betaFeedback

WorksPhysics and Astronomy

Query-Efficient Quantum Multi-Objective Optimization for Materials Configuration Search via Bayesian Weight Selection

Abstract

Materials design is a multi-objective problem, and recent advances in quantum optimization make it possible to bring quantum solvers to bear on such problems. The realistic near-term win, though, is not raw quantum advantage but making an expensive, noisy quantum solver usable by spending its calls wisely. Quantum Approximate Multi-Objective Optimization (QAMOO) traces a Pareto front by solving a family of scalarized Ising problems, one per objective-weight vector, with QAOA — but chooses those weights non-adaptively, sampling them at random, blind to what it has already found. We recast weight selection as multi-objective Bayesian optimization: our method, BW-QAMOO (Bayes-Weighted QAMOO), picks each next weight by expected hypervolume improvement (EHVI) under a Gaussian-process (GP) surrogate over the weight-to-objective map. On a frustrated alloy-configuration benchmark across m∈{3,4}m\in\{3,4\} objectives and n∈{10,11,12}n\in\{10,11,12\} sites, BW-QAMOO reaches the same Pareto hypervolume as QAMOO's random sampling with a median 33% fewer scalarized solves across the 90/95% targets and 60–74% fewer than a uniform-grid baseline), on a genuine (simulated) QAOA solver; a paired Wilcoxon test confirms the saving is significant, and the advantage grows with the objective count m as the weight simplex expands and blind sampling wastes more of the budget. This is a query-efficiency result, and we close with the open challenges on the path to real alloy data and hardware.

Introduction

Designing a material means navigating trade-offs, not optimizing a single number. An alloy is chosen for a compromise among phase stability, cost, density, and chemical disorder, and the design variable is combinatorial: which element sits on which lattice site. This configuration-search problem maps cleanly onto an Ising model, E(σ)=∑ihiσi+∑i<jJijσiσjE(\boldsymbol{\sigma}) = \sum_i h_i\sigma_i + \sum_{i<j} J_{ij}\sigma_i\sigma_j, with spins σi∈{−1,+1}\sigma_i \in\{-1, +1\} encoding site occupancy and couplings JijJ_{ij} the effective interactions [3]. This is exactly the class of problem quantum optimizers are built to attack, which is what has made materials design one of the most active near-term targets for quantum optimization (Appendix A reviews the Ising model and this mapping for the unfamiliar reader).

Quantum optimization for materials, and why it needs help. The evidence for that interest is concrete. Quantum annealers have searched crystal elemental configurations [4] and optimized high-entropy alloy lattices [5], and gate-model devices attack the same energies with the Quantum Approximate Optimization Algorithm (QAOA) [2]. Two facts temper this promise, however. First, on today’s noisy hardware these methods rarely beat good classical solvers outright, so raw quantum advantage is not yet the right thing to claim. Second, and more practically, each solve is expensive: a single QAOA run couples an iterative classical parameter search to repeated circuit sampling and, on real devices, to cloud-QPU queue latency, which together cost seconds to minutes and real dollars per solve [2]. Taken together, these two facts point to a different near-term payoff. The goal is not to outrun classical solvers but to make the quantum tool usable, by wrapping it in classical AI that spends its scarce, costly calls wisely. We therefore treat the quantum solver as an expensive oracle and let classical learning decide what to ask it.

The multi-objective bottleneck, and QAMOO. Deciding what to ask is harder than it first appears, because a material is rarely chosen to minimize one quantity. The same configuration must do well on several competing energies at once, and improving one typically worsens another. This is a poor fit for the bulk of quantum optimizers, which are built for a single objective. The right target is therefore not a single minimizer but the Pareto front: the set of configurations for which no objective can improve without sacrificing another. Quantum Approximate Multi-Objective Optimization (QAMOO) [1] traces this front by scalarization. A weight vector w\mathbf{w} on the simplex collapses the mm objectives into a single Ising cost Ew=∑kwkEkE_{\mathbf{w}} = \sum_k w_k E_k; QAOA minimizes it; and the solution is recorded as one Pareto point (Figure 1). Sweeping w\mathbf{w} over the simplex then produces the front. Our benchmark makes this concrete with m∈{3,4}m \in\{3,4\} competing energies, a deliberately simplified stand-in for the objectives a crystal or alloy designer trades off (formation energy, elastic mismatch, chemical short-range order, and cost), each expressed as its own Ising energy (Appendix D).

Figure 1. The QAMOO pipeline and where BW-QAMOO’s Bayesian weight selection enters. (a) A materials-design instance under several competing configurational energies. (b) Occupancy is encoded as Ising spins σi∈{−1,+1}\sigma_i \in\{-1, +1\}; a weight vector w\mathbf{w} scalarizes the mm energies into a single Ising cost, minimized by one QAOA run. (c) QAMOO chooses the weights non-adaptively, sampling them at random (red). (d) BW-QAMOO instead fits a Gaussian-process surrogate over the weight simplex and picks the next weight by an expected-hypervolume-improvement acquisition (bright regions indicate high acquisition values), so the query goes to a bright, still-unsampled region (explore) rather than the well-sampled cluster (dark, exploit).

The catch is that each scalarized solve is a separate costly quantum query, and QAMOO chooses the weights non-adaptively. Contrary to a natural reading, QAMOO does not grid the simplex; it samples at random, “uniformly at random from the feasible set,” without enumerating an exponential grid [1]. Either way the choice is oblivious to what has already been discovered. This oversight grows more costly as the problem scales: for mm objectives the simplex is (m−1)(m-1)-dimensional, so blind sampling wastes an increasing share of the budget as mm grows. QAMOO itself names better weight selection, “non-uniform sampling of weights or other scalarization strategies,” as an explicit open direction [1], and it is exactly this gap we set out to fill. We ask a simple question: can we reach the same front with fewer quantum queries by choosing the weights adaptively?

Our answer is yes, with a precise scope. We first implement QAMOO on a materials-design instance, then make it query-efficient through a novel integration with multi-objective Bayesian optimization (BO). We treat weight selection as a sequential decision problem and choose each next weight by an expected-hypervolume-improvement acquisition (EHVI [11]) under a Gaussian-process surrogate of the weight-to-objective map. We call the resulting method BW-QAMOO (Bayes-Weighted QAMOO). Across m∈{3,4}m \in\{3,4\} and n∈{10,11,12}n \in\{10,11,12\}, BW-QAMOO reaches the 90/95% hypervolume targets with a median 33% fewer solves than QAMOO (random), and 61–74% fewer than QAMOO (uniform grid) at m=4m=4, on a genuine (simulated) QAOA solver, and the advantage grows with the objective count mm.

From alloy configuration to a (multi-objective) Ising problem. Encode each of nn lattice sites’ binary occupancy by a spin σi∈{−1,+1}\sigma_i \in\{-1,+1\}. In the cluster-expansion picture, each objective’s configurational energy is a sum of on-site preferences and pairwise interactions — an Ising Hamiltonian — and a weight vector w\mathbf{w} on the simplex (wk≥0w_k \ge0, ∑kwk=1\sum_k w_k = 1) scalarizes the mm per-objective energies EkE_k into a single cost:

Ek(σ)=∑ihi(k)σi+∑i<jJij(k)σiσj,Ew(σ)=∑kwkEk(σ)=∑ihˉiσi+∑i<jJˉijσiσj,(1)E_k(\boldsymbol{\sigma}) = \sum_i h_i^{(k)}\sigma_i + \sum_{i<j} J_{ij}^{(k)}\sigma_i\sigma_j,\qquad E_{\mathbf{w}}(\boldsymbol{\sigma}) = \sum_k w_k E_k(\boldsymbol{\sigma}) = \sum_i \bar{h}_i\sigma_i + \sum_{i<j}\bar{J}_{ij}\sigma_i\sigma_j, \tag*{(1)}

the w\mathbf{w}-weighted sums of the per-objective fields and couplings (the effective cluster interactions fitted from first principles in practice [3, 4]). Crucially, the scalarized cost EwE_{\mathbf{w}} is again of Ising form, so every scalarized subproblem is itself a valid single-objective QAOA instance (each σi\sigma_i maps to a qubit Pauli-Z, EwE_{\mathbf{w}} to a diagonal cost Hamiltonian HCH_C; Appendix § 7), and QAMOO [1] traces the front by solving arg⁡min⁡σEw\arg\min_{\boldsymbol{\sigma}} E_{\mathbf{w}} for many w\mathbf{w}. The only quantity our method changes is how many, and which, weights are solved.

Related work. Choosing where to sample next to trace a Pareto front with few expensive evaluations is the subject of multi-objective BO. Scalarization-based methods such as ParEGO [10] draw a weight each iteration and fit a surrogate to the scalarized objective — QAMOO’s random weight sampling is exactly this draw step, and our loop replaces the random draw with one chosen to maximize front progress; hypervolume-based methods select by expected hypervolume improvement (EHVI) [11], the standard multi-objective BO acquisition, now available in differentiable parallel form [12], and connect to Pareto active learning and multi-objective bandits [13]. Our acquisition (2) is EHVI over the weight simplex; our contribution is to port this BO machinery to the weight-selection layer of QAMOO [1], where each “evaluation” is a full scalarized QAOA solve [2].

BW-QAMOO: Bayesian-Optimized Weight Selection

BW-QAMOO casts weight selection as BO: it treats each scalarized QAOA solve as an expensive black-box evaluation and, at each step, chooses the weight w\mathbf{w} whose solve is expected to enlarge the discovered Pareto front the most. Like any BO loop it has two modular parts: a surrogate that predicts a candidate’s objective vector without spending a query, and an acquisition that scores each candidate’s expected “front progress” under that surrogate. Let Pt\mathcal{P}_t be the objective vectors discovered after tt queries and HV⁡(⋅)\operatorname{HV}(\cdot) the hypervolume relative to a fixed reference point r\mathbf{r}. The next weight maximizes the expected hypervolume improvement (EHVI) — the standard acquisition of multi-objective BO, and the reward signal of a hypervolume bandit:

wt+1=arg⁡max⁡w E[HV⁡(Pt∪{σw})−HV⁡(Pt)],(2)\mathbf{w}_{t+1} = \underset{\mathbf{w}}{\arg\max}\ \mathbb{E}\left[\operatorname{HV}\left(\mathcal{P}_t \cup\{\boldsymbol{\sigma}_{\mathbf{w}}\}\right) - \operatorname{HV}(\mathcal{P}_t)\right], \tag*{(2)}

where σw\boldsymbol{\sigma}_{\mathbf{w}} is the (unknown until queried) solution the solver returns for weight w\mathbf{w}.

