Dynamic IR drop is the momentary sag in a chip's supply voltage when many cells switch at once, and it is one of the most expensive things to check during chip design. Machine-learned predictors can replace hours of commercial simulation, but they are trained on very few examples, and those examples are costly to produce. We propose a way to make more training data without trusting a neural network to invent the answers. Each training sample has two parts. The input is a movie of where and when current is drawn across the die. The label is the resulting voltage-drop map. A generative model is trained to produce only the input movie, conditioned on a real design's layout and power-delivery wiring. The label is then computed by a fast numerical solver that applies circuit laws on that same wiring. In short, the generator supplies plausible switching activity, and physics supplies the correct voltages.
This split targets two failures that generative augmentation has shown elsewhere. In medical imaging, synthetic data helps at small mixing ratios but begins to hurt at larger ones, because the downstream model learns the generator's quirks (Synthesizing Diverse Lung Nodules Wherever Massively). In hardware, synthetic samples must also satisfy physical laws exactly, not just look realistic. If the label always comes from a solver, a generated sample can be unusual but its label is never wrong for that input. The claim to test is that this lets synthetic data be mixed in at much higher ratios before it starts to hurt, and that a learned generator of switching patterns adds more useful variety than cheap hand-designed perturbations.
Core Claim and What Is New
The scarce and informative part of a dynamic IR drop sample is the switching pattern over time: which regions draw current together, and in which clock phases. PowerNet showed that the peak transient drop is driven by these time-resolved power maps, and it captured them by taking an element-wise maximum over per-time-step convolutions (PowerNet). PDNNet added a graph model of the power delivery network (the wiring that carries supply voltage from the pins down to the cells), alongside a 3D convolutional branch over current maps (PDNNet). Both are purely supervised on existing simulation labels.
Existing augmentation in this area has three forms. CircuitNet 3.0 rewrites the syntax trees of Verilog source code and reruns full commercial flows, and it reports only static timing and power (CircuitNet 3.0). LMM-IR oversamples real cases twenty times and adds Gaussian noise for static IR drop (LMM-IR). Medical generative models produce the image and its label jointly (MedGen3D).
Our proposal differs from all three in two ways. It generates continuous spatio-temporal current fields rather than code or static maps. It also never lets the generator produce the label. For dynamic analysis the solver integrates the standard network equation of the power grid,
$$G\,v(t) + C\,\frac{dv(t)}{dt} = i(t),$$
where $G$ is the conductance matrix of the extracted power-network graph, $C$ the node capacitances, $i(t)$ the generated per-node current at time $t$, and $v(t)$ the node voltage drops. With backward Euler time stepping, one sparse factorization serves every time step and every synthetic sample on a given design. That is why labels are cheap even for many samples.
Research Program
The generator. The generator is a conditional diffusion model, meaning a model that learns to turn random noise into samples through gradual denoising. It produces a sequence of current maps. Its conditions come from one real design: the cell-density and power-network-density maps, plus a graph embedding of the power-network graph built the way PDNNet builds its graph. The denoiser factorizes space and time, alternating spatial convolutions with temporal attention. This is the same strategy MedGen3D uses to keep 3D generation within GPU memory.
Consistency with the design. The cross-modal gap enters here as a constraint on generated data rather than as a fusion module. A generated current movie has to agree with the other descriptions of the same design. Current may appear only where cells are placed. Each frame's total current must match a total-power value drawn from the range observed for that design. Both rules are enforced by projecting each sample onto them after generation.
Steering toward hard cases. A scalar "simultaneity" condition, the peak fraction of a region switching within one time step, lets us oversample the rare worst-case patterns. This mirrors how 3D MCGAN conditions on nodule size and how CircuitNet 3.0 deliberately targets the tail of hard timing cases.
Calibrating the solver. The solver uses a simplified grid model, so its labels differ from those of Cadence Voltus, the commercial sign-off tool behind CircuitNet's labels. We first measure this gap on real CircuitNet samples by running the solver on real current maps and comparing its output with the Voltus labels. Training then has two stages: pretrain on solver-labeled synthetic data, then fine-tune on real Voltus labels. We keep the predictor's architecture fixed at PDNNet's, so any gain can be attributed to the data rather than to model design.
Baselines
Naive reference. PDNNet trained only on the real training split, with standard flips and rotations. This shows what the data alone supports.
Strongest existing method. PDNNet itself. It reports "up to 39.3% reduction in prediction error" over CNN methods on CircuitNet, with a 545x speedup over the commercial tool (PDNNet). No absolute PDNNet error value was found on Lacuna, so we will retrain it ourselves. As an external anchor, CircuitNet reports a U-Net with ROC-AUC 0.94 and PR-AUC 0.83 on its IR drop task (CircuitNet). We also run PowerNet's maximum-CNN as a second architecture (PowerNet), to check that any gain is not specific to one predictor.
