Differentiable Programming is a hybrid perspective between statistical learning and numerical optimization. This mix is uniquely suited to deliver predictive optimization for supply chains, which also require a mix of learning and optimization.