Li Fudong, a senior expert in artificial intelligence and digital transformation

Ø Bachelor of Engineering from Tsinghua University, Master of Engineering from Beijing University of Posts and Telecommunications

Ø Founder/CEO of the first batch of digital innovation startups of China Unicom Group

Ø Former General Manager of Digital Strategy and Engineering Consulting Project of China Unicom Group

Ø Former digital product expert of Oracle Company (the world's largest enterprise software company)

Ø Former Tianrongxin Company (Network Security Leader) Network Security Product Architect

Summarize
Data is the mapping and recording of the physical world in digital space, and the model is the abstraction and precipitation of the physical world in digital space.
The basic model is trained on a large amount of data to learn the general pattern, and then it can be applied to various business scenarios through fine-tuning.
This paper studies the use of basic models in the digital twin environment of Cyber Physics System (CPS), explores its potential in improving the efficiency and functional effectiveness of digital twin creation, and discusses the challenges of using basic models in a broader environment.
This article takes the autonomous driving system (ADS) as the representative CPS to explain, and points out the development direction of the effective combination of basic models and digital twins.

Keyword
Digital twin; Cyber Physics System; Basic Model; Big Language Model

I. Basic model and digital twin creation

Basic models are an important progress in the field of machine learning in recent years. These models are trained on a large amount of data and can learn and apply general patterns.

Due to its strong adaptability and flexibility, the basic model has shown a wide application prospect in many fields, especially in the digital twin creation of the Cyberphysical System (CPS). Digital twin is a virtual copy that can reflect the state and behavior of a physical system in real time.

The relationship between physical twins and digital twins in the Cyber Physics System,Basic models refer to machine learning models trained on large-scale data, which can capture and understand complex patterns and relationships in data. Common basic models include large language model (LLM), visual model and multimodal model. These models are widely used in the fields of natural language processing, computer vision and multimodal fusion.

The Autonomous Driving System (ADS) is a typical cyber physics system. Its core functions include environmental perception, decision-making and vehicle control.

Digital twin technology improves the development and testing efficiency of the system by simulating the real driving environment and conditions. The basic model has significant potential in generating and optimizing ADS digital twins, such as simulating complex traffic scenarios, predicting potential risks and optimizing driving strategies.

II. Application environment of CPS digital twin basic model

The digital twin of CPS consists of two parts: the digital twin model and the digital twin ability. The digital twin model is a virtual representation of a physical system, while the digital twin capability refers to the various functions that digital twins can perform.

This paper proposes two main application scenarios: using basic models to generate digital twins and using basic models directly as digital twins. In the former, the basic model generates and optimizes the model and functions of digital twins through automation; in the latter, the basic model itself acts directly as a digital twin through fine-tuning, reflecting the status of the physical system in real time and performing the corresponding functions.

Scenario 1: Generate a basic model of digital twins

At present, the technology of generating digital twin models mainly includes model-based system engineering methods and machine learning technologies. Although these methods improve the creation efficiency of digital twins to some extent, they still require a lot of manual intervention. The introduction of basic models is expected to significantly reduce manual work and automatically generate more realistic digital twin models.

Using basic models to generate digital twins faces many challenges, such as the usability and quality of model training data, the fidelity and validity of the model, etc. However, this also provides researchers with a lot of exploration opportunities, especially in the combination of field knowledge and model fine-tuning.

Taking the autonomous driving system as an example, the basic model can generate realistic driving environment simulation to support the testing and optimization of autonomous driving algorithms. In addition, the basic model can also evaluate the authenticity of the generated scenario to ensure the safety and reliability of the autonomous driving system.

The basic model can generate digital twin models and capabilities in many ways. For example, LLM can be used to generate analog models, or digital twins of environmental conditions (such as roads and weather) can be generated through multimodal models. Specific implementation schemes include model element recommendation system, digital assistant and fully automatic model generator.

Scene 2: Fine-tune the basic model as a digital twin

Fine-tuning the basic model so that it can directly act as a digital twin is an important direction of current research. Through fine-tuning, the basic model can capture the behavior patterns and characteristics of specific CPS and provide real-time monitoring and analysis capabilities.

The fine-tuned basic model can be used as the whole digital twin or part of its functions. For example, LLM can be used to generate operation instructions described in natural language, or simulate and analyze complex scenes through multimodal models.

Although the fine-tuning basic model has great potential as a digital twin, its application still faces many challenges, such as the transparency and interpretability of the model, real-time requirements and data privacy protection.

In the autonomous driving system, the fine-tuning basic model can generate highly realistic driving scenarios, and provide comprehensive support and guarantee for autonomous driving through continuous optimization and improvement through real-time data.

