Abstract for: Data Science Workflow For Scaling Reproducible Model Analysis

Integrating System Dynamics with modern Data Science requires a shift toward high-dimensional, data-driven differential equation discovery. However, the computational burden of these tasks—including stiff solvers and sparse regression scales non-linearly. To meet latest standards, researchers must leverage High-Performance Computing (HPC) architectures to manage the increasing complexity of automated model discovery and optimization. This paper evaluates a framework benchmarking R and Python implementations across foundations, identification, policy analysis, and strategic optimization. Utilizing a university HPC cluster, we deployed Slurm Job Arrays for parallel integration and NVIDIA GPU-accelerated Sparse Identification of Nonlinear Dynamics (SINDy). The methodology focuses on a hybrid pipeline to maximize the strengths of both programming ecosystems. Python-JAX implementations achieved 1,200 trajectories per second using a single GPU, representing a 6.6x speedup over R-deSolve parallelized across 32 CPU cores. While both languages performed identically in linear identification accuracy, Python demonstrated superior robustness in non-linear identification of chaotic systems. R remained the gold standard for high-stakes statistical validation and policy visualization. We recommend a hybridized "Polyglot Workflow" for the ISDC community: utilize Python for high-compute model discovery and strategic optimization, while exporting results to R for rigorous policy communication. Future work includes developing neural surrogates to reduce inference time to microseconds and moving toward "Agentic System Dynamics" for autonomous model orchestration.