SNAPO (Smooth Neural Adjoint Policy Optimization) is a framework for optimal control that embeds a neural policy inside a known, differentiable simulator. It replaces hard constraints with smooth approximations, enabling exact gradient computation via adjoint methods. The key innovation is that a single backward pass simultaneously trains the policy and computes sensitivities of the objective with respect to all policy parameters and inputs, at a cost proportional to one reverse pass regardless of the number of sensitivities. This makes SNAPO highly efficient for both policy optimization and sensitivity analysis.

The framework is demonstrated on three real-world domains: natural gas storage, pension fund asset-liability management (ALM), and pharmaceutical manufacturing. For gas storage, training completes in under a minute and yields 365 forward curve sensitivities at no extra cost. For pension ALM, SNAPO achieves a 6.5x to 200x speedup in sensitivity computation compared to traditional bump-and-revalue methods. In pharmaceutical manufacturing, it computes cross-unit sensitivities through a 4-unit process chain, producing 20 ICH Q8 regulatory sensitivities in just 74.5 milliseconds. These results highlight SNAPO's ability to combine fast training with comprehensive sensitivity analysis, making it a powerful tool for optimal control under uncertainty.