Physics-Informed Neural Networks and Causal Inference for Dynamic Modeling and Multi-Objective Optimal Control of Greenhouse Tomato Growth
Accurate greenhouse control requires forecasting how climate interventions alter plant growth while respecting conservation laws, action support and physiological safety limits. We formulate tomato cultivation as a partially observed, intervention-dependent dynamical process and present CAUSE-GROW, a shared physically constrained neural flow for prediction, supported counterfactual reasoning and risk-aware Pareto model predictive control. Its dynamics combine explicit external exchanges, donor-limited nonnegative transfer fluxes, and neural corrections restricted to a conservation-compatible viability cone. A sequential intervention operator couples posterior state inference with action-overlap diagnostics and hidden-confounding sensitivity; a trajectory-level distributional ambiguity model supports tail-risk screening of candidate controllers. We further derive conservation and positive-invariance conditions and specify compartment-disjoint and cross-season evaluation protocols for four public tomato datasets. The experimental analysis is organized around forecasting accuracy, causal intervention error, structural ablation, yield--resource trade-offs and calibration under weather shift. This formulation separates correctness of the physical transition law, interpretability of causal claims and feasibility of greenhouse control within a common dynamical system.
Authors
- Fang Wei (ORCID: https://orcid.org/0000-0002-0469-3567)
- Junli Li
- Yicheng Liu
- Xiufang Tan
- Qi Dong
Publication Details
- Journal
- Research Square
- Published
- 2026-11-20
- DOI
- https://doi.org/10.21203/rs.3.rs-11411539/v1
- Primary Topic
- Greenhouse Technology and Climate Control
- Type
- preprint