Surface litter manure can significantly increase the risk of diseases in broiler brooding houses, such as coccidiosis and colibacillosis. Indoor air quality can be deteriorated, due to the release of ammonia, hydrogen sulfide, and methane. However, existing double-crank shoveling-throwing mechanisms of surface manure cleaning have suffered from low efficiency, performance, and excessive vibration, because the key structural parameters are determined empirically without systematic multi-objective optimization. In this study, a multi-objective optimization was developed for the double-crank shoveling-throwing mechanism using an improved NSGA-II algorithm. Thereby, better performance was achieved to improve the shoveling efficiency, energy consumption, and stability. A kinematic and dynamic model of the double-crank shoveling-throwing mechanism was first established to reveal the influence of structural parameters on the shoveling trajectory and force behaviors. A coupled ADAMS-EDEM simulation model was then constructed to simulate the interaction between the mechanism and manure particles. As such, 60 sets of variable samples were generated after the optimal Latin hypercube sampling. A Gaussian process regression (GPR) surrogate model was constructed to map the relationship between six variables and the shovel throwing quality Q. The relative error between the prediction and simulation was 3.85%, indicating high prediction accuracy. An improved NSGA-II algorithm was proposed to overcome the limitations of the standard NSGA-II algorithm—namely, significant dimensional differences among the three objectives, highly nonlinear parameter-performance mapping, and discontinuous parameter space. Three improvements were introduced: (1) an adaptive normalization mechanism to eliminate dimensional effects; (2) a hybrid optimization framework with global search (NSGA-II) and local refinement (sequential quadratic programming, SQP) for the high convergence accuracy; and (3) an improved crowding distance and solution selection mechanism for the distribution uniformity of the Pareto front. The improved algorithm was compared with standard NSGA-II, NSGA-III, MOPSO, and MOEA/D, according to three performance indicators: Inverted Generational Distance (IGD), Spacing, and Hypervolume (HV). The results showed that the improved NSGA-II algorithm significantly outperformed the rest. Specifically, the IGD value decreased by 71%, 68%, and 64%, respectively, compared with standard NSGA-II, MOPSO, and MOEA/D. Spacing value decreased by 26%, compared with standard NSGA-II, whereas, the HV value increased by 16%. The better performance was achieved in the high convergence, more uniform distribution, and higher coverage of the true Pareto front. The optimal compromise solution was selected from the Pareto set using the entropy-weighted TOPSIS. The optimal parameters were recommended: l₁=70.4 mm, l₂=88.5 mm, l₃=100.2 mm, l₄=30.3 mm, b=99.7 mm, β=37.4°. A prototype was manufactured for the double-crank shoveling-throwing device, according to the optimal parameters. Field experiments were conducted in three repetitions in a brooding house in Zhouzhi County, Xi’an, Shaanxi Province, China, in December 2025. The experimental conditions were as follows: Litter layer with a thickness of 50–60 mm, chicken manure layer thickness of 10-20 mm, and manure moisture content of 30%–35%. The results demonstrated that the optimal mechanism improved shoveling efficiency by 138.24%, whereas the driving torque and the angular acceleration peak at the shovel end were reduced by 35.12%, and 35.60%, respectively. In addition, the residual rate of surface manure decreased from 18.45% to 8.13%, with a reduction of 55.99%, indicating significantly improved cleaning quality. A multi-objective optimization framework was provided for the double-crank shoveling-throwing mechanism using ADAMS-EDEM simulation, GPR surrogate modeling, and an improved NSGA-II algorithm. The dimensional differences and high nonlinearity were avoided for the low computational cost after engineering optimization. The improved NSGA-II algorithm demonstrated superior convergence and distribution performance, compared with mainstream multi-objective algorithms. The optimal mechanism was achieved to balance shoveling efficiency, energy consumption, and operational stability. The findings can offer a complete technical pathway to enhance the performance of hinge-type multi-bar mechanisms, particularly for manure cleaning equipment in the poultry industry.