Determining optimal inventory replenishment decisions requires balancing the costs of excess inventory with shortage risks. While demand uncertainty has been the focus of stochastic inventory modeling, the effects of transportation cost uncertainty are poorly understood. In practice, transportation modes are prone to disruptions that result in stops and cost increases. While historical disruption data is available, it is difficult for practitioners to understand how replenishment orders must be adjusted. To overcome this gap, we combine mathematical optimization with machine learning to predict cost-optimal replenishment orders using only historical data. The problem is modeled as SIRPDD that minimizes total expected costs. With perfect information, optimal decisions are generated as labels for the supervised learning using features from inventory control and disruption-related information. We propose a new CFML framework that optimizes the costs of applying replenishment policies within hyperparameter tuning instead of the prediction score of the individual decisions. To handle the resulting computational complexity, we develop a genetic algorithm. We present a case study for the SIRPDD with transportation cost uncertainty. This case, based on a chemical company on the river Rhine, considers two suppliers, different lead times, order sizes, and direct deliveries. Relevant features include the inventory position, historical water level, their trends, and predictions. We show that our CFML can reduce costs by 20% compared to the \((s,Q)\)-reorder policy, which represents the industry standard, and 18% compared to classical machine learning frameworks.