With the rapid advance of machine learning techniques and the increased availability of high-speed computing resources, it has become possible to exploit machine-learning technologies to aid in the design of photonic devices. In this work we use evolutionary optimization algorithms, machine learning techniques, and the drift-diffusion equations to optimize a modified uni-traveling-carrier (MUTC) photodetector for low phase noise at a relatively low bias of 5 V. We compare the particle swarm optimization (PSO), genetic, and surrogate optimization algorithms. We find that PSO yields the solution with the lowest phase noise, with an improvement over a current design of 4.4 dBc/Hz. We then analyze the machine-optimized design to understand the physics behind the phase noise reduction and show that the optimized design removes electrical bottlenecks in the current design.