Domain-wall logic computes by pushing magnetic boundaries along nanoscale tracks. A domain wall is the moving boundary between two regions of opposite magnetization. Recent devices are fast: a 50 ps NOT gate and walls moving above 1 km/s. Once gates are wired into a circuit, though, a multi-input gate only works if its inputs arrive inside the same clock window. Superconductor logic has the same requirement, and there the standard fix is path balancing: the design tool inserts delay elements until every converging path takes the same time. Domain-wall logic has a source of delay that superconductor tools never model. Walls stick at material defects, the time they stay stuck is random, and the energy holding them depends on how the defect sits relative to the crystal axes. This proposal treats that orientation dependence as something the designer can choose. Routing a track along one axis or another changes both the average delay of the path and how wide its spread is.
The core claim is that path balancing for pinning-limited wall logic needs two things existing tools lack. First, it needs timing statistics that handle the heavy, skewed delay tails that random pinning produces, which a Gaussian model cannot. Second, it needs an optimizer that picks each track segment's orientation as well as its length. Choosing orientation lets the tool equalize the spread of converging paths and not only their averages. Under realistic pinning that should give higher timing yield for the same track area than length-only balancing. The project builds the delay model, the yield estimator and the optimizer, and checks all three against Monte Carlo simulation on ISCAS-85 circuits. It also includes the experiment that would show the extra machinery is unnecessary.
Why this gap, and what is already known
Three bodies of work frame the problem. Superconductor path balancing supplies the optimization method. SFQmap folds the cost of balancing flip-flops into technology mapping. Delay balancing with clock-follow-data schedules clock arrival with iterative linear programming. The ColdFlux toolchain brings balancing to multi-million-junction designs. All three treat each delay element as a fixed number with bounded random perturbation. ColdFlux handles trapped magnetic flux separately, as a layout rule check, and never as a timing quantity.
Statistical pipeline design motivates the proposal and also limits it. Datta et al. showed that under random variation, perfectly balanced stages maximize the chance that some stage fails timing. They called this the "balanced pipeline paradox." Deliberately unbalancing the stages raised yield by 9% at a 2% area cost. Their analysis rests on Clark's approximation, which assumes the maximum of Gaussian delays is itself Gaussian. We keep their central lesson that the right target is spread and tail, not equal means. We drop the Gaussian assumption.
Defect physics supplies the reason the delay tails are non-Gaussian and orientation-dependent. First-principles calculations on PbTiO3 found that an oxygen vacancy pins a ferroelectric domain wall with about 261 meV when the broken bond lies along the polarization axis, against 97 and 76 meV for the other two orientations. If the time a wall stays pinned is thermally activated, so that it scales as exp(E/kT), the 185 meV gap at room temperature (kT ≈ 25.9 meV) corresponds to roughly a thousandfold difference in waiting time. This is our own illustrative arithmetic. In driven motion the barriers are lowered and the real ratio will be smaller. Continuum dislocation modeling shows a similar effect in a different system: interactions that depend on direction appear only at short range and vanish under naive averaging.
These are ferroelectric and dislocation systems, not the CoGd/Pt ferrimagnet of Ma et al.. We do not borrow their numbers as magnetic facts. We use them as evidence that barrier ratios of a few times are physically ordinary, and we sweep that ratio as a parameter. Ma et al. demonstrate single gates and pinning-resistant skyrmions but do not consider cascades. That is the gap this proposal targets.
Core model
We model a track segment of length $L$ routed at orientation $\theta$ relative to the material's anisotropy axes. Its delay is
$$T(L,\theta) = \frac{L}{v(\theta)} + \sum_{i=1}^{N} \tau_i, \qquad N \sim \mathrm{Poisson}(\rho L), \quad \tau_i \sim \mathrm{Exp}\!\left(\nu\, e^{-E_{c_i}(\theta)/k T_{\mathrm{eff}}}\right).$$
Here $v(\theta)$ is the free wall velocity, $\rho$ is the defect density per unit length, and $c_i$ is the orientation class of defect $i$. The class is drawn from a mixture whose weights depend on $\theta$. $E_c(\theta)$ is that class's effective barrier under the drive current, $\nu$ is an attempt frequency, and $T_{\mathrm{eff}}$ is an effective temperature.
