acceptodds
Under review as a conference paper at ICLR 2027

Chip Protection Area Design: Dataset and Evaluation

Abstract

The protection area surrounding a semiconductor chip guards its circuit against thermo-mechanical defects induced during wafer dicing. Its 2D designs are governed by manufacturing constraints and evaluated using test chips that require production processes with substantial costs. An ideal design should achieve optimal 3D mechanical states while meeting the hard constraints. This creates a highly nonlinear optimization landscape where data-driven approaches had show strong promise in other domains. Given the lack of large-scale semiconductor protection area datasets with 3D thermo-mechanical elasto-plastic evaluation, we propose the first 1.17 TB public dataset, containing 110,170 protection area layouts across 18 metal stacks, each paired with fields from thermo-mechanical elasto-plastic finite element simulations. We demonstrate that the dataset enables two key benchmarks: design-to-fields cross-dimensional forward prediction, and fields-to-design inverse generation. The baseline results establish the difficulty of these tasks and highlight opportunities for data-driven methods to address industrial design bottlenecks. In addition to the dataset, we provide an open-source end-to-end automated pipeline from design generation to simulation for reproducibility and potential extensions.

Primary area:
datasets and benchmarks
Keywords:
chip design, protection area, dataset, chip layout, thermo-mechanical, elasto-plastic, FEM simulation, machine learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ugoUOpenReview ↗
Rating
2(−1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents the first large‑scale chip‑protection‑area dataset of 1.17 TB, consisting of 110 170 2D layout designs paired with 3D thermo‑elastoplastic finite‑element simulation results across 18 metal‑stack configurations. Two complementary benchmarks are defined: the forward design‑to‑fields task predicting 3D physical fields from 2D layouts, and the inverse fields‑to‑design generative task producing DRC‑compliant 2D layouts given target physical‑field descriptors. An end‑to‑end open‑source pipeline from design sampling to simulation is released and validated against industrial EDA software. Baseline experiments demonstrate that generalization across unseen metal stacks poses the primary challenge; accurate local point‑wise predictions do not guarantee reliable full‑field reconstruction. Retrieval‑based baselines outperform autoregressive generative models for inverse design. Though constrained by synthetic technology node, fixed‑load assumptions and rectangle‑only layout patterns, this work fills the gap of large‑scale public thermo‑mechanical semiconductor datasets and offers evaluation infrastructure for data‑driven simulation and inverse engineering design.

