Crystalline silicon solar cells with the passivated emitter and rear cell (PERC) design are currently the mainstream cell architecture in industry. Due to the rather complicated device structure, it has been challenging to understand how variations in manufacturing tools cause the observed scattering of the cells’ performances. This also makes it difficult to optimize PERC cells in fabrication lines. In this article, we report on a method where we use numerical device modeling, machine learning, and statistics for getting a deeper understanding of how process variations influence device performance. For this, we use seven model input parameters that affect PERC device performance the most and perform about 400 numerical device simulations in an expected range of these parameters. We then trained and validated a machine learning model on these 400 simulations, which serve to describe about 15 000 fabricated PERC cells. Because the IV parameters of each cell can be described with different sets of the seven model input parameters, we define Euclidean surroundings of the IV -parameters of each manufactured cell and analyze the behavior of the input parameters in these surroundings. The proposed method is applied to commercial production of PERC cells and requires only the four IV parameters of each cell, measured at the end of production (no dummy wafers or lifetime samples are needed). Still the method gives concrete information for improving PERC cells with a modest amount of numerical modeling and hence in a very short time. The method is generally applicable also to other solar cells than PERC cells.