《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. “Optoelectronic in-sensor computing devices and their integration” is the front of data mining. “Networking theories and key technologies of satellite internet”, “ultra-large-scale silicon-based quantum chips”, “automatic development of software assisted by artificial intelligence”, “systematized gaming and intelligent control for multiagent systems”, “cyber-physical security of industrial control systems”, and “chip-based satellite laser communication terminal” are fronts of expert nomination. The remaining fronts are data mining & expert nomination. Table 1.1.2 shows the number of core papers published from 2017 to 2022 related to each research front.

(1) Theory and technology of large models and their computing systems

Theory and technology of large models and their computing systems refer to the semi-supervised/unsupervised basic learning theory of large-scale pretrained models and efficient computing technology composed of parameter fine-tuning, reinforcement learning and other mechanisms, as well as model parallel computing and distributed systems based on this and their optimization strategies and deployment scheme. Large models are pretrained on large-scale data based on learning mechanisms, such as self- supervision, and have powerful representation and generalization capabilities, usually with a large number of parameters. Large models eliminate the dependence on large amounts of labeled data and can serve downstream applications through model fine- tuning, prompt fine-tuning, and context learning for specific tasks. They possess general intelligence capabilities for tasks in multiple application scenarios. Research on the theory and technology of large models and their computing systems is currently maintaining high-speed iterations, and has quickly penetrated applications such as natural language, smart medical care, multimodal generation, and autonomous driving. The theory and technology of large models and their computing systems provide an impetus and theoretical foundation for improving the performance, efficiency, and generalization capabilities of large models. The main research directions include the theoretical framework of large models, large model structures, and training mechanisms,

《Table 1.1.1》

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

《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

and distributed training and deployment strategies. The theoretical framework of large models is based on information and over- parameterization theories, and it studies the explanation theory of new characteristics such as emergence and homogeneity while modeling the computational complexity of large models. Research on the structure and training mechanism optimizes the computing system of the unified framework of “pretraining + generalization” for large models, designs more efficient self- supervision mechanisms, and balances the fitting ability and complexity for large models. Distributed training and deployment strategies study the scalability improvement scheme of large models, use methods such as serverless computing to distribute training tasks to multiple computing nodes in the cloud, and use parallel training strategies to overcome the limitations of storage and computing resources. The future development of the theory and technology of large models and their computing systems still needs to address three issues: privacy and security, evaluation methods, and deployment efficiency. First, the training and deployment of large models requires more efficient encryption and secure communication technologies. The development of secure and trustworthy computing solutions is also urgent. Second, automated and general evaluation frameworks are required to benchmark the capability bounds, robustness, and deviation correction capabilities of large models. Additionally, the real-time performance and training energy consumption of large models must be further optimized.

(2) Networking theories and key technologies of satellite internet

Satellite internet is the third internet revolution, after fixed and mobile communication networks. Specifically, satellite internet is a wireless network that is exploited as an access network. On one hand, the satellite internet has overcome the coverage range of the ground internet and realizes seamless global coverage. Even though continents account for only 29% of the global area, only partial urban and rural districts are covered by the ground internet, and efficient wireless coverage is unavailable in remote areas, such as broad oceans, deserts, forests, and mountains. However, the construction of new ground networks in these areas incurs significant costs and maintenance difficulties; therefore, a satellite internet that employs the satellite network as the access network is required. The decrease in the costs of satellite manufacturing, launch, and communication makes it possible to construct the satellite internet of things (IoT) and achieve global coverage through satellite networks. On the other hand, satellite internet has reformed the original network structure, shifting from a people-centered connection to the interconnection of all things. In recent years, with economic and societal development, several wireless devices have been accessed via wireless networks. Satellite internet offers a large access capacity, strong survivability, and low impacts on weather conditions, which can provide interconnection of things in the global range.

Based on the different heights of satellite orbits, satellite networks are usually divided into three categories: low, medium, and high orbit. Low-orbit satellite networks are generally used as access networks for satellite internet because their low-orbit characteristics contribute to low transmission delays and losses. For example, the “Starlink” project of the American company SpaceX employs satellite networks with orbit heights of approximately 500 km. However, low-orbit satellites move quickly and the visible duration of one satellite is short. Therefore, to realize seamless global coverage, several low-orbit satellites must be deployed for networking, i.e., low-orbit satellite constellations. In recent years, both academia and industry have conducted extensive research on networking theories and key technologies for satellite internet, and significant progress has been made, which mainly includes inter-satellite communication technology, satellite network protocols, elastic routing protocols, and mobility management. The introduction of mobile satellite communication without ground stations has further improved the networking forms of the satellite internet and significantly enhanced its functionality. In conclusion, the advantages of satellite internet can be summarized as follows: First, the satellite internet has eliminated dependence on ground devices, greatly enhanced mobility, and truly achieved mobile access in a global range. Second, existing mobile phones can be directly connected to the satellite internet, where downward compatibility can be achieved. In addition, the collaboration of satellite internet and ground networks has enabled the integration of space and earth communication.