Critical strengths
  1. Filling Domain Data Gaps: Constructing the first large-scale public dataset in this field (110,170 samples, 1.17TB), accompanied by thermoelastic-plastic finite element simulation 3D physics data; Previously, the field of semiconductor thermomechanics mostly consisted of small-scale, private datasets, lacking public, large-scale resources with elastoplastic effects that could support data-driven simulation and reverse engineering research.
  2. Bidirectional Industry-Oriented Reference Task Design: Design two cross-modal tasks: Layout to Physical Field Forward Prediction and Physical Field to Layout Reverse Generation, aligning with the real industrial design closed loop of chip protection zones; Reverse engineering tasks are not evaluated based on image pixel loss, but instead evaluate physical field metrics obtained through DRC design rule checks + resimulation, making them more closely aligned with engineering realities. They do not simply look at geometric appearance similarity but emphasize functional equivalence.
  3. Complete reproducible open-source pipeline: Provides a complete set of open-source code from layout random sampling, DRC design rule validation to thermo-mechanical finite element simulation; Key modules (DRC checker, finite element solver) have been benchmarked against industrial commercial software such as Cadence Virtuoso and Ansys Mechanical, allowing researchers to reproduce the entire dataset and extend new processes and configurations based on it.
  4. Multi-dimensional rigorous evaluation divisions: Sets three test divisions: IID, OOD-Combo (cross-metal stack), and OOD-Field (atypical mechanical samples), focusing on the model's distribution generalization ability; By distinguishing local probe point error and region B spline field descriptor error, the phenomenon of "single-point prediction accuracy does not equal full-field prediction accuracy," providing important insights for subsequent model evaluation.
  5. Comprehensive Task Evaluation System: Forward tasks support multiple model input formats including grid and meshless implicit fields; Reverse tasks fully consider manufacturing constraints, with DRC legality as a pre-screening condition; Provides results from multiple baseline models, clearly exposing the capability ceilings of existing methods and providing clear directions for improvement in future research.
Critical weaknesses
  1. Limitations in Simulation Assumptions for Processes and Materials: The dataset is constructed based on synthetic fictitious process nodes and is not a real commercial process; Only 5 materials are fixed, with no support for more semiconductor materials; Thermal and mechanical loads are fixed point parameters (cooling at 400-22°C, 20GPa pressure). In real manufacturing, the load size and position are randomly distributed, and simulation conditions differ from real wafer cutting conditions.
  2. Limited Layout Pattern Expression Capability: All generated layouts consist only of axis-aligned rectangles and do not support arbitrary polygonal patterns; In contrast, the real chip protection area may have complex oblique and irregular polygonal structures, and the sampling distribution may have artificial statistical bias, which deviates from the real industrial layout.
  3. High cost for inverse generation task evaluation, limited test set size: Fields-to-design reverse engineering task testing uses only 5,000 samples; Each generated layout requires rerunning finite element simulation for scoring, which is a huge calculation overhead; In OOD-Combo cross-metal stack scenarios, autoregressive generation models can hardly output DRC legal layouts, and the current baseline performance is poor.
  4. Some simplified handling in simulation and evaluation: Before evaluation, the physical field is smoothed by Gaussian smoothing to mask local field changes caused by thin layers; The reference does not have mandatory physical balance verification; Finite element truth values depend on mesh resolution and grid algorithms, so simulation truth values have some numerical errors.
  5. Limited Application Scenarios and Narrow Scope: Focusing only on the chip protection area as a specialized unit cell, it cannot be directly used for more general semiconductor design tasks such as chip layout and routing, circuits, etc.; Simplified by using the periodic cell assumption, this assumption may not be applicable to all edge chip scenarios.
  6. Complete dataset files cannot be downloaded directly: Due to platform policy restrictions, the complete 1.17TB dataset cannot be directly distributed during the submission stage; only a small subset of 500 samples is available; To obtain complete data, researchers need to use open-source pipelines to consume large amounts of computing power and regenerate them themselves, which is very difficult to use.
Questions for authors
  1. The dataset is based on synthetic fictitious process nodes, and the loads also use fixed numerical assumptions. How will future plans be made to adjust simulation configurations to closely match real tape-out random loads and real commercial process parameters, further narrowing the gap between simulation and actual industrial conditions?
  2. Current layout sampling only supports axis-aligned rectangles and cannot generate irregular polygons commonly seen in industry. If you want to expand the polygon layout, what do you think is the biggest engineering challenge in sampling algorithms, mesh segmentation, and simulation computing overhead? Are there plans for future iterations?
  3. Fields-to-design Reverse Task Under the OOD-Combo cross-metal stack setup, the autoregressive generation model can hardly output DRC-legitimate layouts. Besides improving generative model capabilities, are we considering introducing methods such as constraint decoding and hard embedding of design rules to increase the yield rate of legitimate layouts in cross-stack scenarios?
  4. Complete datasets cannot be directly distributed during the review phase; researchers need to consume a large amount of computing power to reproduce them themselves. For small teams with limited computing resources, are there solutions to lower the reproduction threshold, such as releasing preprocessed intermediate representations or lightweight subsets?
  5. Using Gaussian smoothing during the review to process the physical field masks local field abrupt changes caused by thin layers. In future versions of the benchmark, is there a plan to provide raw field metrics without smoothing to evaluate the model's ability to capture sharp local stress/strain abrupt changes?
Reviewer_HBMROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a large-scale dataset for semiconductor chip protection area design, containing 110K layouts across 18 metal-stack configurations and corresponding 3D thermo-mechanical simulation results. The authors propose two benchmark tasks: design-to-fields prediction and fields-to-design generation. An automated design generation and simulation pipeline is also provided to support reproducibility. Baseline experiments highlight the challenges of predicting mechanical states and generating manufacturable designs.

