《1 Engineering research fronts》

1 Engineering research fronts

《1.1 Trends in Top 10 engineering research fronts》

1.1 Trends in Top 10 engineering research fronts

Table 1.1.1 summarizes the Top 10 engineering research fronts in the information and electronic engineering field, which encompasses the subfields of electronic science and technology, optical engineering and technology, instrument science and technology, information and communication engineering, computer science and technology, and control science. Table 1.1.2 shows the number of core papers published from 2016 to 2021 related to each research front.

(1)  Space–air–ground–sea integrated networking theory and technique

The space–air–ground–sea integrated network (SAGSIN) based on the ground network and supplemented and extended by the space, air, and sea networks, can provide ubiquitous, intelligent, collaborative, and efficient information assurance infrastructure for various network services in wide areas. The space network comprises the satellite backbone and access network and supports global coverage, ubiquitous connection, and broadband network access. The air network consists of the high-altitude platform and the unmanned aerial vehicle self-organized network, which can enhance network coverage, enable edge computing service, and provide a flexible network configuration. The ground network in SAGSIN comprises the ground internet and mobile cellular network, which controls the network service in areas with dense requests. The sea network uses maritime mobile and satellite networks to fulfill the communication requirements in maritime activities. Additionally, based on the deep integration of heterogeneous networks, the SAGSIN can use various network resources effectively and comprehensively to perform intelligent network management and information processing to cope with different network service requirements and realize network integration, network function modularity, and service customization. Therefore, the SAGSIN has shown unprecedented prospects in diverse fields, such as wide-area coverage, the Internet of Things (IoT), intelligent transportation, remote sensing and monitoring, and the military. Specifically, technologies of the low-earth-orbit (LEO) satellite constellation play a core role in establishing an all-coverage, all-connected, and all- knowing SAGSIN. SpaceX’s “Starlink” from the USA currently leads the competition for low-orbit satellite constellations. It intends to launch 42 000 satellites to build a broadband satellite communication network capable of providing global

《Table 1.1.1》

Table 1.1.1 Top 10 engineering research fronts in information and electronic engineering

Note: ① AI is short for artificial intelligence; CMOS is short for complementary metal oxide semiconductor;

​② For No.3 and No.4, all the detected papers are adopted as the core papers.

《Table 1.1.2》

Table 1.1.2 Annual number of core papers published for the Top 10 engineering research fronts in information and electronic engineering 

coverage. More than 3 000 low-orbit satellites have already been operational in orbit as of August 2022, with more than 500 000 subscribed to broadband access users worldwide. However, several challenges exist in the SAGSIN, as follows: high-dynamic, extra-heterogeneity, ultra-complexity, and multidimensional requirements. The focus of this study includes network architectural design, protocol design, network resource management and scheduling, efficient transmission technology, network security, and privacy preservation, among others.

(2) Theories and algorithms for trustworthy AI

Trustworthy artificial intelligence (AI) enhances the credibility of complex AI systems and algorithms, such as deep neural networks. Specifically, the concept of credibility includes the following four aspects: ① the interpretability and quantifiability of AI systems concerning knowledge representations; ② the interpretability and quantifiability of AI systems regarding representation capacity, which include generalization, robustness, fairness, and privacy protection, among others; ③ the interpretability of AI systems in learning and optimization; ④ the interpretability of the internal mechanism of various AI algorithms.

Therefore, current studies focus on the following to promote the development of trustworthy AI:  ①  qualitatively or quantitatively interpreting knowledge representations modeled using AI systems, such as visualizing semantic information in intermediate-layer features and quantifying the importance of input variables to final decisions; ② evaluating, interpreting, and improving AI systems’ representation capacity, including generalization, robustness, and fairness, among others; ③ explaining the reasons for the effectiveness of AI system optimization algorithms, and exploring and identifying latent defects in recent empirical optimization algorithms; ④ designing interpretable AI systems to enhance credibility in the system design stage.

Although trustworthy AI has recently received extensive attention, several key bottlenecks are rarely involved and explored, including the following: ① exploring, identifying, and quantifying the essential factors that determine the AI system’s representation capacity; ② unifying and interpreting the internal mechanisms of various empirical AI algorithms, thereby uncovering the common nature behind their effectiveness and grasping the essential parts of previous algorithms; ③ the theory-driven design and optimization of AI systems. Particularly, a few international research institutions and teams, such as the Massachusetts Institute of Technology (MIT) and Shanghai Jiao Tong University, have recognized and particularly considered the abovementioned key issues and conducted some forward-thinking explorations on them.

(3) Silicon-based CMOS terahertz imaging technique

The terahertz (THz) imaging technique involves directing continuous or pulsed THz waves onto a target and collecting signals reflected by or transmitted through the object. Fourier transformation algorithm can obtain information of the intensity and phase of signals from each target point. Subsequently, target imaging can be achieved after spectrum analyses and digital signal processing. The THz frequency band is also located in the electromagnetic spectrum between microwave and infrared waves. It has special characteristics, such as high transmission, low energy, high coherence, and high-transient response. Therefore, imaging techniques at THz frequencies have several incomparable advantages over traditional methods, such as visible light, ultrasonic, and X-ray imaging, and have wide application prospects in national security, safety inspection, biomedicine, and environmental monitoring.