(3) Ultra-large-scale silicon-based quantum chips

Quantum computers are expected to surpass classical computers in bringing higher computational power to humanity with quantum chips responsible for computation and information processing at their core. Quantum chips, which tightly integrate quantum circuits onto a substrate, play a vital role in the development of quantum technologies. Unlike the binary chips in classical chips, quantum chips use quantum properties to significantly increase computational parallelism and the ability to handle complex problems. The fabrication of ultra-large-scale quantum chips has become a critical challenge in realizing universal quantum computers because millions of quantum bits (qubits) are required for error correction. Considering that classical computing chips can currently accommodate billions of transistors and that the integrated circuit (IC) industry possesses mature manufacturing techniques and infrastructure, the realization of large-scale silicon-based quantum chips through complementary metal–oxide semiconductor (CMOS) processes holds a natural advantage in terms of scalability and is gradually becoming an international research hotspot in the field of quantum computing.

The development of silicon-based quantum chips involves diverse approaches. One approach involves encoding quantum information onto electron (hole) spins within gate-defined silicon-based quantum dots or phosphorus nuclear spins embedded in silicon. Presently, single-qubit and two-qubit gates have been realized with fidelity that surpasses error-correction thresholds, and the successful construction of quantum processors containing six qubits has been achieved. In the future, the development of silicon-based spin qubits will specifically address the challenges of long-range coupling and uniform high-fidelity in large- scale setups. Additionally, leveraging silicon-based photonic integration processes can facilitate optical quantum computing and communication. Substantial breakthroughs have already been achieved in areas such as high-dimensional quantum entanglement states, quantum key distribution, and quantum teleportation. The evolution of silicon photonics technology requires the compact integration of photon sources, quantum state manipulation, and single-photon detection onto a single chip while minimizing device losses. These advancements have laid the foundation for the quantum computation and communication on silicon-based materials. Consequently, the trajectory of silicon-based quantum chips is set to continue its progression toward large-scale integration and practical applications.

(4) Photon-integrated lasers for quantum applications

Photon-integrated lasers for quantum applications are optical source devices based on technology of planar light waveguide circuit photonic integration and used in fields of quantum optical detection, sensing, measurement, communication and so on. In recent years, in the field of quantum optics, increasing attention has been paid to research on new methods and technologies based on the interaction of lasers and atoms. Specifically, lasers of 509 nm, 633 nm, 780 nm, 795 nm, 852 nm, 976 nm, 1 064 nm, 1 083 nm, 1 310 nm, and 1 550 nm used in optical atomic clocks, Rydberg detection, quantum magnetic probes, gyroscopes, and quantum communication have become mainstream. The traditional quantum light source devices used to generate laser wavelengths are gas, solid, and fiber lasers. Because of their large volume, complex operation, low energy consumption ratio, and lack of reliability, it is difficult to meet the current needs of communication, sensing, and detection in space, air, Earth, sea, and other complex environments. Semiconductor lasers with small size and high efficiency are suitable for these aforementioned applications. Owing to the advantages of direct photoelectric conversion and compatibility with semiconductor processes, they can be used to realize photonic integrated quantum light source devices.

However, because of the size limitations of the resonator and waveguide, semiconductor lasers have high phase noise and poor beam quality, which makes it difficult to meet the requirements of quantum optics for laser spectral purity, wavelength accuracy, and frequency stability. Therefore, it is necessary to use an external optical frequency selection element to induce an internal laser resonator mode through appropriate optical feedback to achieve laser linewidth narrowing, noise suppression, and spectral purification. For a multispectral quantum light source extending from visible light to near-infrared, it is important to solve not only the epitaxial growth of semiconductor lasers with different materials, grating preparation, waveguide etching, cavity surface coating, and other problems but also the structural design of an external optical frequency selective chip and the mode loss control of optical feedback.

In addition, it is crucial to design a special current/temperature drive control circuit for semiconductor lasers with different materials to ensure wavelength accuracy and frequency stability. In quantum optics, the laser frequency standard is the most commonly used light source. To achieve higher frequency stability, it is feasible to use a high-precision drive control circuit and an atomic gas chamber to build a feedback frequency stabilization system that locks the output frequency to the energy level of the atomic or molecular transition. However, the volume of the gas chamber used to provide the transition atoms is large and incompatible with the semiconductor process. Therefore, the realization of an integrated feedback frequency stabilization system for semiconductor lasers is a key problem that must be solved.

(5) Extra-large-scale and ultra-wideband antenna array communication theory and technologies

Extra-large-scale and ultra-wideband antenna array communication is a technology that simultaneously uses an extra-large- scale antenna array and ultra-wideband technology. It enhances the channel capacity and subsequently increases information transmission rates by increasing the number of antennas and expanding the bandwidth. Centimeter waves, millimeter waves, and terahertz frequency bands can provide hundreds of megahertz-level or even gigahertz-level ultra-wide bandwidths. Because of the shorter wavelengths of the carriers within these frequency bands, the antenna sizes are smaller, enabling a significant increase in the number of base station antennas, thus forming an extra-large-scale array. Extra-large-scale antenna array and ultra-wideband technology complement each other, and both are effective means to meet the information transmission rate requirements of 6G.

The spectrum is an important resource in mobile communication systems. With the advancements in communication technology, the bandwidth and frequency bands of communication systems have gradually increased. Research on millimeter-wave frequency bands for 5G has led to the development of mature communication-system models and transmission schemes. Early research on 6G focuses on the terahertz frequency bands above 100 GHz to provide an ultra-wide bandwidth. Because of characteristics such as high path loss and nonstationary channel space, most studies have focused on reducing the complexity and improving the accuracy of terahertz channel modeling. However, as the frequency increases, the path loss becomes more severe, limiting the coverage performance. In 2022, the 3GPP RAN#96 Conference officially defined the 6 425–7 125 MHz band as the U6G licensed spectrum and approved the Release 18 project for the full 6 GHz spectrum (5 925–7 125 MHz). In 2023, the Ministry of Industry and Information Technology released a new version of the “Radio Frequency Allocation Regulations of the People’s Republic of China,” which allocates the U6G frequency band to IMT (including 5G/6G) systems. This frequency band has relatively low path loss, strong electromagnetic wave diffraction, penetration capabilities, and excellent wireless coverage performance. Because of the limited attention paid to U6G, both academic and industrial communities have lacked substantial research on it. The channel model is unclear and is the focus of attention at this stage.

