Line Edge Roughness in Maskless Lithography: Understanding LER/LWR Sources and Getting Them Under Control

What LER and LWR actually measure

LER is conventionally the 3-sigma standard deviation of one resist edge position sampled along a line, over a stated evaluation length. LWR applies the same statistic to the width between the two edges of a feature. For uncorrelated edges, LWR is roughly the root-sum-square of the two edge roughnesses—about 1.4 times the single-edge value—so a 4 nm LER and a 6 nm LWR can describe the same wall. Neither number means anything without its conditions: evaluation length, line count, resist, measurement tool and image-analysis convention all move the reported value by tens of percent.

Roughness matters because lithography's job is not to make one good line but to make every line the same. In photonics, edge roughness couples light out of waveguides and directly inflates propagation loss. In MEMS and NEMS, gap-width variation shifts resonant frequencies and pull-in voltages across a die. In sensors and quantum devices, critical junction dimensions inherit the roughness floor as device-to-device spread. In all of these, the roughness component of variation is invisible in a nominal CD number and only appears in statistics—until it appears in yield.

Where roughness comes from in DMD maskless exposure

The first floor is statistical. A resist edge is a molecular decision boundary decided by a finite number of absorbed photons interacting with a finite number of polymer molecules; the randomness of both shows up as edge position fluctuation. The magnitude scales inversely with the dose and the chemical gradient at the edge, which is why low-dose, high-resolution processes pay a roughness penalty and why resist chemistry—contrast, molecular size, acid diffusion length—sets a floor no tool can beat.

The second floor is the digital mask itself. A DMD imaged through an objective renders the pattern on a pixel grid; a target edge that does not land exactly on the grid is approximated, and any rendering that quantizes the edge position adds step-like roughness correlated with the pixel pitch. Because the projection pixel size differs between objectives and systems, the same GDS file can expose with different edge quality on different configurations. Sub-pixel grayscale rendering, dithering and orientation-aware placement of critical geometry push this contribution below the statistical floor; ignoring it leaves the mask as the dominant roughness source for CDs close to the pixel size.

The third family is process condition. Dose error and focus error both convert smoothly into edge degradation: under-dose edges are noisy and poorly anchored, out-of-focus edges get footed and wider with ragged walls. Bake quality decides whether standing waves scallop the sidewall, and development conditions decide how aggressively the developer attacks the rough edge. Finally, stitching at write-field boundaries produces displacement that is periodic and localized—not roughness in the statistical sense, but easily mistaken for it unless the analysis separates the two.

Measuring LER/LWR so the numbers are worth something

SEM is the workhorse: image the line at a magnification where the pixel size in the image is well below the roughness scale of interest, extract the edge profile with a fixed threshold convention, and compute the standard deviation over the defined length. AFM, especially CD-AFM, complements SEM where sidewall access matters or where charging makes SEM difficult on some substrates. Whatever the tool, three disciplines keep the number honest: fixed imaging and threshold settings, enough lines and locations for the statistic to mean something, and a repeated reference measurement to bound session-to-session drift.

Spectral analysis turns a single number into diagnosis. The spatial frequency content of the edge profile separates the sources: white-noise-like fluctuation points at photon and resist statistics, peaks at the pixel pitch point at mask rendering, long-period content points at dose or focus drift across the field, and narrow localized spikes at known positions point at stitching. Without the spectrum, engineers routinely average all of these into one LER value and then optimize the wrong parameter.

For qualification and purchase, the requirement is that the supplier and the process owner agree on the measurement plan in advance: which structures, how many lines, which evaluation length, which tool and convention, and what repeatability demonstration accompanies the numbers. An LER specification without a metrology plan is not a specification; it is an aspiration that will be renegotiated at the first discrepancy.

• SEM pixel size at least 5–10x finer than the roughness scale
• Fixed threshold and edge-extraction convention across all measurements
• Enough lines, fields and locations for a meaningful 3-sigma
• Spectral analysis to attribute roughness before optimizing
• Reference line set re-measured every session

Reducing roughness in a maskless process, in order

Start by separating the contributors, because the order of attack matters. First qualify the dose and focus windows with a matrix that records CD and roughness together—then hold the process at the roughness-optimal point inside both windows, not at the CD-nominal point that happens to sit on a roughness cliff. Second, check the pixel-grid contribution by exposing the same critical geometry at different orientations or with different rendering settings; anisotropy or a dependence on rendering is a mask-side signal that data preparation can fix. Third, optimize bake and development with cross-sections, since sidewall profile is the most direct read of what the chemistry is doing to the edge.

Only after these is resist selection the lever. Higher-contrast chemistries and formulations with larger effective smoothing length reduce the statistical floor, at the usual price in sensitivity, resolution or etch resistance. In a maskless workflow the resist choice interacts with the available wavelength and the dose the tool can deliver economically, so resist optimization belongs after the tool-side parameters are already controlled—not first, where trial-and-error resist changes are most expensive and least diagnostic.