The harvest and conversion of energy is of crucial importance for human civilization. Today, the fast growth in energy consumption, together with the environmental problems caused by fossil fuel usage, calls for renewable and clean energy supply, such as solar, wind, geothermal, and tidal energy. However, such energies are not consistent in both time and location, bringing energy storage on request. Intensive research has been focused on the development of electrochemical energy storage (EES) devices.

Among these EES devices, hydrogen fuel cells and metal-air batteries have attracted the special attention because of their high theoretical energy densities. Yet, one major issue lies in the sluggish oxygen reduction reaction (ORR) that takes place at the cathodes. For example, the theoretical voltage of a hydrogen-oxygen fuel cell is 1.23 V (standard condition). However, the voltage output obtained under a meaningful current density is only about 0.7 V, where the voltage loss is primarily caused by the overpotential in the cathodes. Developing efficient electro-catalysts, which can lower the overpotential of ORR, is indispensable for achieving high performance devices.

The state-of-the-art ORR electro-catalysts are generally based on platinum, which is limited by cost and scarcity. Developing electro-catalysts based on earth abundant metal elements is critical for large-scale application of fuel cells and metal-air batteries. Among the non-precious-metal catalysts (NPMCs) explored in recent decades, pyrolyzed iron-nitrogen-carbon (Fe-N-C) catalysts is widely regarded as the most promising candidate for replacing platinum due to their high activity. However, the traditional method for preparing Fe-N-C catalysts involves high-temperature pyrolysis of the precursors, which is a highly complex and unpredictable process. As-prepared Fe-N-C catalysts usually contain mixed chemical phases (e.g., Fe-based nanoparticles, Fe-N coordination site and various nitrogen species), as well as carbon scaffolds with random morphology. Such complexity makes it difficult to identify the active site and control the porous structure. Though progress has been made in improving their performance through delicate selection of precursors, such process is largely based on test-and-trial method, shedding little light on the understanding of the material.

In this dissertation, we designed a novel "post iron decoration" synthetic strategy towards efficient Fe-N-C catalysts, which de-convolutes the growth of iron and nitrogen species, enables the rational design of the catalyst structure, and provides a series of effective model materials for active site probing. Specifically, liquid iron penta-carbonyl was used to wet the surface of mesoporous N-doped carbon spheres (NMC), whose porous structure is determined by the template used for preparation. The obtained Fe(CO)5/NMC complex was then pyrolyzed to generate the Fe/NMC catalysts. Through comparative study and thorough material characterization, we demonstrated that the pyridinic-N of NMC anchors the Fe atoms to form Fe-Nx active sites during pyrolysis, while the graphitic-N remains ORR active. The excessive Fe atoms were aggregated forming fine nanoparticles, which were subsequently oxidized forming amorphous-iron oxide/iron crystal core-shell structure. All the composing elements of Fe/NMC catalysts are uniformly distributed on the NMC scaffold, whose porous structure is shown to be not affected by Fe decoration, guaranteeing the effective exposure of active sites. The best performing Fe/NMC catalysts exhibited a high half-wave potential of 0.862 V, which is close to that of the benchmark 40% Pt/C catalyst. Such high activity is primarily attributed to the Fe-Nx active sites in the catalysts. While the surface oxidized Fe crystallites though not being the major active site, is revealed to catalyze the reduction of HO2-, the 2e ORR product, facilitating the 4e reduction of oxygen. Finally, such synthetic strategy is successfully extended to prepare other Me-N-C materials.

Based on the established understanding of the active sites, we then complexed the active Fe(CO)5 molecules with a N-rich metal-organic framework (ZIF-8) to form a precursor, which was subsequently pyrolyzed to form Fe-NC catalysts. During the pyrolysis, Fe(CO)5 reacts homogeneously with the ZIF-8 scaffold, leading to the formation of uniform distribution of Fe-related active sites on the N-rich porous carbon derived from ZIF-8. The zinc atoms in the crystalline structure of ZIF-8 serves as thermo-sacrificial template, resulting in the formation of hierarchical pores that provide abundant easily accessible ORR active sites. In virtue of these advantageous features, the best performing Fe-NC catalyst exhibited a high half-wave potential of 0.91 V in rotating disk electrode experiment in 0.1 M NaOH. Furthermore, zinc-air battery constructed with Fe-NC-900-M as the cathode catalyst exhibited high open-circuit voltage (1.5 V) and a peak power density of 271 mW cm-2, which outperforms those made with 40% Pt/C catalyst (1.48 V, 1.19 V and 242 mW cm-2), and most noble-metal free ORR catalysts reported so far. Finally, such a synthetic method is economic and easily-scalable, offering possibility for further activity and durability improvement.

The successful application of ConvNets and other neural architectures to computer vision is central to the AI revolution seen in the past decade. There have been strong needs for scaling vision architectures to be both smaller and larger. Small models represent the demand for efficiency, as the deployment of visual recognition systems is often on edge devices; large models highlight the pursuit for scalability - the ability to utilize increasingly abundant compute and data to achieve ever-higher accuracy. Research in both directions are fruitful, producing many useful design principles, and the quest for more performant models never stops. Meanwhile, the very fast development pace in the literature can sometimes obscure the main mechanism responsible for certain methods' favorable results.

In this dissertation, we will present our research from two aspects in this area: (1) developing intuitive algorithms for efficient and flexible ConvNet model inference; (2) studying baseline approaches to reveal what is behind popular scaling methods' success. First, we will introduce our work on one of the first anytime algorithm for dense prediction. We will then examine the effectiveness of model pruning algorithms by comparing them with an extremely simple baseline, and argue their true value may lie in learning architectures. Finally, We present our work on questioning whether self-attention is responsible for Transformer's recent exceptional scalability in vision, by modernizing a traditional ConvNet with design techniques adapted from Transformers.

Background

Methods

Results

Conclusions