Supplementary Dataset: Serialized System Dynamics Models and Semantic Search Benchmark Logs

This repository contains the supplementary materials and experimental data supporting the research article: "Semantic Search for System Dynamics Models using Vector Embeddings in a Cloud Microservices Environment". The dataset consists of two primary components: 1. Model Corpus (File: system_dynamics_models_corpus.json) This archive contains the source code and serialized JSON structures of 63 System Dynamics models used as the test environment. The models cover diverse domains including economics, biology, epidemiology, and project management. They serve as the "Search Space" for the reported experiments. 2. Experimental Benchmarks This section contains the raw execution logs for five distinct test scenarios (A–E). These files document the comparative analysis between three retrieval strategies: PostgreSQL Keyword Search (Baseline 1: Exact/Substring matching) Apache Lucene BM25 (Baseline 2: Probabilistic relevance ranking) Proposed Semantic Search (ONNX/nomic-embed-text-v1.5) Scenario A: Dynamics of cash flows and resources (File: benchmark_results_scenario_A.csv) Focus: Validation of Precision. Key Finding: Verifies the reported 100% Precision of the Semantic Search method. Both lexical baselines struggled with domain ambiguity: BM25 achieved only 40% precision, while the SQL baseline yielded a 60% noise rate. The semantic approach effectively filtered structural isomorphism (e.g., distinguishing "economic resource flow" from "grain inventory"). Scenario B: Rosenzweig-MacArthur model stability (File: benchmark_results_scenario_B.csv) Focus: Zero-Shot Retrieval & Recall. Key Finding: Illustrates a vocabulary mismatch where the SQL baseline failed completely (0 results) due to missing exact terminology. In contrast, BM25 and the Semantic Search method successfully retrieved the target 'Rosenzweig MacArthur Predator-Prey Model' (Rank 1). Scenario C: Climate influence on viral transmission (File: benchmark_results_scenario_C.csv) Focus: Contextual Filtering. Key Finding: Highlights significant noise reduction (90% Precision vs ~16% for both Baselines). Both BM25 and SQL were heavily affected by the generic term "influence," retrieving irrelevant ecological models. The semantic approach successfully prioritized "viral transmission" contexts (pathogens, antibodies). Scenario D: Weighted gene regulatory network modeling (File: benchmark_results_scenario_D.csv) Focus: Semantic Disambiguation. Key Finding: Demonstrates robustness against substring matching errors. While the baselines suffered from false positives (e.g., matching "gene" within "Generation" in power grid models), the semantic search maintained 100% precision, correctly distinguishing biological gene regulation from linguistic noise in energy and waste management domains. Scenario E: Population dynamics with seasonal migration patterns (File: benchmark_results_scenario_E.csv) Focus: Conceptual Mapping of Complex Dynamics. Key Finding: Verifies the ability to retrieve models based on behavioral descriptions. The semantic engine successfully retrieved models exhibiting "seasonal migration" (100% Precision), whereas BM25 and SQL achieved low precision (~38%), failing to link the query to structurally analogous stock-and-flow dynamics when explicit "migration" terminology was variable or absent.