使用大语言模型实现与岩土数据集GeotechSet的对话---DFN试验
ChatGPT和GPT-4最具吸引力的一个特性是可以集成用户自定义的数据,不过如在【使用自定义的岩土数据(GeotechSet)增强ChatGPT的回答能力】中的讨论,目前API的费用还很高,因此如何高效地使用开源大预言模型(LLM,Large Language Model)是目前大预言模型研究的一个方向【LangChain---大型语言模型(LLM)的标准接口】。本文对此进行了进一步讨论和试验。
本文使用了Langchain和HuggingFace内置的Embeddings实现相似性查询。第一步使用Langchain的加载文档类导入自己的数据:from langchain.document_loaders import TextLoader, UnstructuredFileLoader
#from langchain.document_loaders.base import BaseLoader
loader = TextLoader('slope.txt')
documents = loader.load()
from langchain.text_splitter import CharacterTextSplitter
text_splitter = CharacterTextSplitter(chunk_size=500, separator=".", chunk_overlap=4)
docs = text_splitter.split_documents(documents)
第三步调用HuggingFaceEmbeddings from langchain.embeddings import HuggingFaceEmbeddings
embeddings = HuggingFaceEmbeddings()
from langchain.vectorstores import FAISS
db = FAISS.from_documents(docs, embeddings)
最后一步进行相似查询。当输入一个查询时,系统给出与查询句子最为相似的段落。
3 DFN试验
使用了论文[(2011) Stability Analysis of a Large Fractured Rock Slope Using a DFN-Based Mass Strength Approach]作为输入数据集。查询句子为"What is Discrete Fracture Network (DFN)"。此外,GPT建议的相似问题如下:(1) Can you explain how DFN-based methods can improve the stability analysis of deepening open pits?(2) What are the current limitations of using DFN realisations in modelling?(3) How can we improve the characterisation of a fractured rock mass?(4) Can you explain how the DFN-based method works?(5) What are the factors that affect mass strength determination?(6) How can we improve the efficiency of coding and computer speeds?(1) The influence of discontinuity-control or representation of the Discrete Fracture Network (DFN) needs to be improved in the stability analysis of the deepening open pits. For example, the simulation of step-path failure at a large-scale has been included in recent research by several authors.(2) Only a limited number of DFN realisations were used in the current modelling. Ideally this needs to be extended to cap ture the full range of the likely strength envelope. This requires improved efficiencies in both coding and computer speeds for more routine use. More importantly there needs to be significant advances in the characterisation of a fractured rock mass, to provide a reliable foundation for a DFN-based approach such as that which is discussed in this paper.离散断裂网络(DFN)是用于建立岩石裂隙网络模型的一种方法,可以在地质工程和石油工业等领域中应用。DFN方法提供一种显式地建立裂隙网络的途径,为建立现实几何裂隙模型提供帮助。为了更可靠地使用DFN方法,需要在岩石裂隙网络特征描述方面有重大进展。在岩石裂隙网络特征描述方面,存在不同的方法,包括完全随机模型、完全随机模型和已知裂隙模式模型。在这方面还需要进一步的研究。总之,为了更好地应用DFN方法,需要进一步研究岩石裂隙网络特征描述、提高编码和计算速度的效率。同时,还需要在建立可靠的DFN模型方面进行更多的研究。
(3) This signifies the importance of strain magnitude relative to the failure environment; with a greater influence of DFN variation in low strain surroundings. Comparison of two sections, with similar fracture intensities. Comparison of strength data from different axial strains within two sections, which have similar fracture intensities, with non downgraded/original properties.(4) In this case, fracture network geometry, strain and con finement were particularly important. The DFN-based method also demonstrated that fracture intensity has a greater influ ence on mass strength in low strain environments. These factors all have relevance for the determination of ma ss strength by any method.4 后处理
把上面得出的4个结果导入GPT-4和Claude,"提炼出最重要的内容,作一个完整详细的总结,并用中文输出"。在深部露天矿山的稳定性分析中,需要提高对离散断裂网络(DFN)的不连续性控制或表征的影响,加强DFN在深部露天矿稳定性分析中的作用,例如最近几位作者的研究已经包括了大规模阶梯路径破坏的模拟。目前的模型仅使用了有限数量的DFN实现,理想情况下需要扩展这些实现来捕捉更完整的强度包络线范围。这需要在编程和计算机速度方面提高效率,以便更常规地使用。更重要的是,需要在破碎岩体的表征方面取得重大进展,为基于DFN的方法提供可靠基础。
这突显了应变幅度相对于破坏环境的重要性;在低应变环境中,DFN变化具有更大的影响。通过比较具有相似断裂强度的两个断面以及不同轴向应变下的强度数据,可以发现断裂网络几何、应变和约束等因素尤为重要。基于DFN的方法还表明,在低应变环境中,断裂强度对整体强度的影响更大。所有这些因素对于通过任何方法确定质量强度都具有相关性。
5 参考文献
[1] (2008) Numerical analysis of pit wall deformations induced by block-caving mining. A combined FEM/DEM-DFN synthetic rock mass approach【在海平面下采矿---神奇的Island Copper露天矿】
[2] (2011) Stability Analysis of a Large Fractured Rock Slope Using a DFN-Based Mass Strength Approach【ChatGPT---学术文献引用的反向校准】[3] (2013) An integrated approach for rockfall analysis with drapery systems【芦山6.1级地震诱发的地震灾害】[6] (2018) Overhanging rock slope by design: An integrated approach using rock mass strength characterisation, large-scale numerical modelling and limit equilibrium methods【工程岩体分类RMS(Rock Mass Strength)】