KGML-N: a knowledge guided machine learning modeling framework for efficient simulation of N-yield response
Applying the right amount of nitrogen (N) fertilizer in corn farming is a major challenge. Current process-based models (PBMs) meet efficiency bottlenecks in large scale applications, while data driven models’ reliability are questioned as they are not interpretable. This thesis presents a solution that combines the best of both worlds by developing a knowledge guided machine learning (KGML) surrogate model—an efficient and interpretable machine learning tool to model N-yield response. Our model KGML-N was well designed to include necessary intermediate variables (IMVs) for N-yield response modeling, and then trained to mimic an advanced PBM ecosys. We then tested our model to see if it could capture environmental impacts on N-yield response and maintain consistency with empirical knowledge under out of distribution scenarios. This work validates our surrogate model as a powerful, fast, and scientifically sound tool for developing more precise nitrogen management recommendations.