Ho provato ad interrogare Perplexity/Kimi K2 sullo stato attuale dell'utilizzo degli LLMs in ingegneria geotecnica. Risulta evidente la necessità di supervisione umana. Tuttavia, ho notato che i modelli migliorano significativamente ogni nuovo aggiornamento.
Research shows that large language models are actively being applied to solve practical geotechnical engineering problems, with several key challenges and applications emerging in recent studies.
## Actionable Problems Being Addressed
**Design and Analysis Tasks**
- **Slope stability assessment**: LLMs analyze multimodal data combining textual descriptions with image analysis to evaluate landslide risks[1]
- **Foundation design**: Specialized frameworks like GeoLLM handle bearing capacity and settlement predictions for single piles, though models with 1.8-72 billion parameters show diminished performance compared to 100+ billion parameter models[2]
- **Reliability programming**: Automating code generation for reliability algorithms including First Order Reliability Method (FORM), subset simulation, random field simulation, and Bayesian updating using Gibbs sampling[3]
**Site Investigation and Characterization**
- **Automated site planning**: LLM-empowered systems extract requirements from regulations and create subsurface geological cross-sections while accounting for stratigraphic uncertainty[4]
- **Geological interpretation**: Multimodal integration of borehole logs, site photos, and technical reports for subsurface characterization[5]
- **Landslide investigations**: Agentic AI systems automate key components of landslide event reconstruction[6]
**Risk Assessment and Compliance**
- **Seismic microzoning**: Combining textual and image data for seismic risk assessment[1]
- **Code compliance checking**: Automated verification against standards like API RP 2A[5]
- **Liquefaction analysis and tunnel safety evaluation**: Real-time information extraction and decision support[7]
## Main LLMs and Approaches
**Commercial Models**
- **ChatGPT/GPT-4**: Most frequently cited, showing strong multimodal interpretation capabilities[1][5]
- **Copilot and Gemini**: Evaluated in state agency pilot programs[8]
**Specialized Frameworks**
- **GeoLLM**: Custom framework with hybrid prompt engineering for geotechnical design, outperforming general-purpose LLMs in domain-specific tasks[2]
- **RAG-based systems**: Retrieval-Augmented Generation using domain-specific databases of research papers, technical documents, and software documentation[9]
## Key Challenges and Limitations
**Technical Barriers**
- **Mathematical accuracy**: LLMs struggle with complex geotechnical calculations and maintaining computational precision[2]
- **Data scarcity**: "Small data" challenges in geotechnical risk assessment require specialized language model architectures[10]
- **Interpretability**: Models lack transparent reasoning for engineering validation[11]
**Implementation Issues**
- **Validation deficit**: LLMs cannot inherently validate data or adhere to design codes without human oversight[12]
- **Standardization gap**: Absence of standardized practices for AI integration in geotechnical workflows[7]
- **Expertise dependency**: Prompt engineering effectiveness relies heavily on user expertise and iterative experimentation[5]
**Performance Constraints**
- **Parameter sensitivity**: Models with 1.8-72 billion parameters show significantly lower accuracy than 100+ billion parameter models, even with domain adaptation[2]
- **Context limitations**: Difficulty interpreting specialized geotechnical content and maintaining engineering judgment[9]
Current research emphasizes that while LLMs demonstrate promising capabilities, they function best as augmentation tools rather than replacements for engineering expertise, requiring rigorous validation and human-machine collaboration frameworks.[11][7]
[1](
https://arxiv.org/pdf/2312.12411.pdf)
[2](
https://www.sciencedirect.com/science/article/abs/pii/S0266352X24007882)
[3](
https://arxiv.org/abs/2506.19536)
[4](
https://www.sciencedirect.com/science/article/abs/pii/S0926580525001438)
[5](
https://arxiv.org/pdf/2507.05613.pdf)
[6](
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5340542)
[7](https://colab.ws/articles/10.1016%2Fj.geoai.2025.100036)
[8](
https://ascelibrary.org/doi/10.1061/9780784485989.006)
[9](
https://ascelibrary.org/doi/10.1061/9780784485989.001)
[10](
https://www.siam.org/publications/s...ll-data-challenges-with-language-models/)
[11](
https://www.construction.cam.ac.uk/...al-intelligence-geotechnical-engineering)
[12](
https://flocode.substack.com/p/effective-use-of-large-language-models)
[13](
https://www.geoprofessional.org/new...g-applications-of-large-language-models/)
[14](
https://www.geoprofessional.org/wp-content/uploads/2024/10/AI-in-Geotechnical-October-2024.pdf)
[15](
https://www.sciencedirect.com/science/article/pii/S3050483X25000358)
[16](
https://www.arxiv.org/abs/2312.12411)
[17](
https://www.tandfonline.com/doi/full/10.1080/17499518.2024.2381026)
[18](
https://www.ivysci.com/articles/312...roblemsolving_with_large_language_models)
[19](
https://www.sciencedirect.com/science/article/pii/S3050483X25000486)
[20](
https://www.fomlig2024.com/En/Menu/49f13769-d30c-4a62-add4-51b3d442581c)
