Full-chip curvilinear ILT (inverse lithography technology) aims to calculate mask patterns across an entire design while accounting for how lithography will print them on the wafer. The approach addresses a long-standing trade-off in advanced semiconductor manufacturing: ILT can produce useful pattern corrections, but the computation required to apply it chip-wide has historically made turnaround time a barrier. A reported memory-manufacturing evaluation examines how GPU acceleration and a stitchless workflow can address that constraint, alongside mask writing and wafer results.
Why inverse lithography technology matters
Conventional optical proximity correction modifies mask shapes to compensate for the limitations of lithography. ILT takes a more direct inverse-design approach: starting with a desired wafer pattern and models of the lithographic process, it calculates mask shapes intended to produce that target. Because the optical system is band-limited, the resulting solutions can be curvilinear rather than purely rectilinear.
That flexibility matters as feature dimensions shrink and process requirements become more demanding. The source study describes prior work in which unconstrained curvilinear ILT produced a larger process window than alternative ILT mask patterns for the tested contact arrays. A process window is the range of conditions—such as focus and exposure dose—under which printed features remain within the specified requirements.
ILT has often been applied to selected hotspots rather than entire masks. The key obstacle has not been the concept of optimizing mask shapes, but the amount of computation involved. Full-chip ILT requires substantially more computation than traditional OPC, and conventional workflows add extra steps to make results compatible with mask-writing processes.
How partitioning creates a runtime challenge
A common way to distribute a large computation is to divide the chip into smaller regions, process them in parallel, and reassemble the results. In a partitioned ILT workflow, however, each region is optimized separately. Features near a boundary may be affected by neighboring features that were processed in another partition, so the two sides can produce inconsistent results.
These boundary discrepancies, known as stitching errors, need to be corrected after the partitions are merged. Correction can require additional optimization passes around the boundaries, which may introduce new boundaries and further work. The workflow described in the study can also include mask-rule cleanup and Manhattanization—the conversion of curvilinear solutions into rectilinear shapes suitable for variable-shaped-beam (VSB) writing—followed by re-optimization to recover wafer performance.
The result is a significant runtime burden. Partitioning may accelerate individual calculations, but it does not eliminate the contextual effects that make seams difficult to resolve. This helps explain why ILT use has often been limited to smaller, high-risk areas of a design.
Full-chip curvilinear ILT with GPU acceleration
The evaluated approach takes a different route: instead of dividing the chip into independently optimized regions, it is designed to optimize full-chip data as a whole. Its computing platform combines multiple CPU and GPU pairs, but the hardware and software are designed to behave like a single, large processing system for the full-chip calculation.
GPUs are suited to workloads that apply similar operations across many data points, a computing pattern known as single instruction, multiple data (SIMD). ILT operates on pixel-domain data and repeatedly evaluates lithography and mask effects, making this kind of parallel processing a natural fit. The system described in the study was developed specifically for full-chip ILT, with hardware, software, models and verification considered together.
The practical aim is to avoid partition boundaries and the corrective work associated with them. Each optimization iteration updates the chip as a whole, allowing proximity effects to be considered across the design rather than only within separate regions. The paper reports that the commercial implementation can generate a full-chip ILT solution in a day; this is a claim about the evaluated system, not a universal runtime guarantee for all designs or manufacturing flows.
Why multi-beam writing supports curvilinear masks
Curvilinear ILT also depends on the ability to write the resulting mask patterns. A conventional VSB mask writer uses a variable rectangular beam, so curving shapes may need to be approximated by many small shots. That can make complex patterns time-consuming to write.
Multi-beam mask writers use an array of individually controlled beams and operate in the pixel domain. As described in the source, this makes write time less dependent on the complexity of the shapes, allowing curvilinear patterns to be produced without first converting them to Manhattan geometry. This removes a manufacturing constraint that previously limited the practical use of curvilinear ILT.
Mask-rule compliance remains essential. The evaluated ILT workflow integrates curvilinear mask rules into optimization so that the resulting patterns can meet mask requirements, rather than relying solely on unconstrained mathematical optimization.
What the memory-manufacturing evaluation found
The evaluation used a memory design, a production free-form illumination source, a multi-beam mask writer and the manufacturer’s process of record. The study reports that the written curvilinear patterns had high pattern fidelity and that wafer contacts printed consistently across the tested array, including different pitches and orientations.
Wafer results were also compared with OPC at multiple focus and dose conditions. In the reported contact-array comparison, ILT prints were more consistent across the sampled conditions, while some OPC results showed necking, uneven printing or contacts that did not print. A separate cut-layer comparison used measured critical dimensions (CDs) to define the process window. Under the study’s stated acceptance criteria, the authors report that the ILT process window increased by over 100% compared with OPC.
These findings are specific to the designs, models, mask process and wafer conditions evaluated. They demonstrate the potential of combining mask and wafer optimization, GPU-accelerated computation and multi-beam writing; they do not establish the same improvement for every layer or production flow.
Practical implications for lithography teams
For process integration and mask teams, the central implication is that ILT’s value depends on more than the optimization algorithm. Runtime, mask rules, writer capabilities and wafer validation all affect whether an approach can be used across a full design. Stitchless optimization targets the overhead of partition-based workflows, while multi-beam writing makes complex curvilinear output more practical to manufacture.
Full-chip curvilinear ILT therefore offers a route to applying inverse mask design beyond isolated hotspots. The reported evaluation connects that computational approach to written masks and printed wafers, including a larger measured process window in the tested comparison. For engineering teams assessing ILT, the practical next step is to evaluate performance against their own design patterns, mask constraints and process-of-record requirements.