The sum of exponential waits has a closed-form moment-generating function. The per-path delay distribution can therefore be computed numerically to high accuracy without sampling. A gate with converging inputs succeeds when the spread of arrival times fits inside its sampling window $w$:
$$Y_g = P\left(\max_j A_j - \min_j A_j \le w\right),$$
where $A_j$ is the arrival time of input $j$. Circuit yield is the probability that every gate succeeds. Paths that share upstream segments are handled by conditioning on the shared segment's delay. Paths with disjoint segments are treated as independent.
The point of the model is that $\theta$ acts on spread differently from how $L$ does. Adding length moves the mean of a path while its variance grows only slowly. Reorienting a segment can shift which defect class dominates, and with it the size of the tail. Two converging paths with equal means but very different tails fail the window constraint. Matching orientations, or deliberately mismatching them as Datta's paradox suggests, addresses the problem that length alone cannot.
Research program
The work has three components. The first is calibration. Using micromagnetic simulation of single tracks with planted pinning sites, in the style Ma et al. used for velocity, we check that escape times are close to exponential. We also measure how strongly barriers depend on drive. If escape times turn out far from exponential, the model switches to an empirical per-class escape distribution. The rest of the pipeline is unchanged.
The second is a yield estimator built from the equations above. It is validated against brute-force Monte Carlo. The estimator matters only if it is much faster than Monte Carlo while staying accurate in the tails.
The third is the optimizer. It takes a domain-wall gate netlist and assigns each segment a length and one of a small set of routable orientations. The objective is to minimize total track area subject to a target circuit yield. Orientation choices are relaxed and then rounded, following the clock-follow-data iterative linear programming approach. After rounding, a sensitivity-driven refinement step adjusts the design, following Datta's area–delay sensitivity ratio but with tail-aware yield gradients. Netlists come from ISCAS-85 circuits (c2670, c3540, c5315, c6288, c7552), mapped to a generic clocked domain-wall gate library. We state the clocking abstraction, gates that sample inputs inside a window, as an explicit assumption, since the cited device work does not specify one.
Baselines
The naive reference point is the design with no balancing beyond the routed lengths the mapper produces. It is evaluated by Monte Carlo yield at the target clock period, and it shows how bad pinning-induced skew is before any intervention.
The first real baseline is length-only full path balancing. Every converging path is extended until mean arrivals match, and margins are set with Clark's Gaussian model. This is the direct translation of how buffer insertion is done in superconductor flows. It runs on the same netlists and the same pinning parameters.
The strongest existing method is clock-follow-data scheduling by iterative linear programming from Aviles et al.. We adapt it so that clock arrival per gate and segment lengths are its variables, and pinning variation is absorbed through margins. In its home technology it reports a 1.48× average area–delay product improvement over full path balancing and 2.07× over multi-phase clocking. On c2670 it absorbs 20% random timing variation with a 7.5% average delay-line increase, reaching a 26.5 ps clock period against a 30 ps target. Those are superconductor numbers. There is no reported result found on Lacuna for path balancing in domain-wall logic, so all domain-wall comparisons are our own runs.
The adapted scheduler gets the same netlists and the same Monte Carlo yield evaluation. It also gets the same optimization budget: wall-clock time, number of linear programming iterations, and refinement passes. Its margins are tuned by grid search to its best yield-area tradeoff, so it is not run at defaults. We also run Datta-style statistical unbalancing with Gaussian sensitivities as a third comparison.
Beating these baselines means reaching the same Monte Carlo timing yield, for example 90%, with measurably less total track length, or higher yield at equal track length. The improvement must hold across the barrier-ratio sweep and not only at one favorable setting.