The channel characteristics of extra-large-scale antenna array systems have not been fully explored in different frequency bands, making channel measurement and modeling one of the primary research directions for extra-large-scale antenna array systems. Extra-high-dimensional channels exhibit spatial nonstationarities. Specifically, when wireless signals emitted by users reach the extra-large-scale antenna array, they form spherical wavefronts. In addition, the channel energy is concentrated only in a portion of the subarrays with significantly reduced dimensions. Most research on tasks, such as channel estimation, precoding techniques, and transceiver design, is based on this property. Furthermore, because of the large number of antennas, radio frequency, power consumption, and complexity in extra-large-scale antenna array systems cannot be ignored. Future research should focus on exploring low-cost system architectures and low-complexity transmission solutions.

(6) Optoelectronic in-sensor computing devices and their integration

Optoelectronic in-sensor computing devices and their integration refers to the integration of sensing, memory, and computing functions into optoelectronic devices and further large-scale integration. The optoelectronic in-sensor computing system addressed the speed and power limitations caused by the traditional von Neumann architecture and integrated optoelectronic sensing functions, thereby facilitating the development of more intelligent and energy-efficient computing systems. As a new intelligent device, the all-in-one optoelectronic fusion memory device can mimic the working mode of the human retina and brain. The device has highly adjustable conductivity and optical responsivity parameters because of the introduction of optoelectronic materials and the use of photons to control the transport characteristics of the carriers and ions. Image data processing speed and energy efficiency can be highly improved by integrating functions such as optical perception, information storage, and logical computing.

Traditional machine vision systems have high energy consumption and latency owing to the repeated movement of data between sensing, memory, and processing units, making it difficult to meet the real-time processing requirements for massive amounts of visual information. With the background of cloud computing, AI, and IoT, optoelectronic in-sensor computing technology has ushered in a huge development opportunity. It has a wide range of applications in automatic driving, wearable electronics, smart homes, and other areas, allowing for more efficient machine vision and brain-like computing.

Although the application prospects of optoelectronic in-sensor computing technologies are broad, there are many challenges in terms of performance, accuracy, and efficiency. For example, developing new functional composite materials to a achieve wide spectral response and constructing high-quantum-efficiency sensing and memory devices, exploring wafer-level processing techniques to achieve high-density integration of devices, and developing an intelligent optoelectronic in-sensor computing system to complete high-level information tasks.

(7) Automatic development of software assisted by artificial intelligence

The automatic development of software aided by artificial intelligence (AI) is a cutting-edge research field that uses AI techniques to assist, accelerate, and optimize the software development process. Its core objective is to reduce the workload of developers and improve the efficiency and quality of software development using intelligent methods. In recent years, the main research directions include the following:

1) Automated requirement analysis: This method uses machine learning and natural language processing techniques to automatically transform and analyze natural language requirements provided by users into requirement models that computers can directly understand and analyze. This can help developers more accurately understand and capture user requirements and reduce errors in understanding requirements.

2) Automated design and coding: This method uses machine learning and natural language processing techniques to automatically generate designs, or code fragments, functions, and even entire modules. This can help reduce the workload of manual design or coding and accelerate the development process.

3) Automated testing: This method uses AI techniques to automatically generate test cases, defect detection capabilities, and improve software testing coverage, thereby improving the software quality.

4) Automated integration and deployment: This method automatically integrates manually written code and automatically generated code by developers, and deploys it into the production environment, thereby improving software delivery efficiency and stability.

5) Intelligent recommendation system: This method recommends code, tools, and techniques that are suitable for the current development context based on the development-history data and project requirements of developers, thereby improving their development efficiency.

Generally, there are several development trends:

1) Intelligent coding: Code generation will become more intelligent and in line with developer intentions, reducing subsequent adjustments and modifications. It is also possible to automatically select the most suitable development strategy and tools based on the characteristics and needs of different projects, thereby providing more flexible and efficient intelligent development services.

2) Adaptive operation and maintenance: The ability to continuously optimize and improve the software operation process and quality of operation based on user feedback and changes in operation and maintenance data.

3) Collaborative development: Natural language processing and intelligent dialog techniques enable real-time interaction and communication with more developers, testers, and domain experts, promote cross-domain cooperation, provide a more friendly and efficient development experience, and create more comprehensive and optimized solutions.

4) Concealment of ethical and security issues: With the improvement of development automation, developers must pay more attention to the underlying ethical and security issues, ensuring that the automatically generated code and decisions are reliable and secure.

