PhD in Causal Machine Learning for Earth Observation
Natural Hazards (NH)
The Artifical Intelligence group is looking for an aspiring PhD candidate to research causal machine learning and uncertainty quantification for Earth Observation time-series.
Currently, predictive AI in Earth Sciences relies heavily on correlation-based machine learning. When an agricultural system fails due to compounding climate extremes - like a simultaneous heatwave, drought, and ozone pollution spike - standard models can forecast the failure (if they can), but they cannot explain why it happened or calculate the exact contribution of each individual stressor. Furthermore, these models often fail and make overconfident predictions when presented with unprecedented climate anomalies. We need a trustworthy AI for high-stakes environmental decisions.
In this research, you will learn how to move machine learning beyond pure correlation into answering counterfactual questions. Using remote sensing multimodal time-series data and Earth foundation model embeddings, you will design and develop causal machine learning models tailored for dynamic, spatiotemporal Earth systems. Your primary focus will be time-series causal attribution: learning how to answer retrospective counterfactual questions to isolate the exact fraction of crop failure caused by specific stressors (e.g., ozone versus heat).
Importantly, your causal models will only be as reliable as the foundation model representations they rely on. Because extreme Earth system events are highly unpredictable, standard foundation models often output overconfident, flawed embeddings when faced with unprecedented climate anomalies. To solve this, you will also develop rigorous Out-of-Distribution (OOD) detection and Uncertainty Quantification (UQ) methods for Earth Foundation Models. By mathematically flagging when incoming data represents a never-before-seen anomaly, you ensure the foundation model does not pass "hallucinated" data to your causal model. This explicitly connects your uncertainty metrics to your causal outputs, ensuring the final system knows exactly when it is safely attributing causes and when it is extrapolating into the unknown.
You will apply your research directly to real-world compounding climate stress data from the Po Valley, working within the newly funded European project PROTEUS. Ultimately, the causal reasoning and uncertainty algorithms you build will serve as the quantitative engine for the "Copernicus Agent," an AI assistant designed to give European policymakers verifiable, data-driven post-mortem analyses of agricultural disasters.
For more information about the position, please contact Vassilis Sitokonstantinou, Assistant Professor, via email: vassilis.sitokonstantinou@wur.nl.
Questions about the procedure? Get in touch with Noorien Abbas, Corporate Recruiter, via vacaturemeldingen.psg@wur.nl .
Ready to apply?
You can apply directly using the apply button on the vacancy page on our website which will allow us to process your personal information with your approval. Only applications submitted through our website will be considered. https://www.wur.nl/en/vacancy/phd-position-trustworthy-ai-and-causal-inference-earth-observation
To apply, please send the following documents (max. 3 pages in total for both documents):
- Complete and up-to-date curriculum vitae;
- Motivation letter.
- Sample of your scientific writing (one paper, thesis, or report, that you have written yourself)
The maximum length of the documents must not exceed 3 pages. If it exceeds, applications will not be considered. Additional files such as grades and transcripts are not required during this stage and will not be considered.
You can apply up to and including May 11th, 2026. The first interviews are scheduled for May 26th 2026.