Here we propose a reinforcement learning enabled physical autonomous manufacturing system (AMS) that is capable of learning the manufacturing process parameters to autonomously fabricate a complex-geometry artifact with desired performance characteristics. The poor sample efficiency of traditional RL algorithms challenges real-world manufacturing decision making due to a high variable cost from raw material, machine utilization, and labor costs. To make decision making sample efficient, we propose to leverage a first-principles based source task for training, transfer effective representations from trained knowledge, and then use these representations to interact with the physical system to learn a probabilistic model of the target reward function. We deploy this idea to a novel dataset obtained from a custom physical AMS machine that can autonomously manufacture phononic crystals, a complex geometry artifact with spectral response as performance characteristic. We demonstrate that our method uses as low as 25 artifacts to model the interesting part of the target reward function and find an artifact with high reward. This task typically requires manual design of phononic crystals and extensive empirical iterations on the order of hundreds.