ExoPrompt: Transformer-based greenhouse climate forecasting with structured conditioning and physics-based simulation

Accurate greenhouse climate forecasting is essential for optimizing crop yield, resource use, and climate control. Traditional physics-based simulators offer interpretability but often rely on simplified assumptions and fixed configurations, while data-driven models struggle with limited generalization due to scarce and localized sensing data. These limitations pose an obstacle to realizing intelligent and adaptive greenhouse systems. Digital twins (DTs), virtual replicas that mirror physical systems, offer a promising paradigm to overcome these constraints and enable real-time decision-making. Addressing this challenge is crucial towards enabling DTs with accurate, adaptive models operating across diverse conditions. We present a simulation-informed framework for robust greenhouse climate forecasting leveraging structured conditioning and synthetic data pretraining. Central to our approach is ExoPrompt, a lightweight conditioning mechanism that encodes environmental, structural, and crop-level attributes into prompt representations, enabling adaptation across diverse scenarios. The model is first pretrained on synthetic data generated using the GreenLight simulator, then fine-tuned on limited sensor data collected under varying lighting conditions. We evaluate our approach on synthetic and real-world datasets across multiple setups. Results show that simulation-based pretraining improves forecasting performance, yielding up to a 46.31% reduction in RRMSE for relative humidity and an 84.94% improvement in CO2 prediction over simulator-only baselines. Conditioning on exogenous parameters further reduces RRMSE by up to 18.79% compared to vanilla models. Finally, we design a controlled simulation experiment varying a single exogenous parameter to isolate and validate the robustness of ExoPrompt under distributional shifts, achieving up to 49.20% RRMSE reduction for CO2 on ground truth data.