MIT Theses
https://hdl.handle.net/1721.1/7582
2022-01-23T09:49:21ZData-Driven Supply Regulation to Improve Farmers’ Income in Agricultural Markets
https://hdl.handle.net/1721.1/139612
Data-Driven Supply Regulation to Improve Farmers’ Income in Agricultural Markets
McCombs, Morgan Jane
Millions of smallholder farmers live in persistent poverty. To improve farmers’ livelihood and connect physically distant markets, the Karnataka state government in India launched the Unified Market Platform (UMP) in 2014. The UMP is an online agri-platform used to facilitate commodity auction markets. In this research we assume the role of the state government and propose to regulate the daily auction supply for sale under the objectives of increasing farmers’ revenue and reducing day-to-day price fluctuation. Using data collected via the UMP, we estimate a model on the relationship between market supply, demand, and realized daily price, then employ Dynamic Programming (DP) to determine the government’s optimal inventory control policy. We characterize the optimal solution and revenue, and how the supply and price model parameters affect the solution. We first consider a benchmark, deterministic formulation, then model extensions that incorporate inventory holding cost and stochastic demand. In the deterministic setting, supply regulation yields an average annual revenue improvement of 0.1%-4.3%; price variance is reduced by an average of 34%-99%. Comparable results are achieved with the model extensions across most commodities and markets. We conclude there is significant benefit to implementing an inventory regulation scheme.
With the prospect of further improving farmers’ revenue, we also consider the combinatorial problem of merging two or more markets. To gain structural insights, we first analyze the special case of potentially merging two identical markets. The theoretical results indicate that merging is optimal if the difference between daily supply quantities is relatively small and the supply quantities are initially small. Using UMP data, our counterfactual empirical analysis shows that merging markets yields additional revenue gains of 0.3%-3.8%. Finally, we numerically analyze potential revenue gains when merging actual markets from the UMP data. Most analyses yield comparable results, however certain settings produce mixed results for different parameters.
2021-06-01T00:00:00ZUncovering Perovskite Degradation Equations Using Scientific Machine Learning
https://hdl.handle.net/1721.1/139611
Uncovering Perovskite Degradation Equations Using Scientific Machine Learning
Naik, Richa Ramesh
Many important materials are metastable or unstable under certain operating regimes. The degradation mechanisms can be varied and complex, making the discovery of underlying differential equations (DEs) through a first-principles approach challenging. This invites the application of data-science methods to infer root causes. Traditionally, machine learning (ML) applied to materials research has focused on optimization and regression over a limited training set. Inferring physical laws directly from data may allow the extraction of more generalizable scientific information that enables one to understand underlying mechanisms. In this study, we apply scientific ML — a blend of traditional scientific mechanistic modeling (differential equations) with machine learning methodologies — to identify differential equations governing the degradation of methylammonium lead iodide perovskite (MAPI), a material with known instability under environmental stress. We explore scientific ML applied to simulated and experimental datasets, obtaining equations that describe the temperature- and time-dependencies of MAPI degradation. Our method of choice is sparse regression method PDE-FIND (Rudy, Samuel H., et al. "Data-driven discovery of partial differential equations." Science Advances 3.4 (2017): e1602614). We find that the underlying DE governing MAPI degradation corresponds to the Verhulst logistic function, often used to describe autocatalytic or self-propagating kinetics. This thesis demonstrates the application of scientific ML in practical materials science systems, highlighting the promise and challenges associated with ML-aided scientific discovery.
2021-06-01T00:00:00ZChange Point Detection in Time Series via Multivariate Singular Spectrum Analysis
https://hdl.handle.net/1721.1/139610
Change Point Detection in Time Series via Multivariate Singular Spectrum Analysis
AlAnqary, Arwa
The objective of change-point detection (CPD) is to estimate the time of significant and abrupt changes in the dynamics of a system through multivariate time series observations. The setup of CPD covers a wide range of real-world problems such as quality control, medical diagnosis, speech recognition, and fraud detection to name a few. In this thesis, we develop and analyze a principled method for CPD that combines a variant of multivariate singular spectrum analysis (mSSA) approach with the cumulative sum (CUSUM) procedure for sequential hypothesis testing. In particular, we model the underlying dynamics of multivariate time series observations through the spatio-temporal model introduced recently in the mSSA literature. The change points in such a setting correspond to a change in the underlying spatio-temporal model. As the primary contributions of this work, we develop a CUSUM-based algorithm to detect such change points in an online fashion. Further, we extend the analysis of CUSUM statistics, traditionally done for the setting of independent observations, to the dependent setting of (multivariate) time series under the spatiotemporal factor model. Specifically, we analyze the performance of our algorithm in terms of the average running length (ARL) – a common metric used traditionally in sequential hypothesis testing to measure the trade-off between the delay in a true detection and the running time until a false detection. We formally establish that for any given detection parameter h > 0, on average, the algorithm detects a change point with a delay of 𝑂(h) time steps, while in the case of no change it takes at least Ω(exp(h)) time steps until it makes a false detection. Finally, we empirically show that the proposed CPD method provides state-of-the-art performance across synthetic and benchmark datasets.
2021-06-01T00:00:00ZEnhancing surrogate models of engineering structures with graph-based and physics-informed learning
https://hdl.handle.net/1721.1/139609
Enhancing surrogate models of engineering structures with graph-based and physics-informed learning
Whalen, Eamon Jasper
This thesis addresses several opportunities in the development of surrogate models used for structural design. Though surrogate models have become an indispensable tool in the design and analysis of structural systems, their scope is often limited by the parametric design spaces on which they were built. In response, this work leverages recent advancements in geometric deep learning to propose a graph-based surrogate model (GSM). The GSM learns directly on the geometry of a structure and thus can learn on designs from multiple sources without the typical restrictions of a parametric design space.
Engineering surrogate models are often limited by data availability, since designs and performance data can be expensive to produce. This work shows that transfer learning, through which training data of varying topology, complexity, loads and applications are repurposed for new predictive tasks, can be used to improve the data efficiency of surrogates, often reducing the required amount of training data by one or two orders of magnitude. This work also explores new potential sources for training data, namely engineering design competitions, and presents SimJEB, a new public dataset of simulated engineering components designed specifically for benchmarking surrogate models. Finally, this work explores the emerging technology of physics-informed neural networks (PINNs) for structural surrogate modeling, proposing two new heuristics for improving the convergence and accuracy of PINNs in practice. Combined, these contributions advance the generalizability and data efficiency of surrogate models used in structural design.
2021-06-01T00:00:00Z