Transforming Drilling Textbooks into Semantic Tools for Design Verification and Knowledge Extraction

Abstract In drilling engineering, a vast body of domain knowledge is encoded in textbooks, handbooks, and industry standards. However, this knowledge is typically expressed in natural language and rarely structured in a way that enables machine understanding or direct integration into engineering workflows. This paper addresses this challenge by building a semantic foundation that bridges textbook-level drilling knowledge and automated design verification. Specifically, we present a pipeline that (1) constructs a large-scale domain QA corpus from "Drilling Engineering Theory and Technology" (textbook) and the "Drilling Manual" (industry standard), (2) applies a two-stage fine-tuning strategy—first performing contrastive learning for domain alignment, then applying triplet loss with hard negative mining to sharpen semantic boundaries—and (3) validates the resulting embedding model on real drilling design documents through a semantic retrieval and knowledge verification workflow. The QA corpus comprises 8,253 high-quality question–answer pairs covering geological conditions, drilling rigs and tools, drilling fluids, well trajectory design, pressure control, and operational parameters. The fine-tuned model, achieves a Recall@1 score of 0.69 (up from 0.50) with a corresponding increase in MRR, indicating more accurate and consistent retrieval of technical passages. In hard cases, the Safety Margin improves from 0.00 to 0.09 and the Confusion Risk decreases significantly, demonstrating stronger robustness and clearer separation between closely related engineering concepts. The scenario-level validation further shows that the fine-tuned model provides stable and operationally usable retrieval for both extraction and verification scenarios. In tasks such as casing program checks, shoe depth validation, and pressure window confirmation, the retrieved evidence aligns closely with expert reasoning patterns, supporting language-driven, automated design verification in practical engineering workflows. These results demonstrate that the fine-tuned model provides a reliable semantic foundation for intelligent, language-driven drilling design verification and sets the stage for broader integration of domain knowledge into automated engineering workflows.