Algorithm 1 BW-QAMOO: Bayesian-optimized weight selection

  1. Input: budget BB, solver Q(⋅)Q(\cdot), reference r\mathbf{r}, pool size PP

  2. S←\mathcal{S} \leftarrow simplex corners {e1,…,em}\{\mathbf{e}_1,\ldots,\mathbf{e}_m\}; query σw←Q(w)\boldsymbol{\sigma}_{\mathbf{w}} \leftarrow Q(\mathbf{w}) for each; P←{σw}\mathcal{P} \leftarrow\{\boldsymbol{\sigma}_{\mathbf{w}}\}

  3. while ∣S∣<B|\mathcal{S}| < B do

  4. hcur←HV⁡(P;r)h_{\mathrm{cur}} \leftarrow\operatorname{HV}(\mathcal{P};\mathbf{r}); draw candidate pool W∼Dirichlet⁡(1)×P\mathcal{W} \sim\operatorname{Dirichlet}(\mathbf{1})^{\times P}

  5. for each w∈W\mathbf{w} \in\mathcal{W} do

  6. μ(w),σ(w)←\boldsymbol{\mu}(\mathbf{w}),\boldsymbol{\sigma}(\mathbf{w}) \leftarrow GP posterior from S\mathcal{S}; σ^(w)←μ(w)−κσ(w)\hat{\boldsymbol{\sigma}}(\mathbf{w}) \leftarrow\boldsymbol{\mu}(\mathbf{w})-\kappa\boldsymbol{\sigma}(\mathbf{w}) ▷ optimistic (UCB) point, κ=1.5\kappa=1.5

  7. a(w)←HV⁡(P∪{σ^(w)};r)−hcura(\mathbf{w}) \leftarrow\operatorname{HV}(\mathcal{P}\cup\{\hat{\boldsymbol{\sigma}}(\mathbf{w})\};\mathbf{r})-h_{\mathrm{cur}}

  8. end for

  9. w⋆←arg max⁡wa(w)\mathbf{w}^{\star} \leftarrow\operatorname*{arg\,max}_{\mathbf{w}} a(\mathbf{w}); σ⋆←Q(w⋆)\boldsymbol{\sigma}^{\star} \leftarrow Q(\mathbf{w}^{\star}) ▷ the one expensive query

  10. S←S∪{w⋆}\mathcal{S} \leftarrow\mathcal{S}\cup\{\mathbf{w}^{\star}\}; P←P∪{σ⋆}\mathcal{P} \leftarrow\mathcal{P}\cup\{\boldsymbol{\sigma}^{\star}\}

  11. end while

  12. return non-dominated subset of P\mathcal{P}

Surrogate. To score (2) without spending a query, BW-QAMOO keeps a surrogate over w↦σw\mathbf{w}\mapsto\boldsymbol{\sigma}_{\mathbf{w}}, refined online as real queries arrive: a Gaussian process, one GP regressor per objective (Matérn-5/2 kernel, an anisotropic length scale per simplex coordinate), giving a posterior mean and variance for each unqueried weight. The acquisition scores each candidate by an optimistic (upper-confidence-bound) estimate of its hypervolume gain: it forms the optimistic objective vector σ^(w)=μ(w)−κσ(w)\hat{\boldsymbol{\sigma}}(\mathbf{w})=\boldsymbol{\mu}(\mathbf{w})-\kappa\boldsymbol{\sigma}(\mathbf{w}) (per-objective posterior mean minus κ\kappa standard deviations, κ=1.5\kappa=1.5; the objectives are minimized, so subtracting is optimistic) and evaluates the hypervolume improvement that point would add, HV⁡(Pt∪{σ^(w)})−HV⁡(Pt)\operatorname{HV}(\mathcal{P}_t\cup\{\hat{\boldsymbol{\sigma}}(\mathbf{w})\})-\operatorname{HV}(\mathcal{P}_t). The κσ\kappa\sigma term supplies genuine exploration — the selector can prefer a weight whose outcome is uncertain over one whose predicted mean gain is merely high — realizing the explore/exploit trade-off of the EHVI acquisition of multi-objective BO [11, 12] in a cheap, deterministic (one hypervolume evaluation per candidate) form, ported to the weight simplex (Appendix § 8). Algorithm 1 summarizes the loop.

Continuous weight selection. Because QAMOO samples weights continuously from the simplex, BW-QAMOO optimizes the acquisition over the continuous simplex too: each step draws a fresh dense Dirichlet candidate pool, scores Eq. (2), and takes the argmax. This converges to the continuous maximizer as the pool grows and keeps the comparison against QAMOO’s random sampler on equal, grid-free footing.

Experimental setting

A materials-configuration benchmark. Our benchmark instantiates the multi-objective alloy-configuration problem of Eq. (1) on n∈{10,11,12}n\in\{10,11,12\} lattice sites with m∈{3,4}m\in\{3,4\} competing energies (a frustrated Ising model). Concretely, E1E_1 is a configurational energy with random signed couplings Jij∈{−1,+1}⋅U(0.5,1.5)J_{ij}\in\{-1,+1\}\cdot U(0.5,1.5) at interaction density 0.5 and site fields hi∼U(−0.3,0.3)h_i\sim U(-0.3,0.3); the remaining E2,…,EmE_2,\ldots,E_m are competing “mixing-energy” objectives with fixed distinct coupling matrices, standing in for rival ordering tendencies. All are minimized; the goal is their Pareto front. This is a deliberate synthetic model of crystal design: each EkE_k plays the role of a real design objective, mapped to the same Ising form of fields and couplings but instantiated with random couplings rather than values fitted from first principles — Appendix § 9 gives the correspondence and is explicit about what the idealization does and does not capture. We use random couplings so the exact front is enumerable for scoring (§ 4); fitted cluster interactions are the natural next step (§ 5). Sweeping mm from 3 to 4 is deliberate — the weight simplex grows from two- to three-dimensional, exactly the growth non-adaptive sampling handles poorly.

