Conference Agenda

At BASF, we create chemistry for a sustainable future. Our ambition: We want to be the preferred chemical company to enable our customers’ green transformation. Through science and innovation we provide our customers, active in nearly all sectors, with products to meet the current and future needs of business and society. BASF Antwerp is the largest integrated production site in Belgium and the second largest of the BASF Group worldwide.
The chemical industry worldwide and in Europe in particular is sailing through rough waters. The combination of high uncertainty due to the geopolitical situation, structurally high costs, regulatory burden and climate change prove a truly challenging cocktail.
For a production site like BASF Antwerp, one of the key levers to master this challenge is the smart use of developments in digitalization, automation and AI to make sure our production plants operate under the most optimal conditions.
However, despite the hype that often accompanies new developments, generating a significant and sustainable impact often proves harder than expected. We will discuss some of the typical challenges that have to be overcome. We will present a number of real life realizations: i) optimizing production and productivity using data analytics, artificial intelligence and model based optimization & control, ii) developing digital twins in legacy plants, iii) streamlining workflows using agile cocreation and iv) introducing drones and robots to facilitate daily operations. Through these real life examples we will illustrate some of the lessons we learned so far on our journey and hint at some of the future opportunities we see.

In the context of Chemical Processing Industry (CPI), a wide range of sensor technologies and data collection methods are available to use. These sources provide valuable insights into the monitoring of physical and chemical phenomena occurring throughout the process, the status of equipment, prevailing process conditions, attributes related to raw materials and product quality, emissions data, logistic issues, and more (Ye et al., 2020). Despite the intensive use of sensors in industrial settings, they often miss critical process information. Critical issues such as leaks, corrosion, and insulation degradation may escape the observational space of industrial sensing devices and go undetected.

Therefore, it becomes imperative to seek alternative sources of information for acquiring insights into the state and health of industrial processes. One such alternative is industrial text data available in operators reports, alarm tags, process memorandums, etc. Despite the existence of numerous text processing methods, there is a notable lack of studies exploring their applicability in the chemical processing industry (CPI) context. This work endeavors to investigate the efficacy of text processing techniques for information retrieval from industrial text data, specifically, for process safety and containment event (PSCE) prediction.

To assess the value of information contained in the industrial text data, two scenarios were considered, one being simulated using GPT-3 model from OpenAI, and another being real industrial data. The tested NLP approach (Reimers and Gurevych, 2019) has proven to be efficient regarding information retrieval from simulated text data. On other hand, the extraction of information present solely in industrial text data presented some challenges, such as, specific vocabulary and incomplete information with respect to the topic of analysis.

To address the challenges associated with extracting information from industrial text data, one potential solution involves fine-tuning NLP models for specific production contexts. This approach requires a high-quality dataset representative of the targeted manufacturing process. However, this method poses limitations in terms of generalizing to other manufacturing processes and sites. In this case, the limitation is related to the site/process specific vocabulary used. Additionally, it is still not able to cope with incomplete information. An alternative approach is integrating industrial text data with available numerical (sensor) data (Strelet et al., 2023), which can mitigate the inherent limitations of text data and enhance scalability across different production environments.

References

Ye, Z., Yang, J., Zhong, N., Tu, X., Jia, J., & Wang, J. (2020). Tackling environmental challenges in pollution controls using artificial intelligence: A review. Science of The Total Environment, 699, 134279. https://doi.org/10.1016/j.scitotenv.2019.134279

Strelet, E., Peng, Y., Castillo, I., Rendall, R., Wang, Z., Joswiak, M., Braun, B., Chiang, L., & Reis, M. S. (2023). Multi-source and Multimodal Data Fusion for Improved Management of a Wastewater Treatment Plant. Journal of Environmental Chemical Engineering, 111530. https://doi.org/10.1016/j.jece.2023.111530

Reimers, N. and Gurevych, I. (2019). Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3980–3990, Hong Kong, China. Association for Computational Linguistics

As large language models (LLMs) have shown remarkable capabilities in conversing via natural language [1], the question arises in which way LLMs could potentially assist scientists and engineers in research and industry with domain-specific tasks [2]. Existing approaches of leveraging LLMs in science have been employed in a multitude of different domains [3] such as ChemCrow for autonomous chemical synthesis execution [4]. Since state-of-the-art LLMs have shown remarkable capabilities in code generation in various programming languages [5], LLMs could assist scientists with using tools such as modeling environments by converting textual information into structured domain-specific languages [6].
We generate dynamic chemical reactor models in Modelica code format from textual descriptions as user input. Firstly, we fine-tune Llama 3 8B Instruct on synthetically generated Modelica code for different reactor scenarios. The supervised fine-tuning procedure is conducted using the parameter efficient finetuning technique low rank adaptation [7]. Secondly, we compare the performance of our fine-tuned model to the baseline Llama 3 8B Instruct model as well as GPT-4o. A human trained in the chemical engineering domain assesses the models’ predictions manually with regards to syntactic and semantic accuracy of the generated dynamic models.
Our initial findings show that the fine-tuned model is able to follow the syntax of the Modelica language more accurately than the respective base model Llama 3 8B Instruct and GPT-4o. However, the fine-tuned model reveals shortcomings with respect to the semantic accuracy of the generated systems of equation compared to GPT-4o and Llama 3 8B Instruct. We expect that adapting training and inference settings in successive investigations will significantly improve the fine-tuned model’s Modelica code generation capabilities and thus, in the long run, enable chemical engineers to save time when performing dynamic simulations.

