On-chip photonic computing increasingly relies on large MZI arrays, where silicon Mach–Zehnder modulators constrain accuracy and efficiency due to phase-shift linearity, insertion loss, drive efficiency, and uniformity. Conventional parameter-sweep design flows struggle to balance multiple objectives (insertion loss, Vπ·L, bandwidth, phase-shift linearity) under limited computing resources. To address this, we propose a compact inverse design framework that reduces the dimensionality of parameters by pruning the design space to performance-sensitive variables, such as waveguide structure, electrode length, and coupler geometry. With 50-Ω impedance matching and 10 GHz bandwidth, a genetic algorithm achieves multi-objective optimization, simultaneously improving modulation efficiency, insertion loss, and phase-shift linearity. System-level simulation on small MZI arrays show reduced normalized computation error in matrix-multiplication tasks. In addition, tolerance analysis under laser linewidth, wavelength misalignment, and thermal drift confirms robustness. Therefore, our scheme provides a practical, scalable solution for silicon MZM inverse design in photonic computing.