Figure 2. Hypervolume reached versus number of scalarized solves on the simulated QAOA solver, for m=4m=4 objectives (mean over 10 seeds, ±1\pm1 standard-error band). Each curve is the fraction of the exact Pareto hypervolume secured after 1, 2, 3, . . . queries, for QAMOO (uniform grid) (grey), QAMOO (random) (red), and BW-QAMOO (teal); dotted horizontal lines mark the 90/95% targets. The vertical dashed lines drop from where the mean BW-QAMOO (teal) and QAMOO (random) (red) curves first reach the 90% target, with that crossing query annotated on the axis; their separation is the query saving. Because these crossings are read off the averaged curve, they differ slightly from the per-instance mean runs-to-target in Table 1 (which averages each instance’s own crossing). Up-and-to-the-left is better: BW-QAMOO crosses 90% well left of QAMOO (random) (e.g. 10 vs. 17 queries at n=11n=11) with a tighter band, while QAMOO (uniform grid) trails. (m=3m=3 shows the same ordering, see Table 1.)

Solver and strategies. Each weight is solved by minimizing the scalarized Ising with a genuine gate-based QAOA circuit on a shot-based state-vector simulator (Qiskit Aer [16]); the circuit construction, angle optimization, and read-out are detailed in Appendix § 7. We compare three weight-selection strategies, all sharing the same QAOA solver: (i) QAMOO (uniform grid), which walks a granularity-6 uniform simplex grid in order (a naive non-adaptive reference; 28 weights at m=3m=3, 84 at m=4m=4); (ii) QAMOO (random), i.i.d. Dirichlet sampling of the weights (QAMOO’s actual method [1]); and (iii) BW-QAMOO (Algorithm 1), with its GP-EHVI surrogate (seeded from the mm corners, budget cap B=40B=40). Both BW-QAMOO and QAMOO (random) operate grid-free, so their comparison is exact; QAMOO (uniform grid) is included as the naive baseline. As evaluation metric, we use runs-to-target: the number of scalarized QAOA solves a strategy needs to reach 90% and 95% of the exact Pareto hypervolume. Fewer runs to the same target means a more query-efficient strategy. Every experiment averages over 10 random seeds (instances) per cell.

Results

Figure 2 and Table 1 are the central result. Against QAMOO’s own random-sampling method, BW-QAMOO reaches the same hypervolume with a median 33% fewer scalarized QAOA solves at the 90/95% targets, winning all 12/12 cells, and the saving is significant by a paired Wilcoxon test (Table 1). The advantage grows with mm: the median saving rises from 29% at m=3m=3 to 34% at m=4m=4 (range 19–48% across the m=4m=4 cells), where the weight simplex is three-dimensional and random sampling wastes more of its budget — the teal curve crossing each target line well left of the red one in the bottom row, with a far tighter standard-error band (BW-QAMOO is not just faster but more reliable). Against QAMOO (uniform grid) the gap is larger still (61–74% at m=4m=4), because a fixed-resolution grid grows combinatorially — 84 weights at m=4m=4 versus 28 at m=3m=3 — and spends most redundantly, “the gridding is wasteful” failure made concrete. Every run is one expensive query, so these are proportional cost reductions. The QAOA result here is a simulation, not a hardware run (see limitations).

m=3m=3m=4m=4
90%90\%95%95\%90%90\%95%95\%
nnBWRGBWRGBWRGBWRG
108.510.717.214.825.721.712.417.546.221.226.258.8
115.06.212.58.723.816.211.215.635.519.137.050.8
128.212.916.519.224.732.311.819.244.722.937.559.1
sig.***********************

Table 1. QAOA queries to reach a target hypervolume fraction (runs-to-target, mean over 10 seeds; fewer is better). Columns are BW-QAMOO (BW), QAMOO (random) (R), and QAMOO (uniform grid) (G). The “sig.” row gives one-sided paired Wilcoxon tests that BW-QAMOO needs fewer solves than the baseline in that column (R or G), pooling 30 instances per (mm, target) (*** p<10−3p<10^{-3}, ** p<10−2p<10^{-2}, * p<0.05p<0.05).

Discussion, limitations, and open challenges

The saving is consistent and, in this setting, sizable, and it grows precisely where it matters: as the objective count rises and the weight simplex expands, blind sampling wastes an increasing share of the budget while an adaptive selector concentrates queries on the weights that still enlarge the front. Since each query is one expensive quantum solve, these are direct, proportional reductions in the cost of tracing a front on a quantum device.

We read this as evidence for a broader and, we think, promising direction: the mature machinery of classical machine learning and BO can substantially improve how near-term quantum optimizers are used, without any change to the quantum subroutine itself — the solver stays an expensive oracle, and a classical surrogate decides what to ask it. The same idea should transfer to other quantum-optimization pipelines that sweep a family of related problems.

