A single-gate invertible XOR from purely dipolar-coupled low-barrier nanomagnets at room temperature: design with LLG and micromagnetic checks (simulation study)
Year: 2026
Type: preprint
Abstract: Probabilistic bits (p-bits) built from low-barrier nanomagnets sample a classical Boltzmann distribution, and a network whose ground states encode a truth table is an invertible gate. In demonstrations with stochastic magnetic tunnel junctions the weights are computed electronically, and invertible gates coupled only by the stray field have so far been designed in a two-state model and corrected afterwards. In this simulation study (classical probabilistic computing, not quantum computing) we design directly in a continuous-angle macrospin model, with the cross-entropy of the truth-table rows as objective, and find an invertible XOR in a single gate at 300 K: four CoFeB ellipses of 80 × 75.8 × 2 nm (E_b = 4 k_BT) and four high-barrier bias magnets in a 400 × 400 nm box, coupled only dipolarly. The converged continuous model gives P_table = 0.820 and worst-row ratio W = 0.696; the stochastic Landau–Lifshitz–Gilbert equation (LLG) with the same readout gives 0.820 ± 0.0005 / 0.696 with the bias magnets frozen and 0.815 ± 0.0004 / 0.712 with all magnets dynamic. A micromagnetic check finds a nearly uniform ellipse (|⟨m⟩| = 0.996) that reverses coherently over a barrier of 3.85 ± 0.1 k_BT, pair couplings within 1 %, and, for the whole gate with renormalised cell magnetisation and frozen bias magnets, P_table = 0.830 ± 0.007 and W = 0.69 ± 0.1, consistent with the macrospin gate. The gate is narrow and fragile: ±1 nm lithographic errors spread the barriers over 2.2–6.0 k_BT and lower the mean W to 0.54, and ±5 % errors swap an easy axis in 37 % of chips. We also report negative results (two layers, perpendicular magnets), a study of rewritable bias magnets, and a free mediator that can invert an effective coupling. The design is released openly. The computations, the code, the analysis, the text and the five referee reports were produced by AI agents (large language models), carried out within the A.U.R.O.R.A. Project, on the workstation that hosts A.U.R.O.R.A., an open-source AI system (Apache-2.0), doi:10.5281/zenodo.23650857. The referees are AI agents, not human reviewers. The author chose the questions and the scope, made the decisions and is responsible for the content. The archive contains every script, the data behind every number, the logs, the figures and their script (python3 paper/make_figures.py regenerates all figures, identically, from the stored data).
Author: Vicari, Francesco
Language: English
Subfield: Condensed Matter Physics
Field: Physics and Astronomy
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7237335247
Cites:
Cited by:
Related to:
