Many of the most interesting materials in a combinatorial search, including high-entropy alloys, doped cathodes and solid solutions, have no fixed atomic arrangement. Elements sit on lattice sites at random. To compute such a material's energy with density functional theory (DFT, the standard quantum-mechanical method for computing material energies from the electron density), one must simulate a finite periodic box of a few dozen to a few hundred atoms. Any such box is one random draw of the arrangement, so its energy carries sampling noise. Averaging many boxes is expensive because DFT cost grows roughly as the cube of system size. Cosmologists face the same structure: an N-body simulation box is one random realization of the universe. Mitigating the noise of DESI mocks using analytic control variates removes most of that noise by running a cheap analytic model (the Zeldovich approximation) on the identical initial conditions. It then subtracts the model's realization-specific fluctuation, which reduces power-spectrum noise 5–10 times without bias.
We propose the materials analogue. For each random alloy box, we evaluate a cheap surrogate on the exact same atomic arrangement. The surrogate's average over all random arrangements is known exactly, either analytically or from thousands of cheap evaluations. We then correct the DFT energy by the surrogate's deviation from that average. The resulting estimate of the configurational-average energy stays unbiased even when the surrogate is wrong. Its variance shrinks by a factor of $1/(1-\rho^2)$, where $\rho$ is the correlation between DFT and surrogate across arrangements. The payoff for combinatorial discovery is cheaper and better-characterized labels for disordered compositions. Current active-learning pipelines such as Scaling deep learning for materials discovery largely avoid these compositions, because they search over ordered prototypes.
Core Claim
Let $E_i$ be the DFT energy per atom of random box $i$ at a fixed composition, and let $S_i$ be a surrogate energy for the same box. Let $\mu_S$ be the exact expectation of $S$ over the random-arrangement ensemble at that box size. The estimator
$$\hat{E}_{CV} = \frac{1}{n}\sum_{i=1}^{n}\left[E_i - \beta\,(S_i - \mu_S)\right]$$
is unbiased for the finite-box ensemble mean of $E$ for any fixed $\beta$. Its variance is minimized at $\beta = \mathrm{Cov}(E,S)/\mathrm{Var}(S)$, where it equals $(1-\rho^2)$ times the naive variance. This is the standard control-variate identity, and it is where the expected gain comes from.
The specific claim we test goes further. We use a cluster-type surrogate, an energy written as a weighted sum of pair and triplet site-occupation correlations. For an ideal random alloy, the expected value of each such correlation is a closed-form product of concentrations. So $\mu_S$ costs nothing, exactly as the Zeldovich mean does in the cosmology case. This reframes the special quasirandom structure (SQS, a single box hand-searched so that its correlations match the random-alloy values). SQS is the "force $S_i - \mu_S = 0$" special case. The control-variate version regresses out residual correlation deviations from many unconstrained random boxes, and it also yields an honest error bar.
The cosmology paper also gives a warning that we turn into a measurable question. There, surrogate correlation decays at small scales where non-linear collapse dominates. Here we expect $\rho$ to drop where local chemistry, magnetism or relaxation dominates. We will map where that happens.
Research Program
The program builds and characterizes the estimator using two surrogate families. The first is an analytic-mean surrogate: pair and triplet correlations with coefficients fitted on a held-out set of boxes. The second is a universal machine-learned interatomic potential (M3GNet, as benchmarked in Matbench Discovery - Can machine learning identify stable crystals?). For the potential, $\mu_S$ is estimated from about $10^3$ cheap relaxed evaluations. That estimate adds a small, quantifiable variance term, which makes the second variant a two-level Monte Carlo scheme.
To avoid the $O(1/n)$ bias that arises from fitting $\beta$ and $\rho$ on the same few DFT samples, we cross-fit. The coefficients for box $i$ come from the other boxes or from a prior batch. Methods for learned and multiple control variates, Neural Control Variates for Variance Reduction and Control Variates for MCMC, supply the estimation machinery. Our contribution is not new statistics. It is identifying controls whose ensemble mean is exactly known for alloy configurations, and showing when they are strongly correlated with relaxed DFT energies.
One limitation must stay explicit. The estimator removes sampling variance within a fixed box size. It does not remove systematic finite-size bias from periodic images, and it does not remove functional error. We measure the box-size bias separately at 32, 64 and 128 atoms so the two effects are not conflated.
The program then puts the estimator inside a discovery loop over a five-element composition simplex. Each composition's label is the mixing energy, together with its per-composition control-variate standard error. That standard error enters a Gaussian-process surrogate as heteroscedastic noise, meaning noise that differs from point to point. An acquisition rule decides whether the next DFT box goes to a new composition or to refining an uncertain one near the stability threshold. Kinetics and vibrational entropy remain outside scope, as the GNoME paper notes for 0 K hull screening.
