Large models empower the chemical industry with a new development paradigm

1. Liu Zhongmin, Academician of the Chinese Academy of Engineering (CAE), Director of Dalian Institute of Chemical Physics, Chinese Academy of Sciences
2. Ye Mao, Researcher of the Dalian Institute of Chemical Physics, Chinese Academy of Sciences, and Deputy Director of Low Carbon Catalytic and Engineering Research

In the chemical industry, the development of new technologies is hindered by prolonged cycles and high risks. It is remarkable that an R&D engineer can make a technical achievement and witness its practical application within their lifetime. To leverage AI in accelerating technological R&D and driving it to mass production is a topic well worth exploring.

The chemical industry, a key player of process manufacturing, has consistently grappled with the challenge of protracted technological development and application. Typically, an initial small-scale experiment must be conducted, followed by a medium-scale experiment. Eventually, the scale can be gradually expanded to mass production and application.

The chemical industry adopts a phased technological development and application strategy due to the inherent complexity of chemical reaction processes. Even an angstrom-level (Å, unit of length, equal to 0.1 nanometer) change in chemical bonds can result in significant 10-meter-level changes in the reactor. The entire chemical plant occupies a vast area of several square kilometers, making it theoretically challenging to accurately describe such a complex system.

Therefore, the chemical industry is restricted by physical constraints, limited time, huge costs, and extremely high risks of failure. These things have become a bottleneck to the development of new chemical technologies. For example, it took 30 years for methanol-to-olefin technology to go from the lab to industrial application, undergoing three phases of technological development and experiment.

Even if the development of a new technology is complete, it may become obsolete during final implementation due to how long it took to make and changes to market demand and technical environments. Many technologies die in the medium-scale experiment phase, either in the lab or in industrial applications. Therefore, we need to change our mindset and explore new approaches to R&D.

The chemical industry has been focusing on the development of intelligent technologies in scientific research for many years. Now, we should focus on deploying intelligent technologies in industrial applications to solve the problems of slow technological development and difficult implementation, shifting from “AI for science” to “AI for engineering.”

Today, AI technologies represented by large models are growing rapidly, changing the traditional R&D mode of the chemical industry and providing a new data-driven development pattern for complex chemical systems. We can use AI and digital twin technologies to simulate the amplification process from labs to factories, and learn and improve the models in virtual plants to accelerate R&D. At the same time, we need to partner with technology companies like Huawei to develop innovative large models. We can contribute our chemical industry expertise and experience and cooperate with partners to develop professional vertical application models with moderate numbers of parameters. In this way, we can reduce the cost and threshold of large model applications, benefiting the entire chemical industry.

By further evolving professional large models into intelligent twins, we can build a virtual plant that connects to and interacts with the physical plant. The models can be applied in physical plants to verify assumptions and accumulate experience. Physical plants provide a large amount of data for model training and improvement so that the virtual plant can simulate the physical plant in all aspects. Eventually, the results of small-scale tests are verified directly in the virtual plant, greatly shortening the development cycle that used to take more than a decade to complete.

If all chemical plants can connect such a system, specific processes can be optimized and controlled, and chemical industry planning will become more efficient and agile. As a result, it no longer takes three years to plan for the next five years. Such a transformation will undoubtedly boost the competitiveness and growth of the chemical industry in China.

The Dalian Institute of Chemical Physics of the Chinese Academy of Sciences and Huawei have jointly developed a large model for the chemical industry. The model includes basic modules such as data processing and professional knowledge graphs, as well as key modules like automatic generation of reaction dynamics and automatic generation of process flow diagrams. For example, with the automatic reaction kinetics generation module, the large model can extract reaction mechanism knowledge from massive literature, generate a possible reaction network by converting the mechanism into a reaction rule, and automatically recommend an experimental solution. Finally, a robot performs a high-throughput experiment to construct reaction kinetics for industrial process development. This chemical large model is now being iterated from version 1.0 to version 2.0.

We are currently developing a comprehensive model application ecosystem for the chemical industry, aimed at propelling the intelligent and low-carbon transformation of the sector and securing a dominant position in the industry's transformation process. The objectives are outlined below:

• Eliminating the bottleneck of cascaded amplification: The large model is used to simulate and predict the behavior of chemical reactions in reactors of different scales, thereby reducing the number of medium-scale experiment processes and accelerating the R&D of new technologies.
• Shortening the R&D period: The large model is used to process and analyze massive amounts of data and quickly identify the optimal path, significantly shortening the R&D cycle of new technologies.
• Reducing R&D investment: The large model is used to reduce the number of experiments, improve R&D efficiency, and reduce new technology R&D costs.
• Enabling industry users: Intelligent tools are provided for industry users to optimize the production process and improve efficiency as well as product quality.
• Supporting intelligent decision-making and planning: The large model provides accurate data analysis and prediction, which helps management departments make more intelligent and effective decisions and plans.

The integration of AI in the chemical industry has already demonstrated promising outcomes, while also presenting a plethora of challenges, particularly in the realms of engineering and process development. Chemical processes are characterized by a complex, multi-scale system, wherein the development is driven by both empirical data and theoretical frameworks. Moreover, the continuous and dynamic nature of chemical processes renders it more challenging to establish a digital twin, as compared to discrete systems.

To develop chemical large models and intelligent twins, it is essential to cultivate a comprehensive understanding of cross-disciplinary knowledge, technologies, and expertise, foster cross-disciplinary collaboration, and develop AI talent with expertise in the chemical industry. Furthermore, it is also vital to strengthen the synergy and cooperation between industry, academia, and research institutions.