This dissertation studies topics of international economics, such as the long-run relationship between international capital flows and economic growth, the gravity of consumption risk sharing between countries, and the spacial consumption risk sharing across 50 states in the United States. Each chapter of the dissertation approaches one of these three topics.The first chapter examines the long-run patterns of net private and public capital inflows empirically and theoretically. Using data for 83 developing countries over the sample period 1980--2019, the empirical results show a robust positive correlation between net private capital inflows and GDP per capita growth and a robust negative correlation between net public capital inflows and economic growth. Further, this paper finds that the patterns of private and public capital flows remain when human capital is controlled in the regression model. Based on these findings, I provide a theoretical framework that explains the long-run joint behavior of private and public capital flows in a small open economy. In the balanced growth model, the benevolent government spends money on human capital investment, which is the key source of growth. The government increases its expenditure to sustain a higher balanced growth rate, which leads to an increase in public savings in the international reserve accumulation. The second chapter studies how frictions in bilateral economic linkages shape the consumption pattern across economies. Using state-level data from the US, we find that the degree of bilateral consumption risk sharing across states decreases in geographic distance. To explain this novel fact, we develop a DSGE model that incorporates trade, migration, and finance as channels of risk sharing which are subject to frictions that covary with distance. Calibrated to the US data, the model not only enables us to quantify the magnitude of the frictions in each channel, but also allows us to examine the interplay among the channels and disentangle their effects on the level, volatility, and comovement of consumption across states. Counterfactual analyses based on the model provide guidance for the design of macroeconomic policies that aim to reduce cross-region consumption disparity. The third chapter presents new evidence that trade costs impede cross-country consumption risk sharing. Our analysis exploits cross-sectional and time-series variations in trade costs across country pairs. Using the data for a large panel of countries over the period 1970--2014, we find that bilateral risk sharing improves once a pair of countries become partners under a regional trade agreement. Moreover, we establish a gravity model of consumption risk sharing by finding that countries that are more geographically distant from each other exhibit weaker bilateral risk sharing. The effect is more pronounced in the absence of RTAs, which suggests that trade-promoting policies mitigate the impact of geographic distance on risk sharing. Furthermore, we build a causal relationship between trade and risk sharing by using instrumental variables. These results point to the importance of the trade channel for international consumption risk sharing. Based on these findings, lifting trade barriers will benefit countries by reducing consumption fluctuations.

This dissertation presents three empirical studies on topics in regional growth, international finance, and global commodity prices. The first chapter analyzes the labor market reallocation following the reduction of transportation costs. The second chapter discovers developing countries' external debt holdings abilities with different exchange rate regimes, and the third chapter reveals how commodity prices are affected by information from the news in high-frequency trading.

The first chapter discovers how labor is reallocated across cities as they become more connected to the High-Speed Railway (HSR) network in China. Specifically, by adapting graph theory from computer science, this chapter constructs a continuous connectivity index that captures the changes of 285 cities' indirect connections to the HSR network over ten years. Additionally, using China City Statistic Yearbook that covers these cities' labor market outcomes from 2003 to 2017, this chapter finds that increasing the indirect connectivity to the HSR network has insignificant effects on total employment. However, there are heterogeneous effects on employment in different industries. This chapter suggests that cities with higher indirect connectivity to the HSR network have more employment in skilled and non-service industries but less employment in service industries. Moreover, with or without direct connections to HSR generates different effects on labor reallocation. Cities have HSR contributes to the increase in employment in skilled industries, while cities near HSR has more increase in non-service employment.

The second chapter examines how exchange rate regimes affect external debt holdings for developing countries, both at normal times and during default episodes. In order to uncover the connection of more floating exchange regimes and external debt holdings, this chapter collects 57 countries' external debt level and exchange rate regimes from 1970 to 2007, including 232 default episodes. Using GMM regressions, the results reveal that compared to countries with more fixed exchange rate regimes, countries with more floating exchange rate regimes hold less external debts, especially during default episodes. Moreover, switching from fixed exchange rate regimes to floating regimes further reduces external debt holding, this can be driven by the abandonment of fixed exchange rate regimes commitment.