Recently, the THz imaging technique has attracted broad research interest worldwide with the continuous upgrading of silicon-based technology and the great improvement of its radio frequency (RF) performance. The CMOS-assisted technique is characterized by its small size and low power consumption, among others, which can meet the commercial demand for high integration density and low cost. Additionally, the silicon-based CMOS THz imaging technique has achieved several technological breakthroughs in resolution. For example, Cornell University developed a 220 GHz imaging system with a 2 mm lateral resolution and a 2.7 mm range resolution based on the 55 nm bipolar complementary metal oxide semiconductor (BiCMOS) technology. However, overcoming the diffraction limit and further improving the imaging resolution remain an important research focus. Furthermore, the complex parasitic and coupling effects of silicon-based technology, integrated circuit (IC) distribution effects, and source synchronization technology are also the research focus in this field.

(4)   Silicon-based AI photonic computing chip theory and design

AI is a strategic technology leading the future, and computing power is a solid foundation to support AI’s rapid development. However, as the performance of microprocessors increases slowly, Moore’s law is starting to fail, and conventional electronic computing circuits cannot meet the growing demand for AI computing power because of the existence of a “power wall” and “memory wall”. Additionally, photons possess inherent advantages, such as low latency, low power consumption, high throughput, and parallelism, over electrons. The silicon-based AI photonic computing chip realizes linear analog calculations using the physical transmission characteristics of light in the silicon-based waveguide to provide powerful photonic chips for AI applications by leveraging the silicon photonics integration technology.

Considerable efforts have recently been directed to studying silicon-based AI photonic computing chips. Currently, these study areas include limited to image matrix convolution photonic computing chips, integral and differential photonic chips, complex-domain Fourier transform photonic chips, reservoir photonic computing chips, photonic neuromorphic computing (brain-like computing) chips, heuristic algorithm solvers for NP-complete problems, and spiking neural network photonic chips, among others.

Silicon-based photonic computing is envisioned as a potential solution for pushing the limits of electronic computing in the post-Moore era. Interestingly, with the continuous improvement in silicon-based optoelectronic integration, photonic computing chips can greatly accelerate AI algorithms’ processing speed and create possibilities for new-processor architectures. Furthermore, combining photonic analog computing and electronic digital logic operation will provide integrated optoelectronic computing with complementary advantages that would change the existing computing system model and build a novel computing basic system with high computing capacity and low power consumption. Therefore, it is an inevitable development trend in the future.

(5)  High-precision space-based gravitational wave detection technology

Space-based gravitational wave detection refers to using multiple spacecraft to form a giant laser interferometer in space to detect gravitational waves. Space-based gravitational wave detection mainly focuses on the millihertz frequency band, which contains gravitational wave signals characterized by large magnitude and long duration and also enjoins the rich types, large numbers, and diverse spatial distribution of gravitational wave sources. These factors make the millihertz frequency band the gold mine in gravitational wave detection, with significant implications for astrophysics, cosmology, and fundamental physics research.

Particularly, space-based gravitational wave detection involves two core technologies. The first involves establishing “probe heads” for gravitational wave detection, using a group of objects in nearly ideal inertial motion to provide reference points in space position for measuring the distance change caused by gravitational waves. The corresponding technology is known as space inertial reference, in which one is faced with the problems of high-precision inertial sensing, micro-Newton-level propulsion, and high-precision drag- free control, among others. The second involves developing a “ruler” for gravitational wave detection, using a laser to measure the distance change between inertial reference points on different satellites. The corresponding technology is known as inter-satellite laser interferometry. Here one faces problems of the ultra-stable optical bench, long life satellite-borne frequency-stabilized laser, and weak light phase locking, among others.

Furthermore, space-based gravitational wave detection necessitates a novel concept of spacecraft development. For example, the thrusters, which originally belonged to the satellite bus, are currently a key element in constructing the gravitational wave “probe heads”. Additionally, the structure and thermal stability of the satellite bus have become key factors that can affect successful gravitational wave detection. Therefore, in designing and developing spacecraft for gravitational wave detection, the boundary between the satellite bus and the payload must be broken down and considered as a whole.

Moreover, any scientific and technological powers that may be capable of space-based gravitational wave detection face a great challenge. After approximately 30 years of planning, the European Space Agency is seeking to launch the first space- based gravitational wave detector in the 1930s, possibly with some minor contribution from the USA. Japan has been relentless in the pursuit of its own space-based gravitational wave detection mission. China is actively studying space- based gravitational wave detection and is striving to play an important role in this field. In 2020, the Chinese Ministry of Science and Technology launched the key R&D plan for “gravitational wave detection”, with the most funds going to support the key genetic technologies of space-based gravitational wave detection.

