acceptodds
Under review as a conference paper at ICLR 2027

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

Abstract

Open time-series forecasting (TSF) benchmarks cover retail, energy, weather, and traffic, but supply-chain logistics remains underserved. We introduce ISOMORPH, the first public digital twin of a multi-echelon logistics network with fully interpretable, user-configurable parameters and modular topology, demand process, and control rules. The simulator advances a directed routing graph in discrete time: demand arrives at the destination, is served from stock or recorded as backlog, and triggers replenishment through the network. By construction, the state vector tracks per-node on-hand inventory together with outstanding orders, in-transit shipments, and a smoothed demand estimate across the network, so the dynamics close as a Markov chain on a high-dimensional but tractable state space whose transition kernel acts linearly on the empirical distribution of the state. The released data reproduces the bullwhip effect—the standard empirical signature of supply-chain dynamics—at magnitudes consistent with industry evidence, and three conservation laws structurally encoded in the Markov chain serve as verification tools when users extend the simulator with their own control rules. We release datasets at two catalogue scales ( and ) with a -rollout scenario library at , exhibiting dynamics largely absent from fixed TSF benchmarks: variance amplification, cascading bottlenecks, regime shifts, and cross-channel coupling through shared macro shocks. Zero-shot evaluation of three foundation models (Chronos, Moirai, TimesFM) against three in-domain-trained baselines (ARIMA, ETS, PatchTST) spans four target observables: demand and three network-dependent quantities: backlog, fill rate, and edge utilization. Comparison with ETTh1, Electricity, and Weather under a common protocol shows that the ISOMORPH targets introduce forecasting regimes that differ from those represented in standard real-world TSF benchmarks. This positions ISOMORPH as a complementary, regenerable logistics-domain benchmark with explicit network-state targets and controlled variation in the underlying data-generating system. The same pairing produces forecast confidence bands across an ensemble of scenario configurations, providing a form of forward UQ from parameter uncertainty. This demonstrates how ISOMORPH can serve as a configurable digital-twin testbed for evaluating and developing foundation models' uncertainty quantification under controlled variation in the underlying system configuration. Code (MIT) is released as anonymized supplementary material.

Primary area:
datasets and benchmarks
Keywords:
Benchmark Datasets, Simulation-based Data Generation, Digital Twin, supply-chain simulation, Time Series Forecasting, Foundation Models

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_UfQXOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper aims at a simulator and forecasting benchmark. It provides clear support for simulation coherence (RQ1) and model comparisons under controlled conditions (RQ4). For behavioral fidelity (RQ2) and benchmark novelty (RQ3), it shows relevant supply-chain patterns and different forecasting challenges, but further evidence is needed to establish how well these reflect actual operations and what they add to existing benchmarks. The scenario analysis (RQ5) illustrates how forecasts vary with system settings, though this alone does not establish reliable uncertainty estimates. Its applicability to real-world operational decisions remains uncertain. The reported results do not show operator-specific calibration or whether using these forecasts reduces stockouts, costs or service failures. This limits the practical claims, but the simulator can still be a useful research testbed.

Critical strengths
  • Mechanistically specified simulator with internal consistency checks. The released setup models a 13-node network with 3 sources, 9 warehouses, and 16 edges. Its state tracks inventory, outstanding orders, shipments, and demand estimates. The paper proves three pathwise conservation laws (Proposition C.1; §§2–3.2, App. C).
  • Benchmark includes network-state targets and varied operating scenarios. The release includes 7,300-day trajectories, C=50 and C=200 datasets, and a 33-rollout scenario library. Forecast targets include demand, backlog, fill rate, and edge utilization; scenarios vary demand, inventory, capacity, lead times, and disruptions (§4; App. G).
  • Broad, target-specific forecasting evaluation. The paper compares three zero-shot foundation models with ARIMA, ETS, and PatchTST across four targets and multiple forecast horizons. For example, TimesFM performs best among the foundation models on demand, while Moirai performs best on backlog (Table 2; Table 6).
Critical weaknesses
  • Limited validation against real operations. The SupplyGraph comparison covers one company, 41 SKUs, and 221 days. Only 15 of 39 reported statistics fall within the simulator’s scenario range; the paper also notes differences in demand intermittency and bullwhip ratios (App. K).
  • Narrow operational assumptions and release scope. The released datasets use one 13-node network. Runs hold the control policy fixed; the base configuration assumes unbounded source supply and deterministic edge transit. The paper identifies alternative replenishment policies and other extensions as future work (App. B; App. F).
  • Uncertainty results do not establish decision value. The 20-scenario ensemble is a designed set, not a probability distribution over system conditions. Backlog and fill-rate intervals are often poorly calibrated, and the paper does not test whether forecasts improve inventory, service, or cost outcomes (App. E, Tables 7–8; App. F).
Questions for authors

Clarification requests

  • Beyond the output-statistic comparison with SupplyGraph, did any operator data inform the network structure or parameter choices, such as capacities, lead times, or replenishment rules? This would affect how strongly the released simulator can be interpreted as representative of real operations.
  • Were the disruption responses compared with observed operational disruption records? The paper’s comparison uses a synthetic dataset and different service measures; evidence from real events would affect the assessment of disruption realism.
  • Could the authors clarify whether the reported results were generated with the supplementary scripts as provided? In particular, the edge-utilization code appears to use shipment counts instead of item volumes and to assign capacity use to different days than the simulator; the MASE confidence intervals and TimesFM CRPS calculations also appear inconsistent with the stated metrics. Corrections or clarification of their effect on reported results would affect reproducibility and confidence in those results.

