Machine Unlearning for Industrial AI Models

Industrial AI models are increasingly used to replace simulation tools for predicting key performance indicators (KPIs). They offer faster predictions but may introduce errors that can affect critical decisions. Due to the complexity of industrial data, it is often difficult to identify the input conditions under which a model performs poorly. Clustering is therefore used as a preliminary step to detect groups of demands with high prediction errors. This paper investigates a selective machine unlearning strategy applied to these weak groups. Instead of retraining the model from scratch or discarding large portions of data, the approach focuses on removing or reweighting only harmful patterns and preserving useful information. Experiments on a large industrial dataset show that this strategy reduces errors in weak groups and maintains stable performance across the remaining data. This improvement supports the use of AI models as reliable surrogates for industrial decision-making.