(6) Manufacture of integrated circuits at the atomic scale

The term “atomic scale” in IC technology generally refers to the thickness of a single atomic layer. This thickness scale depends on the size of atoms and the crystal structure and is typically in the range of 0.1 nm, such as 0.2–0.5 nm. With the development of ICs beyond the 10 nm technology node, critical physical dimensions, tolerance of deviations in the critical dimensions, and the precision of measurement tools all enter the range of atomic scale. Transistors feature an increasing number of critical layers that are only a few atoms thick or wide. Such critical layers include gate dielectrics, metal gate layers for work function engineering, and fins of the fin field-effect transistor (FinFET), all with a feature size less than 10 atomic layers thick. Interestingly, as the common process for IC manufacture, atomic layer deposition (ALD) produces a film of 0.03–0.07 nm per cycle, far below the atomic scale thickness. Besides the absolute feature size, the mass production of ICs emphasizes the tolerance of thickness or width control, which benefits the yield rate. For example, the work function metal gate layers variation should be less than one atomic layer thick; otherwise, the transistors would suffer from unacceptable deviations in threshold voltage and performance, causing failure of the IC chips integrated with billions of devices. Therefore, the measurement tools applied in IC manufacturing have high- precision of approximately 0.01 nm, which is lower than a single atomic layer, to ensure the precise measurement of the above dimensions. Additionally, two-dimensional (2D) and oxide channel materials have been widely applied in academic research to fabricate transistors and simple circuits, providing a new solution to realizing atomic scale IC manufacture in the future.

(7) Brain–machine interfaces and their clinical applications

The brain–machine interface (BMI) is an intelligent system that establishes a direct and bidirectional communication pathway between the brain and external devices. The BMI either controls the external device using brain signals or modulates the brain state using external devices, aiming to monitor the brain state, treat brain disorders, and enhance brain functions. Since the 1970s, when the concept of “BMI” was first proposed, BMI technology has been continuously advancing, with an explosive development trend in the past decade. The key technologies in BMI are as follows: ① electrode design, fabrication, and minimally invasive implantation technology to obtain large-scale neural signals; ② neural decoding technology to understand the brain activities from complex, large-scale neural signals; ③ electrical, magnetic, and optical technology to stimulate neural populations; ④ closed-loop neural modulation technology based on real-time neural signals; ⑤ high-performance, low power neurochip technology, which integrates neural signal storage, decoding, stimulation, and modulation.

BMIs have many applications in the diagnosis, treatment, rehabilitation, and other aspects of mental/neurological diseases. For example, first, BMIs will restore motor and sensory functions. Such BMIs use neural signals to decode the brain’s movement intentions to drive external devices to complete the intended movements; simultaneously, sensory feedback is provided. Interestingly, these BMIs will provide new approaches to treating motor disabilities, such as paralysis. Recently, this BMI type will be extended to explore the restoration of more precise motor and sensory functions, such as human speech decoding and vision recovery. Second, BMIs will enhance cognitive functions. Such BMIs employ external devices to reconstruct or facilitate the communication between brain regions, aiming to repair or enhance specific cognitive functions. One example is the development of memory prostheses to explore the enhancement of impaired memory functions in patients. Third, BMIs will facilitate the treatment of neurological and neuropsychiatric disorders. Such BMIs use neural signals to guide the external devices in delivering optimal brain stimulation in real-time for a precise intervention. Notably, these BMIs have demonstrated great potential in treating neurological and neuropsychiatric disorders, such as Parkinson’s disease, epilepsy, and depression.

Although BMI technology has broad clinical applications, its performance, accuracy, efficiency, and safety still necessitates resolving several major challenges. Some challenges are as follows: ① developing neural recording and stimulation hardware that is stable in the long term, biocompatible with brain tissue, and has high spatial– temporal resolutions; ② developing accurate and stable BMI decoding algorithms to achieve the fine-grained control of various complex external devices; ③ developing precise and robust BMI modulation algorithms to effectively and safely modulate various brain states; ④ studying the ethics and data security of BMI technology.

(8)  Humanoid robot behavioral developmental learning and cognitive technology

Humanoid robots should interact with the surrounding physical world in a developmental learning manner to strengthen their behavioral capabilities, enhance their level of human-like cognition, such as movement, manipulation, understanding, memory, and reasoning, and exhibit more intelligent behavioral actions. The related technology is known as humanoid robot behavioral developmental learning and cognitive technology. The research directions are as follows: ① autonomous behavioral development; ② embodied intelligence (the evolutionary process of ontological structure and intelligence of robots performing various tasks in actual physical environments); ③ affordance learning (potential behaviors between robots and environments and the effects of these potential behaviors); ④ robot learning platforms (simulation software or real physical machines). Therefore, novel algorithms for brain-like constructs with learning and cognitive capabilities, particularly memory and learning, should be developed. The development of behavioral cognitive systems will enable robots to be more human-like and perform active, intrinsically motivated, lifelong learning and development in motor skills and behavioral cognition. Perceptual representations generalizable in different environments and tasks and intertwined with multimodal perceptual joint learning are crucial. Subsequently, AI is considered a physical entity to study, and the changes in robotic bodies in response to natural selection are studied in a simulation environment. This is a completely different paradigm from viewing AI as only an algorithm. Additionally, studying the affordance of robots is necessary to rescue and exploration tasks. Therefore, analyzing the potential behaviors between the robot and the environment and the effects of these potential behaviors can enable the robot to perform better in the unknown environment. Finally, we must upgrade or develop robotic platforms that will enable a better analysis of the functional interaction between robots, humans, and the environment.

