Binary-Coded Track Offsets for Mid-Frequency-Aware Polishing Path Planning

When a small polishing tool sweeps an optical surface in parallel passes, the material removed is the sum of overlapping copies of the tool's removal footprint, one copy per pass. This sum is effectively a convolution of the footprint with the path. If the passes are spaced badly, the overlap leaves periodic ripples called mid-spatial-frequency (MSF) errors. These ripples scatter light and limit what high-end optics can achieve. Current practice picks one uniform spacing for the whole part, found by a one-dimensional search. We propose a different approach: plan the position of every individual pass as a discrete optimization. Each pass's offset from a nominal grid is written as a short binary number, so a pass placed among $K$ candidate positions costs $\log_2 K$ yes/no variables instead of $K$.

The central claim connects the two fields. For a smooth, bell-shaped tool footprint, the leftover surface error is close to a quadratic function of those binary digits. Positions coded with powers-of-two weights are linear in the bits, and removal is nearly linear in small position shifts. The logarithmic code is therefore a natural fit for the physics, not a lossy compression. For sharp or "W-shaped" footprints this near-linearity breaks down. The same footprint property that makes ripples hard to suppress should then also make compact encodings lose their advantage. This prediction can be tested, and it determines when logarithmic encoding is worth using in manufacturing planning.

Background and the gap

He et al. model overlapping removal ribbons as a convolution of the tool influence function (TIF, the removal profile of one stationary or one-pass tool contact) with the path. They show that W-shaped and trapezoidal profiles produce much worse ripple than Gaussian ones: at 1 mm spacing, the W-shape gives 32.56 nm RMS and the Gaussian 17.39 nm. They also find that RMS error is "serrated" in path spacing, so a wider spacing can beat a narrower one. Their iterative search over a single uniform pitch reaches below 2 nm RMS on fused silica at 0.4 mm and 1.33 nm RMS on single-crystal silicon at 0.36 mm. This paper motivates our work and supplies our forward model. Our main departure is that we let pitch vary pass by pass across the part.

Song et al. reshape a rigid conformal lap into a Gaussian-like TIF. They also show that the Preston removal coefficient falls at high tool speed, with an optimum near 300 RPM on silicon carbide. For us this result is a constraint. Removal is not linear in speed outside a band, so per-pass removal rates must stay within a bounded feed range rather than being free continuous variables.

On the encoding side, three papers show that binary index codes beat one-hot codes. One-hot means one variable per option with an "exactly one" constraint. Parmentier and Vidal's OCEAN uses logarithmic binaries in a mixed-integer program and cuts the formulation from about 163,000 to about 18,000 nonzeros on Adult. It solves in minutes where earlier exact formulations ran past 5 hours. Dulac-Arnold et al. factor reinforcement learning over action-code bits for a 12.5x speedup on 100-action Mountain Car. Majeed and Hutter bound surrogate state spaces polynomially in $\log|A|$.

These works also share a limitation that both sets of authors flag. Their codes are arbitrary, and nearby options may receive distant bitstrings. Both point to "informed codes" as open. Our proposal supplies such a code for a physical setting, and argues that the right notion of "informed" here is linearity of position in the bits, not small Hamming distance between neighbors. We do not use reinforcement learning. Our forward model is known and static, so planning is an optimization problem, not a learning problem. The RL papers motivate the encoding but do not supply our method.

Core Claim

Consider a cross-section with $T$ passes. Pass $t$ sits at position $x_t$ with removal weight $w_t$, set by feed rate. The residual surface is

$$r(x) = z(x) - \sum_{t=1}^{T} w_t\, f(x - x_t),$$

where $z$ is the target removal map and $f$ is the TIF. We write $x_t = x_t^0 + \delta \sum_{j=0}^{d-1} 2^j b_{tj}$, where $x_t^0$ is a nominal grid point, $\delta$ is the offset resolution, $b_{tj}\in\{0,1\}$ are the bits, and $d=\log_2 K$. We choose $K$ to be a power of two, so no bit pattern is left unused.

Expanding $f$ to first order around $x_t^0$ makes the band-limited MSF error $\|r\|^2$ quadratic in the bits and the weights jointly. The error of this surrogate is controlled by $\Delta^2 \max|f''|$, where $\Delta = K\delta$ is the offset range; this is the standard Taylor remainder. For a Gaussian of width $\sigma$, $\max|f''|\sim 1/\sigma^2$, so offsets up to a fraction of $\sigma$ are well modeled. A W-shaped or volcano-shaped footprint has much larger curvature, so the usable range shrinks sharply.

This gives two predictions. First, the binary-coded mixed-integer quadratic program, re-linearized in an outer loop, should match or beat one-hot formulations in solve time at equal solution quality, with an advantage that grows with $K$. Second, that advantage should shrink as the TIF departs from Gaussian, because more re-linearizations are needed and the surrogate misleads the solver. A corollary runs against common practice. Gray coding, the usual fix for "Hamming cliffs," should hurt here, because Gray-to-position decoding is nonlinear in the bits.