(8) Systematized gaming and intelligent control for multiagent systems

Systematized gaming and intelligent control for multiagent systems refers to the process by which intelligent agents adjust their behavior and optimize system parameters using game theory, interactive strategies, and intelligent control methods. This technology balances individual and group interests within the system. This topic faces challenges of complex system structures, uncertain game environments, incomplete decision-making information, and uninterpretable results. In this regard, current research hotspots focus on the following. ① Multiagent institutionalized and systematized game theory model. This study explores the game evolution law of multiagent systems by using and combining several AI learning algorithms. ② Modeling of multilevel, multiscale, multimode, nonlinear, and uncertain time-varying dynamic systems and analysis, simulation, prediction, optimization, and control of multiagent systems. ③ Multiagent autonomous navigation and swarm cooperation. This study addresses the issues of autonomy, intelligence, and scalability of multiagent systems, considering uncertain environments, incomplete decision information, and limited communication.  ④  Decision-making processes in multiagent systems. This includes cooperative negotiation, resource allocation, and task assignment. Achieve effective collaboration and decision making among agents. ⑤ Robustness analysis framework for multiagent algorithms and models. Reduce the complexity of cooperative decision- making algorithms and address model biases between data-driven methods and real scenarios.

Overall, several problems require further investigation in the future. ① Enhance the interpretability and controllability of multiagent systems. Make system behavior and decision making more understandable and adjustable and improve system reliability and security. ② Integrate game theory, multiagent learning, and control theory. Promote game theory and intelligent control methods and improve overall system performance and intelligence. ③ Application of interdisciplinary research on game theory, learning, and control in emerging fields, such as intelligent transportation and logistics management.

(9) Cyber-physical security of industrial control systems

An industrial control system (ICS) is composed of various automated acquisition, monitoring, and control components for the automated operation and supervision of industrial infrastructure. ICSs include supervisory control and data acquisition (SCADA) systems, distributed control systems (DCS), and programmable logic controller (PLC) systems. Currently, ICSs are the nerve and operation center of national key infrastructure such as industrial production, smart grid, and smart transportation. The ICS has become a top target of adversaries because of their importance and openness. With the continuous improvement of attackers’ vulnerability discovery capabilities and attack techniques, the cyber-physical security problem of ICS is becoming increasingly severe. It mainly refers to the security risks caused by attackers leveraging the characteristics of the tight integration of cyber space and physical space of ICS to launch coordinated attacks in the cyber and physical domains, leading to unobservable and uncontrollable dilemmas. For example, attackers can bypass defense methods to enhance cyber security, such as isolation and intrusion detection, and breakthrough physical security protection mechanisms, such as device redundancy. With these capabilities, attacks can penetrate across the ICS monitoring and control layers to destroy the physical process of the ICS. Such cyber-physical threats can use the characteristics of the physical processes of ICS to design malicious data tampering mechanisms that operate with strong collaboration and high concealment. This poses a significant challenge to ICS security.

Presently, studies related to the cyber-physical security of ICS have mainly focused on the following three aspects. ① Attacker capability modeling/system vulnerability analysis. These studies designed attack strategies from the attacker’s perspective based on the available knowledge of the ICS, such as its architecture, protocols, and control algorithms, thereby analyzing the vulnerabilities of ICSs. ② Attack detection. This method can detect malicious attacks by passively collecting dynamic data from the system or actively adding dynamic authentication information and then building a normal mode of the system. ③ Attack defense.

This method includes defense strategies such as attack isolation, resilient control, and moving target defense. In the future, researchers will conduct vulnerability analysis and defense strategy design while considering the characteristics of large-scale, multilevel, and strong coupling of ICSs to enhance their cyber-physical security of ICSs.

(10) Chip-based satellite laser communication terminal

Satellite laser communication has the advantages of large bandwidth, high speed, concentrated emission energy, and strong anti-interference and anti-interception capabilities. It is currently the dominant position in the field of space network technology in which countries globally are vying for. The chip-based satellite laser communication terminal is a microdevice that realizes the function of satellite laser communication through optoelectronic integration technology. It integrates or partially integrates functional components, such as optoelectronic communication devices and electronic control units, into a single chip, resulting in a small size, lightweight, and low-power consumption. It can satisfy the diverse requirements of high-speed communication, data transmission, and satellite networking. In recent years, researchers have mainly focused on the following aspects. ① Heterogeneous optoelectronic integration technology. To improve chip performance by optimizing integrated chip design and manufacturing processes on silicon-on-insulator (SOI), silicon-based thin-film lithium niobate (LNOI), silicon nitride (Si3N4), indium phosphide (InP), and other material platforms. ② High-speed laser communication technology to solve key issues such as modulation, demodulation, codec, and signal processing of laser communication to improve the communication rate and transmission distance. ③ Optimize the reliability of the laser terminal to improve the signal anti-interference ability and environmental adaptability, such as temperature and radiation, and improve the stability and life. In the future, chip-based satellite laser communication terminals will be further integrated to achieve higher transmission rates, longer communication distances, higher reliability, and smaller and lighter terminals with lower power consumption. With the aim of multifunctional and networked communication, the compatibility and interoperability of laser terminals will be improved to promote the application and further development of chip-based satellite laser communication terminals in areas such as navigation, relay communication, Earth observation, deep-space exploration, and low-orbit internet constellations.

《1.2 Interpretations for three key engineering research fronts》

1.2 Interpretations for three key engineering research fronts

1.2.1 Theory and technology of large models and their computing systems

Large models and their computing systems learn the feature representation of the data from large-scale unlabeled data and pretrain the model with numerous parameters. It has the advantages of strong generalization, wide application scenarios, and low dependence on labeled data. However, because of the poor explain ability of the current large model, high dependence on training data, and high training and deployment costs, the theory and technology of large models and their computing systems urgently require breakthroughs in three aspects: theoretical framework, model structure, and training and deployment strategies.

