Chemists often make functional nanomaterials by sacrificial templating. They grow an ordered porous crystal first, typically a metal–organic framework (MOF, a lattice of metal ions joined by organic linkers). They then heat, etch, or chemically convert it into a "daughter" material that inherits the parent's shape and porosity: hollow nickel–cobalt hydroxide cages, defect-rich copper oxide, doped porous carbon. The scientific output of such a campaign is usually a claim about a synthesis knob. Examples are "pyrolysis at 400 °C maximizes glucose sensitivity" or "a 1:1 Ni:Co ratio maximizes capacitance." The trouble is that the expensive measurement that defines the target cannot be run on every sample. Full electrochemical testing, long cycling runs, and realistic-humidity sensing are all too costly. Campaigns therefore lean on cheap stand-ins: surface area, X-ray crystallinity, a single quick voltammetry scan, or an ML prediction. Fairness researchers call this proxy label bias. When you optimize or infer against an observable stand-in for an unobservable target, you learn the stand-in's distortions. The best-known case is a hospital algorithm that used healthcare cost as a proxy for need (Dissecting racial bias in an algorithm used to manage the health of populations).
This proposal brings the statistical repair tools from that literature into templated synthesis. The tools are design-based estimators that pair abundant proxy measurements with a small, deliberately sampled set of gold-standard measurements. We use them to estimate synthesis-parameter effects on daughter-material performance with confidence intervals that actually hold. The specific claim is about the failure mode. In sacrificial templating, proxy error is not random noise: it depends on the very synthesis parameter being studied. Proxy-only conclusions can therefore get the shape of the dose–response curve wrong, not merely blur it. A two-phase design that concentrates gold measurements where the template transformation changes regime should recover valid effects at a fraction of the measurement cost.
Core Claim
The source papers already show that proxy error in templated synthesis is tied to the synthesis parameter.
In the hollow NiCo layered-double-hydroxide cages made by etching ZIF-67, a cobalt imidazolate MOF (Fabrication and electrochemical properties of hollow cage-like nickel cobalt layered hydroxides with porous structure), surface area rises with Ni content: 233.3 m²·g⁻¹ for Ni₁Co₀ versus 182.6 m²·g⁻¹ for Ni₁Co₁. Charge-transfer resistance also rises with Ni, and the capacitance optimum sits at Ni₁Co₁ (801 F·g⁻¹ at 0.5 A·g⁻¹). Surface area as a proxy would therefore point the composition effect the wrong way.
In Cu-MOF-derived CuO/C (MOF derived core-shell CuO/C with temperature-controlled oxygen-vacancy for real time analysis of glucose), the framework begins to collapse at 450 °C. Any proxy calibrated on intact-framework samples will be miscalibrated exactly past that temperature.
In MOF-derived CuO/rGO CO₂ sensors (Exploring MOF-Derived CuO/rGO Heterostructures for Highly Efficient Room Temperature CO2 Sensors), dry-lab response is a proxy for field performance. Performance degrades above 40% relative humidity.
Each case matches the structure that Fairness Evaluation in Presence of Biased Noisy Labels formalized for arrests versus offenses. There, a small group-dependent gap (α = 0.07 on COMPAS) was enough to dissolve an apparent calibration result. In our setting the "group" is a synthesis condition.
We frame the estimand explicitly. Let $x$ be the synthesis parameters, $Y$ the gold-standard daughter property, and $\hat Y$ the cheap proxy. The target is a parameter-effect functional, such as the coefficient vector $\beta$ solving $E[\psi(Y, x; \beta)] = 0$ for a regression of performance on synthesis parameters. The proxy-only analysis solves the same equation with $\hat Y$ in place of $Y$. It is biased whenever $E[\hat Y - Y \mid x]$ varies with $x$, which the examples above suggest is the norm. Design-based estimators correct this using a measured subset with known inclusion probabilities. They do so without assuming the proxy is accurate.
Relation to Prior Work
Three statistical papers supply the method. Using Imperfect Surrogates for Downstream Inference: Design-based Supervised Learning for Social Science Applications of Large Language Models builds bias-corrected pseudo-outcomes from LLM labels plus a probability sample of human codes. Across 18 datasets, surrogate-only regressions covered the truth only 40% of the time at nominal 95%, while design-based supervised learning held nominal coverage. M-estimation under Two-Phase Multiwave Sampling with Applications to Prediction-Powered Inference adds adaptive multiwave sampling with inverse-probability weighting, reporting up to 1.8x effective sample size gains at nominal 90% coverage. Towards the Efficient Inference by Incorporating Automated Computational Phenotypes under Covariate Shift gives cross-fitted doubly robust estimators with a no-harm guarantee relative to gold-only analysis. It reports confidence intervals up to 60% shorter than standard prediction-powered inference.
