11:15 AM - 11:30 AM
[SCG50-06] Preliminary investigation of the application of large language models to geotechnical problems
Keywords:Large language models, Geotechnical engineering
This presentation reports a preliminary investigation into the potential application of large language models (LLMs) like GPT in geotechnical engineering, a critical field within civil engineering focused on the behavior of earth materials and the integrity of ground structures. Despite advancements in computational methods, geotechnical engineering faces challenges like unpredicted subsurface conditions and increasing project complexity, often leading to costly overruns and safety risks. This study explores the integration of LLMs, known for their ability to understand and generate language with human-like proficiency, as a solution to enhance predictive capabilities, optimize design processes, and streamline decision-making in geotechnical engineering.
The presentation will begin with a brief introduction to LLMs and their main extensions to cover various practical usages. It is well-known that LLMs, trained on extensive natural language datasets, can comprehend, summarize, and interpret text, offering potential as auxiliary tools in geotechnical engineering for processing technical documents, extracting data, predicting outcomes, and generating design concepts. In contrast, this study aims to assess the applicability and effectiveness of LLMs in geotechnical tasks by exploiting the LLMs' ability to extract informative features from text, perform multimodal modeling, and provide explanable predictions.
We organized a workshop dedicated to this study and completed four case studies to demonstrate LLMs' role in simplifying complex problems and enhancing decision-making. The four geotechnical tasks include slope stability assessment, microzoning by seismic risk, parameter recommendation for liquefaction simulation, and site similarity prediction. In additional to the benefits from LLMs, our study revealed that the probabilistic nature of LLMs and their reliance on word relationships necessitate expert oversight and tailored inputs for accurate solutions in complex engineering tasks. Fine-tuning LLMs for specialized responses in geotechnical engineering remains a challenge, underscoring the need for effective interface design for seamless integration with other systems.
In conclusion, the integration of LLMs signifies a shift towards more efficient, data-driven approaches in geotechnical engineering. Their potential to shape the future of the field is underscored. The study exemplifies how even beginners in LLMs and data science can rapidly integrate these tools into their workflows, driving innovation and enhancing efficacy in this foundational engineering realm.
The presentation will begin with a brief introduction to LLMs and their main extensions to cover various practical usages. It is well-known that LLMs, trained on extensive natural language datasets, can comprehend, summarize, and interpret text, offering potential as auxiliary tools in geotechnical engineering for processing technical documents, extracting data, predicting outcomes, and generating design concepts. In contrast, this study aims to assess the applicability and effectiveness of LLMs in geotechnical tasks by exploiting the LLMs' ability to extract informative features from text, perform multimodal modeling, and provide explanable predictions.
We organized a workshop dedicated to this study and completed four case studies to demonstrate LLMs' role in simplifying complex problems and enhancing decision-making. The four geotechnical tasks include slope stability assessment, microzoning by seismic risk, parameter recommendation for liquefaction simulation, and site similarity prediction. In additional to the benefits from LLMs, our study revealed that the probabilistic nature of LLMs and their reliance on word relationships necessitate expert oversight and tailored inputs for accurate solutions in complex engineering tasks. Fine-tuning LLMs for specialized responses in geotechnical engineering remains a challenge, underscoring the need for effective interface design for seamless integration with other systems.
In conclusion, the integration of LLMs signifies a shift towards more efficient, data-driven approaches in geotechnical engineering. Their potential to shape the future of the field is underscored. The study exemplifies how even beginners in LLMs and data science can rapidly integrate these tools into their workflows, driving innovation and enhancing efficacy in this foundational engineering realm.