Chemical Process Design Approaches Leveraging Process Modularization, Intensification, and Large Language Models
Abstract
This dissertation investigates critical aspects of chemical engineering, including reactor intensification, process expansion, distributed natural gas desulfurization, and large language models. The key findings of this work highlight the complex interplay among reactor intensification, separation design, production costs, and environmental considerations, emphasizing the necessity for tailored approaches. Economic analysis confirmed triazine-based absorption and SourCatTM to be cost-effective desulfurization methods. Moreover, In one of our study we underscored the limitations of general-purpose language models and advocated for the use of domain-specific training data to enhance reliability in process design applications. As a case study, the first study discusses the critical effects of reactor intensification on downstream processes and the impact of different process expansion approaches on the cost and environmental effects of ethylene oxide production. In this study, the production of ethylene oxide was simulated at a production capacity of 100 kt/yr using a conventional reactor and reaction kinetics. Subsequently, the impact of reactor intensification was investigated using microfibrous-entrapped catalysts on separation section design. Second, three expansion approaches, namely, the Brownfield, retrofitting, and Greenfield processes, were explored with an emphasis on the implications of their cost and environmental impact. Despite offering cost reduction, the limited separation capacity of the Brownfield process limits the potential increase in production achievable through reactor intensification. Although retrofitting reduces production costs, it introduces complexities at a capacity of >130 kt/yr, necessitating additional separation columns. The Greenfield process, with the lowest carbon emissions among the three investigated expansion approaches, is cost-effective only at production rates above 160 kt/yr. The retrofitted process involves a moderate increase in carbon emissions, whereas the Brownfield process leads to a substantial increase in carbon emissions with capacity expansion. These findings emphasize the complex interplay among reactor intensification, separation train design, production costs, environmental considerations, and capacity limitations in chemical processes, underscoring the importance of formulating tailored approaches to suit specific requirements and carbon emission goals. The second section discusses natural gas as a reliable, clean, and abundant energy source for heat and electricity generation. A vital aspect of using natural gas as an energy source is the removal of water, CO2, and H2S before pipeline transmission. Desulfurization is a crucial and routinely used purification step for removing H2S from natural gas in conventional reservoirs. With increasing energy demand, stranded/distributed natural gas resources, which remain unused for economic reasons (small scale and remote from the market), are being increasingly considered at present. Investigating low-cost distributed natural gas desulfurization is essential as a preprocessing step for utilizing stranded gas. In this study, three desulfurization processes, namely, triazine-based absorption, a liquid redox system (LOCAT®), and solid-bed oxidation (SourCatTM), were analyzed to study their economics and capability to remove H2S from small-scale distributed natural gas resources. Besides considering the impact of modularizing these processes, which typically use standardized equipment sizes owing to the transmission difficulties of stranded gas resources, the economics of the modularized processes were compared with those of traditional-design processes. Each process was simulated using Aspen Plus at different sulfur concentrations (500–2500 ppm) and natural gas capacities (1–100,000 MSCFD), and the corresponding desulfurization costs were estimated. In addition, a sensitivity analysis was used to identify the parameters influencing the desulfurization cost. The results revealed that the triazine-based absorption process is economical for processing capacities lower than 500 MSCFD, whereas SourCatTM is cost-effective for processing capacities between 500 and 100,000 MSCFD. Furthermore, the raw material, pump utility, and sorbent costs substantially impact the triazine-based absorption, LOCAT®, and SourCatTM processes, respectively. An analysis of the applicability of modularized desulfurization units in gas fields is expected to provide valuable insights into the quantitative assessment of the modularization effectiveness and its potential application limitations. In our third study explored the usefulness of the large language models (LLMs)in process design context-based evaluation of, specialized metrics to assess chemical process design-specific query evaluation, problem-solving abilities, and correct terminology usage. This work highlights the general-purpose language model limitations of using general training sets to convey process engineering concepts, emphasizing the need for training on design diagrams such as process flow diagrams. The general-purpose model hallucinates in a specific process design task by either over- or under-assuming. This research reiterates the age-old phenomenon of "garbage in, garbage out” in system modeling, demonstrating that models trained on general language data exhibit lower accuracy in specialized contexts. Each engineering term was heavily contextualized to address the issues of ambiguity and polysemy in process design terminology to reduce model hallucinations. Subsequently, words and phrases were transformed into vector representations using techniques like word-to-vector. By calculating the Euclidean distances between these word vectors, terms that were close in meaning were identified within the process design context. Clustering algorithms, such as k-means, group these vectors into clusters of contextually similar terms, effectively disambiguating polysemous words based on their usage in process design documents. Graphs were then used to find relationships and contextualize technical understanding using process design diagrams such as flowsheets, flow diagrams, and graphs. Furthermore, maintaining consistent distances and clustering while training the language model ensured that the language model learned specifically in the specific flowsheet context and provided consistent answers to specific questions, thereby enhancing proficiency and maintaining consistency for each query. By analyzing the impact of training data quality on LLM performance, the integration of domain-specific texts was found to enhance model reliability. These findings suggest that targeted training approaches, such as the strategy described in this work, for different types of training data other than text are essential for developing LLMs capable of effectively operating within technical fields, such as process design engineering.