Multi-Modal Fusion for Yield Optimization: Integrating Wafer Maps, Metrology, and Process Logs with Graph Models

Yield optimization in advanced manufacturing rarely proceeds as a tidy pipeline; it arises from the gradual convergence of evidence across spatial wafer patterns, multivariate metrology, and asynchronous process and equipment events that interact in ways that are only partially observable. Prior studies often separate these modalities, assigning convolutional encoders to wafer maps, sequence models to metrology, and template based encoders to logs, an arrangement that can perform well locally yet struggles to sustain cross-modal alignment or to reason over the hierarchy that links defects to steps and equipment. Building on these observations, we introduce a manufacturing semantics oriented framework that embeds lots, wafers, dies, steps, equipment, and recipes in a heterogeneous graph, and uses cross modal attention gating to reconcile image, time series, and event representations while performing relation aware message passing. The research was not frictionless; time synchronization required iterative windowing, spatial normalization exposed orientation drift, and naive imputation inflated variance in rare steps, which motivated temperature controlled gating and a lightweight contrastive warm-up. On two production lines the approach improves, to some extent, standard classification metrics and stabilizes top k attribution under feasible latency. Alternative explanations remain possible, including benefits from stricter leakage control or product specific distributions. The work makes explicit the structural link among defects, process, and equipment, and points toward auditable, engineer actionable analytics; further research is needed on long term stability, cross site generalization, and the joint optimization of accuracy, cost, and energy.