Research Desk
Selected publications and papers related to AI forecasting, planning, digital twins and supply-chain risk. Source findings are separated from practical commentary.
Check the linked source for its full methods, limitations and current publication status before relying on it. A preprint has not necessarily completed peer review. Research findings do not guarantee the same result in every organization.
A collaborative analytics framework for AI-based supply-chain demand forecasting
Source finding: The study compares a manufacturer statistical baseline, a manufacturer-only machine-learning forecast and a joint forecast using retailer warehouse-outbound data. In its manufacturer–retailer case, the joint model reduced test-set absolute error relative to manufacturer-only AI; the authors report the size and confidence interval in the source.
Practical commentary: The result is a useful prompt to test whether downstream sell-out or inventory data improves decisions. It does not mean every partner should share every data field; governance, incentives, data latency and privacy need to be designed alongside accuracy.
Open source ↗Agentic digital twins: bridging model-based and AI-driven decision support
Source finding: The paper develops a conceptual framework for agentic supply-chain digital twins, connecting agents, simulation and model-based optimization, and discusses industry cases and a research agenda.
Practical commentary: Use this as a framework for architecture questions: which decisions remain governed by optimization, where an agent can help explore options, and how human approval and audit trails fit. Conceptual frameworks are not evidence of a guaranteed performance gain.
Open source ↗Multi-scale global supply-chain spatiotemporal demand forecasting using Mamba
Source finding: The paper proposes a multi-scale forecasting model combining a state-space backbone, evolving spatial relationships and event-aware inputs, then compares it with selected forecasting baselines in its reported experiments.
Practical commentary: A complex model can be a research candidate where cross-region relationships and long histories exist. Validate performance with time-based and region-based holdouts, simpler baselines and forecast-value measures before operational use.
Open source ↗Large Language Models for Supply Chain Decisions
Source finding: The authors describe using LLMs to help planners understand optimization outputs, explore scenarios and update mathematical models. The source reports faster decision workflows in its described applications.
Practical commentary: Treat this as a design direction, not proof that an unconstrained chatbot can safely make planning decisions. Ground outputs in trusted planning data, keep approvals explicit, and measure decision time and quality in a bounded pilot.
Open source ↗LLMs for Supply Chain Management
Source finding: The authors introduce a retrieval-augmented supply-chain LLM and report experiments on standardized SCM tasks and supply-chain games. The paper reports gains from retrieval augmentation in its tested setup.
Practical commentary: The evaluation is task- and environment-specific. Before adoption, test retrieval accuracy, unsupported answers, data freshness and decision performance against a non-LLM baseline using your own workflows.
Open source ↗Understanding the Supply Chain and Risks of Large Language Model Applications
Source finding: The paper assembles a dataset of LLM applications and their dependencies, then analyzes dependencies and vulnerability reports across models, datasets and libraries.
Practical commentary: For a business AI deployment, maintain an inventory of model, data, package and hosting dependencies. The paper studies the LLM application ecosystem; it is not a direct measurement of supply-chain planning outcomes.
Open source ↗Machine Learning for Master Production Scheduling: Combining probabilistic forecasting with stochastic optimisation
Source finding: The study combines machine-learning probabilistic demand forecasts with stochastic optimization for capacity-constrained master production scheduling. It evaluates 17 datasets covering 303 products and discusses flexible demand distributions.
Practical commentary: This supports testing forecast distributions inside planning decisions rather than judging models only by point error. The dataset and planning setup bound how far the reported results transfer to another operation.
Open source ↗