Manufacturing-Aware Benchmark for Additive Manufacturing Design Optimization

Additive manufacturing (AM) allows for complex, lightweight designs but introduces process-specific constraints that traditional benchmarks often overlook. We introduce a manufacturing-aware benchmark designed for PLA/FDM that includes nine design variables covering geometry, process, and orientation; six competing objectives (material weight, build time, structural efficiency, support volume, surface quality, and total cost); and 20 physics-based manufacturability constraints grouped into eight categories. The framework combines feasible-set exploration using Latin Hypercube Sampling (LHS) (n=10,000), multi-method sensitivity analysis, sample-based Pareto front extraction, and standardized algorithm evaluation under equal budgets for differential evolution (DE), particle swarm optimization (PSO), Bayesian optimization (BO), simulated annealing (SA), and a hybrid algorithm (HA). Statistical comparisons employ nonparametric tests and effect sizes, while multi-objective performance is analyzed using dominance relations and hypervolume. Sample-based feasibility is roughly evenly split (49.8% feasible), with violations mainly caused by print speed and thermal/support limits; print speed, support density, and build orientation are the main factors affecting feasibility and performance; build time and cost are nearly collinear; and improving surface quality generally requires slower speeds and thinner layers. Feasibility-based Pareto analysis shows significant reductions in weight and support volume while maintaining strength, revealing clear trade-offs and knee solutions. Boundary checks show 96.7% interior solutions, indicating the problem is not boundary-driven, and the few layer-height violations are minor and correctable. Algorithmically, scalarized results favor the hybrid approach (with PSO ≈ DE), while multi-objective indicators highlight complementary strengths in front coverage. Performance differences are highly significant (Kruskal–Wallis H = 179.23, p 0,001; η² = 0.717; ω² = 0.711); the hybrid achieved the best scalarized outcome with tight uncertainty (95% CI −4.9163, −4.7979, CV = 4.48%), and BO attained the highest mean hypervolume (0.847). The benchmark offers a reproducible, realistic testbed that bridges algorithmic evaluation with manufacturing feasibility and can be expanded with higher-fidelity models and materials.