Production ramp-up is the period of time during which a manufacturing process is scaled up from a small laboratory-like environment to high-volume production. During this scale-up, the firm needs to overcome the numerous discrepancies between how the process is specified to operate as written in the process recipe and how it actually is operated at large volume. The reduction of these discrepancies, a process that we will refer to as learning, will lead to improved production yields and higher output. In addition to its learning effort, however, the firm also attempts to change the process recipe itself, which can be in direct conflict with the learning objective. We formalize this intertemporal tradeoff between learning and process change in the form of a dynamic optimization problem. Our model explains the idea of a "copy-exactly" ramp-up, which freezes the process for some time period, i.e., does not allow for any change in the process. Mathematically, this corresponds to a process improvement policy which delays process changes, thereby exhibiting a nonmonotone trajectory, which we show to be optimal if the initial knowledge level is low, the lifecycle short and demand growth is steep, and learning is difficult.