All three were validated on text, health records, surveys, and remote sensing. None faced a setting where the experimenter controls the covariates and the proxy's failure is driven by a physical regime change, such as framework collapse. We reuse their estimators, and contribute the physical-regime-aware sampling design and the materials validation.
Two materials papers motivate the gap. Matbench Discovery - Can machine learning identify stable crystals? shows that proxy accuracy and discovery utility diverge: M3GNet has the best F1 (0.58) while MEGNet has the best discovery acceleration factor (2.70). It also shows that errors concentrate near the stability threshold, which the benchmark calls the "triangle of peril." PRO-MOF: Policy Optimization with Universal Atomistic Models for Controllable MOF Generation warns that a reinforcement-learning generator can exploit a universal potential's blind spots to produce artifactually stable MOFs. Both papers diagnose proxy gaming. Neither offers valid inference under it, and neither touches template conversion.
MOF Synthesis Prediction Enabled by Automatic Data Mining and Machine Learning** predicts parent-MOF synthesis conditions from literature. That task is adjacent, but it concerns making the parent, not converting it. Its data also suffers the one-sided censoring problem we return to below.
The closest conceptual precedent is the cost-versus-need correction of Obermeyer et al. (Dissecting racial bias in an algorithm used to manage the health of populations). There, changing the target label shifted Black enrollment in the high-risk program from 17.7% to 46.5%. Our analog is changing the target from "surface area" or "a quick scan" to "the property you actually claim."
Research Program
The work has three parts.
The first part is a fully controlled computational testbed. Here gold labels exist for every item, so coverage can be checked exactly. On the WBM test set of Matbench Discovery, the proxy is an ML-predicted distance to the convex hull, and gold is the DFT value. The estimands are stability rates and regression effects across substitution prototypes and chemical families. These are the "which knob matters" questions that screening campaigns actually report. A second computational arm uses QMOF, with a fine-tuned JMP model (From Molecules to Materials: Pre-training Large Generalizable Models for Atomic Property Prediction) as the proxy for DFT bandgaps. The estimand there is the bandgap shift attributable to metal-node identity. These arms test the machinery where the proxy error is known to concentrate near a threshold.
The second part is the core: a wet-lab campaign on the ZIF-67 to NiCo-LDH system. We run a factorial grid over Ni:Co salt ratio, nitrate etching time, and ethanol water content, with roughly 120 syntheses. Every sample receives cheap proxies: powder X-ray diffraction, a single-rate cyclic voltammetry scan, and a model prediction from synthesis parameters. As a validation investment, every sample also receives the gold measurements: galvanostatic capacitance at 0.5 and 10 A·g⁻¹ and 2000-cycle retention. Having gold on everything lets us subsample "gold" retrospectively under any design and check coverage against the full-data truth. A smaller companion grid on Cu-MOF pyrolysis spans temperatures through the 450 °C collapse point. Its proxy is X-ray-derived crystallinity plus a quick amperometric response, and its gold is calibrated glucose sensitivity. This arm deliberately straddles a regime change.
The third part is the sampling design itself. Standard multiwave sampling allocates gold measurements where the proxy's residual variance is predicted to be large. We add one physically motivated ingredient: an allocation term that upweights synthesis conditions near suspected transformation boundaries. Examples are etching times where the core empties, or temperatures where crystallinity drops. A cheap classifier on the proxy measurements flags these boundaries. Inclusion probabilities stay known and bounded away from zero, so the inverse-probability-weighted estimator stays valid whether or not the boundary guess is right. A wrong guess costs efficiency, not correctness.
Where no gold standard exists at all, as with literature-mined synthesis data, we do not attempt correction. Instead we report Fogliato-style sensitivity bounds: how large a censoring rate of unreported failures would overturn a claimed parameter effect. This is a secondary deliverable.
Baselines
The naive reference point is the surrogate-only analysis. We fit the parameter-effect regression on cheap proxies for all samples, as a campaign that reports surface area or a quick scan would. It costs no gold measurements. We expect it to be biased with poor coverage, mirroring the 40% coverage reported by Using Imperfect Surrogates for Downstream Inference. Beating it means nominal coverage plus a correctly signed composition effect.
The second reference is gold-only analysis on the same gold budget, using a random subset with no proxies. This is the honest small-study approach, and it is the bar the no-harm principle promises never to fall below. Beating it means a narrower interval at equal coverage. We report the gold budget at which the debiased method's interval width matches gold-only at full budget.
The strongest existing method is multiwave predict-then-debias from M-estimation under Two-Phase Multiwave Sampling, run with its residual-variance allocation. It reports up to 1.8x effective sample size gain at 90% coverage on its own benchmarks. There is no reported result on Lacuna for any debiased-inference method on materials or synthesis data. Its performance here is therefore an unknown we measure rather than a number we can quote. We also run design-based supervised learning and the cross-fitted doubly robust estimator under uniform sampling.