First, there are two main trends in the theoretical framework of large models. One is to integrate information and over- parameterization theories in the research of large models, and explore the theoretical limit of the representation ability of large models such as generative pretrain transformers (GPT). Second, to introduce concepts involving graph theory into the theoretical analysis of large model training processes, such as using hypergraphs to explain and improve the stability of large model unsupervised bin-wise pretraining. The main research institutions in this direction include Harvard University, Tsinghua University, Zhejiang University, Hikvision Digital Technology Co., Ltd., and Warsaw University.

Second, an important research trend in large model structure design is the cooperation of large model structures with transfer learning strategies to solve the degradation and bias of large models caused by unfiltered data, thereby improving their predictive ability of large models. The other is to introduce the concept of prompt programming into the model design to reduce the overfitting phenomenon of large models in the bottom and middle layers, and to guide the model to better understand and perform specific tasks through well-designed prompts. The main research institutions in this direction include the Facebook AI Research Institute, Seoul National University, and the National Kaohsiung University of Applied Sciences.

In addition, in terms of training and deployment strategies for large models, the use of new hardware such as McDRAM to accelerate the real-time performance of large model inference and the energy consumption ratio of the deployment process is a potential technology for deploying large models on edge devices. In addition, using approximate computing technology to accelerate the inference process or designing customized binarization strategies and quantization methods for large models may effectively improve their inference speed and save computing resources in the future. Major research institutions in this area include the University of Bremen, the National University of Singapore, Huawei Technologies Co., Ltd., and the University of Sydney.

Table 1.2.1 shows the distribution of the main output countries of core papers in the engineering research front of “theory and technology of large models and their computing systems”. The USA has the highest number of core papers globally, accounting for approximately one-third of all papers. China is second only to the USA, but the average publication year of papers is newer, indicating a state of rapid catching up. China’s international cooperation partners are mainly the USA and Australia (Figure 1.2.1). Four of the Top 10 output institutions (Table 1.2.2) are from China, with the rest located in the USA, Australia, Singapore, and other countries. In terms of institutional cooperation (Figure 1.2.2), the three Chinese institutions have relatively close cooperation with the University of Sydney, while the three domestic institutions have relatively close cooperation. In terms of the number of citing papers (Table 1.2.3), China ranked first (accounting for 44.56%), followed by the USA, and the remaining countries accounted for less than 10%. Except for Harvard University, which ranks seventh, all the Top 10 institutions that produce citing papers are from China (Table 1.2.4), reflecting the high attention of Chinese scientific research institutions to the theory and technology of large models and their computing systems.

Over the past five years, many research results in the theory and technology of large models and their computing systems have been achieved. However, in terms of the overall development process of the research field, their application and research are still in their infancy, and many key bottlenecks must be resolved urgently. Figure 1.2.3 shows the key development directions for the next 5–10 years.

First, model compression and distributed training. Currently, large models represented by ChatGPT contain more than 10 billion parameters, necessitating huge computing and storage resources. Future development directions include more efficient model compression and acceleration technologies to reduce the parameter scale and computational cost of the model, improve its deployment efficiency on edge devices, such as laptops and mobile phones, and realize real-time inference and decision making in various computing scenarios. Conversely, as the scale of models grows, distributed training and collaborative learning of large models will be critical, and new distributed training strategies and techniques will help speed up model training while maintaining model performance.

Second, automated and smarter model design. Currently, large enterprises such as Google and Microsoft rely on their powerful computing resources to design multi-input multiple-output large-scale model structures, such as T5 and Kosmos. However,

《Table 1.2.1》

Table 1.2.1 Countries with the greatest output of core papers on “theory and technology of large models and their computing systems”

《Figure 1.2.1》

Figure 1.2.1 Collaboration network among major countries in the engineering research front of “theory and technology of large models and their computing systems”

《Table 1.2.2》

Table 1.2.2 Institutions with the greatest output of core papers on “theory and technology of large models and their computing systems”

《Figure 1.2.2》

Figure 1.2.2 Collaboration network among major institutions in the engineering research front of “theory and technology of large models and their computing systems”

《Table 1.2.3》

Table 1.2.3 Countries with the greatest output of citing papers on “theory and technology of large models and their computing systems”

《Table 1.2.4》

Table 1.2.4 Institutions with the greatest output of citing papers on “theory and technology of large models and their computing systems”

《Figure 1.2.3》

Figure 1.2.3 Roadmap of the engineering research front of “theory and technology of large models and their computing systems”

methods such as reinforcement learning and evolutionary algorithms can be used to automatically design efficient model structures suitable for specific tasks. Therefore, automated neural architecture search and model optimization technologies will be developed in the field of large models in the future. In addition to Huawei’s Pangu, Baidu’s ERNIE Bot, Zhiyuan’s WuDao, and other general-purpose pretraining models, the future will inevitably require more pretraining models for specific fields such as medical care and finance.

Third, privacy and security. With the increasing number of applications based on large models, such as ChatGPT and Dell-E, the privacy of large models has gradually attracted more attention. Future research directions include maintaining high performance of the model while protecting user data privacy. In terms of security, the robustness and fairness of large models will be the focus of future research, including how to make the model perform well on diverse data while avoiding bias against different groups.