Suggestions for future work

  • Validate the simulator against multi-echelon operational and disruption data, including service and inventory outcomes.
  • Evaluate whether forecasts or scenario analysis improve decisions—such as inventory, service levels, or costs—and compare them with current operational methods.
  • For probabilistic UQ claims, define and justify scenario probabilities; otherwise, present the results explicitly as coverage across a designed scenario set.
Reviewer_q8MsOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents ISOMORPH, a supply chain simulator for generating time series datasets. It models inventory, demand, transportation, and replenishment in a multi-echelon network. The authors release datasets with different configurations and evaluate several forecasting foundation models. They also study uncertainty quantification under different scenarios.

Critical strengths
  1. The paper addresses an underexplored domain in time series forecasting. The simulator is configurable and provides network-state variables beyond demand, such as backlog and edge utilization.

  2. The simulator includes several useful dynamics, such as inventory replenishment, transportation delays, and capacity constraints. The released code and datasets can be useful for future research.

Critical weaknesses
  1. Limited validation with real-world data. The authors compare the generated data with SupplyGraph, but only 15 out of 39 statistics fall within the simulated ranges (Appendix K). Some important properties, such as demand variability, autocorrelation, and bullwhip ratios, show large differences. I am not convinced that the generated data can represent real supply chain dynamics well.

  2. Limited multivariate modeling and evaluation. The simulator includes shared demand shocks and transportation constraints, but several important dynamics are missing or simplified, such as adaptive inventory control, shared production resources, and state-dependent transportation delays. Also, the experiments mainly evaluate separate forecasting targets, without showing how much the models benefit from cross-variable dependencies. This limits its value as a multivariate forecasting benchmark.

Questions for authors
  1. Given the differences from SupplyGraph, how can the authors justify that ISOMORPH generates realistic supply chain time series? Can the simulator be calibrated using real data?

  2. Can the authors provide experiments showing whether modeling cross-variable dependencies actually improves forecasting performance, for example, by comparing multivariate and channel-independent forecasting?

Reviewer_CkEhOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces ISOMORPH, a configurable supply-chain simulation framework for generating synthetic time-series datasets and benchmarking forecasting models. The framework models multi-echelon logistics networks through a discrete-time Markov formulation that captures inventory levels, outstanding orders, in-transit shipments, demand, and replenishment decisions. The authors establish three conservation laws to ensure consistency of the simulated material flows. Using a multi node logistics network, the framework generates datasets with multi-product categories and multiple scenarios capturing supply-chain dynamics, including the bullwhip effect, demand fluctuations, transportation bottlenecks, and cascading disruptions. The authors evaluate three pretrained time-series foundation models alongside traditional and deep-learning forecasting baselines on demand, backlog, fill rate, and edge utilization. They further investigate uncertainty quantification under variations in simulation parameters. The experimental results suggest that the generated datasets introduce forecasting challenges not adequately represented in conventional benchmarks, positioning their approach as a complementary, reproducible benchmark for supply-chain forecasting research.

Critical strengths
  1. Existing time-series forecasting benchmarks largely focus on demand observations without explicitly capturing the underlying logistics network. ISOMORPH addresses this limitation by providing a configurable, multi-echelon simulator that jointly models demand, inventory, transportation, replenishment, and network-level dependencies. The release of complete network-state trajectories, together with interpretable parameters and modular components, constitutes a potentially valuable resource for developing and evaluating forecasting methods under controlled operational conditions.

  2. The paper provides a well-structured simulator with explicit state-space representations, pathwise conservation laws, and configurable operating scenarios. The conservation laws offer useful consistency checks, while the demonstrated bullwhip effects, cascading disruptions, and bottleneck dynamics illustrate the framework's ability to generate meaningful supply-chain behaviors. The open-source release and extensive documentation further strengthen its potential utility as a research benchmark.

From a personal standpoint, I have always waited for a dataset like this. It is realistic enough and addresses a critical gap in the available benchmarks. Proprietary/Contractual obligations restrict industry from sharing their operational data, and thus it has led to a serious scarcity of available supply chain (time-series) data. This simulator is a first step in breaking that negative trend. People working on time-series modeling, root cause analysis, supply chain analytics, time-series causal discovery, etc., can benefit significantly from such a simulator.

Critical weaknesses
  1. The proposed framework does not adequately establish its characterization as a digital twin. While ISOMORPH provides a configurable simulator of multi-echelon supply-chain operations, the paper does not demonstrate the defining physical–digital integration typically associated with digital twins. Specifically, there is no demonstrated mechanism for ingesting real-time IoT or operational measurements, synchronizing simulated states with those of an actual supply-chain network, or enabling bidirectional feedback between the digital and physical systems. The released framework instead generates synthetic trajectories using predefined stochastic processes and operational rules. Consequently, it is unclear what distinguishes ISOMORPH from a conventional discrete-time stochastic supply-chain simulator.

  2. Despite emphasizing multi-echelon dependencies and network-state information, the forecasting evaluations primarily employ univariate predictions without explicitly leveraging the available network topology, inventory relationships, or transportation constraints. Consequently, the experiments demonstrate forecasting difficulty on synthetic time series but do not establish whether the proposed network representation provides meaningful advantages for developing supply-chain forecasting models.

Questions for authors

Please address the weaknesses.

Comments on the paper

1 comment

Thank you for your response. All of my remaining concerns are addressed. I will retain my positive score.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.