Research Program

The planning objective follows He et al.'s framing: maximize productivity, measured as fewest passes or shortest total path, subject to MSF RMS below a target. We solve it as an outer search over the pass count $T$ and an inner binary-coded quadratic program over offsets and bounded weights. Each inner solve updates the linearization point to the current positions and repeats until positions stop changing. A final exact convolution simulation reports the true residual.

The work has three stages. The first is a simulator. It uses He et al.'s three profile families (Gaussian, W-shape, trapezoid) and a continuous family between them, generated by mixing a Gaussian with a ring term to mimic Song et al.'s volcano shape. Errors are evaluated in the MSF band. The second stage covers one-dimensional cross-sections with uniform and non-uniform target maps. The third stage is a two-dimensional freeform extension: passes are spiral or raster tracks on a surface whose TIF width varies with local curvature, so the best pitch changes across the part. This is where a single uniform pitch should fail most clearly, and where a pass now interacts with many neighbors because the TIF is several pitches wide.

Physical validation on fused silica, with Zygo interferometry, is a desirable later stage. It depends on access to a sub-aperture machine and is not assumed for the core claims.

Baselines

The naive reference is a single uniform pitch chosen by dense one-dimensional sweep, reproducing He et al.'s procedure in our simulator. We first confirm that the simulator recovers their serrated RMS curve and the qualitative Gaussian-versus-W gap at 1 mm. The reported numbers to beat come from He et al.: RMS below 2 nm on fused silica at 0.4 mm pitch, and 1.33 nm RMS on silicon at 0.36 mm. In simulation, beating this baseline means one of two things. Either we reach the same RMS target with at least 10% fewer passes or shorter path, or we reach a lower RMS at equal pass count, on targets where the optimal pitch is not uniform. On uniform flat targets we expect at most a tie, and we report it.

The strongest exact competitor is the same mixed-integer quadratic program with one-hot offsets, which uses $K$ binaries per pass and a special-ordered-set constraint. We give it the same solver, time limit, re-linearization loop and warm start. Modern solvers handle such sets well, so this is a real contest. Beating it means lower time-to-target-gap at $K\ge 32$ with no loss in final RMS.

For one-dimensional problems we add exact dynamic programming over pass positions. Its state is the last $m$ positions, where $m$ is the number of passes within one TIF support. This is the exact reference wherever $K^m$ is tractable, and we use it to measure our optimality gap.

We also include multistart continuous optimization of positions and weights with quasi-Newton steps. Finally, an ablation isolates the key ingredient by holding everything else fixed and replacing the dyadic binary code with a Gray code and with a random code.

Why not just…?

Why not just keep a single uniform pitch and tune the tool shape? Song et al. and He et al. show that a Gaussian TIF helps a great deal, yet He et al. still need careful pitch selection to go from 17.39 nm at 1 mm to 1.33 nm. Their uniform pitch cannot adapt to curvature-varying TIF width on freeforms. The two-dimensional freeform experiment directly measures how much a uniform pitch gives up there. If the loss is small, the proposal reduces to a negative result, and we say so.

Why not just solve a nonnegative least-squares dwell problem on a fine fixed lattice? This approach is convex and standard, and it is the most serious threat to the proposal. Its weakness is that it returns dense solutions with many passes at fine pitch, the opposite of the productivity goal. Feed-rate bounds tied to the Preston velocity effect also limit how freely the weights can vary. We include a sparsity-penalized version (an L1 penalty on the weights) with a final pruning step as an additional baseline. If it matches our pass counts at equal RMS, the discrete formulation is unnecessary for that regime.

Why not just use one-hot variables or dynamic programming, which are simpler and exact? Both are run as baselines. Dynamic programming is exact but its state grows as $K^m$, and $m$ is large whenever the TIF is several pitches wide. One-hot formulations grow linearly in $K$ per pass. OCEAN's sparsity gains suggest the binary code should scale better, but that evidence comes from tree ensembles, not convolution objectives. The scaling study settles the question either way.

Why not just use reinforcement learning, following CHEQ? Cramer et al. learn impedance along a fixed path and leave path generation open. Our forward model is known, so learning adds sample cost without adding information. Our planned paths could serve as the input that a controller like CHEQ then tracks.

Proposed Experiments

The decisive experiment sweeps TIF shape from Gaussian to W-shape and $K$ from 8 to 256, on one-dimensional and freeform targets. For each setting we record three quantities. The first is surrogate error, measured as linearized versus exact residual. The second is solve time to a 1% gap for binary, one-hot, Gray and random codes. The third is final MSF RMS against pass count, compared with uniform-pitch and sparse least-squares baselines.

The proposal is supported if three things hold together. Binary coding must beat one-hot in time-to-target at large $K$ for near-Gaussian footprints. That advantage must shrink as the footprint becomes W-shaped. Gray and random codes must underperform dyadic binary.

If binary loses to one-hot even for Gaussian footprints, the transfer from OCEAN fails for convolution objectives. If sparse least squares matches our pass counts, the discrete planner is unnecessary. Both outcomes are reportable, and either would show practitioners when not to reach for compact encodings.