All methods share fairness constraints. They see identical proxy features and identical cross-fitting folds. They use the same proxy models, retuned under a fixed hyperparameter budget, and the same gold budgets of 10%, 20%, and 40% of samples. Results are averaged over 500 retrospective resamples of the fully measured grid.
The key ablation removes the transformation-boundary term from our allocation, leaving everything else fixed. If it adds no efficiency over plain residual-variance allocation, the physical ingredient is not earning its place, and the contribution reduces to a careful application study. That is still useful, but it is a smaller claim than the one we make.
Why not just…?
Why not just use a better proxy, such as higher-level DFT or a fuller characterization suite? Better proxies shrink the error but do not make it independent of the synthesis parameter. That dependence is what breaks inference. Using Imperfect Surrogates for Downstream Inference documents coverage collapse even with fairly accurate surrogates when errors correlate with covariates. The collapse at 450 °C will defeat any proxy calibrated mostly on intact frameworks. Our design also absorbs better proxies directly: a stronger proxy simply yields narrower debiased intervals. The computational arm tests this by swapping proxy models of differing accuracy and checking that surrogate-only bias persists while the debiased estimator tightens.
Why not just run Bayesian optimization or active learning? Bayesian optimization answers a different question: which single condition is best. It does not ask how a parameter affects performance, and how sure we should be. Its adaptively chosen points carry no inclusion probabilities, so a post hoc regression on them has no coverage guarantee. Matbench Discovery also found that the Bayesian-optimization wrapper BOWSR did not beat plain MEGNet on discovery acceleration. Still, a chemist who only wants the optimum may reasonably prefer it. We therefore report, as a side comparison, whether each method's selected optimum on the NiCo grid matches the full-data optimum.
Why not just fit a calibration curve from proxy to gold on the measured subset and plug it in? This is the real threat. With a random subset and a flexible model, regression calibration can approach the debiased estimators' point accuracy. It fails in two places. Calibration fitted mostly on intact-regime samples extrapolates poorly across the collapse boundary. And its naive intervals ignore calibration-model error. The fully measured grids settle the question directly. We report plug-in calibration's bias and coverage, with bootstrap intervals, beside the debiased estimators. If plug-in calibration matches them in coverage, we will say so, and the contribution narrows to the sampling design.
Why not just measure the gold standard on every sample? For a 120-sample grid we do exactly that, but only to validate. Long cycling and realistic-humidity tests are what limit real campaigns to 3 temperatures or 5 compositions in the source papers. The point is to let future campaigns explore wider grids at the same measurement cost.
Decisive Checks
The proposal succeeds if three things hold on the fully measured NiCo-LDH grid. First, the debiased estimators hold nominal 95% coverage for the Ni:Co and etching-time effects at a 20% gold budget, while surrogate-only analysis with surface area or a single CV scan does not. Second, the surrogate-only analysis misplaces the composition optimum relative to the full-data truth, consistent with the published surface-area versus capacitance tension. Third, boundary-aware allocation achieves a measurably larger effective sample size than residual-variance allocation on the Cu-MOF pyrolysis grid, which spans the collapse temperature. Equal efficiency there would falsify the physical-design claim.
The computational arms must reproduce the coverage pattern with exact ground truth. They should show surrogate-only bias concentrated in near-threshold families, where Matbench Discovery locates its triangle of peril. If these checks pass, the deliverable is a measurement protocol that templated-synthesis labs can adopt: which samples to fully characterize, and how to report parameter effects with honest uncertainty.
References
- Dissecting racial bias in an algorithm used to manage the health of populations
- Fabrication and electrochemical properties of hollow cage-like nickel cobalt layered hydroxides with porous structure
- MOF derived core-shell CuO/C with temperature-controlled oxygen-vacancy for real time analysis of glucose
- Exploring MOF-Derived CuO/rGO Heterostructures for Highly Efficient Room Temperature CO2 Sensors
- Fairness Evaluation in Presence of Biased Noisy Labels
- Using Imperfect Surrogates for Downstream Inference: Design-based Supervised Learning for Social Science Applications of Large Language Models
- M-estimation under Two-Phase Multiwave Sampling with Applications to Prediction-Powered Inference
- Towards the Efficient Inference by Incorporating Automated Computational Phenotypes under Covariate Shift
- Matbench Discovery - Can machine learning identify stable crystals?
- PRO-MOF: Policy Optimization with Universal Atomistic Models for Controllable MOF Generation
- MOF Synthesis Prediction Enabled by Automatic Data Mining and Machine Learning**
- From Molecules to Materials: Pre-training Large Generalizable Models for Atomic Property Prediction