Assessment of material supply risks in make-to-order manufacturing based on the TabPFN neural network model

Abstract

Relevance: Growing global instability and the vulnerability of supply chains lead to increasing risks in material procurement for industrial enterprises. The problem of forecasting and assessing material supply risks is particularly relevant in make-to-order manufacturing and high-mix, low-volume production. Such manufacturing systems, characterized by product diversity and low buffer inventories, are highly sensitive to delays in material delivery. Traditional risk management systems, designed for stable mass production, are often unable to react quickly enough to the dynamics of order-driven, high-mix environments. The purpose of this study is to improve risk assessment of production materials procurement through the use of machine learning for prediction of material delivery delays and development of an risk evaluation method on the base of such forecasts. To achieve this goal, the following tasks were solved: a model for forecasting delivery delays was developed and integrated supply risk index, based on probability of a delays and their impact introduced.  The interpretability of the classifier model is also investigated, and an example of its integration into the resource planning system of the customer enterprise is provided. Methods: To obtain delay predictions, a multi-class classifier based on the state-of-the-art TabPFN model is used. The TabPFN model is based on a transformer neural network architecture adapted for tabular data analysis; a set of historical procurement data obtained from the customer's information systems was used for its training. Results: In the study, the TabPFN-based classifier model demonstrated a high degree of result reliability even when using default parameters, ordinal consistency of results, and a low tendency to underestimate supply delays, which indicates the obtained model's orientation toward risk minimization. The proposed risk assessment methodology allows for considering the impact of the forecasted supply delay on production depending on the current production program and warehouse inventories. Novelty: Combining multi-class delay classification based on business-defined thresholds with the state-of-the-art TabPFN model enables accurate delay forecasting even on limited datasets. Another novel aspect is the proposed approach to calculating a comprehensive risk index, which combines predicted delay probabilities with an evaluation of their impact on the production schedule. The combination of machine learning-enabled delay prediction and the proposed risk index allows for automated risk assessments on a near-real-time scale. Conclusions: The research demonstrates that the application of machine learning methods can enhance the accuracy and timeliness of supply risk assessment, enabling near-real-time monitoring and more informed decision-making regarding the provision of materials for production.