Oxides, not layouts, still set the ceiling (09/10/26)

TL;DR This review argues that scalable superconducting processors are now limited by uncontrolled surfaces, buried interfaces, and chemically ambiguous defects rather than circuit layout, and that no platform yet quantitatively links a specific atomic structure to a measured coherence change. For cQED hardware, the actionable core is the materials gap at metal–dielectric surfaces and Josephson junctions, plus the call for atomistic decoherence mechanisms, high-throughput proxies of device performance, and characterization tools built for quantum materials.

• Main claim: Across superconducting circuits, solid-state defects, quantum photonics, and 2D heterostructures, coherence and scalability are bottlenecked by poorly controlled surfaces, buried interfaces, and defects whose atomic identities remain incomplete.
• Superconducting focus: The superconducting-qubit section organizes the materials problem around surface chemistry/interface engineering, buried Josephson-junction interfaces, fabrication-driven loss, and the lack of predictive materials metrics that track device $T_1$/$T_2$ at scale.
• Cross-platform diagnosis: A shared structure–coherence problem recurs: light-element chemistry in disordered or buried environments correlates with loss and noise, yet no platform closes the loop from a known atomic configuration to a quantified coherence shift.
• Open needs: The authors prioritize three deliverables—atomistic mechanisms of decoherence, high-throughput proxy metrics predictive of device performance, and characterization methods designed for quantum materials—as the path to better coherence, yield, and integration.
• Hardware relevance: For multi-qubit cQED, the review reframes materials process control and junction/surface metrology as first-order device engineering, not secondary materials science.

TL;DR NEGF-DFT transport through equal-thickness Al/${\mathrm{Al}}_2{\mathrm{O}}_3$/Al barriers gives mean $E_J/h=2.78$ GHz ($\sigma =4.67$ GHz) for a 10-member amorphous ensemble versus $0.73$ GHz for crystalline ${\mathrm{Al}}_2{\mathrm{O}}_3$, with amorphous values spanning nearly two decades; Al-rich percolation pathways drive that spread, while a $200\times 200$ ${\mathrm{nm}}^2$ junction averages $>2\times 10^4$ such patches—directly tying oxide microstructure to the transmon $E_J$ scale.

• Main result: Ambegaokar–Baratoff $E_J$ from Fermi-level transmission yields amorphous mean $E_J/h=2.78$ GHz with $\sigma =4.67$ GHz, versus $0.73$ GHz for the crystalline reference at the same barrier thickness.
• Variability: Individual amorphous $E_J$ values span nearly two orders of magnitude, set by stoichiometric inhomogeneity rather than uniform barrier thickness.
• Mechanism: Scattering-state analysis shows pure tunneling; Al-rich, low-barrier regions form percolation-like paths that strongly boost conductance.
• Device scale: A realistic $200\times 200$ ${\mathrm{nm}}^2$ junction self-averages over $>2\times 10^4$ microscopic regions, suppressing single-patch extremes in wafer-scale $E_J$.
• Method: First-principles NEGF-DFT device modeling compares one crystalline ${\mathrm{Al}}_2{\mathrm{O}}_3$ barrier to ten melt–quenched amorphous ${\mathrm{Al}}_2{\mathrm{O}}_3$ barriers of identical thickness.

TL;DR On BSC's five-qubit Red segment, a corrected depth-1 QAOA cost layer yields certified bipartite entanglement (tomographic negativity $0.077$, BCa 95% CI $[0.065,0.091]$) only on the good coupler; a pre-registered CZ-amplitude sweep then shows that drive strength causally sets that negativity via a peaked $|\zeta |$ response, while depth-3 certificates and optimisation both fail.

• Main result: Across six calibrations, two-qubit tomographic negativity of the corrected depth-1 QAOA cost layer on the good coupler is $0.077$ (BCa 95% CI $[0.065,0.091]$), excluding zero every day; the $J=0$ twin stays separable and the degraded coupler is consistent with zero.
• Causal device physics: On a fixed Bell preparation (not the QAOA layer), reducing CZ pulse amplitude lowers certified negativity along a positive, well-fit slope ($R^2=0.83$–$0.96$), with every pre-registered rung-level interval excluding zero after Holm and Benjamini–Hochberg correction.
• Non-monotone CZ response: The negativity peaks at conditional phase $|\zeta |=\pi$; the nominal amplitude overshoots that peak, so the linear slope is a local linearisation of a peaked response rather than a global monotone law.
• Coherence controls: Constant fringe visibility plus the separate $J=0$ off-state twin support a coherent coupling effect; average gate fidelity is non-monotone in drive and is reported only as a diagnostic.
• Depth limit: Entanglement certification collapses by depth three, and a pre-registered sixteen-instance depth-3 ensemble shows no optimisation signal (one-sided Wilcoxon $p=0.15$); the authors claim no advantage on classically trivial instances.