III. Problems to be solved

Although the basic model provides an opportunity to develop various forms of digital twins, their application, especially in security-critical and mission-critical areas, has raised great concerns about its inherent uncertainty.

For example, Huang Yuheng (transliteration) and others have empirically studied the different quantitative methods of LLM cognitive uncertainty (Ref. =2), aims to understand whether uncertainty estimation can and how to help describe the ability of LLM to perform different tasks.

Tanneru and others proposed a new method to quantify the uncertainty in the interpretation of LLM natural language (Ref. =3), the goal is to understand or even quantify the uncertainty of the large model.

It is foreseeable that the application of reliable and trustworthy basic models in reality still has a long way to go. For the CPS field that pays great attention to security, security, etc., we should carefully ensure that the deployment of the basic model is safe and reliable, such as conducting a thorough risk assessment and complying with strict verification processes.

For any solution, we need to thoroughly consider its cost-effectiveness. First of all, the development of basic models may involve high computing resources, data collection and preprocessing, model training, fine-tuning, etc. These costs may vary in different CPS fields.

Secondly, the deployment of certain basic models requires specialized hardware resources (such as GPU, FPGA), especially considering that the fine-tuning basic model (scene 2) needs to communicate with CPS, which requires high-throughput and low-latency communication and computing. In addition, the basic model of fine-tuning must be maintained regularly, which will increase the maintenance cost.

However, the application of the basic model will bring many advantages, such as reducing the manual workload required to build a traditional digital twin model. Therefore, it is necessary to compare the cost-effectiveness of using basic models with traditional solutions through empirical research, so as to prove that the use of basic models is reasonable.

In addition to customizing large basic models for digital twins, in some cases, lightweight dedicated basic models are also required. This is performed on dedicated hardware with low computing power for digital twin applications, and it is also applicable to key application scenarios that cannot access large basic models through the cloud.

In addition, these dedicated basic models will be applicable to digital twins with high requirements for reasoning time, because the reasoning time of lightweight models is shorter and more suitable for real-time digital twin applications. At the same time, because these models can be deployed on dedicated resources, there is no need to send data to the cloud, etc., thus improving security and privacy protection.

There are many challenges in using the basic model to generate CPS digital twins. For example, how to deal with illusions (i.e. content that is not based on factual information generated by the basic model) when generating digital twins or using them as digital twins. At present, there are many coping strategies to solve hallucination problems.

Another challenge is how to evaluate the fidelity of digital twins. There are two aspects. First of all, in terms of using the basic model to generate digital twins, the fidelity of digital twins can be accessed as usual. Secondly, when the basic model is used as a digital twin, it opens up a new research direction for evaluating its fidelity, because it will require the definition of new indicators and methods.

Finally, the basic model brings many moral and legal challenges, and the use of digital twins to generate basic models is bound to face the same problem. For example, using basic models to generate digital twins may generate copyrighted models, or use private data for training, and whether discriminatory and biased decisions will be made.

IV. Summary and Prospect

The basic model provides new methods and tools for the creation and optimization of CPS digital twins. This paper proposes two main methods of using basic models to generate and fine-tune digital twins, and discusses their advantages, challenges and future research directions. We believe that with the continuous development of basic model technology, the application of CPS digital twins will be more extensive and in-depth.

This paper discusses the application of the basic model in the digital twin of the Cyber Physical System (CPS), and focuses on analyzing its potential in improving the efficiency and functional effectiveness of digital twin creation. Through research, we found that the basic model has significant advantages in generating and optimizing digital twin models and capabilities, which can reduce manual work and improve the realism and accuracy of the model.

Taking the autonomous driving system (ADS) as an example, the new scheme demonstrates the practical application of the basic model in generating complex traffic scenarios, predicting potential risks and optimizing driving strategies. At the same time, we also discussed the challenges faced when using the basic model, including the usability and quality of model training data, the transparency and interpretability of the model, and the protection of data privacy.

Looking to the future, the application of the basic model in CPS digital twins will continue to expand and deepen. With the development of technology, the ability and scope of application of the basic model will continue to improve, making it play a role in more fields and application scenarios.

In order to make full use of the potential of the basic model, researchers need to continue to explore how to train and fine-tune the model more effectively to ensure its accuracy and reliability in specific applications.

At the same time, solving the transparency and explanatory problems of the model will be an important direction for future research, especially in security-critical and mission-critical CPS applications. In addition, with the increasing attention to data privacy and security, the development of basic models that can protect data privacy will also become an important research field.

In a word, the application of the basic model in CPS digital twins has a broad prospect, but it also needs to deal with a series of technical and ethical challenges. The author believes that with the gradual resolution of these problems, the basic model will provide strong support for the further development of industrial intelligence and CPS.