This is a deliberately narrow study: a frustrated-Ising benchmark at n≤12n \le12 with random couplings, kept small so the exact Pareto front is enumerable for scoring, and — most importantly — run on a noiseless QAOA simulation rather than real hardware. The decisive next step is a deployment on an actual quantum device, where gate noise, readout error, and queue latency all bear on the per-query cost the method sets out to reduce; there, the GP’s calibrated variance could do double duty, steering queries toward the weights a noisy device solves reliably. Beyond that, a proxy reference point would push the measurement past the enumerable regime to larger nn and mm; a scalarization that combines the objectives non-linearly, rather than by the weighted sum we use, would reach the non-convex parts of the front a weighted sum provably cannot; a policy meta-learned across instances could carry savings between related alloy compositions; fitted cluster interactions in place of random couplings [4, 5] would close the remaining gap to a real materials-design workflow; and extending the site variable from two species to the multi-component alloys (ternaries, high-entropy alloys) where the field’s interest concentrates would broaden the applicability — a change to how each objective is encoded into qubits, not to the weight-selection loop, which is agnostic to the encoding beneath it (Appendix § 10).

Background: the Ising model and its link to our problem

The whole paper — the benchmark, the scalarization, and the QAOA solver — rests on the Ising model, so we state it here for readers coming from the materials or machine-learning side rather than physics.

The model. An Ising model is an energy defined over nn binary variables, the spins σi∈{−1,+1}\sigma_i \in\{-1,+1\}. Its energy is a sum of on-site fields and pairwise couplings,

E(σ)=∑ihiσi+∑i<jJijσiσj,σ=(σ1,…,σn)∈{−1,+1}n,(3)E(\boldsymbol{\sigma})=\sum_i h_i\sigma_i+\sum_{i<j}J_{ij}\sigma_i\sigma_j,\qquad\boldsymbol{\sigma}=(\sigma_1,\ldots,\sigma_n)\in\{-1,+1\}^{n}, \tag*{(3)}

where the field hih_i biases spin ii toward +1+1 or −1-1, and the coupling JijJ_{ij} rewards or penalizes a pair for being aligned (σiσj=+1\sigma_i\sigma_j=+1) or anti-aligned (σiσj=−1\sigma_i\sigma_j=-1). Minimizing EE over the 2n2^n spin configurations is the combinatorial problem of interest; it is NP-hard in general [3], and no more than a quadratic (degree-two) polynomial in the spins is needed to encode a remarkable range of hard problems [3]. A change of variable σi=1−2xi\sigma_i=1-2x_i turns Eq. (3) into an equivalent QUBO (quadratic unconstrained binary optimization) over bits xi∈{0,1}x_i\in\{0,1\}; the two forms are interchangeable and we use whichever is convenient.

Frustration — what makes it hard. If all couplings could be satisfied at once the ground state would be trivial. Frustration is the situation where they cannot: on a loop of spins with competing coupling signs, no configuration satisfies every bond simultaneously, so the energy landscape develops many competing local minima. This is precisely the regime our benchmark targets (random signed couplings; Appendix § 9), and it is what makes the configurational ground state non-trivial to find.

Why it fits our problem. In an alloy or crystal, the design variable is which of two species occupies each of nn lattice sites — a natural binary spin σi∈{−1,+1}\sigma_i\in\{-1,+1\}. To leading order in a cluster expansion (the standard way configurational properties of alloys are represented [7, 8, 4]), each configurational property is a constant plus on-site terms plus pairwise terms — exactly the Ising form of Eq. (3), with hi,Jijh_i,J_{ij} the fitted effective interactions. So a materials-configuration objective is an Ising energy; Appendix § 9 makes this correspondence explicit for each objective we use.

Why it fits quantum optimizers. Each spin σi\sigma_i maps to a qubit through the Pauli-ZZ operator, and Eq. (3) becomes a diagonal cost Hamiltonian HC=∑ihiZi+∑i<jJijZiZjH_C=\sum_i h_iZ_i+\sum_{i<j}J_{ij}Z_iZ_j whose ground state is the minimizing configuration. Both quantum annealers [6] and gate-model methods such as QAOA (Appendix § 7) are built to seek that ground state, which is why the Ising encoding makes materials-configuration search a direct target for quantum optimization. A single weighted sum of several Ising energies, Ew=∑kwkEkE_w=\sum_k w_kE_k (Eq. (1)), is again of the form Eq. (3) — the property BW-QAMOO relies on, since every scalarized weight is then itself a valid Ising (hence QAOA) instance.

Background: QAOA and QAMOO

QAOA. The Quantum Approximate Optimization Algorithm [2] is a variational gate-model method for minimizing an Ising cost such as Eq. (1). The cost is written as a diagonal cost Hamiltonian HC=∑ihiZi+∑i<jJijZiZjH_C=\sum_i h_iZ_i+\sum_{i<j}J_{ij}Z_iZ_j (each spin σi\sigma_i becomes a qubit Pauli-ZiZ_i), and a mixer HM=∑iXiH_M=\sum_iX_i drives transitions between bitstrings. Starting from the uniform superposition ∣+⟩⊗n|+\rangle^{\otimes n}, QAOA applies pp alternating layers,

∣γ,β⟩=∏ℓ=1pe−iβℓHMe−iγℓHC∣+⟩⊗n,(4)|\gamma,\beta\rangle=\prod_{\ell=1}^{p}e^{-i\beta_\ell H_M}e^{-i\gamma_\ell H_C}|+\rangle^{\otimes n}, \tag*{(4)}

parameterized by 2p2p angles (γ,β)(\gamma,\beta). A classical optimizer (here COBYLA) tunes the angles to minimize the measured expected energy ⟨HC⟩\langle H_C\rangle; the final state is sampled (4096 shots) and the lowest-energy bitstring observed is returned as the solution. Depth pp trades circuit expressivity against noise; we use p=2p=2. Each such solve — angle optimization plus sampling — is one “expensive query” in the main text. We run QAOA on the Qiskit Aer simulator with shot-based readout and

no noise model: it confirms the algorithm behaves as intended but omits the device noise, readout error, and queueing of real hardware.

