Sequential KDE‑guided zero-shot regression under process changes across materials

• Sequential self-training enables zero-shot regression under domain shift. • KDE score thresholds select source-process samples near target-process samples. • A generator adds pseudo-target-process variables. • A regressor learns from observed data and pseudo-augmented data. • Automatic reuse of informative source data reduces the need for new experiments. This study presents a zero-shot regression framework that enables product quality prediction for a target process–material combination with no operating history. The framework facilitates scale-up and line transfer while minimizing experimental effort. A kernel density estimator sequentially selects source-process samples whose material features and product quality are similar to those on the target process, so that only relevant source data are reused. For each selected source-process sample, we predict operating conditions on the target process that would realize the same product quality for the same material, assuming that the material properties and product quality remain the same across processes. We then build a prediction model on a combined dataset consisting of the observed target-process samples and the selected source-process samples paired with their predicted target process operating conditions. A case study with two processes and seven materials demonstrates that the proposed method achieved consistent prediction accuracy; the median root mean squared error between prediction and measurement was 0.066 when using only target-process samples, 0.086 when combining target- and source-process samples without sample selection, and 0.055 with the proposed framework. The framework can successfully predict the behavior of new materials on the target process without additional experiments while suppressing negative transfer by automatically selecting and reusing existing data.