Can a physics-guided transformer predict electromigration stress in multi-segment interconnects?

The model predicts stress profiles, not just single points

Electromigration is a reliability failure mode in on-chip interconnects: as technology scales to smaller nodes, rising current densities make it one of the dominant reliability challenges. Identifying wires susceptible to EM degradation requires accurate transient stress analysis, but running physics-based solvers across many interconnects stays computationally expensive.

The paper's answer is a physics-guided transformer that predicts hydrostatic stress in multi-segment interconnect lines. Each line is converted into an ordered sequence of geometry- and DC-aware segment tokens; transformer attention captures line-level context; and a lightweight query decoder returns stress at selected locations and time instants. The output is therefore a stress value at a queried (x, t), and the training objective also optimizes full stress profiles as line-level objects.

Stress at a point depends on the whole line

The paper frames the prediction problem around a physical argument. Within a multi-segment interconnect line, stress at a queried point depends not only on local segment properties such as geometry and electrical quantities, but also on the surrounding segment sequence and terminal behavior. Effective EM prediction therefore needs models that capture both local segment-level features and broader line-level context.

That is the role of the transformer encoder: it learns contextual representations of all segments in the line, so a query at one location can draw on the rest of the line rather than only the segment it sits in. The paper describes EM stress evolution through Korhonen's stress diffusion model, in which the EM driving term and the opposing back-stress term balance to determine net atomic flux and stress buildup. The architecture is designed to respect that line-level structure rather than treat each segment independently.

Physics terms constrain how the model is trained

The training objective combines three ingredients: a normalized supervised regression term, a linewise relative-L2 term, and physics-guided continuity and terminal-flux terms. The normalized MSE term fits pointwise predictions to reference labels. The linewise relative-L2 term optimizes full stress profiles as line-level objects, with a floor that prevents near-zero-stress lines from dominating the objective.

The two physics-guided terms encode boundary behavior. The continuity term penalizes stress mismatch at internal segment interfaces, comparing the right endpoint of one segment with the left endpoint of the next. The terminal-flux term regularizes the predicted stress slope at the two blocking terminals, using each terminal segment's current density and length; in implementation the derivative is approximated with a short finite difference near each endpoint. These terms are the mechanism by which the model is pushed toward physically consistent profiles rather than only pointwise agreement.

Test relative-L2 error stays below 8% on all six benchmarks

Evaluation uses interconnect lines extracted from IBM power grid benchmarks, with stress labels generated by the reference solver and accuracy averaged over three reliability time instants: t = 5, 10, and 20 years. Across all six benchmarks, test relative-L2 error remains below 8%. The strongest relative accuracy appears on IBM5 (3.7%) and IBM6 (4.2%), while IBM2 (7.0%) and IBM3 (7.8%) are the most challenging cases, which the paper attributes to larger structural variation and a wider stress-scale range.

The paper also reports absolute deviations in pascals, and these tell a different story from the relative metric. IBM3 has a test RMSE of 2.72 x 10^5 Pa and MAE of 6.73 x 10^4 Pa, much larger than IBM5's RMSE of 3.18 x 10^1 Pa and MAE of 1.86 x 10^1 Pa, even though their relative errors are comparable. Relative-L2 measures normalized stress-profile error; RMSE and MAE report absolute stress deviations. A low relative error on a benchmark with a wide stress range can still come with large absolute deviations, so the two metric families should not be read interchangeably.

Train and test errors stay close on most benchmarks

For supervised learning, the authors select a compact representative subset from each IBM benchmark, chosen to represent the benchmark's line population while reducing the reference data required. The remaining extracted lines are held out and used only for evaluation. The Train/all column in the dataset statistics ranges from 15% (IBM4) to 62% (IBM5), so the supervised fraction differs substantially across benchmarks.

Train and test relative-L2 values are close for most benchmarks, which the paper reads as evidence that the model does not simply memorize the supervised subset. The gap is small in most cases: IBM1 goes from 7.3% train to 6.8% test, IBM4 from 5.2% to 5.7%, and IBM5 from 4.4% to 3.7%. This is the paper's own interpretation of the comparison, and it is bounded by the fact that the held-out lines come from the same benchmarks and the same reference solver.

Speedup ranges from 368x to 2459.68x across benchmarks

Runtime is compared on the 10 largest extracted lines of each IBM benchmark, with the table reporting average runtime over those cases. For the proposed method the measured time is a forward pass of the trained model; the reference time is computing the reference stress solution for the same line. Training and hyperparameter search are offline one-time costs and are excluded from this per-line comparison.

The largest acceleration is on IBM2, at up to 2459.68x (0.139 s/line for the reference solver versus 5.65 x 10^-5 s/line for the model). The smallest is IBM1, at 367.95x, which the paper notes still leaves the model more than two orders of magnitude faster than the reference solver. The other benchmarks fall between: 406.24x on IBM3, 712.86x on IBM4, 1675.76x on IBM5, and 706.53x on IBM6. The discussion reports an average speedup of 1054.84x over the matrix exponential solver. Because the comparison uses the 10 largest lines per benchmark, the speedup figures describe that worst-case-oriented setting rather than every line in each dataset.

The evidence covers lines, not interconnect trees

The evaluation is limited to multi-segment interconnect lines extracted from IBM power grid benchmarks, with stress labels from the reference solver and accuracy averaged over t = 5, 10, and 20 years. The paper's own framing explains why this decomposition is reasonable: in Cu dual-damascene technology, diffusion barriers confine atomic transport within individual metallization layers, suppressing mass exchange across adjacent layers, so large networks such as power grids can be analyzed layer by layer, and within each layer the routing topology can be decomposed into multi-segment lines of wire segments and vias.

That scope also marks the boundary. The framework is presented for multi-segment interconnect lines, and the authors state that future work will extend it to interconnect-tree topologies and investigate adaptive selection of samples for label generation. The reported errors and speedups therefore describe line-level prediction on these benchmarks, not tree-level or full-network analysis.

Profile fidelity at scale matters more than peak speedup

The headline number is the 2459.68x speedup on IBM2, but the paper's own results suggest the more decision-relevant comparison is between relative and absolute accuracy across benchmarks. IBM3 keeps a test relative-L2 of 7.8% while carrying the largest absolute deviations in the study (RMSE 2.72 x 10^5 Pa, MAE 6.73 x 10^4 Pa), and the paper links IBM2 and IBM3's difficulty to larger structural variation and a wider stress-scale range. A reader comparing deployment candidates should therefore look at which benchmark resembles the target lines in structural variation and stress range, not only at the best speedup figure.

The intended deployment case is reuse: the trained model is meant to evaluate many new interconnect lines, with training and hyperparameter search treated as offline one-time costs. Within that setting, the paper's strongest claim is that a compact supervised subset per benchmark can support test relative-L2 below 8% across all six benchmarks while preserving the dominant stress peaks and localized variations near segment transitions, as shown for a representative 110-segment line. The priority this paper establishes is profile-level fidelity across structurally varied lines, with speedup as the enabling condition rather than the primary result.

Sources used in this answer