Fourth, modeling and reasoning of cross-modal knowledge. Existing technology has enabled the correlation and generation of multimodal data such as images, texts, and audio. However, when faced with more complex and abstract reasoning scenarios, such as protein structure, real-time automatic driving, and multiparty games, existing large models, such as BAI-Chem and Pangu, can only answer questions based on inductive associations in a large amount of training data. To make large models closer to human intelligence, future research will focus on the following: ① quantitative modeling and expression of knowledge in large cross-modal models; ② exploring the relationship between reasoning decisions and knowledge representation in large models; ③ reasonability and interpretability of model multimodal knowledge.

1.2.2 Networking theories and key technologies of satellite internet

Satellite internet is a new revolution in the internet field with global coverage, on-demand access, on-demand service, secure communication, and reliable communication. It provides strong support for the prospect of global interconnection; therefore, it has attracted the interest of many countries. In the 1990s, Motorola Company in the US established the Iridium Satellite System, which consists of 66 low-orbit satellites with inter-satellite links and onboard processing capabilities. Meanwhile, Laura and Qualcomm Companies in the US have built a global satellite system that includes 48 low-orbit satellites, each of which employs transparent forwarding. These two satellite internet systems have undergone bankruptcy restructuring for commercial reasons. After a period of downtime, the satellite internet has recently set off its second craze on an even larger scale. In 2017, the British company OneWeb proposed the OneWeb project, which plans to launch 1 980 satellites to constitute a low-orbit constellation with global coverage. Furthermore, the American company SpaceX plans to launch 42 000 low-orbit satellites to form star links capable of supporting high-speed mobile communication globally. In recent years, China has accelerated the construction of the satellite internet. For example, the “Hongyan Constellation” designed by the China Aerospace Science and Technology Corporation includes 324 satellites, whereas the “Hongyun Constellation” designed by the China Aerospace Science and Industry Corporation includes 156 satellites.

Although satellite internet has grown rapidly, many challenging issues still exist, including network architectures, routing protocols, inter-satellite communication, and mobility management, which are presented as follows.

(1) In terms of network architecture, there are currently two main architectures. The first type is the nonterrestrial network (NTN), which is led by 3GPP, the international organization for the standardization of mobile communication. This open architecture is compatible with existing terrestrial cellular networks and is an integral part of the entire 6G network. The other architecture was specifically designed for the Starlink project by SpaceX Company, which has a closure property. The main research institutions in this area include the Harbin Institute of Technology, the Chinese Academy of Sciences, and Waseda University, Huawei.

(2) In terms of routing protocols, new elastic routing protocols must be developed because satellite networks are highly dynamic, satellite positions constantly vary, and ground network routing protocols are inapplicable. The basic idea is to use the regularity of the constellation motion to map real satellite nodes to virtual nodes. When satellites move or the ground terminal switches, the routing table between virtual nodes is exchanged between physical nodes; thus, the routing information exchange can be completed. The main research institutions in this direction include the University of Surrey, SpaceX, Beijing University of Posts and Telecommunications, Xidian University, the National University of Defense Technology, and Beijing Institute of Technology.

(3) In terms of inter-satellite communication, there are mainly two development trends. The first is inter-satellite microwave communication, which is highly technologically mature and currently has broad applications. However, microwave communication has difficulty meeting the requirements of high-speed communications and has limitations such as limited frequency band capacity and severe co-frequency interference. The second is inter-satellite laser communication, which is widely used in the new- generation satellite internet. Compared with microwave communication, laser communication has several advantages, including large bandwidth, low communication payload, strong anti-interference ability, and good confidentiality. The main research institutions in this direction include the University of Surrey, SpaceX, Chinese Academy of Sciences, Beijing University of Posts and Telecommunications, University of Electronic Science and Technology, and Zhejiang University.

(4) In terms of mobility management, one option is to use centralized mobility management, in which local agents manage terminals. Each time a terminal initiates a location update, the messages are transmitted to local agents. Another option is to use distributed mobility management, in which the Earth is divided into multiple zones, where terminals can register to a virtual gateway composed of satellite clusters covering that zone. Therefore, large-scale satellite network mobility management can be achieved. The main research institutions in this direction include Peng Cheng Laboratory, Southeast University, University of Luxembourg, Huawei, and ZTE Corporation.

Table 1.2.5 shows the main countries that output core papers in this cutting-edge field. China has an obvious advantage, ranking first in the world in terms of the number of core papers with over 67% proportion, and cooperates with Japan, the UK, Saudi Arabia, Republic of Korea, Canada, Australia, and Norway (Figure 1.2.4). The Chinese Academy of Sciences, and Harbin Institute of Technology jointly rank first among the Top 10 institutions that output core papers (Table 1.2.6). Seven institutions are from China, and the rest are from Saudi Arabia, Japan, and Luxembourg. In terms of institutional cooperation (Figure 1.2.5), King Abdulaziz University, Waseda University, and Xi’an University of Posts and Telecommunications have close cooperation. In terms of the number of cited core papers (Table 1.2.7), China still ranks first, with more than 50%, while other countries account for less than 10%. The Top 10 institutions that output cited core papers (Table 1.2.8) were mostly from China, except Waseda University, which ranked sixth, reflecting the strong research strength of China in this field.

