acceptodds
Under review as a conference paper at ICLR 2027

Topology-Preserved Neural SDF Reconstruction for Simulation-Ready Meshing in Advanced Packaging

Abstract

For integrated IC to integrated system, STCO extends DTCO from chip-level to system-level co-optimization; therefore, multiphysics modeling is the core-driven force for the STCO. However, mesh generation is essential for multiphysics simulations in advanced packaging. Traditionally, these meshes require tedious manual CAD repair; it sometimes can be even slower than the simulation itself. Alternatively, a learned signed distance function (SDF) can bypass CAD repair, while reconstruct meshes directly from the easily available point scans. Whereas, it often fails to capture critical internal features, such as vias and narrow gaps between stacked layers. To address this issue, we propose a new CAD-preserved SDF with the following two improvements: 1) We design a new topology-constrained algorithm before training by pre-computing constraints from the input data. This certificate is a checkable sufficient condition that, when it holds, preserves the topology of the learned field. 2) We introduce a discretization certificate, a second checkable condition under which the topology is still preserved after discretization. Meanwhile, we add a gradient-consistency regularizer to remove stripe artifacts from flat surfaces. To be specific, we verify that every measured through-path remains open on the TSV package. The meshes generated with these improvements are then used directly for steady-state thermal simulations in a commercial multiphysics solver, without manual geometry editing. Under a shared-field vertex-placement protocol, the resulting pipeline ranks first on all six metrics for the six chip-package models, each carrying its own free-space certificate pre-computed from its own scan; on three Thingi10K figurines, which have no internal corridors, the same placement rule leads on most metrics and isolates its geometric benefit from the topology claim.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
point cloud reconstruction, signed distance function, surface reconstruction, mesh generation, topology preservation, mesh simplification, chip packaging, simulation-ready meshes

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_SR9BOpenReview ↗
Rating
3
Confidence
1
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

In System-in-Package, multiphysics simulations require clean, simulation-ready meshes with proper material regions and open vias. Scans of advanced packaging often have missing surfaces or noise in narrow inter-layer gaps and via interiors. This paper presents a CAD-preserved, topology-aware neural signed distance function (SDF) reconstruction framework designed to generate simulation-ready meshes directly from point cloud scans of advanced chip packages.

Critical strengths
  1. Completely eliminates the labor-intensive manual CAD repair phase, allowing extracted meshes to directly import, volume-mesh, and solve in commercial multiphysics simulators.

  2. Curvature-adaptive vertex placement and clamped post-processing produce clean, lightweight meshes with high aspect-ratio quality and dramatic face-count reductions

Critical weaknesses
  1. In practice, a single localized field error drives the global inequality to fail, even when 99.5%+ of samples pass and through-paths remain physically open. The theoretical bound relies on an estimated Lipschitz constant (LL) derived empirically from Eikonal residuals rather than a strict, provable Lipschitz bound on the neural network.

  2. Free-space sample generation requires GPU ray-casting based on predefined through-hole axes and mid-surfaces. This makes the method dependent on domain knowledge or heuristics to identify where vias and gaps lie before training

  3. The pre-computation and cubical-complex evaluations is only shown on a fixed 5123512^3 reference grid. Geometric features or gaps thinner than a single reference voxel cannot be certified or reliably measured. However, it seems most circuit thermal mechanical simulation requires much higher resolution (in the unit of nm). Hence the scale of this paper presented is very different from the practical usage. This is the main questions whether this method is practical for me.

Questions for authors

Please see my Critical Weaknesses for questions.

Reviewer_qXerOpenReview ↗
Rating
3(+1)
Confidence
1(−1)
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies simulation-ready mesh reconstruction from point clouds of chip packages, focusing on the unintended closure of narrow through-holes and inter-layer gaps. The method uses a neural SDF with precomputed free-space samples and clearance constraints to discourage erroneous filling of these regions. Output meshes are generated through curvature-adaptive meshing or dense surface extraction followed by geometric post-processing. The paper also proposes sufficient conditions for keeping designated channels open in the continuous field and preserving them through discretization and post-processing.