Ansatz, gates, and the Pauli cost Hamiltonian (exact circuit run). We use the standard QAOA ansatz (Qiskit’s QAOAAnsatz) with p=2p=2, so the circuit carries 2p=42p=4 variational parameters (γ1,β1,γ2,β2)(\gamma_1,\beta_1,\gamma_2,\beta_2). Each objective is scalarized to an Ising energy, converted to an equivalent QUBO in binary variables xi∈{0,1}x_i \in\{0,1\} (via σi=1−2xi\sigma_i=1-2x_i), and the QUBO is re-expressed as a diagonal cost Hamiltonian using xi=12(1−Zi)x_i=\frac{1}{2}(1-Z_i). For a QUBO ∑iaixi+∑i<jbijxixj\sum_i a_i x_i+\sum_{i<j}b_{ij}x_i x_j this yields

HC=(12∑iai+14∑i<jbij)⏟constant offsetI−∑i(12ai+14∑j≠ibij)Zi+∑i<j14bijZiZj,(5)H_C=\underbrace{\left(\frac{1}{2}\sum_i a_i+\frac{1}{4}\sum_{i<j}b_{ij}\right)}_{\text{constant offset}}I-\sum_i\left(\frac{1}{2}a_i+\frac{1}{4}\sum_{j\ne i}b_{ij}\right)Z_i+\sum_{i<j}\frac{1}{4}b_{ij}Z_iZ_j, \tag*{(5)}

i.e. single-qubit ZiZ_i terms (a local field) and two-qubit ZiZjZ_iZ_j terms on every coupled pair, plus an identity offset that shifts the energy but not the minimizer. Concretely the cost layer e−iγℓHCe^{-i\gamma_\ell H_C} compiles to an RZ(θ)=e−iθZ/2R_Z(\theta)=e^{-i\theta Z/2} rotation for each ZiZ_i term and an RZZ(θ)=e−iθZiZj/2R_{ZZ}(\theta)=e^{-i\theta Z_iZ_j/2} interaction (a CX–RZR_Z–CX gadget) for each ZiZjZ_iZ_j term; the mixer layer e−iβℓHMe^{-i\beta_\ell H_M} with HM=∑iXiH_M=\sum_i X_i compiles to an RX(2βℓ)R_X(2\beta_\ell) on every qubit. The initial state ∣+⟩⊗n\lvert+\rangle^{\otimes n} is prepared by a layer of Hadamards. During angle optimization the expectation ⟨HC⟩\langle H_C\rangle is evaluated on the Aer EstimatorV2 primitive; COBYLA (max. 100 iterations) tunes the angles from a seeded random start γ,β∼U(0,π)\gamma,\beta\sim U(0,\pi); the tuned circuit is then measured with the SamplerV2 primitive at 4096 shots, and among the sampled bitstrings we return the one of lowest QUBO energy. This “tuned” readout isolates the query-efficiency question from sampler noise; every scalarized weight is one such circuit solve. Figure 3 shows the resulting gate-level circuit on a representative 3-qubit slice.

QAOA ansatz circuit at depth $p=2$ on a representative 3-qubit slice

Figure 3. The QAOA ansatz actually run (depth p=2p=2), on a representative 3-qubit slice with two coupled pairs. After a Hadamard layer prepares ∣+⟩⊗n\lvert+\rangle^{\otimes n}, each of the pp layers applies the cost unitary e−iγℓHCe^{-i\gamma_\ell H_C} — an RZR_Z per single-ZZ term and an RZZR_{ZZ} per coupled pair of Eq. (5) — followed by the mixer e−iβℓHMe^{-i\beta_\ell H_M}, one RXR_X per qubit. The 2p=42p=4 angles (γ1,β1,γ2,β2)(\gamma_1,\beta_1,\gamma_2,\beta_2) are tuned by COBYLA; the final state is measured at 4096 shots.

QAMOO. Quantum Approximate Multi-Objective Optimization [1] extends QAOA from a single objective to a Pareto front over mm objectives. Its mechanism is scalarization: for a weight vector w\mathbf{w} on the simplex it forms the single Ising cost EwE_{\mathbf{w}} of Eq. (1), solves that one problem with QAOA, and records the returned configuration as one candidate Pareto point. Sweeping w\mathbf{w} across the simplex and keeping the non-dominated configurations traces an approximation to the front. Crucially, QAMOO chooses the weights by random sampling: rather than discretizing the simplex on a grid, it samples w\mathbf{w} uniformly at random from the feasible set, “without having to enumerate an exponential number of grid points” [1]. Because every EwE_{\mathbf{w}} is itself a valid Ising problem, no change to the quantum subroutine is needed — only a choice of which weights to solve. This paper replaces the random draw with an adaptive, acquisition-driven selector (Algorithm 1) while leaving the per-weight QAOA solve untouched.