Currently, satellite internet is highly valued by major countries and is developing at an unprecedented speed. The development roadmap for this frontier is shown in Figure 1.2.6. The following paragraphs outline its development trends in four areas: network architecture, routing protocols, inter-satellite communication, and mobility management. In terms of network architecture, the NTN architecture is currently in the discussion and standard-setting stage and is expected to be fully developed and commercialized by 2027. In terms of routing protocols, the currently used fixed routing protocols have inherent flaws that seriously constrain the scale and performance of networks, whereas elastic network protocols are gradually improving and are expected to mature by approximately 2026. In terms of inter-satellite communication, microwave communication currently used has low

《Table 1.2.5》

Table 1.2.5 Countries with the greatest output of core papers on “networking theories and key technologies of satellite internet”

《Figure 1.2.4》

Figure 1.2.4 Collaboration network among major countries in the engineering research front of “networking theories and key technologies of satellite internet”

《Table 1.2.6》

Table 1.2.6 Institutions with the greatest output of core papers on “networking theories and key technologies of satellite internet”

《Figure 1.2.5》

Figure 1.2.5 Collaboration network among major institutions in the engineering research front of “networking theories and key technologies of satellite internet”

《Table 1.2.7》

Table 1.2.7 Countries with the greatest output of citing papers on “networking theories and key technologies of satellite internet”

《Table 1.2.8》

Table 1.2.8 Institutions with the greatest output of citing papers on “networking theories and key technologies of satellite internet”

《Figure 1.2.6》

Figure 1.2.6 Roadmap of the engineering research front of “networking theories and key technologies of satellite internet”

communication rates and strong interference and is gradually being replaced by inter-satellite laser communication, which is expected to be completed by 2026. In terms of mobility management, the existing centralized management is transitioning to distributed management because of its high signaling overhead and transmission latency. The development trend of satellite internet is summarized as follows. First, the satellite internet has shifted from focusing on communication to integrating communication, navigation, and remote sensing, allowing for a multifunctional satellite internet. Second, satellite internet has developed from ground-devices-assisted access to mobile satellite communication without a ground station, achieving the goal of on-demand access. Third, has developed from satellite internet to satellite IoT to achieve the vision of the interconnection of all things. In terms of the application scenarios of satellite internet, has been used in emergency communication, remote area communication, and logistic communication. With the rapid development of the satellite internet, it will be widely used in various fields such as the economy, society, and military in the future.

1.2.3 Ultra-large-scale silicon-based quantum chips

A quantum chip is a device designed and manufactured based on the principles of quantum mechanics for applications in quantum computing and communications. Unlike classical chips, which use binary bits, quantum chips use quantum bits (qubits) as their fundamental information units. Qubits can exist in superpositions 0 and 1, and they can share information among multiple qubits through entanglement. This characteristic endows quantum chips with significant advantages in addressing specific problems, allowing them to solve certain computationally challenging tasks, such as large-integer factorization, quantum simulation, and optimization, at faster speeds than classical computers. Fabricating and manipulating quantum chips are a challenging because of the fragility of quantum bits, which are susceptible to destruction by uncontrollable environmental influences. A universal quantum computer often requires a substantial number of qubits to implement a quantum error correction and achieve the desired quantum advantage. Nevertheless, in terms of the number of qubits alone, various physical implementations, including superconducting circuits, diamond nitrogen-vacancy centers, and trapped ion technologies, are yet to meet these demands. Consequently, the production of large-scale quantum chips containing millions of qubits has emerged as a critical challenge for the realization of universal quantum computing. Given the successful manufacture of chips containing billions of transistors within the traditional semiconductor industry, the integration of silicon-based electronics with quantum technology holds the promise of establishing a quantum computing platform for the existing semiconductor manufacturing infrastructure. The compatibility of silicon-based quantum chips with traditional semiconductor processes provides advantages in terms of production cost, scalability, and integration. This has propelled large-scale silicon-based quantum chips into a focal point of research in the field of quantum computing, offering a promising pathway for the feasibility and scalability of quantum computing.

Silicon-based quantum chips present various development routes. One approach involves encoding quantum information into electron (hole) spins confined in gate-defined silicon quantum dots or implanted phosphorus nuclear spins embedded in silicon. In 1998, theoretical physicists predicted that spin states within silicon nanostructures could serve as carriers of quantum information, marking the inception of an experimental research race. Single-qubit operations can be achieved by manipulating and measuring the individual electron spins. Leveraging the exchange interaction between the two spins enables the implementation of two-qubit gate operations. Initial experiments used III–V semiconductor materials; however, unavoidable hyperfine interactions hindered their further advancement. In 2013, several research groups simultaneously reported breakthroughs in silicon-based spin qubits. This was facilitated by the isotopic purification of silicon to suppress hyperfine interactions, resulting in a significant enhancement in fidelity. Currently, single-qubit and two-qubit gates with fidelity far exceeding the error-correction thresholds have been realized. Furthermore, because of the nanoscale physical dimensions of silicon quantum dots and dopant atoms and their compatibility with modern IC technology, they can be scaled up to large-scale qubit arrays with reasonable chip footprints. Quantum processors comprising six qubits have already been demonstrated as larger quantum dot platforms in one-dimensional and two-dimensional configurations.

Another approach is to use silicon-based photonic integration processes to achieve quantum entanglement between photons and quantum-state manipulation, thereby enabling quantum computing and communications. Photonic quantum technology stands out because of its advantages, such as longer decoherence time, multiple degrees of freedom, no need for vacuum, and low temperature. In contrast to the bulky, unstable, and poorly scalable traditional optical instruments, silicon-based photonic chips fabricated using CMOS nanomanufacturing techniques offer high integration, stability, controllability, and scalability. Currently, integration of several hundred optical components on a single chip has been achieved, and it is expected that various core photonic quantum functionalities, including quantum light sources, quantum control pathways, and single-photon detectors, can be integrated on a single chip. In recent years, silicon-based photonic chips achieved made significant breakthroughs in the fields of boson sampling, multiphoton high-dimensional quantum entangled states, quantum key distribution, and quantum teleportation.