The third chapter is a group project with Yifei Sheng and Yunxiao Zhang. To uncover the news impact on the price of WTI crude oil futures, the third chapter applies supervised and unsupervised machine learning algorithms to conduct news sentiment and topic analysis. With the assumption that the crude oil futures market is efficient enough to respond quickly to new information, this chapter obtains high-frequency price and news from the Bloomberg terminal. Using results from logistic regression and K-means clustering, this chapter defines the positive score and topic for each news article as inputs for the final logistic regression. The regression results show that the "World Crude Oil" news is more positively correlated with price increase than other topics. Moreover, the "WTI Crude Oil" news has the highest correlation with the price increase as the positive score increases.

Often deemed as a "wonder material", graphene has exhibited remarkable promises in a broad range of research fields thanks to its exceptional electronic, thermal, and mechanical properties. However, issues such as the inevitable existence of defects and the complex microstructures of graphene-based materials stand as a bottleneck in realizing its full potential in real-life applications. With the fast growth of big data, machine learning has been widely applied in many fields such as finance, biology, and healthcare. The advent of machine learning approaches also offers solutions to learning patterns from complex data in material design and discovery, reducing the need for expensive, time-consuming, and tedious laboratory experiments or numerical simulations. In the present thesis, machine learning-assisted simulation and design approaches for functional nanomaterials are demonstrated, with a focus on the graphene family. Molecular dynamics simulations are conducted to numerically investigate the mechanical behavior of graphene-based materials such as graphene, graphene oxide and graphene aerogel, and various machine learning techniques including kernel ridge regression, Gaussian process metamodels, and deep reinforcement learning are used in the predictive and generative modeling of these materials. Finally, the concept and the promise of machine learning interatomic potentials in achieving efficient and accurate simulations for metal-organic framework materials are presented. The research constituting the present thesis may shed light on some new possibilities of simulating and designing functional nanomaterials, which may further improve the performances of applications such as stretchable electronics, supercapacitor devices, carbon sequestration technologies, among others.

Engineering composites, often simply referred to as composites, are materials made from two or more drastically different constituent materials. The concept of digital materials has been a popular tool among researchers to mathematically describe the structural constitution of engineering composites, by viewing them as collections of small base voxels. In this dissertation, the term "digital material" is used synonymously with "composite"; however, the former emphasizes its mathematical representation, while the latter pertains more to a material classification. Digital materials (composites) can be carefully tailored to provide mechanical properties that don't exist in traditional homogeneous materials, allowing them to play a crucial role in a wide variety of engineering applications, including aerospace, automotive, marine, construction, and sports equipment. However, this performance flexibility also poses challenges in the material design stage as experimental trial and error composite design search is too costly and inefficient. Therefore, numerical simulations such as finite element method (FEM) have been widely used in the preliminary analysis of digital materials, owing to the well-established theory of solid mechanics and numerical solutions to differential equations. Besides, recent advancement in machine learning (ML) science significantly enhances human beings' ability to manipulate big data, allowing further revolution in the classical computational frameworks on simulation and design of digital materials. A more comprehensive background introduction to the existing literature on engineering composites can be found in the first chapter.

This dissertation is composed of three major components where the first one starts by discussing supervised surrogate modeling for composite materials. Typical numerical simulations possess a computational expense on the order of $N^2-N^3$ where $N$ represents the amount of spatial discretization (mesh). As a result, simulations on large material systems with complicated mesh can take days or months to accomplish. To mitigate the expensive data cost, ML models are trained to surrogate the actual numerical simulation with faster prediction and decent accuracy. We compare some popular ML models' (e.g. linear regression, tree-based model, neural networks. etc.) performance on predicting material displacement field, where prediction time is reduced by orders of magnitudes with decent accuracy.

Despite the outstanding accelerating capability of surrogate models, obtaining sufficient high-fidelity data has been the main bottleneck for most learning problems. In the second part, we show the possibility of training ML models without the supervision of ground truth data. This can be realized by forcing the ML model to respect natural physics laws, creating what is known as physics-informed neural networks (PINNs). Here, we illustrate how weak formulated differential equations can address non-smooth or non-differentiable fields in PINNs. In addition, PINNs can be further combined with supervised learning to approach complex systems whose complete governing equations are not accessible. This physics-based training framework substantially reduces the need for ground truth data.

