Line-edge roughness (LER) is a dominant contributor to critical-dimension (CD) variability as lithographic features approach the nanometer scale. While SEM is the conventional workhorse for roughness monitoring, SEM-based LER suffers from noise bias, digitization artifacts, and limited spatial-frequency bandwidth. Atomic force microscopy (AFM) offers a geometry-based alternative by directly measuring topography, but AFM-based LER is itself sensitive to analysis choices.

This work proposes a practical, reproducible framework for quantitative AFM-based LER metrology of lithographic line/space patterns. The framework addresses four AFM-specific sources of variability:

1. Edge definition — two complementary definitions are compared:

   • a height-threshold method evaluated at the same relative height used for CD extraction, and
   • a model-based definition placing the edge at the center of an error-function (erf) transition (the point of maximum gradient).

   We show analytically that the height-threshold method at t = 0.5 is geometrically consistent with the erf center for approximately symmetric transitions, and confirm this experimentally for the PMMA patterns studied.

2. Effective resolution — pixel-size convergence study; measured LER plateaus below ~1–2 nm/pixel for this dataset (used as an experimental convergence criterion, not a universal probe-resolution prescription).

3. Probe convolution — tip-shape effects on the measured edge.

4. Statistical confidence — confidence intervals derived from an effective sample size estimated from the correlation length, accounting for correlated samples along the edge.

The framework is validated on electron-beam-lithographed PMMA line/space patterns (250 nm and 100 nm half-pitch, two PMMA molecular weights: 192 kDa and 904 kDa), and cross-validated against an independent in-house workflow (LACERM).

The data/ directory is a placeholder for the AFM measurement data used in the manuscript. LER, PSD, correlation length, and roughness exponent in the paper were computed with the LACERM software from binary edge maps produced by the edge-detection methods described in the manuscript. Processed data supporting the findings are available from the corresponding author on reasonable request.