Baselines
The naive reference is the plain mean of $n$ independent random boxes per composition, with its sample standard error. Every method gets the same total DFT budget, the same box sizes, and the same functional (PBE) and convergence settings. Surrogate compute is counted in wall-clock terms.
The strongest standard practice is SQS. We test both a single SQS per composition and the mean of several SQS. This is the direct competitor because it already attacks configurational noise by matching correlations. We beat it if control variates give lower mean-squared error against the reference ensemble mean at equal DFT calls. Mean-squared error counts both residual bias and variance, so SQS's hidden bias from unmatched long-range or higher-order correlations is included in the comparison.
The strongest learned competitor uses the surrogate directly. One version takes the cluster model's prediction at the random-alloy correlations. The other is Δ-learning, which fits a correction from potential to DFT across compositions. These predictors can be cheap and low-variance but biased. We beat them if our estimator's error, which is unbiased by construction, is smaller once bias is measured against the reference.
For the universal potential alone, the file's reference numbers come from crystal stability screening rather than alloy averages. On Matbench Discovery, M3GNet reaches an F1 of 0.58 and $R^2$ of 0.59 (Matbench Discovery). For our target quantity, no reported result was found on Lacuna for variance or error in DFT configurational-average energies of disordered supercells. Generating the reference is therefore part of the work.
Two ablations isolate the key ingredient. The first evaluates the surrogate on different random boxes from the DFT boxes. This keeps the surrogate information but destroys the pairing, and the gain should vanish. The second fixes $\beta = 1$ instead of using cross-fitted optimal $\beta$.
Why not just…?
Why not just use SQS, which already removes configurational noise? SQS matches only the few correlations it targets and gives no error estimate. Our method contains SQS as a special case, so the experiment is direct. If a single SQS already lands within the reference ensemble's standard error at every tested composition, the proposal's value collapses to error bars alone. We would report that outcome as a real result.
Why not just run bigger boxes or more random boxes? Larger boxes cost roughly cubically more. In cosmology, matching the control-variate gain required 1.4–3.2 times more simulation volume (Mitigating the noise of DESI mocks). More boxes reduce variance only as $1/n$, while the control variate multiplies the effective $n$ by $1/(1-\rho^2)$. The naive-mean baseline makes this trade explicit in DFT-hours.
Why not just trust a universal ML potential, or multi-fidelity Δ-learning, and skip DFT? This is the real threat when the potential is very accurate for a chemistry. Matbench Discovery shows such models misclassify exactly near the stability threshold, the "triangle of peril", which is where labels matter most. The control variate keeps DFT as the anchor and uses the potential only for its fluctuations. The direct-surrogate baseline shows per chemistry whether the potential's bias is small enough to make the anchor unnecessary.
Proposed Experiments
The first experiment builds a reference. We choose eight compositions spanning fcc and bcc high-entropy alloys and one magnetic, Fe/Co/Ni-rich family, where we expect $\rho$ to degrade. For each, we relax 200 random 64-atom boxes with DFT, plus smaller sets at 32 and 128 atoms. This gives a reference finite-box mean and the box-size bias. By resampling subsets of size $n = 2$ to $16$, we measure each estimator's bias and mean-squared error. We also measure the effective-sample-size gain, the naive variance divided by the estimator variance at equal cost.
The decisive check has three parts. Paired control variates must give at least a 3× effective-sample-size gain on non-magnetic alloys, with bias statistically indistinguishable from zero. The unpaired ablation must show no gain. And the measured gain must track $1/(1-\rho^2)$ computed from held-out boxes. If $\rho$ for relaxed energies falls below about 0.8 in most systems, the idea fails, and we would say so. Relaxation is the likely culprit, since the analytic surrogate sees unrelaxed occupations while DFT relaxes positions.
The second experiment runs the discovery loop on a five-element simplex at a fixed DFT budget. It compares labels from naive averaging, SQS and control variates. The metric is the fraction of compositions correctly classified as stable or unstable against competing Materials Project phases, scored against a dense reference computed at the end. The working hypothesis is that control-variate labels reach the same classification accuracy with substantially fewer DFT calls, with the gain concentrated near the threshold.
As an external sanity check, we compare mixing energies with the experimental single-phase or multi-phase labels in the 1,252-alloy set of Machine-learning informed prediction of high-entropy solid solution formation. Its physics rule reaches 73% accuracy, 81% with size misfit, and its Gaussian-process classifier reaches 93%. We treat this only as a plausibility check, because experimental phase formation involves kinetics that 0 K energies cannot capture.
References
- Mitigating the noise of DESI mocks using analytic control variates
- Scaling deep learning for materials discovery
- Matbench Discovery - Can machine learning identify stable crystals?
- Neural Control Variates for Variance Reduction
- Control Variates for MCMC
- Machine-learning informed prediction of high-entropy solid solution formation: Beyond the Hume-Rothery rules