Modeling Material Flow Dynamics in Industrial-Natural Systems: Machine Learning and Causal Analysis for Resilience Evaluation and Sensor Minimization
This dissertation builds a unified toolkit for analyzing coupled industrial-natural systems through four integrated advances. First, it extends Sparse Identification of Nonlinear Dynamics (SINDy) by adding higher-order input derivatives, allowing parsimonious surrogate models for both chemical transesterification and watershed hydrology while retaining physical interpretability. Second, it introduces a hybrid strategy that injects SINDy-derived error terms into established Cardinal Temperature Models, sharply improving microalgal growth forecasts without sacrificing mechanistic meaning. Third, it develops a causality-guided sensor-minimization method that combines Liquid Time-Constant neural networks with perturbation analysis to identify the smallest measurement set that still reconstructs system states; the approach is validated across mechanical, chemical, and ecological testbeds. Fourth, the work integrates these node-level surrogates into a material-flow network simulator to evaluate resilience of a soybean-biodiesel supply chain under RCP 4.5 and 8.5 climate scenarios, revealing nonlinear production failures, recovery dynamics, and tipping thresholds linked to farm size and climate forcing. Collectively, the framework delivers interpretable equations for planners, lightweight soft sensors for operators, and quantitative resilience metrics for decision-makers—providing a transferable blueprint for sustainability assessment in complex material-flow networks.
Funding
DGE-1842166
NSF-FMRG Eco 2229250
NSF-CBET 1805741
History
Degree Type
- Doctor of Philosophy
Department
- Agricultural and Biological Engineering
Campus location
- West Lafayette