In recent years, there have been numerous significant achievements in the field of large-scale silicon-based quantum chips. The countries and institutions contributing to the key research papers are outlined in Tables 1.2.9 and 1.2.10. The USA, the UK, and Netherlands rank among the top three nations in terms of the number of key research papers published. Prominent contributing institutions include Delft University of Technology, University of Bristol, and Peking University. Furthermore, many of these key research papers resulted from collaborations between various research institutions across different countries, as illustrated by the collaboration networks between leading countries and institutions in Figures 1.2.7 and 1.2.8. Tables 1.2.11 and 1.2.12 list the main countries and institutions responsible for citing key papers in this field. The Chinese Academy of Sciences and University of Science and Technology of China rank the top two, reflecting China’s notable interest and engagement in this direction.

From the perspective of the overall development trajectory within the field, ultra-large-scale silicon-based quantum chips are still in their nascent stages and many key problems must be solved. As depicted in Figure 1.2.9, future development of ultra-large-scale silicon-based quantum chips will focus on the following key directions:

(1) Silicon-based spin quantum chips

1) Large-scale arrays: The practical realization of spin quantum chips requires increasing the fidelity of the initialization, manipulation, and readout modules for each qubit to sufficiently high levels. Large-scale fabrication of high-fidelity qubits using CMOS technology is difficult because the properties of spin qubits, such as valley splitting, spin-orbit coupling, and tunneling coupling between quantum dots, are very sensitive to atomic-level defects. Therefore, the quality of the material growth is crucial. Rapid detection and automated control of each qubit’s parameters are also essential.

2) Long-range coupling: Currently, most silicon-based spin qubits rely on nearest-neighbor coupling, which requires proximity between quantum dots or phosphorus-doped atoms, limiting the layout of the dense arrays. Developing methods for long-range coupling can enable the separation of qubits to over larger distances. Several experimental approaches have been explored, including floating gates, microwave cavities, superconducting resonators, and electron shuttling.

《Table 1.2.9》

Table 1.2.9 Countries with the greatest output of core papers on “ultra-large-scale silicon-based quantum chips”

《Table 1.2.10》

Table 1.2.10 Institutions with the greatest output of core papers on “ultra-large-scale silicon-based quantum chips”

《Figure 1.2.7》

Figure 1.2.7 Collaboration network among major countries in the engineering research front of “ultra-large-scale silicon-based quantum chips”

《Figure 1.2.8》

Figure 1.2.8 Collaboration network among major institutions in the engineering research front of “ultra-large-scale silicon-based quantum chips”

《Table 1.2.11》

Table 1.2.11 Countries with the greatest output of citing papers on “ultra-large-scale silicon-based quantum chips”

《Table 1.2.12》

Table 1.2.12 Institutions with the greatest output of citing papers on “ultra-large-scale silicon-based quantum chips”

《Figure 1.2.9》

Figure 1.2.9 Roadmap of the engineering research front of “ultra-large-scale silicon-based quantum chips”

(2) Silicon-based integrated photonic quantum chips

1) Large-scale integration: Integrating quantum light sources, quantum-state manipulation pathways, and single-photon detectors on a single chip while meeting all the key performance metrics remains a challenge. It is critical to reduce the losses caused by the interaction between photons and the surrounding medium within the chip. In addition, to realize practical applications, the complexity of the multiphoton high-dimensional entangled states that photonic quantum chips can generate must be continuously improved to achieve a sufficiently large state space.

2) Multichip interconnects: As the number of qubits increases, it becomes increasingly difficult to integrate more optical components on a single chip. Future developments may leverage the advantages of optical communication to achieve interconnectivity between multiple chips, thereby enabling the construction of large-scale quantum processors through distributed quantum computing. However, achieving high performance interconnects between silicon-based integrated photonic quantum chips remains a technical challenge, thereby necessitating the development of ultralow-loss interconnect technologies to enhance the fidelity of quantum-state transmission between chips.

Ultra-large-scale silicon-based quantum chips have tremendous potential in various applications. Both silicon-based spin quantum chips and silicon-based integrated photonic quantum chips are poised to become pivotal components of future quantum technologies, with profound implications for solving complex problems and enhancing information security. In the future, silicon- based quantum chips will continue to progress toward greater scalability and broader applications owing to their compatibility with traditional CMOS technology.

《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 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. “Light-controlled phased-array technology”, “control technology of unmanned systems based on brain–computer interface”, and “artificial- intelligence-based fault diagnosis and detection” are fronts of data mining. The remaining fronts of expert nomination. The annual number of core patents published for the Top 10 engineering development fronts from 2017 to 2022 is listed in Table 2.1.2.

(1) Light-controlled phased-array technology

A light-controlled phased-array antenna (LCPAA) mainly includes transmitting-receiving (TR) modules, an antenna array, light- sensitive components, and a light-controlling module, in which the radiation characteristics of the LCPAA are engineered by manipulating the electromagnetic response of light-sensitive materials (and/or components) in the space-/time-/frequency-/ spectrum-domains through light. Without loss of generality, the controlling parameters can be the intensity, wavelength, beam

《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