In the final section, we employ different optimization algorithms on the trained surrogate models to aid composite material design. We showcase the application of ML-assisted genetic algorithms in designing lattice structures and airfoils within large parameter spaces. For a digital material shear homogenization problem, the design optimization process can be harmoniously integrated with the solution process using a common objective function within the PINN framework. Apart from the above-mentioned surrogate model-based optimization schemes, complete black-box digital material design optimization is realized via reinforcement learning exploration.

The increased use of microwave technology, encompassing electromagnetic waves in the gigahertz range for applications such as wireless communication, defense, and weather radar, has highlighted the importance of effectively controlling wave responses—transmission, absorption, and reflection— in engineering and science. Absorbing unwanted waves is crucial for applications like radar stealth, reducing electromagnetic interference, and protecting against electromagnetic pollution. While carbon-based and ferrite-based composites offer a simple solution for wave absorption, they are limited in bandwidth and absorption strength due to their fixed properties post-fabrication. Incorporating engineered structures alongside these materials presents a more effective solution, enhancing bandwidth, absorption strength, and additional functions such as lightweight and load- bearing capabilities. This thesis demonstrates the design of multifunctional electromagnetic wave- absorbing structures through various methods, including simulation-based, experimental-based, and data-driven approaches. Diverse structures such as octet-truss, octet-foam, and Kelvin foam— mechanical metamaterials known for their lightweight and load-bearing properties—are explored for their broadband absorption capabilities and multifunctionality. Additionally, a tunable crisscross structure inspired by a chameleon’s color-changing strategy is showcased, utilizing machine learning- based prediction and optimization.

This dissertation presents three empirical studies on topics in labor market and globalcommodity prices. The first chapter analyzes how consumption, spending and labor market decisions vary across the life-cycle, and their implications for the aggregate economy. The second chapter discovers the dynamic adjustments of 49 world commodity prices in response to innovations in the nominal and real shocks, and the third chapter reveals how commodity prices are affected by information from the news in high-frequency trading.

The first chapter studies the effect of aging on service consumption quantity and pricein the United States. In particular, I ask, do older households consume more service goods? Can aging explain the rise of service price? I investigate the impact of age profile on service consumption and service price in the U.S. using household survey data and metropolitan statistical areas (MSAs) level variations, respectively. The results show that after controlling for household income, as people age, they consume progressively more health insurance and medical services and less food away from home. Furthermore, there is a positive correlation between service price and old-age dependency ratio1 in the long run across MSAs. I then build a two-period two-sector OLG model with an education entry cost to explain these results. In this model, age affects service price through two channels: 1) on the demand side, labor markets with older populations tend to demand more services; 2) on the supply side, the education friction would cause worker mobility from the service sector to the manufacturing sector, which further increases the service price.

The second chapter is a group project with Professor Hyeongwoo Kim. We studydynamic adjustments of 49 world commodity prices in response to innovations in the nominal exchange rate and the world real GDP. After we estimate the dynamic elasticity of the prices with respect to these shocks, we obtain the kernel density of our estimates to establish stylized facts on the adjustment process of the commodity price toward a new equilibrium path. Our empirical findings imply, on average, that the law of one price holds in the long-run, whereas the substantial degree of short-run price rigidity was observed in response to the nominal exchange rate shock. The real GDP shock tends to generate substantial price fluctuations in the short-run because adjustments of the supply can be limited, but have much weaker effects in the long- run as the supply eventually counterbalances the increase in the demand. Overall, we report persistent long-lasting effects of the nominal exchange rate shock on commodity prices relative to those of the real GDP shock.

The third chapter is a group project with Yifei Sheng and Zijing Zhu. To uncover thenews impact on the price of WTI crude oil futures, the third chapter applies supervised and unsupervised machine learning algorithms to conduct news sentiment and topic analysis. With the assumption that the crude oil futures market is efficient enough to respond quickly to new information, this chapter obtains high-frequency prices and news from the Bloomberg terminal. Using results from logistic regression and K-means clustering, this chapter defines the positive score and topic for each news article as inputs for the final logistic regression. The regression results show that the "World Crude Oil" news is more positively correlated with price increase than other topics. Moreover, the "WTI Crude Oil" news has the highest correlation with the price increase as the positive score increases.