Research Hardware & Infra·arXiv cs.LG·May 6

Full-chip CMP modelling based on Fully Convolutional Network leveraging White Light Interferometry

Researchers propose a deep learning approach to accelerate Chemical-Mechanical Polishing simulation in semiconductor manufacturing by combining White Light Interferometry and Atomic Force Microscopy data. The work targets a critical bottleneck in IC design verification, where traditional Density Step Height modeling demands expensive calibration and computational overhead. By training fully convolutional networks on surface metrology data, the method could compress layout manufacturability checks from weeks to hours, directly reducing time-to-market for chip design teams. This represents a practical application of computer vision and deep learning to solve a high-stakes manufacturing constraint that affects the entire semiconductor supply chain.

Modelwire context

Explainer

The paper doesn't propose a novel network architecture. Instead, it demonstrates that surface metrology data from two different measurement modalities (White Light Interferometry and Atomic Force Microscopy) can be combined as training inputs to a standard FCN, compressing what traditionally required weeks of iterative calibration into a forward pass. The novelty is in the data fusion strategy and its application to a specific manufacturing constraint, not in the model itself.

This connects to the DeepONet work from May 1st, which also extended operator learning to handle geometric flexibility without retraining. Both papers solve a similar structural problem: how to build ML systems that generalize across input variations (here, different chip layouts and their resulting surface topographies) without requiring domain-specific parameterization or repeated calibration. The CMP work is narrower in scope but more immediately deployable, targeting a concrete manufacturing pain point rather than a general physics operator.

If semiconductor design teams at major foundries (TSMC, Samsung, Intel) report adopting this approach within 18 months and publish case studies showing actual time-to-tapeout reductions, that confirms the method works at production scale. If the paper remains confined to research clusters without vendor integration by end of 2027, it likely solved a problem that was already being addressed through other means (faster simulators, better calibration automation) and didn't offer enough practical advantage to displace them.

Coverage we drew on

Learning the Helmholtz equation operator with DeepONet for non-parametric 2D geometries · arXiv cs.LG

This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.

MentionsFully Convolutional Network · White Light Interferometry · Atomic Force Microscopy · Chemical-Mechanical Polishing · Density Step Height

Read full story at arXiv cs.LG →(arxiv.org)

Modelwire Editorial

This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.

Our mission How we write

Modelwire summarizes, we don’t republish. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.