The GP-EHVI surrogate in detail

The GP-EHVI surrogate. We fit one Gaussian-process regressor [14] per objective (Matérn-5/2 kernel, an anisotropic length scale per simplex coordinate), giving a posterior mean and variance for each unqueried weight rather than a point estimate. The acquisition is an optimistic (upper-confidence-bound) surrogate for the expected hypervolume improvement in Eq. (2): for each candidate weight it forms the optimistic objective vector σ^(w)=μ(w)−κσ(w)\hat{\boldsymbol{\sigma}}(\mathbf{w}) = \boldsymbol{\mu}(\mathbf{w}) - \kappa\boldsymbol{\sigma}(\mathbf{w}) (κ=1.5\kappa=1.5; the objectives are minimized, so subtracting the posterior standard deviation is the optimistic direction) and scores the hypervolume gain that this single point would contribute, HV⁡(Pt∪{σ^(w)})−HV⁡(Pt)\operatorname{HV}(P_t \cup\{\hat{\boldsymbol{\sigma}}(\mathbf{w})\}) - \operatorname{HV}(P_t) — one deterministic hypervolume evaluation per candidate, no Monte-Carlo sampling. The κσ\kappa\boldsymbol{\sigma} term now supplies genuine exploration, so the selector can prefer a weight whose outcome is uncertain over one whose predicted mean gain is merely high — the explore/exploit balance at the heart of Bayesian optimization [15]. This is a cheap upper-confidence-bound realization of the GP-EHVI acquisition of multi-objective Bayesian optimization [11, 12], ported to the weight simplex; a full Monte-Carlo EHVI (averaging the gain over posterior draws from the same surrogate) is a drop-in alternative at higher per-step cost.

Beyond this benchmark. The GP’s advantages should widen further where this benchmark’s assumptions break: on noisy or stochastic solvers (real hardware), where its calibrated variance can route queries to weights the device solves reliably; under non-smooth or multi-modal weight-to-objective maps (e.g. across a ground-state phase transition); and under non-linear scalarizations — which combine the objectives by something other than a weighted sum, so as to reach non-convex parts of the front, at the price of a less smooth weight-to-objective map that the surrogate must then learn. These are left to future work (§ 5).

From crystal-design objectives to Ising energies

Our benchmark is a synthetic instantiation of crystal/alloy design, and this appendix makes the correspondence explicit and states plainly where it is idealized. The design variable is which of two species occupies each of the nn symmetry-distinct sites of a fixed lattice, encoded as a spin σi∈{−1,+1}\sigma_i \in\{-1,+1\} (species A/B). Every objective below is a function of the occupation pattern σ\boldsymbol{\sigma}, and each is naturally quadratic in the spins — a constant, plus on-site fields hiσih_i\sigma_i, plus pairwise couplings JijσiσjJ_{ij}\sigma_i\sigma_j — which is exactly the Ising form Eq. (1) that QAOA minimizes. This quadratic-in-occupation structure is not incidental: it is the leading order of a cluster expansion [7, 8], the standard way configurational properties of alloys are represented, so the mapping to Ising is the physically motivated one.

Candidate objectives and their Ising terms.

  • Configurational / formation energy (E1E_1 here). The effective interactions between occupied sites: hih_i a site preference, JijJ_{ij} a pair interaction (negative JijJ_{ij} favors like neighbors, i.e. clustering/phase separation; positive favors unlike neighbors, i.e. ordering). Frustration — competing signs on a loop that cannot all be satisfied — is exactly what makes the configurational ground state non-trivial, which is why E1E_1 uses random signed couplings.

  • Elastic / strain-mismatch energy. Size mismatch between the two species stores elastic energy that depends on which neighbors are unlike; to leading order this is again ∑ijKijelσiσj\sum_{ij} K_{ij}^{\mathrm{el}}\sigma_i\sigma_j with KijelK_{ij}^{\mathrm{el}} set by the bond stiffness and the atomic size difference. Minimizing it favors arrangements that relieve strain.

  • Chemical short-range order (disorder). A target degree of mixing is written as a penalty on the count of like-neighbor pairs, ∑ijJijsroσiσj\sum_{ij} J_{ij}^{\mathrm{sro}}\sigma_i\sigma_j (up to a constant); tuning its sign drives the

configuration toward ordering or toward a solid solution.

  • Cost / scarcity. The price or supply risk of the chosen composition is linear in occupancy, ∑iciσi\sum_i c_i\sigma_i (a pure field term hi=ci/2h_i = c_i/2, no couplings): a cheaper species on more sites lowers it. Density and other composition-linear properties enter the same way.

Because each objective is a weighted sum of fields and couplings, a scalarization ∑kwkEk\sum_k w_kE_k is itself an Ising energy (its fields and couplings are the w-weighted sums; Eq. (1)), which is what lets a single QAOA solve handle any weight w.

What the synthetic model does and does not capture. Our E2,…,EmE_2,\ldots,E_m use fixed, distinct random coupling matrices rather than the hi,Jijh_i,J_{ij} one would fit from DFT or a measured cluster expansion, so the numbers here reflect the structure of the problem — several competing quadratic objectives over a frustrated configuration space, with a weight simplex whose dimension grows with mm — but not any specific material. The query-efficiency result depends only on that structure (a family of related scalarized Ising solves swept over w), not on the couplings being physical, so it should carry over once real interactions are substituted; conversely, we make no claim that the resulting fronts correspond to real materials trade-offs. Replacing the random couplings with fitted cluster interactions — and moving beyond two species on a single sublattice (Appendix § 10) — is the natural next step (§ 5).