Oversampling baseline. LMM-IR-style twenty-times oversampling with Gaussian noise, applied to PDNNet.
Ablations that isolate the key ingredient. First, the solver baseline: the fast solver alone, run on real current maps at test time with no learning. Second, solver pretraining on real stimuli only, which adds new labels without adding new stimuli. Third, heuristic stimuli: real current maps with region-wise random time shifts and shuffles, labeled by the solver. Fourth, joint generation: the same diffusion model extended to emit the IR drop map as an extra channel, with no solver.
Fair comparison. Every arm uses the same design-level train and test split, the same predictor, the same number of training steps, and the same hyperparameter search budget. Every augmentation arm is matched on the number of synthetic samples.
What counts as winning. We beat PDNNet if hotspot F1 on unseen designs improves in the low-data regimes, with the gain exceeding the variation across three seeds. We beat heuristic stimuli if the learned generator wins at matched sample count and matched solver labels.
Why not just…?
Why not just perturb real current maps by hand and label them with the solver? This is the most serious threat. If it matches the learned generator, the useful contribution is solver-labeled augmentation, and the diffusion model is unnecessary. The heuristic-stimulus arm answers this directly. We will report the result whichever way it comes out. If the heuristic arm ties, the narrower claim (solver labels allow high mixing ratios) still stands and is cheaper to adopt.
Why not just generate inputs and labels together, as medical-imaging work does? Joint generation is exactly the regime in which 3D MCGAN saw gains at a 1x mixing ratio and a dip at 3x (Synthesizing Diverse Lung Nodules Wherever Massively). Our ratio sweep compares the joint arm with the solver-labeled arm at ratios of 1x, 3x, 10x and 30x. The prediction is that the joint arm degrades earlier.
Why not just oversample with noise? Noise does not change which regions switch together, and that co-switching is what creates hotspots. LMM-IR, trained with noise-based oversampling, still reaches only an F1 of 0.58 on static hotspots (LMM-IR). Here this is an explicit baseline arm rather than an assumption.
Why not just run the fast solver instead of a learned predictor? If the simplified solver tracked Voltus well, no predictor would be needed. The solver baseline measures this directly. A strong solver result would be informative in its own right. A weak one explains why we use solver labels only for pretraining.
Why not just fix the fusion instead of adding data? FindRec's distribution-alignment module gave a 5.7% NDCG@5 gain on recommendation data (FindRec). LMM-IR's cross-attention gained ten F1 points over its image-only ablation. However, neither addresses the shortage of switching patterns. Because we hold the architecture fixed, our results stay compatible with these methods. A cross-attention PDNNet variant, with and without our augmentation, tests whether the two gains add up.
Proposed Experiments
All experiments use CircuitNet's dynamic IR drop task, with splits made by design so that test designs are never seen in training. We train on 5%, 10%, 25% and 100% of the training designs. Metrics are mean absolute error on the IR drop map, Pearson correlation, and hotspot F1 and PR-AUC, with hotspots defined at 90% of the peak drop as in the ICCAD 2023 setting used by LMM-IR. If CircuitNet 2.0 provides dynamic IR drop labels, a secondary test trains on 28nm designs and evaluates at 14nm (CircuitNet 2.0).
Decisive checks. Four results would decide the question.
The first is whether solver-labeled learned stimuli beat real-only PDNNet in hotspot F1 at 10% data.
The second is whether learned stimuli beat heuristic stimuli at matched counts.
The third is whether the solver-labeled arm keeps improving, or holds steady, up to 10x mixing while the joint-generation arm declines.
The fourth is the measured agreement between solver and Voltus. If the solver's Pearson correlation with Voltus is low and pretraining on its labels hurts, the approach fails. In that case we would report how much solver fidelity the approach needs before it becomes useful.
References
- Synthesizing Diverse Lung Nodules Wherever Massively: 3D Multi-Conditional GAN-Based CT Image Augmentation for Object Detection
- PowerNet: Transferable Dynamic IR Drop Estimation via Maximum Convolutional Neural Network
- PDNNet: PDN-Aware GNN-CNN Heterogeneous Network for Dynamic IR Drop Prediction
- CircuitNet 3.0: A Multi-Modal Dataset with Task-Oriented Augmentation for AI-Driven Circuit Design
- LMM-IR: Large-Scale Netlist-Aware Multimodal Framework for Static IR-Drop Prediction
- MedGen3D: A Deep Generative Framework for Paired 3D Image and Mask Generation
- CircuitNet: an open-source dataset for machine learning applications in electronic design automation (EDA)
- FindRec: Stein-Guided Entropic Flow for Multi-Modal Sequential Recommendation
- CircuitNet 2.0: An Advanced Dataset for Promoting Machine Learning Innovations in Realistic Chip Design Environment