Membrane separation is a well-established technology that is already widely used in industry for liquid and gas treatment. Its main advantages are compactness, modularity and flexibility, low production costs, and low energy requirements, since energy consumption is mainly limited to the compression or vacuum equipment. Nevertheless, to obtain these benefits, the membrane separation system must be carefully designed, that is, an optimal (or high-quality) configuration and its operating conditions must be determined. This problem can be formulated as a Mixed-Integer Non-Linear Programming (MINLP) model that includes cubic equality constraints arising from the discretization of the differential equations governing membrane physics. The resulting problem is difficult to solve with off-the-shelf solvers; we therefore designed an ad hoc solution strategy that provides high-quality solutions for several relevant industrial case studies.
Certifying the adversarial robustness of a neural network amounts to proving that no perturbation within a given input region can change its prediction. As the certification problem is NP-complete, it is in practice mostly solved with branch-and-bound methods using a bound propagation relaxation at each node. The tightness of the bounds is crucial as it determines the size of the branching tree, which is exponential in the number of unstable neurons in the worst case. Semidefinite programming (SDP) relaxations give tighter bounds than bound propagation ones but suffer from a large computational cost. To address this limitation, we propose a dynamic bundle method based on a Lagrangian relaxation that enables a significant speed-up compared with existing SDP-based verifiers.
We study mixed-integer nonlinear programs in which on/off decisions switch nonlinear equations, the variables split into blocks linked by a few continuous quantities, and storage couples consecutive periods. Such problems arise in operating water, gas and heating networks. We propose a two-part method. First, each block is modeled with piecewise-linear relaxations, and consensus ADMM coordinates the blocks. Inside each block, a feasibility funnel tests every candidate schedule. It discards infeasible schedules with no-good cuts, and a schedule is accepted only if the exact model yields a point that passes an independent checker. When a single block holds the whole objective, the same loop can prove optimality. Second, a neighborhood search over the blocks and a warm-started monolithic solve improve the solution and certify it when possible. On a standard pump-scheduling benchmark of 40 instances, the method finds a feasible schedule for every instance within one hour and proves 19 of them optimal within three hours. We also show that the best published results for one network of the benchmark come from a heuristic that switches pumps within a period, so they are not solutions of the fixed-period model.
Decision-Oriented Approximation Frameworks For Wind Farm Layout Optimization
Many real-world optimization problems involve computationally expensive objective functions, making direct optimization impractical. Approximation techniques can reduce this burden, but their suitability for optimization is often overlooked. This work investigates optimization-oriented approximation frameworks for expensive blackbox optimization, with a focus on Wind Farm Layout Optimization (WFLO). It studies how approximation, optimization, and data acquisition can interact to efficiently guide the search toward high-quality solutions. Particular attention is given to iterative correction strategies, including additive and multiplicative corrections, where an existing approximation is progressively adjusted using expensive evaluations within the optimization loop [1]. The research investigates the trade-off between approximation accuracy and optimization tractability, together with sampling strategies and stopping criteria, with the aim of improving the approximation where it is most relevant to the optimization process. The long-term goal is to develop optimization-approximation frameworks that make effective use of limited expensive evaluations and scale to challenging problems such as large-scale WFLO x