First Journal Paper Published - ExoPrompt in Computers and Electronics in Agriculture

Thrilled to share that my first journal paper as a PhD student has been published!

ExoPrompt: Transformer-Based Greenhouse Climate Forecasting with Structured Conditioning and Physics-Based Simulation is now out in Computers and Electronics in Agriculture.

This work started from a simple but important question: how do you build a climate forecasting model that actually generalizes across different greenhouses, crops, and environmental conditions – when real-world sensor data is scarce? Our answer is ExoPrompt, a lightweight conditioning mechanism inspired by prompt tuning in NLP. It encodes structural, environmental, and crop-level attributes into learnable prompt representations, allowing the model to adapt to diverse scenarios without heavy architectural changes.

The pipeline combines the best of both worlds: synthetic data generated from the physics-based GreenLight simulator for pretraining, followed by fine-tuning on limited real-world sensor data.

Highlights

Simulation-based pretraining reduces real-world prediction error by up to 84.94% for CO2 over simulator-only baselines
Conditioning on exogenous parameters further reduces RRMSE by up to 18.79% compared to vanilla models
Controlled experiments validate ExoPrompt’s robustness under distributional shifts, achieving up to 49.20% RRMSE reduction for CO2
ExoPrompt lays a foundation for predictive digital twins in smart greenhouse control

Highlights

Links