest.
We test our thinking against experienced mycologists and specialist groups
We keep investing in research and development to get around the limits of our current model:
- Factors we do not model: the health of host trees, competition between species, micro-variations in the soil
- Accuracy: 70-80% against 95% for cutting-edge Big Data models
- Generalisation: more local adaptation is needed
- Scientific transparency: open code the community can check, no hidden magic

5The future of Cepio

With a hybrid approach between knowledge-driven and data-driven models, we will be able to combine the best of both worlds.

Our future models will be strengthened by data: hybrid models.
The hybrid approach could give us the ideal platform for anyone hunting mushrooms who is not afraid of using a bit of technology now and then, to:
- Find the right window to go out searching
- Find a place to go searching
- Find a particular mushroom species
We will support beginners and seasoned foragers alike:
- Asking questions and learning about the weather, the biotopes and the biological cycles that mushrooms need in order to grow
- Sharing what you know by giving foraging tips on a local scale

Club Cepio

Cepio today

Who are we?

Thierry V.

Hugues L.