References
[1] Sébastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, Harsha Nori, Hamid Palangi, Marco Tulio Ribeiro, and Yi Zhang. Sparks of artificial general intelligence: Early experiments with gpt-4. March 2023. doi: 10.48550/ARXIV.2303.12712.

[2] Artur M. Schweidtmann. Generative artificial intelligence in chemical engineering. Nature Chemical Engineering, 1(3): 193–193, March 2024. ISSN 2948-1198. doi: 10.1038/s44286-024-00041-5.

[3] Lukas Schulze Balhorn, Jana M. Weber, Stefan Buijsman, Julian R. Hildebrandt, Martina Ziefle, and Artur M. Schweidtmann. Empirical assessment of chatgpt’s answering capabilities in natural science and engineering. Scientific Reports, 14(1), February 2024. ISSN 2045-2322. doi: 10.1038/s41598-024-54936-7.

[4] Andres M Bran, Sam Cox, Oliver Schilter, Carlo Baldassari, Andrew D White, and Philippe Schwaller. Chemcrow: Augmenting large-language models with chemistry tools. April 2023. doi: 10.48550/ARXIV.2304.05376.

[5] YueWang, Hung Le, Akhilesh Deepak Gotmare, Nghi D. Q. Bui, Junnan Li, and Steven C. H. Hoi. Codet5+: Open code large language models for code understanding and generation. May 2023. doi: 10.48550/ARXIV.2305.07922.

[6] Pieter Floris Jacobs and Robert Pollice. Developing large language models for quantum chemistry simulation input generation. September 2024. doi: 10.26434/chemrxiv-2024-9g2w2.

[7] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. June 2021. doi: 10.48550/ARXIV.2106.09685.

The future of Process Systems Engineering (PSE) lies in greater automation, enabling faster and more accurate operational decision-making [1]. This research investigates whether Large Language Model (LLM) Autonomous Agents (LAAs) can be a driving force in this transformation. LAAs are AI-driven systems capable of executing complex, multi-step tasks autonomously, with minimal human intervention [2]. LAAs have demonstrated potential in automating processes in other industries; for instance, ChatDev within software engineering: a framework that leverages task-specific LLM agents to manage software design, coding, and testing [3]. This case study illustrated the potential of using multiple LLM agents to streamline complex processes, automate decision-making, and significantly boost overall performance by dividing the workload among interacting agents [3]. Building on the success of ChatDev, this project seeks to explore whether similar innovations could be beneficial to PSE.

To explore the potential of LAAs within PSE, a framework was developed to automate sustainable operational decision-making in a system of Gas-Oil Separation Plants (GOSPs). Since PSE relies on a combination of models, simulations, and tools, the case study was structured to encompass a variety of techniques. The objective is to determine the optimal sustainable operation of a GOSP system given a production target from upper management. A concise prompt relays this target, initiating the automated workflow: following the Mixture of Experts (MoE) framework [4], a series of expert LLM agents collaborate to complete the task. The workflow integrates multiple tools, such as interfacing with a HYSYS flowsheet, solving multi-objective optimisation problems, and analysing cost-emissions trade-offs from a Pareto frontier. It emulates a realistic series of analyses for operational decision-making, leveraging quick low-fidelity simulations for initial assessments and high-fidelity modelling for refinement. Subsequently, two specialised agents—one focused on economic objectives and the other on environmental considerations—evaluate the Pareto front to negotiate an optimal operating point.

Although the case study is framed within the sustainable operation of the oil and gas industry, the broader purpose of this research is to explore how LAAs can be leveraged in PSE. If successful, by automating multi-step processes, LAAs could offer significant improvements in efficiency and flexibility in operational decision-making, extending their benefits beyond just decision-making in PSE. Ultimately, this research aims to demonstrate how LAAs can play a pivotal role in the industry's transition towards the ‘plant of the future’, where diverse models and technologies are seamlessly integrated into a unified, intelligent framework.

References
[1] Gamer, T., Hoernicke, M., Kloepper, B., Bauer, R. and Isaksson, A.J. (2020) 'The autonomous industrial plant – future of process engineering, operations and maintenance', Journal of Process Control, 88, pp. 101–110. doi: 10.1016/j.jprocont.2020.01.012.

[2] Liu, Z. et al. (2023) 'Bolaa: Benchmarking and orchestrating LLM-augmented autonomous agents', arXiv preprint, arXiv:2308.05960.

[3] Qian, C. et al. (2024) 'Chatdev: Communicative agents for software development', in Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, 1, pp. 15174–15186.

[4] Wang, J., Wang, J., Athiwaratkun, B., Zhang, C. & Zou, J., 2024. Mixture-of-Agents enhances large language model capabilities. arXiv preprint, [online] Available at: https://arxiv.org/abs/2406.04692