More than two species per site

A binary spin σi∈{−1,+1}\sigma_i \in\{-1,+1\} names exactly two species, so the encoding of Appendix § 9 models a binary alloy. Multi-component alloys — ternaries and, in the extreme, high-entropy alloys [5] — put q>2q > 2 species on each site, and a single binary variable per site can no longer name the choice. We stress at the outset that this is a change to the per-objective encoding only: BW-QAMOO treats each scalarized solve as a black box and is indifferent to how many qubits and terms that box contains, so its weight-selection loop, and the query-efficiency result, are unaffected. What changes is the qubit count and structure of each solve. There are three standard routes, each with a concrete cost.

One-hot encoding. Give each site qq binary variables xi,k∈{0,1}x_{i,k} \in\{0,1\}, one per species, and enforce “exactly one species per site” with a penalty λ(∑kxi,k−1)2\lambda\left(\sum_k x_{i,k}-1\right)^2. The penalty is quadratic in the bits, so the problem remains an Ising/QUBO instance solvable by the same QAOA subroutine [3]; the price is nqnq qubits rather than nn and a penalty weight λ\lambda that must be tuned (too small admits infeasible one-hot patterns, too large drowns the objective). This is the most common choice and the one that most directly preserves our pipeline.

Binary / domain-wall encoding. Name the qq species with ⌈log⁡2q⌉\lceil\log_2 q\rceil bits per site, using exponentially fewer qubits. The objective in these bits generally acquires terms of degree >2> 2 that must then be quadratized (via ancillas) to return to Ising form; domain-wall encodings [9] keep the required couplings local and are a practical middle ground between qubit count and Hamiltonian complexity.

Potts / higher-spin objectives. Physically, q>2q > 2 discrete states per site is a Potts model rather than an Ising one, the natural generalization for multi-component ordering; to run it on gate-model QAOA it is reduced to qubits by one of the two routes above. In every case the scalarization ∑kwkEk\sum_k w_kE_k stays a single problem of the same (encoded) form, so QAMOO’s sweep — and our adaptive selection of which weights to sweep — applies unchanged.

References

  1. [1]Quantum Approximate Multi-Objective Optimization. Nature Computational Science 5 (2025). https://doi.org/10.1038/s43588-025-00873-y; arXiv:2503.22797.
  2. [2]E. Farhi, J. Goldstone, and S. Gutmann. A Quantum Approximate Optimization Algorithm. arXiv:1411.4028 (2014).arxiv.org/abs/1411.4028
  3. [3]A. Lucas. Ising formulations of many NP problems. Frontiers in Physics 2:5 (2014).DOI
  4. [4]Optimal elemental configuration search in crystal via quantum annealing (2025). arXiv:2503.09356.
  5. [5]Quantum annealing-assisted lattice optimization for high-entropy alloys. npj Computational Materials 10 (2024). https://doi.org/10.1038/s41524-024-01505-1.
  6. [6]T. Kadowaki and H. Nishimori. Quantum annealing in the transverse Ising model. Physical Review E 58(5):5355–5363 (1998).DOI
  7. [7]J. M. Sanchez, F. Ducastelle, and D. Gratias. Generalized cluster description of multicomponent systems. Physica A 128(1–2):334–350 (1984).DOI
  8. [8]G. Ceder. A derivation of the Ising model for the computation of phase diagrams. Computational Materials Science 1(2):144–150 (1993).DOI
  9. [9]N. Chancellor. Domain wall encoding of discrete variables for quantum annealing and QAOA. Quantum Science and Technology 4(4):045004 (2019). arXiv:1903.05068.DOI
  10. [10]J. Knowles. ParEGO: A hybrid algorithm with on-line landscape approximation for expensive multiobjective optimization problems. IEEE Transactions on Evolutionary Computation 10(1):50–66 (2006).DOI
  11. [11]M. T. M. Emmerich, K. C. Giannakoglou, and B. Naujoks. Single- and multiobjective evolutionary optimization assisted by Gaussian random field metamodels. IEEE Transactions on Evolutionary Computation 10(4):421–439 (2006).DOI
  12. [12]S. Daulton, M. Balandat, and E. Bakshy. Differentiable Expected Hypervolume Improvement for Parallel Multi-Objective Bayesian Optimization. In Advances in Neural Information Processing Systems (NeurIPS) (2020). arXiv:2006.05078.arxiv.org/abs/2006.05078
  13. [13]M. Zuluaga, A. Krause, and M. Püschel. e-PAL: An active learning approach to the multi-objective optimization problem. Journal of Machine Learning Research 17(104):1–32 (2016).
  14. [14]C. E. Rasmussen and C. K. I. Williams. Gaussian Processes for Machine Learning. MIT Press (2006).DOI
  15. [15]B. Shahriari, K. Swersky, Z. Wang, R. P. Adams, and N. de Freitas. Taking the Human Out of the Loop: A Review of Bayesian Optimization. Proceedings of the IEEE 104(1):148–175 (2016).DOI
  16. [16]A. Javadi-Abhari et al. Quantum computing with Qiskit. arXiv:2405.08810 (2024).arxiv.org/abs/2405.08810

Paper details

vlnc:W_a2du3tsy
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents