Semiconductor Technology from A to Z

When the Mask No Longer Prints What It Shows

In the ideal case, a photomask contains exactly the geometry that should later appear on the wafer: a rectangle on the mask yields a rectangle in the resist. This ideal, however, only holds as long as the smallest features on the mask are considerably larger than the wavelength of the light used. Once feature size and wavelength enter the same order of magnitude, diffraction effects dominate the imaging behavior — edges are no longer imaged sharply but are systematically distorted by the optical system.

This relationship is described by the so-called k1 factor, which relates the achievable resolution to the wavelength and aperture of the exposure system. Modern processes operate at k1 values well below what classical, uncorrected optics can still image sharply — the exposure wavelength (still 193 nm in many nodes) is larger than the printed feature width itself. In such a "sub-resolution" process, simply transferring the desired geometry 1:1 is no longer sufficient.

Optical Proximity Correction: Pre-Distorting the Mask

Optical Proximity Correction (OPC) solves this problem by inverting it: rather than trying to make the optics more perfect, the mask geometry is pre-distorted so that the known, predictable imaging errors of the optical system produce the desired result on the wafer. A corner that would round off without correction is already over-drawn on the mask, so that it appears sharp again after imaging.

OPC is therefore not an after-the-fact repair step but a fixed part of the mask data flow: before actual mask fabrication, every modern layout passes through automated OPC software that adjusts the geometry chip-wide based on a calibrated model of the exposure process. Without this step, most of today's feature sizes simply could not be manufactured with the available exposure wavelength.

The diagram below shows the basic problem using a simple example: the desired layout on the left, and the result that would actually print without correction on the right.

Corner Rounding and Line-End Shortening

The effects shown in the previous section are the most immediate consequence of diffraction: sharp, right-angled corners in a mask structure become rounded contours in the resist, because the optical system no longer transmits the high spatial-frequency components that sharp edges would require. At convex corners, the rounding "eats into" the structure; at concave corners (for instance the inner corner of an L-shaped layout), it instead fills in the notch.

A related phenomenon occurs at the end of a trace: line-end shortening. Because the optical system images the end of a line not as a sharp edge but as a gradually falling intensity, the resist that actually develops at the line end pulls back relative to the mask specification — a line meant to reach exactly to a certain contact may, in print, end a few nanometers short of it.

The Proximity Effect: Same Width, Different Result

The effect that gives Optical Proximity Correction its name, however, is a third, more subtle one: two lines drawn with exactly the same width on the mask print at different widths depending on how many other structures sit in their immediate neighborhood. An isolated line receives its diffraction pattern solely from its own edges. A line in the middle of a dense line array, by contrast, has its diffraction pattern overlap with those of its neighbors — the resulting intensity profile in the resist differs systematically from that of the isolated line, even though the mask geometry is identical.

This so-called iso-dense bias varies in magnitude across the pitch spectrum (center-to-center spacing of features) and is non-linear — it must be characterized separately for each process, typically by systematically exposing and measuring test structures with varying pitch. The result of this characterization feeds directly into the correction model introduced later in this chapter.

The diagram below shows the proximity effect on an isolated line compared to a line in the middle of a dense array — both with identical mask width but different print results.

Mask Bias: Correcting the Dimension

The simplest form of correction addresses the proximity effect from the previous section directly: if it is known that a line in a dense array systematically prints narrower than intended, it is simply drawn somewhat wider on the mask from the outset — the so-called mask bias. Since the magnitude of this deviation depends on the local pitch, the bias is not a fixed constant but is determined separately for each neighborhood situation in the layout, based on the previously characterized iso-dense curve: densely packed regions receive a different bias value than isolated lines or intermediate pitch ranges.

Mask bias thus corrects only the line width as a whole — it does not change the shape of a corner or a line end. These local, geometrically confined errors require a second, more targeted correction mechanism.

Serifs and Hammerheads: Targeted Corner Reinforcement

To counteract corner rounding, a small additional square structure is added directly to the mask at particularly sensitive corners — above all at convex outer corners — a so-called serif. The serif itself remains below the resolution limit and does not print as a standalone structure; however, it locally alters the light intensity at the edge of the main structure so that the corner appears sharper after imaging than it would without this reinforcement.

At the end of a line, an enlarged variant of the same principle corrects line-end shortening: a so-called hammerhead — a local widening of the line end perpendicular to the line direction — compensates for the systematic pullback of the resist at the line end, so that the line still reaches its intended position even after optical imaging.

Both techniques — serifs and hammerheads — can be formulated as simple geometric rules: "add a square of size Y at every convex corner with interior angle smaller than X" or "widen every line end by Z nm." This rule-based approach is the simplest form of OPC and will be contrasted with the model-based variant two sections from now.

The diagram below shows the layout geometry from Section 1 after adding serifs and a hammerhead, along with the substantially improved print result this produces.

The Problem: Isolated Lines Behave Differently

Serifs and mask bias correct the shape and width of a structure directly at its own contour. For a more fundamental problem, however, this is not enough: an isolated line receives its diffraction pattern — as described in Section 2 — solely from its own edges, while a line in a dense array benefits from the diffraction patterns of its neighbors. This difference in diffraction conditions affects not only the printed width but also how tolerant the respective structure is to focus and dose variations in the exposure process — its so-called process window. Isolated structures typically have a considerably smaller process window than densely packed ones, even if both print at exactly the same width after correction.

Serifs and bias alone cannot fix this deficit, since they only alter the structure itself, not its optical surroundings.

Scatter Bars: Neighbors That Never Print

Sub-resolution assist features (SRAFs), colloquially also called scatter bars, solve this problem by artificially giving the isolated line neighbors: additional, very narrow lines are drawn on the mask parallel to the actual structure — narrow enough to remain below the resolution limit of the exposure system themselves and therefore not appear as a standalone structure in the resist. They do, however, alter the diffraction pattern at the location of the main structure so that it more closely resembles that of a densely packed environment — the isolated line now optically "looks" almost like a line within an array, with a correspondingly larger, more stable process window.

Placing SRAFs is subject to strict geometric rules: the spacing to the main structure must be large enough to avoid merging but small enough to still have an effect; the width of the scatter bars themselves must remain safely below the printing threshold, even under unfavorable process conditions. Software-assisted SRAF placement today generates these assist structures automatically based on the characterized diffraction behavior of the respective process, typically as the last step before the actual serif and bias correction.

The diagram below shows an isolated line without and with scatter bars — the assist structures themselves do not appear in the printed resist.

Rule-Based OPC: Fast but Coarse

The corrections introduced in the previous sections — mask bias, serifs, scatter bars — can be organized as a lookup table: for a given line width and a given spacing to the nearest neighboring structure, the table returns a previously characterized correction value. This rule-based OPC is computationally cheap, since only a single table lookup is needed per edge, with no elaborate simulation.

The limits of this approach show up with complex, two-dimensional geometry: near a corner, a line end, or an unusual neighborhood configuration, several proximity effects overlap simultaneously, for which no single table row can still provide a fitting answer. Rule-based OPC always treats an edge as a whole: the entire straight edge segment is shifted by the same amount, regardless of whether an additional structure near one end creates a different correction need there than at the other end.

Model-Based OPC: Simulated Segment by Segment

Model-based OPC takes a fundamentally different approach. Instead of looking up table values, it uses a calibrated numerical model of the entire imaging process — optical model and resist model combined — to predict the actually expected printed contour for a given mask geometry. To do this, every edge of a polygon is subdivided into many short segments, which can then be shifted independently of one another.

The correction algorithm works iteratively: it simulates the print of the current mask geometry, compares the result against the target contour, shifts each segment individually toward a better match, and repeats this cycle until the simulated contour agrees with the target within a defined tolerance or a maximum number of iterations is reached. Because each segment responds individually, the correction near a complex 2D situation — such as near a corner — can turn out considerably stronger than further away on the same edge, where simple rule-based correction would only know a single, averaged value.

This level of detail comes at a cost: model-based OPC requires a lithography simulation across thousands of segments simultaneously, chip-wide, for every iteration — a computational effort orders of magnitude beyond that of the rule-based variant. In practice, modern OPC flows therefore combine both approaches: rule-based correction for simple, well-characterized situations, model-based refinement wherever the geometry demands it.

The diagram below compares both approaches on the same edge near a neighboring structure: on the left, the uniform shift of rule-based correction; on the right, the segment-by-segment shift of model-based correction, adapted to the local geometry.

Why Correction Alone Is Not Enough

An applied OPC correction is initially just a claim: the correction model predicts that the adjusted mask geometry will print the desired result under the assumed process conditions. Whether this prediction actually holds across the entire chip design — often repeated billions of times — and whether it survives realistic variation in the manufacturing process, must subsequently be checked separately. This OPC verification runs technically similar to the actual correction: a full-chip lithography simulation, but this time not to adjust the geometry, but to check the already corrected layout.

The Process Window: More Than Just the Ideal Case

Exposure tools do not hold focus and dose perfectly constant — both vary within a certain range from wafer to wafer, across the wafer surface, and even within a single exposure field. A correction that only works at the exact target focus and target dose would be worthless in practice. OPC verification therefore simulates not just the nominal operating point but an entire process window of focus and dose combinations — typically the four extreme corners plus several intermediate points — and checks whether the structure stays within the allowed tolerance at every single one of these points.

Well-executed OPC noticeably widens this process window compared to the uncorrected layout: structures that previously only printed cleanly within a narrow focus-dose range remain within specification after correction even under larger deviations. The wider window is thus a direct, measurable quality metric for the correction itself — in addition to plain agreement at the nominal operating point.

Hotspots: Where the Correction Falls Short

Locations where the simulated contour violates a design rule at one or more points of the process window — for instance because two neighboring lines nearly touch under unfavorable focus (bridging) or a line narrows so much that it nearly breaks (necking) — are flagged as hotspots. Every hotspot found must be fixed before tape-out: through a targeted local re-correction, or if necessary through a manual layout change at that location. Verification then reruns at the affected location until even the worst-case point in the process window stays within tolerance.

The diagram below shows the principle using a process window diagram: the axes represent focus and dose deviation from the target value, the shaded area marks the combinations at which the structure still prints correctly — before and after OPC compared, with a hotspot that only falls inside the window after correction.

The Computational Cost of Full-Chip OPC

A modern chip contains several billion polygons. If every edge is subdivided into dozens of segments for model-based correction and each segment is simulated across multiple iterations, the resulting computational effort poses a serious challenge even for large data centers. Full-chip OPC runs today routinely spread across thousands of CPU cores in parallel and still often require many hours to individual days of compute time — per mask layer, and a modern process has several dozen of those.

This data volume has also changed the mask data format itself: the classic GDSII format, originally designed for comparatively simple, uncorrected layouts, runs into practical limits with the heavily fragmented, serif-rich geometries that result from OPC. The OASIS format was specifically developed to store the substantially larger data volumes after correction more compactly, among other things through more efficient compression of repeating geometry patterns.

EUV: New Wavelength, New Effects

With the transition to EUV lithography (extreme ultraviolet, 13.5 nm wavelength), the starting situation shifts fundamentally: since the wavelength now lies far below the printed feature size, the k1 factor described in Section 1 is considerably more relaxed for many structures than with 193 nm immersion lithography — the classic proximity effects appear weaker in many cases, and the correction demand from mask bias and serifs decreases accordingly.

In exchange, EUV brings its own, novel effects that require their own correction treatment. Since no transmissive lenses exist at this wavelength, EUV lithography operates exclusively with reflective mirror optics, and the light strikes the mask not perpendicularly but at an angle — this gives rise to so-called mask 3D effects, in which the finite thickness of the mask absorber layer itself casts shadows and distorts the image asymmetrically, depending on a structure's orientation relative to the angle of incidence. Added to this is a lower photon count per exposure, leading to statistical fluctuations — so-called stochastic effects — which cannot be fully mastered by classical, deterministic OPC alone.

OPC therefore does not disappear with EUV but transforms: the fundamental principles introduced in this chapter — mask bias, serifs, SRAFs, model-based segment correction, process-window verification — remain valid but are extended with additional model components for mask-3D and stochastic effects. For structures that fall below the resolution limit of a single exposure even with EUV, multi-patterning is added on top: a layout is split across multiple masks whose OPC corrections must be coordinated with one another — a topic beyond the scope of this chapter, but one that builds on exactly the foundations introduced here.