Several real-world challenges can be represented as large optimization problems with hard constraints (e.g., physical laws, business logic). Although deep neural networks may outperform traditional optimizers as “approximate solvers” for unconstrained optimization problems, they often generate infeasible solutions when hard constraints exist. In this context, Donti et al. propose DC3, a framework that embeds differentiable operations into neural network training to enforce feasibility. They implement the DC3 algorithm for both convex and non-convex optimization tasks and show how it might be applied to optimize power flows on the electrical grid.