Paradigms of Scientific Machine Learning: From Model-Based Inference in NMR Spectroscopy to Predictive Automation in Agricultural Robotics

Data-driven modeling is crucial for extracting insights from complex scientific systems. Modern machine learning provides two powerful but philosophically distinct paradigms: model-based inference, which aims to understand physical mechanisms, and predictive modeling, which optimizes for forecasting accuracy. This dissertation explores and constrast these two paradigms through their application in two different domains: Nuclear Magnetic Resonance spectroscopy (NMR) for biochemical analysis and computer vision for agricultural robotics.

The first section presents two distinct Bayesian inference models I developed to analyze 1D NMR spectra. The first is a nonparametric model which enables robust spectral deconvolution and automatic peak detection of general 1D NMR spectra which we apply to flourine-19 ($^{19}$F), while the second is a parametric model used to quantify the kinetics of chemical exchange. Both models provide not only estimates of spectral parameters but also a rigorous quantification of their associated uncertainty.

The second section focus on a deep learning approach I developed to address the challenge of automated feed monitoring in precision agriculture. I describe a Convolutional Neural Network (CNN) which was trained to accurately predict leftover feed quantities from images; a crucial step in the development of personalized feeding schedules.

Ultimately, these distinct lines of inquiry converge on a central theme: the choice between inferential and predictive modeling is dictated by the scientific objective, whether that objective is achieving deep physical insight or enabling robust automation.