(9)  Quantum circuits and chips

The quantum circuit model is a universal language for describing quantum algorithms, with each computation process comprising a sequence of quantum gates, measurements, and other possible operations. Most quantum algorithms, including Shor’s algorithm, Grover search algorithm, and HHL algorithm, are presented via the quantum circuit model. Additionally, the quantum circuit model is widely used to simulate quantum physical and chemical systems. Currently, quantum computing has entered the era of noisy intermediate-scale quantum systems. Given the current technology and engineering level of physical hardware, there are inherent limitations on the scale, depth, and qubit number of quantum circuits. Therefore, the degree of simplification of quantum circuits directly affects the scope of quantum computer applications. Designing quantum circuits for various practical computation problems with as few gates as possible, as shallow as possible, and as few qubits as possible is an important research direction in quantum computation. Besides, it is an important research direction to characterize the computation power of quantum circuits under various resource constraints and the separation between quantum and classical circuits.

Quantum chips are compact quantum circuits that are integrated on various physical platforms. The realization of quantum chips is an inevitable path toward the practical commercialization of quantum communication and computation. Additionally, the technical routes of quantum chips can be categorized into the following, based on the dependence of quantum circuits on the different physical platforms: superconducting quantum chips, semiconductor quantum dot (QD) chips, and optical quantum systems. Among these different implementations, superconducting circuits outperform other platforms in the scaling of the qubits that can be integrated; semiconductor QDs have good integrability and scalability, making them strong candidates for realizing solid-state quantum computing; the realization of optical quantum chips benefits from the techniques using traditional photonic ICs and classical optical communication. The major challenges of quantum chip implementations include improving the fidelity of quantum gates, coherence time, and decreasing the crosstalk and measurement error. Fault tolerant circuits/chips will play an important role. For future directions, people are devoted to expanding the scale of integration and improving the coherence time, manipulation accuracy and speed, and scalability.

(10)  Future industrial internet architecture and full-element interconnection

The term “industrial internet” was first proposed by General Electric (GE) in 2012. It focuses mainly on predictive maintenance to realize industrial automation and intelligence. After that, European countries, mainly Germany, put forward “Industry 4.0” in 2013. China also announced the “Made in China 2025” plan in 2015, which gave a richer connotation to the “industrial internet” and gradually established the full- element interconnection architecture.

The industrial internet architecture includes the network infrastructure, enabler platform, and security mechanism. The so-called “full-element interconnection” represents connecting people, machines, materials, rules, and the environment through networks and identification systems. The interconnectivity also includes the link between the entire lifecycle management among the value, supply, and industrial chains and R&D, production, and logistics, among others. The related technology expands across four directions as follows: ① network technologies, such as interconnection, deterministic transmission, identifier resolution, and computing and network convergence; ② data-handling operations, such as data collection, cleaning, training, and analyses; ③ intelligent platform and management technologies, such as cyber–physical systems, model and application analyses, supply chain and lifecycle management; ④ security technologies, such as the security of network, data, and physics, among others.

The industrial internet has advanced to the trial and deployment stage from the conceptual consensus. Specific technologies, such as the platform, identifier, and 5G have been applied in the industry. Interestingly, the identifier resolution system has been implemented at the five National Top Nodes in China. The following trends appear on the horizon from the perspective of architecture and technology:

1)  Toward specific scenarios: Derives architectures for guiding the implementation of various scenarios and expands the deployment to small- and medium-sized enterprises.

2)  More integrated technologies: IT (Information Technology), OT (Operations Technology), and CT (Communication Technology) are further integrated to solve the ecological interconnection of cloud networks and improve the efficiency and quality of all production and manufacturing links, by combining virtual reality, digital twin, and deterministic lossless connections.

3)   Trusted and more secure: Privacy-preserving and data- trusted technology will be further adopted to solve the personnel, systems, and equipment security problems.

4)    Systematic promotion: This improves the overall automation system and leverages new communication technologies. This also enables the generation of a full- element interconnected system, connects the business chain, and creates new industrial models.

《1.2 Interpretations for three key engineering research fronts》

1.2 Interpretations for three key engineering research fronts

1.2.1 Space–air–ground–sea integrated networking theory and technique

The SAGSIN comprises the space, air, ground, and sea networks that complement the conventional terrestrial network in coverage, flexibility, and monotonicity and is essential to realize ubiquitous network access and service customization. Because of the non-uniform mechanisms of existing networks, massive variations in resource distribution, complex and dynamic wireless channels, and network security challenge the network operators. The four main research directions of the SAGSIN must cope with the following: network architecture design, protocol design, resource management, and transmission technique. Therefore, we introduce the advances in these aspects as follows:

1)  Two main trends exist in network architecture design. The international standardization organization 3GPP promotes the integration of non-terrestrial networks (NTNs), including all NTNs, such as satellite and aerial networks, and terrestrial cellular networks, where the NTNs are vital in the 5G network and future 6G networks to form an interconnected network. However, the other trend is to design an efficient, globally controllable, and low cost virtualized SAGSIN architecture, based on software-defined networking (SDN) and network function virtualization (NFV). Notably, the main research institutions in this direction include the University of Waterloo, Tsinghua University, and Beijing Jiaotong University, among others.

2)   For communication protocol design, the Consultative Committee for Space Data Systems (CCSDS) protocol can achieve near-lossless multimedia streaming transmission in the SAGSIN, despite payload limitations through the iterative processing of adjacent frames, which greatly expands the supporting exchange capability of space mission information systems. Digital video broadcasting (DVB) series protocols overcome the limitation of traditional uplink power control on the volume of the RF front-end, which effectively improves the spectrum efficiency of satellite communication links, optimizes the space segment of the network, and significantly reduces the cost of satellite-based internet protocol (IP) services. However, these two protocols were previously proposed, and organizations and institutions, including 3GPP, are also exploring new communication protocols.

