Accelerating Scientific Dataset Discovery with Semantic Metadata Search and Large Language Models

The rapid growth of large-scale scientific datasets from experiments, observations, and simulations has brought unprecedented opportunities for data-driven discovery, while also posing significant challenges for efficient and accurate data/metadata retrieval. Self-describing file formats such as HDF5 and netCDF store rich metadata alongside data objects, yet existing search methods are predominantly lexical, lacking semantic expressivity and requiring users to possess detailed knowledge of metadata structures. This limitation hinders effective dataset discovery, particularly in cross-domain and heterogeneous environments. To address these challenges, this dissertation presents a unified research effort that advances semantic metadata search for scientific datasets and benchmarks the energy efficiency of large language model (LLM) inference in retrieval-augmented generation (RAG) systems. First, we propose kv2vec, a distributed representation method for key–value metadata pairs that captures semantic similarity beyond exact lexical matching. Second, we develop PSQS (Parallel Semantic Querying Service), which leverages kv2vec embeddings and parallel processing to construct semantic indexes, enabling efficient and accurate semantic retrieval. Third, we introduce ICEAGE (Intelligent Contextual Exploration and Answer Generation Engine), a novel natural language query system for scientific datasets that preserves metadata context, supports both simple and nested queries, and achieves 0.96 overall accuracy with reduced GPU dependency. Extending beyond retrieval, we identify LLM inference as the dominant performance and energy bottleneck in RAG-based scientific data discovery. To address this bottleneck, we develop TokenPowerBench, a benchmarking framework that decomposes LLM inference energy consumption across system components, models, and workloads, and evaluates optimization techniques such as quantization. Our experiments demonstrate that the proposed methods improve query hit rate by over 15%, recall by over 17%, and throughput by up to 6.06× compared to state-of-the-art baselines, while achieving consistent energy savings of up to 30% in LLM inference. Together, these contributions bridge the gap between lexical and semantic search, enable natural language interaction with datasets, and provide actionable insights for building energy-efficient LLM inference frameworks. The outcomes have broad implications for accelerating scientific discovery, optimizing computational resource utilization, and guiding the sustainable design of AI-driven data management systems.