Soil–Plant–Microbe Embeddings & Uncertainty for Viticulture
Description
Position overview:
Create robust biological representations (embeddings) relevant to regenerative systems—spanning soil biology, plant stress signals, and microbial community patterns—together with calibrated uncertainty. These representations will feed downstream prediction, causal reasoning, and planning modules in the Regenerative AI Engine.
Problem statement (why this exists):
Biology-heavy inputs are high-dimensional, noisy, and strongly confounded (batch effects, site effects, seasonality). If we embed these signals poorly—or without uncertainty—recommendations become brittle and unsafe. This track builds representations that are information-rich and explicitly risk-aware.
Data you will work with:
- Public datasets across soil biology / microbiomes / omics / physiology proxies (curated and preprocessed with leakage controls).
- Optional internal data: Symbioverse field datasets (Spain) may provide real-world validation signals (not required).
Key responsibilities (linked to the innovation project):
- Select and curate suitable public datasets and define preprocessing standards (batch effects, confounders, leakage controls).
- Build feature engineering + embedding models for “biology-heavy” inputs (representation learning).
- Implement uncertainty quantification / calibration (e.g., conformal methods, probabilistic embeddings, calibration curves).
- Demonstrate downstream utility with at least one benchmark task and comparisons vs simpler baselines.
Why your expertise is crucial:
This track needs someone who can treat biological data with statistical seriousness and produce representations that stand up under domain shift.
Key outcomes for the Visiting Talent:
- Publishable-grade ML/bioinformatics artefacts with clear evaluation.
- A reusable library (biofeatures/) with documentation, tests, and examples.
- Practical experience building risk-aware inputs for decision-support systems.
Stream:
- Deep-Tech Talents (research-focused).
Skills
Sign in to read and join the discussion.