Two ablations isolate the ingredients. The first keeps our non-Gaussian yield estimator but fixes all orientations, which separates the value of correct statistics from the value of the orientation knob. The second allows orientation choice but scores yield with Clark's approximation, which shows whether the knob helps only when the tails are modeled.
Why not just…?
Why not just lengthen the shorter tracks? Track length is the obvious delay element in wall logic, and when pinning is weak it may be all that is needed. Our barrier-ratio sweep includes the near-isotropic end, where the length-only baseline should match our method. The proposal predicts a crossover. Above some barrier anisotropy and defect density, length-only designs miss the yield target because extra length adds pinning sites and widens tails instead of narrowing them. If no crossover appears within physically plausible ranges, the orientation idea is wrong for this regime, and we will report that.
Why not just run Monte Carlo inside the existing linear-programming scheduler? This is a real threat for small circuits. With enough samples, Monte Carlo-scored clock-follow-data scheduling could find designs as good as ours. Its weakness is cost. Gate-level yield targets near the tail need many samples for every candidate design, and that cost multiplies across thousands of gates in c6288 or c7552. We test it directly by giving the Monte Carlo-in-the-loop variant the same compute budget. If it matches us at equal budget on the large circuits, the analytic estimator is not a contribution and the claim narrows to the orientation knob alone.
Why not just use Gaussian statistical sizing as Datta et al. did? Their method is the right idea with the wrong distribution. Clark's approximation is known to accumulate error in deep pipelines, and a sum of exponential waits drawn from a mixture of rates is skewed and can be multimodal. The second ablation measures how much yield the Gaussian model mispredicts and whether that misprediction changes the chosen design. If both estimators lead to the same designs, the non-Gaussian machinery is unnecessary.
Why not just use skyrmions, which avoid pinning? Ma et al. show skyrmion bubbles resisting defect pinning, but their demonstrated skyrmions are micrometer-sized, while their fast 50 nm gate uses domain walls. Converging skyrmion paths of different lengths would still need balancing. Skyrmions therefore set a different value of $\rho$, near zero, not a different problem, and our model covers that case as the zero-pinning limit.
Decisive checks
The first check is estimator fidelity. Gate-level and circuit-level yield predicted by the analytic estimator must agree with 10⁶-sample Monte Carlo. We report this agreement separately for the tail region, where yield exceeds 90%, and compare it to the error of Clark's approximation on the same circuits.
The second check is the yield–area frontier. For each ISCAS-85 circuit and each barrier ratio, we plot timing yield against total track length for the naive design, length-only balancing, the adapted clock-follow-data scheduler, Datta-style unbalancing, the two ablations, and our method. The proposal succeeds if its frontier dominates the adapted scheduler's at barrier ratios at or above the ratio implied by the PbTiO3 values, which is about 3.4. It also needs the orientation-only ablation to account for a substantial share of the gain.
The third check is physical plausibility. The micromagnetic calibration must show that escape-time statistics differ measurably between track orientations in a ferrimagnetic stack. If they do not, the orientation knob does nothing in the target material, and the work reduces to a tail-aware balancing tool for wall logic. That would still be useful, but it is a smaller claim than the one stated here.
References
- SFQmap: A Technology Mapping Tool for Single Flux Quantum Logic Circuits
- Delay Balancing With Clock-Follow-Data: Optimizing Area Delay Tradeoffs for Robust Rapid Single Flux Quantum Circuits
- Results From the ColdFlux Superconductor Integrated Circuit Design Tool Project
- Delay Modeling and Statistical Design of Pipelined Circuit Under Process Variation
- First-principles study of oxygen-vacancy pinning of domain walls in<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:msub><mml:mrow><mml:mi mathvariant="normal">PbTiO</mml:mi></mml:mrow><mml:mrow><mml:mn>3</mml:mn></mml:mrow></mml:msub></mml:mrow></mml:math>
- A continuum model for distributions of dislocations incorporating short-range interactions
- Ultrafast and reliable domain-wall and skyrmion logic in a chirally coupled ferrimagnet