WACA-UNet: Weakness-Aware Channel Attention for Static IR-drop Prediction

Accurate spatial prediction of power-integrity issues such as IR drop is critical for reliable VLSI design, yet traditional simulation-based solvers do not scale efficiently to modern designs. We formulate static IR drop estimation as a pixel-wise regression problem over heterogeneous multi-channel layout maps. However, existing CNN-based surrogates implicitly assume uniform importance across input channels, overlooking the inherent imbalance among heterogeneous physical features. As a result, dense feature maps (e.g., hypothetical IR drop (HIRD) maps) tend to dominate channel attention, while sparse but physically critical cues—such as wire and via resistance—are suppressed, leading to degraded hotspot detection. To address this issue, we propose Weakness-Aware Channel Attention (WACA), a recursive two-stage gating mechanism that explicitly enhances underutilized yet informative channels. Unlike conventional attention mechanisms that primarily amplify already dominant features, WACA reuses the weights of standard channel-attention modules (e.g., SE, CBAM) to adaptively rebalance channel contributions without introducing additional learnable parameters. Integrated into a ConvNeXtV2-based Attention U-Net, WACA produces more balanced feature representations and improves localization robustness. On the ICCAD-2023 benchmark, WACA-UNet achieves an MAE of 0.0524 and an F1 score of 0.778, outperforming the contest winner by 61.1% and 71.0%, respectively. These results suggest that modeling channel-wise heterogeneity can serve as a useful inductive bias for hotspot-sensitive IR-drop prediction under the evaluated benchmark setting.