<p>Achieving optimal performance in all-optical logic gates based on Metal–Insulator–Metal (MIM) plasmonic waveguides remains challenging due to their complex, multi-parameter design space of these devices and the high computational overhead of conventional simulation-based optimization. In this paper, we present a surrogate-assisted multi-parameter optimization framework (SAMPO) to effectively tackle these constraints. The plasmonic half-adder is chosen as the reference structure because it is a basic part of computing units that combines two separate logic operations (AND and XOR) into one compact circuit. Moreover, as the half-adder serves as the foundation of the full-adder and provides the computational core of integrated photonic circuits, it is an exemplary option for verifying the proposed optimization methodology. In SAMPO, a feed-forward neural network trained on data generated by Finite Difference Time-Domain (FDTD) simulations serves as a fast surrogate model to map structural parameters to optical responses. The model is integrated with the Starfish Optimization Algorithm (SFOA) to effectively traverse the design space. The optimized design attains contrast ratios (CRs) of 12.43 dB for the AND gate and 10.79 dB for the XOR gate inside a compact area of 0.14 μm². The demonstrated surrogate-driven optimization method validates its robustness and computational efficiency.</p>