Theory — overview

Background on the models and methods used.

Contents

• ADM1da model — description of the 41-state model used here
  (Schlattmann 2011, agricultural extension of ADM1) and its key state indices.
• PINN — the per-window Physics-Informed Neural Network smoother: from
  the intuition for a newcomer through the loss maths and the ADM1-specific engineering to the implementation (pinn.py, pinn_smoother.py).
• Pre-trained observer — the amortised GRU observer: pre-trained
  offline on many simulated scenarios, then near-instant online inference with self-supervised fine-tuning.
• UKF ↔ PINN fusion — combining a UKF and a PINN estimate via
  covariance intersection.

Background reading

The formal observability analysis of ADM1-based models is documented in Observability → Literature review.