TL;DR The paper builds a fault-tolerant factory that cultivates the entangled non-Clifford resource $|CS⟩=CS|++⟩$ directly—rather than synthesizing it from three $T$ states—with a protected branch-label record and a certified fault-order-three logical-error channel. At matched binary logical error rate, the direct route uses fewer operations at lower physical noise and loses that edge only between $8\times 10^{-4}$ and $9\times 10^{-4}$.

• Target state: Magic-state cultivation prepares non-Clifford resources needed for universal fault-tolerant computation; here the target is the two-qubit entangled state $|CS⟩=CS|++⟩$, obtained by applying the controlled-$S$ gate to $|++⟩$.
• Branch record: Two commuting Clifford involutions (self-inverse Clifford operators used as logical projections) select four usable outcome branches related by Pauli-frame updates; a minimal six-bit $[6,2,4]$ classical record protects the branch label against readout errors that would otherwise map one valid record onto another.
• Factory stack: Verified ${\mathrm{CAT}}_7$ ancillas, Steane-code error detection, CZZ-based controlled checks, and immediate Steane-to-surface expansion ($d=5\to 13$) form the factory; optimizing that expansion cuts accepted-output operation count by about $35$–$37\%$.
• Fault certificate: Under an active-location stochastic-Pauli noise model, exact enumeration finds no accepted closed-boundary logical failure through fault order two, while explicit order-three failure mechanisms exist; the decoder uses $263$ operational detector bits (plus $168$ withheld for validation) via low-order exact resolution, a higher-order catalogue, and four-coset BP+OSD.
• Cost vs three-$T$: At fixed binary logical-error rate, direct $|CS⟩$ remains cheaper in operation count at the two lower-noise benchmarks even if the three-$T$ route already has output patches; the ordering reverses between $8\times 10^{-4}$ and $9\times 10^{-4}$—relevant for superconducting surface-code roadmaps where magic-state volume dominates runtime.

TL;DR Shallow phase shadows use a sparse Clifford–IQP (instantaneous quantum polynomial) ensemble of random $H$/$S$/$CZ$ circuits to estimate stabilizer-state fidelities with provable sample efficiency, even though the ensemble is not an approximate relative-error 2-design. On all-to-all connectivity the protocol runs at constant depth with $O(n\log n)$ auxiliaries and mid-circuit feedforward, or at $O(\log n)$ depth without auxiliaries—cutting the usual linear-depth cost of global Clifford shadows.

• Main result: Shallow phase shadows (SPS) estimate fidelities to stabilizer states—states fixed by a commuting Pauli group, including many QEC codewords and graph/cluster states—from randomized measurements that need far less randomness than a full Clifford 3-design.
• Circuit: Each shot applies $U=H^{\otimes n}D$ with sparse diagonal $D$ built from random $S$ and $CZ$ gates; off-diagonal $CZ$s appear with probability $p=(\gamma \ln n)/n$, so sparsity $\gamma$ trades gate count against estimation bias.
• Depth: On all-to-all hardware, commuting $CZ$s plus mid-circuit measurement and classical feedforward yield $O(1)$ depth using $O(n\log n)$ auxiliary qubits (GHZ-like “rails”), or $O(\log n)$ two-qubit depth with no auxiliaries.
• Second-look principle: Because the sparse ensemble is not a relative-error state 2-design, a single post-processing track can leave $O(1)$ bias; measuring a complementary track (e.g., a Hadamard-conjugated observable) and keeping the lower-bias estimate restores vanishing bias for stabilizer fidelities.
• Hardware link: The protocol uses only native long-range $CZ$s and is framed for ions/neutral atoms; for superconducting chips the takeaway is that task-adapted, $CZ$-only shadows plus mid-circuit feedforward can shrink global characterization depth when long-range or bus-mediated connectivity is available.