In simulation-driven design optimisation, the high-fidelity solver dominates the per-evaluation cost, and generative search compounds the load. This paper proposes PISCO (Physics-Informed Surrogate-Critic Optimisation), a workflow that targets two performance criteria jointly — the number of high-fidelity solver calls and the quality of the obtained design — through a strict separation of four components: a physics-informed surrogate of the physical fields, an analytical scoring function reusing the solver's own discrete quadrature, a low-dimensional design generator (analytical for parametric problems, a decoder learned from initial fields for free design spaces), and an active high-fidelity re-evaluation step. Only the surrogate learns inside the loop; surrogate and scoring function form a differentiable critic against which the generator is optimised. Two examples in linear elasticity demonstrate the approach. On a parametric shape problem (cantilever plate with an elliptical cut-out), PISCO matches the compliance of differential evolution at roughly an order of magnitude fewer high-fidelity calls, and outperforms both Bayesian optimisation and the Covariance Matrix Adaptation Evolution Strategy (CMA-ES). On a free-topology problem (L-bracket), it undercuts the Solid Isotropic Material with Penalisation (SIMP) baseline by 13.8% in compliance, well outside SIMP's seed-to-seed variability. An ablation indicates that on these problems a linear Principal Component Analysis (PCA) decoder performs as well as a learned non-linear Variational Autoencoder (VAE). Transfer to other physical domains is supported by the architecture but not validated in the present work, which restricts itself to linear 2D compliance optimisation. Preprint of an article submitted to a peer-reviewed journal.
Andreas Arnold (Sun,) studied this question.