Critical strengths
  1. A large-scale dataset addressing an underexplored problem in semiconductor manufacturing.
  2. The benchmark connects 2D layout design with 3D thermo-mechanical simulation, providing meaningful cross-dimensional learning tasks.
  3. The automated data generation pipeline and validation against commercial tools support reproducibility.
  4. The forward and inverse design benchmarks provide useful starting points for future research.
Critical weaknesses
  1. The simulations use fixed thermal loading and mechanical pressure at only three positions. These assumptions may not reflect the variability of real wafer dicing processes.
  2. The forward task primarily evaluates basic PointNet/CNN architectures, while the inverse task uses retrieval and autoregressive models. Comparisons with more advanced generative methods would better establish the benchmark's difficulty and value.
  3. The inverse-design benchmark evaluates DRC validity and field-descriptor accuracy, but does not directly assess whether generated designs improve crack resistance or reduce manufacturing failure risk.
Questions for authors
  1. How well does the dataset generalize to different technology nodes, material properties, and loading conditions?
  2. Can the inverse-design evaluation incorporate direct measures of mechanical reliability or defect risk beyond field-descriptor matching?
Reviewer_Wy4aOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
3(+1)
Read review
Paper summary

The paper introduces a synthetic dataset of 110,170 chip-protection layouts across 18 metal-stack configurations, paired with three-dimensional thermo-mechanical elasto-plastic simulation states. The proposed pipeline combines rule-aware layout sampling, design-rule checking, and finite-element simulation under thermal loading followed by three mechanical loading cases. Two benchmarks cover layout-to-field prediction and field-descriptor-conditioned layout generation. Forward evaluation combines interface probes with B-spline descriptors of assessment-zone fields; inverse evaluation checks design validity and re-simulates valid candidates. Baselines include PointNet/Conv2D encoders with grid or implicit decoders, nearest-neighbor retrieval, and autoregressive generators. Results indicate substantial remaining difficulties, especially on unseen stacks. The resource and evaluation protocol are the central contributions.

Critical strengths
  1. A distinctive and practically motivated problem at substantial scale: pairing manufacturability-constrained layouts with 3D thermo-mechanical responses (including metal plasticity) across 18 stacks is a genuinely useful resource for surrogate modeling and inverse design.

  2. The evaluation is well thought out: DRC- and simulation-based scoring suits the one-to-many inverse problem, the stack/field splits go beyond random splits, and the same-stack retrieval baseline gives a concrete reference for judging learned generation.

Critical weaknesses
  1. The manuscript claims validation against Virtuoso and Ansys but reports no quantitative agreement or mesh-convergence results, which is a serious gap for a dataset paper since downstream errors cannot be interpreted without an estimate of label accuracy.

  2. The main inverse metric is computed on smoothed, x-averaged fields and conditioned on validity (e.g., ARG-Joint's OOD-Combo NMAE rests on 2 of 5,000 candidates), so it is unclear the benchmark actually measures success on the stated engineering objective.

Questions for authors
  1. Could you quantify how well the simulator agrees with Ansys on representative stacks and loading cases, including assessment-zone fields and localized extrema, and report mesh-refinement sensitivity and DRC agreement? How are non-converged simulations handled?

  2. Does a low descriptor NMAE preserve the ranking of designs by unsmoothed peak or high-quantile S1 / strain energy density? A sensitivity analysis over the smoothing scale and the x-aggregation would help a lot here.

  3. Could you report the fraction of all targets for which a candidate is both DRC-valid and below a stated field-error threshold? Also, the IID ARG validity count is 704 in Table 1 but 741 in Table 9, and its OOD-Field NMAE is 0.4386 in Table 1 but 0.4368 in Table 10. Please reconcile these.

  4. How do the forward models compare against a training-mean predictor under the same probe aggregation and normalization? Several probe NRMSE values are near one, so this baseline is needed to know how much layout-dependent signal the models actually learn.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.