Real-Time Error Feedforward for Write-Field Stitching in Maskless Lithography

Where field placement error comes from

Each write field inherits error from three stacked sources. The stage positions the field center with its own accuracy and repeatability; the substrate holder deforms and expands with temperature; and the deflection optics map the beam with residual scale, rotation and distortion of their own. Stage and holder errors are position-dependent—the same field number lands differently depending on where it sits on the travel—while deflection errors are stable in the beam frame and appear as a fixed distortion pattern in every field.

The composite signature is what stitching sees: two adjacent fields written minutes apart, each with slightly different placement transforms, produce a seam whose mismatch is the difference of their errors. Because the components have different time constants—deflection distortion stable for hours, holder drift following thermal time constants of minutes to hours, stage error following position—the composite error is best attacked piecewise, with each correction channel handling the component whose timescale matches its bandwidth.

Layer one: the static field grid

The foundation is a calibration pass that writes alignment-rich marks at every field boundary or corner, measures the written grid against the ideal, and solves a per-field correction: translation, gain, rotation and, where the optics support it, keystone terms. This grid correction removes everything repeatable, which on a thermally settled tool is usually the majority of stitching error.

Two disciplines make the static layer trustworthy. First, measure the grid under the same thermal state the production job will see—a grid calibrated cold degrades within the first hour of writing as the tool warms. Second, timestamp and archive each grid: comparing successive calibrations reveals which components are drifting and how fast, and that drift rate is the specification the next correction layer must meet.

• Expose a calibration grid covering all field boundaries, not a sampled subset
• Solve translation, scale, rotation and distortion per field with stated uncertainty
• Calibrate in the production thermal state, after the standard warm-up routine
• Archive grids with timestamps; the drift between them sizes the next layer's job

Layer two: encoder-interpolated feedforward during the job

Between calibration and the fields written hours later, the world moves. Encoder-interpolated feedforward closes that gap by correcting each field from live data: as the stage parks at or passes each field position, current encoder readings feed a model that updates the field's placement correction. The model is built from the archived grids—correction value regressed against stage position and holder temperature—so the live reading is mapped to the correction the calibration history says applies.

The critical specification is latency. If encoder data reaches the pattern generator tens of milliseconds before the field starts, position-dependent error is corrected cleanly; if it arrives after the field is written, the architecture degenerates to feedback. The data path—encoder interface, controller, pattern generator—must publish a measured end-to-end latency, and the write sequence must guarantee a stable position sample inside that window before the field trigger fires.

Layer three: deflection feedforward inside the field

Whatever the first two layers leave lands inside the field, where the deflection channel can still act. Per-field correction values—residual offset and distortion after the grid and interpolated updates—drive the beam deflection during exposure, repositioning sub-elements of the pattern at deflector speed rather than stage speed. This is what absorbs the fast tail: stage settle residuals, encoder quantization at field corners, and slow stage motion during write-on-the-fly writing.

The deflection layer has its own error and its own range. Corrections beyond a few percent of the field size stress deflector linearity, and the deflector's own calibration drift must be covered by the static grid. Keep the correction hierarchy explicit: stage and holder errors in the field placement path, deflector errors in the deflection path, and a documented handoff so no error is corrected twice—or not at all.

Qualifying the feedforward chain

Feedforward earns trust the same way any metrology system does: by measured evidence against stated conditions. The qualification set below covers the failure modes that matter—a stale grid, a slow data path, a deflector running out of range—and doubles as the ongoing maintenance check, since drift behavior changes as tools age and environments shift with the seasons.

Run the full set after commissioning, after any change to the data path or optics, and at intervals long enough to expose seasonal drift. A tool that passes all four has a stitching architecture that is understood, not just adjusted.

• Full-substrate stitch exposure: seam error at every adjacent field pair, compared to the metrology floor
• Long-job drift test: stitch error logged versus elapsed time across a realistic production duty cycle
• Latency verification: measured encoder-to-deflection delay against the architecture's requirement
• Correction-range audit: deflection corrections stayed inside the linear range for every field