Title: ISOMORPH: A Supply Chain Digital Twin for Simulation, Dataset Generation, and Forecasting Benchmarks

Original: arXiv:2605.12768v2 Announce Type: replace-cross Abstract: While open time-series forecasting (TSF) benchmarks are well-established for sectors like retail, energy, weather, and traffic, supply-chain logistics has been notably neglected. To address this gap, we present ISOMORPH, the inaugural public digital twin of a multi-echelon logistics network. This tool features interpretable, user-configurable parameters alongside modular topology, demand, and control rules. ISOMORPH operates as a simulator that progresses a directed routing graph through discrete time steps. Within this framework, demand is either satisfied from existing inventory or logged as backlog, subsequently triggering replenishment actions across the network. The system’s state monitors inventory levels, pending orders, shipments in transit, and a smoothed demand estimate, resulting in Markovian dynamics within a manageable state space. The generated data successfully replicates the bullwhip effect at magnitudes consistent with empirical observations, and three conservation laws serve as verification mechanisms for future simulator enhancements. We have released datasets covering two catalogue sizes ($C=50$ and $C=200$), six scenario variations, and 20 perturbations based on Latin-hypercube sampling. These datasets capture dynamic behaviors rarely seen in static TSF benchmarks, such as variance amplification, cascading bottlenecks, regime shifts, and cross-channel coupling driven by shared macro shocks. When evaluated in a zero-shot setting, four foundation models—Chronos, Moirai, TimesFM, and Lag-Llama—achieved MASE scores that surpassed public GIFT-Eval references at low-to-moderate forecasting horizons, suggesting their potential integration into current benchmark frameworks. Additionally, these models generate forecast confidence bands by applying Latin-hypercube perturbations to demand-side parameters, facilitating forward uncertainty quantification (UQ). This capability, which is absent in standard TSF datasets, highlights that foundation models can act as efficient surrogates for digital-twin-based UQ. Code (MIT): https://github.com/tuhinsahai/ISOMORPH. Interactive demo: https://huggingface.co/spaces/HyeminGu/ISOMORPH-demo.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC