On a lens assembly line, the incoming material is given, but each component can be rotated before assembly, so the key decision is which feasible angle to apply to each arriving batch. Historical shop-floor data are sparse and unbalanced: operators reuse angles that worked before, and failed trials are rarely recorded. The archive reflects habits rather than a mapping from material to best setting.

We propose a strategy in which the recommended angle depends on each batch's material labels and measurements. Batches are grouped by material label, balancing distinct groups against enough variety within each to support a model. Within each group, a model treats the angle as a circular variable whose effect varies with the material. The grouped models are used only if they predict better than a single global model; the selected model then recommends the feasible angle with the highest predicted passing rate.

On a real lens line, validation selected seven material groups. The grouped models predicted more accurately than a global model and a regression tree, and their recommendations raised the predicted overall passing rate by about 7.7%, with gains of 5–10% within groups.

This case study shows how material-dependent angle selection turns habit-driven shop-floor data into actionable batch-level recommendations.