3)   Two main trends exist in SAGSIN resource management. One is the AI-driven resource management and scheduling mechanism, which can adapt to the characteristics of several network nodes, large-decision-making space, and heterogeneous resources in SAGSINs, thereby effectively improving the use of network resources. Another trend is the service function chain or network slicing resource scheduling technology, which uses SDN and NFV technologies to abstract and pools the physical network resource to maintain the service isolation between users and the satisfaction of multi- dimensional requirements. The main research institutions in this direction include Tsinghua University, University of Waterloo, Xidian University, and the National University of Defense Technology, among others.

4)    Inter-satellite laser communication is a potential technology for efficient transmission technology to realize high-speed inter-satellite links. Compared with RF based inter-satellite communication, it can achieve a much higher data transmission rate through a smaller antenna size. Simultaneously, the laser beam can provide the inter-satellite laser link with a narrower beam and higher directivity, eliminating interference and promoting network security. Currently, microwave communication remains the main inter-satellite link communication method in engineering. Therefore, the preliminary inter-satellite laser communication test and deployment are expected to be performed by the end of 2023. The main research institutions in this direction include Beihang University, Xidian University, Southeast University, Beijing Jiaotong University, and Northeastern University, among others.

The LEO constellation system construction is also a necessary R&D focus for the SAGSIN. The Iridium system is the earliest planned and deployed global coverage satellite network. In 2015, the SpaceX projected “Starlink” LEO satellite network sparked interest in academia and the industry. SpaceX announced that tens of thousands of low-orbit satellites would be launched to provide high-speed network access globally. Until now, “Starlink” has completed its initial deployment, which can reach the highest download speed of 301 Mbps and has provided network access in dozens of European and American countries. Additionally, China has several LEO satellite communication systems that are expected to be built, including “Apocalypse”, “Hongyan”, “StarRides”, and “China SatNet”, among others. Furthermore, the earliest of them are expected to be deployed by the end of 2023.

Table 1.2.1 presents the countries with the greatest output of core papers on “space–air–ground–sea integrated networking theory and technology”. China has obvious advantages, and the number of core papers is about three times that of second-placed Canada. China’s international cooperation target is mainly Canada; however, it has some collaboration with the UK, the USA, and Japan (Figure 1.2.1). Among the Top 10 institutions with the greatest output of core papers (Table 1.2.2), the University of Waterloo in Canada produces the most papers. Additionally, six of the Top 10 institutions with the greatest output of core papers globally are from China; the rest are distributed in Japan, Norway, and the UK. Regarding institutional cooperation (Figure 1.2.2), five institutions from China collaborate closely with the University of Waterloo and two institutions from China collaborate closely with the University of Surrey, and the Beijing Institute of Technology and the University of Oslo have some cooperation. Concerning the greatest output of citing papers (Table 1.2.3), China ranks first, accounting for 49.62%, followed by the USA. The rest of the countries are all less than 10%. Among the Top 10 citing paper producers (Table 1.2.4), all are from China, excluding the fifth-ranked University of Waterloo, which reflects China’s considerable attention in this direction.

Currently, the theory and technology of space–air–ground– sea integrated communication networking are at different levels of development in China and abroad. However, they are all in the design and initial deployment stage. Figure 1.2.3   is the roadmap of the engineering research front of “space–air–ground–sea integrated networking theory and technique”. From the perspective of technical indicators, the greatest LEO constellation will contain thousands of satellites by 2025. Additionally, it is expected that the scale of a single constellation will reach tens of thousands by 2030. In the next 5 years, the LEO satellite network test rate can reach 500

《Table 1.2.1》

Table 1.2.1 Countries with the greatest output of core papers on “space–air–ground–sea integrated networking theory and technique”

《Figure 1.2.1》

Figure 1.2.1 Collaboration network among major countries in the engineering research front of “space–air–ground–sea integrated networking theory and technique”

《Table 1.2.2》

Table 1.2.2 Institutions with the greatest output of core papers on “space–air–ground–sea integrated networking theory and technique”

《Figure 1.2.2》

Figure 1.2.2 Collaboration network among major institutions in the engineering research front of “space–air–ground–sea integrated networking theory and technique”

《Table 1.2.3》

Table 1.2.3 Countries with the greatest output of citing papers on “space–air–ground–sea integrated networking theory and technique”

Mbps, and the delay can reach 60 ms. Furthermore, the LEO satellite network test rate will reach a minimum of 5 Gbps, with a delay of at least 20 ms from 2027 to 2032. Therefore, the total network throughput of LEO satellite networks is predicted to reach 97 Tbps, 218 Tbps, and 820 Tbps from 2022 to 2024, 2025 to 2028, and 2032, respectively. From the viewpoint of research direction, the main research directions in current “space–air–ground–sea integrated networking

《Table 1.2.4》

Table 1.2.4 Institutions with the greatest output of citing papers on “space–air–ground–sea integrated networking theory and technique”

