A Probabilistic and Explainable Machine Learning Approach for Adaptive Supply Chain Optimization under Disruptions

Abstract

Today's supply chains are facing greater risks from fluctuating demand, transportation problems, geopolitical issues, and changing customer consumption habits. These dynamic uncertainties are often not considered by classic deterministic optimization algorithms, leading to suboptimal decisions and lack of operational resilience. To overcome those challenges, this paper introduces a machine learning framework for optimization of a supply chain with probabilistic models and interpretability in case of disruptions. The proposed framework brings together the fields of probabilistic forecasting, risk assessment and robust optimization within a single decision-support pipeline. The uncertainties of demand and supply are well quantified using Advanced machine learning models which produce both point forecasts and calibrated prediction intervals. These forecasts are based on multi-domain data, such as transactional, logistics, financial and behavioural data, to make more accurate and robust forecasts. The probabilistic outputs are used in stochastic and robust optimization models to optimize inventory management, production scheduling, and distribution planning in uncertain operating environments. Explainable Artificial Intelligence (XAI) methods like SHAP and LIME are used to understand model outputs and the most important factors affecting optimisation decisions to improve transparency and facilitate human-centric decision-making. The proposed framework is tested with extensive simulation studies which include demand surges, supply delays and network disruptions. Experimental results show that the services achieved with the proposed approach have a better service level, cost efficiency, robustness, and operational resilience than traditional deterministic services and non-interpretable services. Moreover, the framework is very adaptive to the changes of operating environments and can be applied to actual supply chain. The proposed methodology is scalable, interpretable, and decision-oriented and integrates probabilistic machine learning, explainable artificial intelligence and strong optimization for intelligent, resilient, and data-driven supply chain management under uncertainty, which will contribute to next-generation adaptive supply chain systems.