My expertise in CAD reconstruction and repair is limited, and I will consider comments from reviewers with relevant expertise.

Critical strengths
  1. The paper addresses an important problem in CAD reconstruction: preserving narrow through-holes and inter-layer gaps that surface reconstruction metrics may fail to capture.

  2. The paper presents an engineering-oriented pipeline combining neural SDF reconstruction, meshing, and geometric post-processing, and evaluates its components on multiple chip-package geometries.

Critical weaknesses
  1. The presentation is difficult to follow because the relationships between the pipeline components, input assumptions, and evaluation protocols are not clearly established. For example, Figure 1 suggests a sequential three-stage pipeline, whereas Section 3.1 describes two alternative meshing routes. The required inputs are also introduced incrementally: Section 3.1 specifies a point cloud, while normals, certified free-space samples, and, in some experiments, GT-SDF supervision appear elsewhere. In addition, Section 3.5 compresses six post-processing operations into a single paragraph without clearly explaining their intermediate outputs or how each operation relates to the displacement bounds. A clearer workflow diagram, an explicit statement of input requirements, and a stage-by-stage algorithm description would substantially improve readability.

  2. The necessity and cost-effectiveness of the neural SDF are unclear. Per-instance optimization takes hours to tens of hours, while the final pipeline still relies on precomputed free-space information and extensive geometric post-processing. It is therefore difficult to determine whether the main benefits come from the neural representation or from these additional constraints and processing steps. Although the paper ablates the free-space loss, this does not establish whether a conventional reconstruction method could achieve comparable results when given the same information and post-processing. The authors should discuss why a neural SDF is needed and, ideally, provide a controlled comparison demonstrating its additional benefits relative to its computational cost.

Questions for authors

Please see weaknesses.

Reviewer_S9VhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a topology-preserving neural SDF reconstruction framework for generating simulation-ready meshes from point clouds. The method introduces free-space constraints to preserve internal vias and gaps, a curvature-adaptive meshing strategy, and a discretization certificate to constrain topology changes during mesh extraction and post-processing. Experiments demonstrate improvements in topology preservation, geometric reconstruction, and mesh quality, with generated meshes successfully used in thermal simulations.

Critical strengths

The paper addresses 3D reconstruction challenge that preserving internal topology while generating meshes suitable for physical simulation. This is particularly important for advanced packaging with complex internal structures.

The proposed free-space constraints and Lipschitz-based certificate provide a useful way to protect internal vias and gaps that are not sufficiently observed in the input point clouds.

The paper integrates neural SDF reconstruction, curvature-adaptive meshing, and mesh post-processing. The improvements in mesh quality and reduction in mesh complexity are encouraging.

Critical weaknesses

The free-space samples are generated using through-hole axes, inter-layer mid-surfaces, and geometric filtering, which may require additional knowledge of the internal structures. It is unclear how these free-space regions can be reliably identified from sparse or incomplete point clouds.

Table 1 uses the authors' own frozen SDF field as the reference and compares against simplified vertex-placement versions of NMC, NDC, and VoroMesh rather than their complete pipelines. Although Appendix D provides additional end-to-end comparisons, these remain limited and use different mesh element budgets. More comprehensive and fair comparisons would strengthen the experimental claims.

The proposed method requires per-scene neural SDF optimization. Although it reduces manual CAD repair, the computational cost remains high. Its scalability to large numbers of packaging designs and scans is unclear.

Questions for authors

Can the authors demonstrate a complete point-cloud-to-mesh pipeline that satisfies the proposed topology-preservation certificate, rather than verifying the learned SDF and post-processing stages separately?

Can the authors provide a direct comparison of physical simulation results between the reconstructed meshes and the original CAD meshes under identical boundary conditions?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.