From Time-Series to Text: Multimodal and Agentic Approaches for Automated Drilling Report Generation

Daily Drilling Reports (DDRs) summarize complex operations by analysing high-frequency sensor data. Automating DDR generation is challenging, requiring joint reasoning over multivariate time series and domain-specific semantics. In this work, we evaluate three methodologies for DDR generation from raw sensor data: (i) a multimodal architecture integrating time series foundation models (TSFMs, e.g., Chronos-2) via cross-modal conditioning, (ii) a vision-language model (VLM) reasoning over visualized sensor plots, and (iii) a tool-augmented Agentic approach utilizing SQL for iterative data exploration. Our results reveal complementary strengths. The Agentic system achieves the highest accuracy in precise numerical extraction due to explicit tool use. In contrast, the TSFM-based multimodal approach excels at capturing temporally extended contextual events, demonstrating stronger implicit temporal understanding. These findings demonstrate that DDR generation requires distinct, complementary approaches, highlighting a clear path for future work: developing hybrid systems that combine the reliable data retrieval of agents with the inductive biases of multimodal TSFMs.