《Figure 1.2.3》

Figure 1.2.3 Roadmap of the engineering research front of “space-air-ground-sea integrated networking theory and technique”

theory and technology” include network construction, inter-satellite laser communication technology, space–air– ground–sea networking protocol designing, multimode fusion terminal design, and potential application development. The construction of the integrated network is currently in the initial stage, which includes the construction of the LEO satellite backbone network and the testing of the system terminal. It is expected that the construction of the backbone will be completed by the end of 2025, and more low- and super-low- orbit satellites will be launched according to application needs by 2032. The current inter-satellite communication technology in the LEO satellite network is also under development regarding inter-satellite laser communication technology, and the current communication technology is primarily microwave communication. Nevertheless, laser communication remains in the test stage, which is expected to be deployed by 2025. Furthermore, concerning satellite multimode fusion terminals, several requirements on mass, volume, compatibility in a heterogeneous network, and application integration must be satisfied to adapt to multi-system, multi-band, multi- network, and multi-application, among others. Recently, the satellite multimode technique is only in the preliminary stage, and the related products are limited to gateways and larger terminals; however, portable terminals are expected to be designed and introduced into the market from 2025 to 2032. Moreover, for potential application development, the current application scenarios for the SAGSIN mainly focus on wide- area broadband access, military communications, IoT, and the Internet of Vehicles, among others. However, in the future, more potential services will be explored to further exploit the potential capabilities of the SAGSIN. Additionally, with the support of international organizations, such as 3GPP and IMT- 2030, the standardization of space–air–ground–sea integrated communication and networking has officially started, and part of the agenda is gradually being conducted. Therefore, relevant technologies, protocols, and index requirements will be further improved in the next 5–10 years.

1.2.2 Theories and algorithms for trustworthy AI

The AI field has been revolutionized by leaps and bounds, considering the great success of complex AI systems, such as deep neural networks. However, these systems are frequently regarded as black-box systems because of their complex architectures and massive parameters. Therefore, people neither understand the internal decision logics of these systems nor explain the advantages and disadvantages of these systems based on representation capacity, such as why neural networks can achieve superior performance and are simultaneously highly vulnerable to adversarial attacks. Given the uninterpretability of AI systems in the decision-making process and the representation and optimization capabilities, the credibility, controllability, and security of these systems are greatly impaired, thereby hindering the widespread application of AI, particularly in intelligent healthcare, self- driving, and other high-risk areas.

Therefore, academia and industry are committed to enhancing the interpretability of AI systems and algorithms to build trustworthy, controllable, and secure AI. Specifically, trustworthy AI aims to enhance the interpretability and quantifiability of AI systems in knowledge representation, representation capacity, optimization, and learning ability, and the interpretability of the AI algorithms’ internal mechanism.

Recently, major research interests of trustworthy AI are as follows:  ① qualitatively or quantitatively interpreting the knowledge representation modeled with AI systems, such as visualizing the semantic information contained in intermediate-layer features and quantifying the importance of input variables to final decisions; ② evaluating, explaining, and improving the representation capacity of AI systems, such as theoretically proving or empirically studying the boundaries of generalization and robustness, explaining the internal mechanisms of neural networks in generalization, robustness, and representation bottlenecks, and developing various methods (e.g., the adversarial training) to improve the robustness, fairness, or to avoid privacy leakage; ③ explaining the internal mechanism of the effective optimization of AI system algorithms, and exploring the latent defects in current empirical optimization algorithms, for example, to explain why optimization methods, such as stochastic gradient descent and dropout are effective and to identify potential mathematical defects in classical optimization operations, such as batch normalization; ④ designing interpretable AI systems and incorporating credibility into the system architecture in the system design stage, for example, with a specifically designed objective function for the convolutional neural network, each filter of the high-level convolutional layer automatically represents specific semantics.

Notably, the field of trustworthy AI has recently received extensive attention, achieving many achievements. Tables 1.2.5 and 1.2.6 present the countries and institutions, respectively, with the greatest output of core papers on this research front. Tables 1.2.7 and 1.2.8 present the countries and institutions with the greatest output of citing papers on this front, respectively. Typical institutions include MIT, Chinese Academy of Sciences, and Shanghai Jiao Tong University, among others, which are mainly in the USA and China. Additionally, many of the core papers are jointly produced by different research institutions in different countries. Figures 1.2.4 and 1.2.5 show the cooperation networks between major countries and institutions, respectively.

Although trustworthy AI has received extensive attention recently, most studies remain at the level of engineering algorithms, such as visualizing neurons of a neural network, estimating the importance of input variables, and evaluating the robustness of neural networks using the accuracy under adversarial attacks. However, several critical and fundamental bottlenecks in trustworthy AI are rarely involved and explored, which include:

1)  Exploring, identifying, and quantifying the essential factors that determine the representation capacity of AI systems: Specifically, many factors, such as network architectures and optimization tricks, affect the representation capacity of AI systems. However, such factors frequently produce complex and diverse effects at the representation capacity level. Therefore, determining the fundamental factors that impact

《Table 1.2.5》

Table 1.2.5 Countries with the greatest output of core papers on “theories and algorithms for trustworthy AI”

《Table 1.2.6》

Table 1.2.6 Institutions with the greatest output of core papers on “theories and algorithms for trustworthy AI”

《Table 1.2.7》

Table 1.2.7 Countries with the greatest output of citing papers on “theories and algorithms for trustworthy AI”

《Table 1.2.8》

Table 1.2.8 Institutions with the greatest output of citing papers on “theories and algorithms for trustworthy AI”

《Figure 1.2.4》

Figure 1.2.4 Collaboration network among major countries in the engineering research front of “theories and algorithms for trustworthy AI”

《Figure 1.2.5》

Figure 1.2.5 Collaboration network among major institutions in the engineering research front of “theories and algorithms for trustworthy AI”

representation capacity is challenging. Moreover, researchers can accurately evaluate and interpret the representation capacity of AI systems only when the decisive factors of the representation capacity are clearly identified.

