KIAFE – Künstliche Intelligenz-Assistenz zur Fließbilderstellung

When planning chemical process plants, the challenge is to find an optimum between effectiveness, acquisition and operating costs. In addition, plants should operate as energy- and resource-efficiently as possible and produce low CO2 emissions. Even experienced process engineers often rely on heuristics and empirical values here. Artificial intelligence (AI) can support plant planning and make it faster and more effective. This has a positive impact on costs in the development and design phase. As a result, companies can lower their operating costs while reducing their CO2 emissions, giving them a competitive advantage. AI can be trained with simulations and expert knowledge to assist the engineer.

Process engineering deals with the technical and economic implementation of all processes in which substances are changed according to type, property and composition. Process engineering is needed in a variety of industries, such as pharmaceuticals, chemicals, waste management, plastics production, drinking water treatment and fuel production. Some of the tasks an engineer has to solve in process engineering are, among others, the planning of processing steps, the selection and development of necessary apparatus and machines, the determination of methods for mass transfer or the definition of necessary measurement and control technology. Finally, the engineer uses software to design a flow diagram for the desired material conversion based on the various parameters. The use of AI assistance can accelerate this process and make it more effective.

Our goal is to develop an AI-based technology for the design of complex energy and resource efficient processes. The project addresses current and future requirements regarding resource use, emissions, and management of energy demand. Companies in the chemical industry have to integrate ecological aspects into their production planning against the background of climate protection and preservation of biodiversity, especially due to new guidelines regarding the reduction of CO2 emissions. For this purpose, the software used by engineers to generate flow diagrams can be supported by AI. In our approach, an engineer creates a design, which serves as input for the AI along with the desired process parameters. The AI, using reinforcement learning and a simulator, calculates several evolved flowsheets that provide an improved solution for the process task. For this, target values are given to the AI with the intention that the process plant performs its function, e.g., under the constraint of being as resource and energy efficient as possible.  The engineer may use technically sound solutions from the AI and augment them with further details. The process can be iterated in tandem between human and machine optimization to find the best solution in the tension between ecological, economic and technical criteria.

ABE fermentation is chosen as the first use case. Here, acetone, butanol and ethanol are produced from fermentation of different substrates. The substances produced in the process can be recycled as biofuels, solvents or animal feed, among other things. Despite the promising process, the design of ABE fermentation plants still involves numerous risks. For example, the product butanol has a toxic effect on the microorganisms responsible for fermentation above a certain concentration, the yield is comparatively low, and the separation of the ABE solvent mixture is very costly in terms of energy and economy. This therefore offers enormous optimization potential for AI and reinforcement learning. The project has the potential to strengthen the AI presence and acceptance in the material changing industry and at the same time, through the early integration of ecological criteria, to make the development and operation not only more economical, but at the same time more environmentally friendly. The applicability in the material-changing industry paves the way for the applicability of the newly developed technology to various sectors and future-relevant issues of the manufacturing industry, such as the use of biogenic or renewable raw materials or the integration of renewable energies.

This is where our reinforcement learning AI brings clear benefits to the industry:

There are multiple cost savings in flowsheet design supported by our AI assistance. Firstly, through optimized design taking into account construction costs and CO2 emissions, resulting e.g. in a solution with fewer components and clever material flows. On the other hand, by optimizing the parameters of plants, which leads, for example, to a reduction in energy requirements and a further reduction in CO2 emissions, thus also lowering operating costs.

In addition, there are the time savings in the otherwise manual flow diagram creation. The engineer saves time in planning. The relief (not replacement) of employees by AI is particularly important against the background of the shortage of skilled workers.

Plant design

When planning a new plant, our AI can optimize the parameters / design of the plant. The AI assistance is trained with experts and thus bundles the knowledge of many specialists. The AI can flexibly keep an eye on many requirements at the same time and find an overall cost-optimal and sustainable solution.

Process optimization

Process optimization with AI and especially reinforcement learning is very powerful. AI can simultaneously consider many requirements and make an overall cost-optimal and at the same time sustainable adjustment of all plant components. Likewise, some transportation can be optimized and, most importantly, it can respond dynamically to changes. Thus, the AI can quickly adapt to changing raw material prices, qualities and availability and find an optimal configuration.

Control design and verification

Reinforcement learning finds optimal behavior in dynamic environments. Thus, optimal control or regulation can be trained with AI. The final trained AI can then be fixed and tested for safety. Thus, an efficient and lean control component can be deployed in the plant. In addition, imitation learning can be used to create a classic, decision tree-based control or closed-loop control system with the AI. This can then be verified propositionally.