Human-guided, AI-accelerated system dynamics via pipeline algebra

Abstract

System Dynamics (SD) gives Systems Engineers (SEs) and Model-based Systems Thinking (MBST) practitioners in general a rigorous way to reason about feedback-rich problems, yet adoption remains low because assembling a causal-loop diagram (CLD), validating its logic, and converting it into an executable model are still labor-intensive and require specialized skills. Human analysts remain central to judgment, but they should spend their time on insight rather than tool wrangling. This dissertation demonstrates that chat-based large language model (LLM) pipelines can remove that bottleneck, automating polarity-reversal checks, loop-dominance mapping, latent embeddings that capture joint structure-behavior signatures for similarity search, missing-loop discovery, and first-cut model synthesis, thereby lowering the entry barrier for SEs and cutting typical SD turnaround from days to minutes. Powered by ChatGPT o3, a transformer pre-trained on internet-scale corpora of prose, source code, and mathematical notation, a single interactive session can (i) read narrative text, (ii) propose syntactically complete CLDs, (iii) diagnose structural anomalies, and (iv) translate diagrams into executable SD code. The pipeline then runs an LLM-generated simulator and, through chain-of-thought prompting, iteratively tunes loop structure and parameters until simulated behavior reproduces a reference mode that the same LLM mined from prose and refined via targeted web search. These feats rest on three enablers: cross-modal pattern learning that maps language to graph and code representations; chain-of-thought prompts that force the model to externalize intermediate reasoning; and an agentic, simulator-in-the-loop refinement cycle that tests and revises its own drafts. The full loop finishes in minutes, far faster than manual workflows, and while domain judgment is still decisive at checkpoints, no specialized SD software expertise is required. Three studies validate the approach. First, the pipeline outperformed forty-three graduate students and matched an instructor benchmark when extracting Janis's groupthink CLD, finishing each run in under fifteen minutes. Second, it converted qualitative CLDs into quantitative executable SD model simulators in minutes through expert-in-the-loop refinement. Third, symbolic routines with no LLM involvement computed generalized loop sets for 678 models and clustered 59K equations from a 1K-model corpus; by clarifying how feedback structures share influence across behavior modes, loop sets provide SEs a principled aid to loop dominance comprehension and are slated for integration with the LLM toolkit. Every transformation, whether an LLM invocation, a symbolic routine, or a shell command, is expressed in Pipeline Algebra (PA), a typed workflow language that serves as an executable notation for thought, records explicit pre- and post-conditions, supports deterministic replay, and aligns naturally with meta-algorithmic control. Forthcoming primitive operators, exposed as terminal functions through the OpenAI functions API, will let GPT itself invoke loop-set analytics, polarity-reversal detection, and higher-order transformations such as map and comap, unifying them within the same declarative fabric and steering control from bespoke orchestration code toward the language model, thereby laying the groundwork for self-optimizing model laboratories that combine formal mathematics with AI-guided pattern discovery. By combining LLM automation with a rigorous workflow backbone, the approach lets SEs exploit SD with far less overhead and far greater throughput. By shifting the analytic burden from tool wrestling to insight building, it invites a wider pool of engineers and decision makers to deploy feedback-based modeling, an advance that can sharpen climate-mitigation policy and other responses to feedback-driven challenges.