2)   Unifying and interpreting the internal mechanisms of various empirical AI algorithms: To address a specific research problem, scholars propose different AI algorithms from different empirical perspectives. Specifically, these algorithms often contain identical or similar internal mechanisms. The common essence of these algorithms may be revealed through the unification and explanation of the internal mechanisms of these different empirical algorithms, and the reliability of these algorithms can be evaluated and compared.

3)   Theory-driven design and optimization of AI systems, particularly neural network systems: The current architecture design and training optimization of AI systems are mostly empirical, which implies that people develop effective architectural design and optimization algorithms based on several experimental observations. However, unified theoretical feedback is required to guide the design and optimization of AI systems so that the representation capacity of AI systems can meet the demands of specific tasks. Therefore, the controllability of AI systems is truly realized.

Specifically, a few international research teams have realized and considered the abovementioned key issues and conducted forward-thinking explorations on them. For example, the team from the Shanghai Jiao Tong University uniformly explained numerous algorithms for improving adversarial transferability, whereas those from University of California, Berkeley proposed that the principles of “self- consistency” and “simplicity” are the cornerstones of AI systems and used these principles to guide AI systems’ design with interpretable representations and training.

In the past decade, much research progress has been achieved for trustworthy AI. However, from the perspective of the development process of the entire field, trustworthy AI remains in its infancy, and many key bottlenecks exist, which need to be urgently solved. Specifically, important research directions for the next 5–10 years include the following aspects (Figure 1.2.6).

First, further improve the explanation for knowledge representations. Most of current explanations stem from heuristics, which lack theoretical reliability. Additionally, these explanations have no ground-truth, making it impossible to verify the reliability of explanations. Therefore, future research priorities may include: ① unifying numerous existing empirical explanations and revealing their common mechanism; ② proposing new explanations with theoretical guaranties; ③ evaluating the reliability of explanations objectively.

Second, develop the explanation and quantification for representation capacity. Future research priorities may include: ① exploring and identifying the essential factors that affect the representation capacity of AI systems; ② uniformly explaining the internal mechanism of various AI algorithms proposed for improving representation capacities; ③ proposing precise quantitative metrics to evaluate the representation capacity of AI systems; ④ explaining and proving properties and defects of representation capacities of AI systems.

Third, conduct theory-driven design and optimization of AI systems. Today, effective architectural design and optimization methodologies are typically developed based on experimental observations. However, trustworthy AI requires a unified theory, which can provide targeted feedback to guide the design and optimization of AI systems. Future research priorities in this area could be ① investigating the connection between network architectures and knowledge representations, ② investigating the connection between the model performance and knowledge representations, and ③ investigating the relationship between knowledge representations and various types of representation capacity, such as generalization, robustness, and fairness.

1.2.3 Silicon-based CMOS terahertz imaging technique

Traditional terahertz (THz) imaging methods are mainly based on pure electronic or photoelectric setups. The former mainly depends on Schottky diodes and III–V devices, whereas the latter focuses on photoconductance, optical rectification, and quantum cascade lasers. In reality, all the devices are expensive and bulky, and some even call for cooling equipment. Additionally, they are incompatible with traditional microelectronics packaging methods, which further increase the difficulty of circuit integration. However, with the rapid development of silicon-based technology, the silicon-based CMOS THz imaging technique can achieve lower power consumption and a smaller size, which can meet the market demand for low cost and high integration density and has gradually become a hot research topic in the field of THz imaging.

Table 1.2.9 summarizes the research status on the silicon-based CMOS THz imaging technique. The top three countries in core papers are the USA, Germany, and China. In terms of citations, China is overtaken by Japan and France and drops to the fifth place. The main research institutions with the greatest output of core papers are shown in Table 1.2.10. The University of Wuppertal and the Vilnius University are the top two institutes with the greatest output of core papers. In China, only Nanjing University breaks into the top 10. Regarding total citations, the top three institutes are Princeton University, University of Wuppertal, and University of Michigan, while Nanjing University

《Figure 1.2.6》

Figure 1.2.6 Roadmap of the engineering research front of “theories and algorithms for trustworthy AI”

《Table 1.2.9》

Table 1.2.9 Countries with the greatest output of core papers on “silicon-based CMOS terahertz imaging technique”

of China ranks around tenth. Based on cooperation among major countries (Figure 1.2.7), the main partners of China is the USA. Germany has developed extensive cooperative ties with countries in Europe, America, and Asia. Based on cooperation among major institutions (Figure 1.2.8), stable cooperation has been established in continental Europe, including the General Jonas Žemaitis Military Academy of Lithuania, Vilnius University, and Institute of High-Pressure Physics of the Polish Academy of Sciences, while Cornell University built partnerships with University of California, Los Angeles and University of Michigan, as listed in the table. Institutions with the greatest output of citing papers on this front are shown in Table 1.2.11. China accounts for over a third of the publications and ranks first worldwide. The USA and Germany rank second and third, respectively. In Table 1.2.12, China dominates the list of institutions, with seven in the world’s top 10, whereas only two are in the USA, and one in Germany.

Three factors are primarily the focus of initial research on CMOS silicon-based THz imaging: high sensitivity, high integration density, and high resolution. Incoherent direct- detection technology was used initially, but its low sensitivity and high input power requirement are the main challenging factors for solid-state electronic design. Based on the 0.13 μm silicon germanide (SiGe) BiCMOS technology, a specific coherent imaging transceiver chip can upgrade the sensitivity by more than 10 times. The array’s scale must be expanded to achieve a higher imaging resolution, but this is impossible for traditional coherent detection arrays because

《Table 1.2.10》

Table 1.2.10 Institutions with the greatest output of core papers on “silicon-based CMOS terahertz imaging technique”

《Figure 1.2.7》

Figure 1.2.7 Collaboration network among major countries in the engineering research front of “silicon-based CMOS terahertz imaging technique”

《Figure 1.2.8》

Figure 1.2.8 Collaboration network among major institutions in the engineering research front of “silicon-based CMOS terahertz imaging technique”

《Table 1.2.11》

Table 1.2.11 Countries with the greatest output of citing papers on “silicon-based CMOS terahertz imaging technique”

《Table 1.2.12》

Table 1.2.12 Institutions with the greatest output of citing papers on “silicon-based CMOS terahertz imaging technique”

of the limitation of the centralized design methods for local oscillator signals. Instead, the 32-element phase-locked dense heterodyne receiver array, based on 65 nm CMOS technology, enables the integration of two interleft 4×4 arrays onto a chip of 1.2 mm2, significantly shrinking the whole receiver array. Based on lateral resolution improvement, the fully integrated ultra-wideband (UWB) inverse synthetic aperture imaging technology in terms of 55 nm BiCMOS technology can achieve a lateral resolution of 2 mm and a range resolution of 2.7 mm.

To date, many technological advances have been achieved to improve THz imaging resolution, but it is still restricted to a spot size in the range of millimeters, because of the diffraction limit in the silicon-based technique. Resolution in the micron range is necessary for biomedical or material characterization. It can be achieved using imaging from a far-field to a near-field method, with a lateral resolution of 10–12 μm.

The silicon-based CMOS THz imaging technique has grown rapidly in popularity over the past decade because of market demand for cheap cost and high integration density. Many investigations were conducted and prompted the development of THz imaging technology. With the ongoing advancement of technology, the development of the THz imaging technique is moving toward high integration density, high-accuracy, and a large array. However, it simultaneously encounters three unavoidable challenges:

1)  At higher frequencies up to the THz range, the effectiveness of active device models is reduced, and the loss of passive devices is increased, which restricts the rapid development of the THz circuit by silicon-based technology. Simultaneously, multilayer metal and medium characteristics in silicon-based technology introduce complex parasitic and coupling effects into the devices operating at THz frequencies.

2)  As the THz wavelength is so short, more circuit integration space is available for the systems. However, THz circuits are more susceptible to distribution effects and surface roughness. Therefore, the system must be integrated according to innovative packaging and interconnection technologies.

3)  When expanding from a single channel to an array chip, to achieve a high angle resolution, it is necessary to ensure the cooperation of multiple channels, which puts a higher requirement on the technology of source synchronization. Finally, more sophisticated calibration systems are needed to ensure the accuracy of signal detection and transmission.

According to the BCC Research forecast, the global market for the main terahertz technology may reach 3.5 billion dollars by 2029. However, the market shares from the silicon-based integrated circuit industry are not included here because of its relatively backward development and maturation. Figure

1.2.9 presents a roadmap for the evolution of the engineering research frontiers of the CMOS silicon-based THz imaging technology. The chip fabrication is expected to finish around 2029, and on-chip testing will be on schedule. By 2032, technology optimization and integration research will be performed with breakthroughs in chip size and resolution. In the next 10 years, the integration of the THz technique by CMOS silicon-based methods will push the THz imaging technology toward a much larger market size.

《2 Engineering development fronts》

2 Engineering development fronts

《2.1 Trends in Top 10 engineering development fronts》

2.1 Trends in Top 10 engineering development fronts

The Top 10 engineering development fronts in the information

《Figure 1.2.9》

Figure 1.2.9 Roadmap of the engineering research front of “silicon-based CMOS terahertz imaging technique”

and electronic engineering field are summarized in Table 2.1.1, encompassing the subfields of electronic science and technology, optical engineering and technology, instrument science and technology, information and communication engineering, computer science and technology, and control science. The annual number of core patents published for the Top 10 engineering development fronts from 2016 to 2021 is shown in Table 2.1.2.

(1)  Super-large-scale digital twin visualization and simulation system

A technical “digital twin” enables dynamic interactions between physical and virtual entities. It allows real-time connection, two-way mapping, simulation analyses,

《Table 2.1.1》

Table 2.1.1 Top 10 engineering development fronts in information and electronic engineering

《Table 2.1.2》

Table 2.1.2 Annual number of core patents published for the Top 10 engineering development fronts in information and electronic engineering