When computational chemists screen catalysts for alkaline hydrogen evolution, the quantity that matters most is the energy needed to split a water molecule on the catalyst surface, called the water dissociation barrier. Machine-learned surrogates now estimate these barriers far faster than quantum-mechanical simulation. To say how much to trust them, practitioners check the surrogate against a set of reference barriers computed with density functional theory (DFT), the standard quantum-mechanical method. The reference set has a hidden flaw. The DFT searches that fail to converge are thrown away, and they fail most often on the rough, hard, high-barrier surfaces. What remains resembles the "healthy volunteer" cohort of a biobank: the reference set over-represents well-behaved reactions. Any error bar calibrated on it will be too optimistic on exactly the candidates a screen should reject.
This note proposes a way to keep the failed calculations instead of discarding them. A failed search still carries information, because physics bounds the barrier it was trying to find. The highest energy along any path that connects the starting and ending states is at least the true barrier. The reaction energy, clipped at zero, is at most the true barrier. A failed calculation therefore gives an interval rather than nothing. We plug these intervals into conformal selection, a family of statistical methods that pick candidates while bounding the fraction of false picks. The key step is to substitute each failed calculation's worst-case label. Doing so restores the statistical guarantee without having to model why calculations fail. The claim is testable, and it is specific to how barrier data are generated.
Core Claim
The obvious combination of these fields is to wrap an existing conformal selection method around a pretrained barrier model and screen for low barriers. That is a straightforward application, and it inherits the flaw above. Selection by Prediction with Conformal p-values introduced conformalized Benjamini–Hochberg (cfBH). The method builds a p-value for each candidate from calibration scores and caps the false discovery rate (FDR) at a target $q$. Its guarantee requires that the calibration and candidate pools be exchangeable, meaning statistically interchangeable, and the authors name distribution shift from selection mechanisms as unresolved. Multivariate Conformal Selection and Confidence on the focal make the same assumption about the calibration data. JOMI, the method in that second paper, corrects for the screening selection: the bias introduced when we report intervals only for the candidates we picked. It does not correct for bias in which labels exist.
Our claim is that, for reaction barriers, label-level self-selection can be neutralized by design rather than by modeling. The procedure has three parts. First, draw the calibration reactions uniformly at random from the candidate pool. Second, attempt DFT on every one of them. Third, treat each non-converged search as an interval-censored label, meaning a label known only to lie within an interval. For a reaction with DFT-relaxed initial and final energies $E_{IS}, E_{FS}$ and any connecting path $\pi$ whose energies we have evaluated with DFT, the true forward barrier $E_a$ satisfies
$$\max(0,\,E_{FS}-E_{IS}) \;\le\; E_a \;\le\; \max_{x\in\pi} E(x) - E_{IS}.$$
The lower bound holds because the saddle point cannot lie below either endpoint. The upper bound holds because the barrier is the minimum, over all connecting paths, of each path's highest point. We test the null hypothesis "the barrier is too high," $E_a > \gamma$ for a target threshold $\gamma$. For that test, the right substitution for a censored calibration label is the value in its interval that makes the calibration point look most null-like. Under the cfBH construction this can only raise the p-values, so FDR control survives at the cost of some power. The cost should be small when the intervals are tight, and many failed searches get most of the way to a saddle before stopping, so tight intervals should be common. The method needs no model of why a search fails. That matters here: convergence failure plausibly depends on properties no covariate captures, such as unmodeled surface reconstruction, spin state, or a rough energy landscape.
Research Program
The pieces are assembled around the Volmer step, $H_2O + e^- + * \to H^* + OH^-$, the water-splitting step that limits hydrogen evolution in alkaline media. Bifunctional interfaces such as Pt/Mo2N-NrGO and Ce single atoms on Ru accelerate this step through dual-site stabilization. These papers motivate direct barrier prediction over scaling-relation shortcuts. They are also the complex surfaces on which searches for the transition state, the highest point on the path, are most likely to fail.
The surrogate is the pretrained GemNet-OC workflow from CatTSunami, which supplies our method. It computes barriers zero-shot, meaning without task-specific training, using the nudged elastic band (NEB). NEB represents the reaction path as a chain of intermediate structures, called images, and relaxes the chain toward the minimum energy path. CatTSunami then checks the result with DFT single-point energies. That paper also documents the practice we target: NEBs that fail or deviate are discarded or sent to fallbacks, with no statistical accounting.
We build three things. The first is a label pipeline that, for every attempted DFT NEB, records either a converged barrier or a certified interval. The upper bound needs care. The maximum over discrete images does not bound the maximum along the continuous path. We therefore add DFT single points at interpolated geometries between the highest images until successive refinements change by less than a tolerance. We report the residual slack as an explicit additive term. We also use DFT-relaxed endpoints for the lower bound, and reactions whose endpoints themselves fail to relax are logged separately.
The second is the censoring-aware selection procedure itself. It is built on cfBH, and a multivariate variant on mCS-dist, the distance-based score from Multivariate Conformal Selection. The multivariate variant jointly requires a low water dissociation barrier and a hydrogen adsorption free energy $\Delta G_{H^*}$ near zero. We expect the model's error on barriers to be systematically one-sided, because saddle geometries have stretched bonds that equilibrium training data rarely cover. We therefore use one-sided or asymmetric scores, following Regression Conformal Prediction under Bias. That paper shows symmetric intervals pay a $2|b|$ width penalty under constant bias $b$.
The third is an audit of how much outcome-dependent dropout actually distorts guarantees in barrier data. This is the self-selection question stated quantitatively: what we learn by comparing reactions with and without converged labels.
Baselines
The do-nothing reference is the practice CatTSunami models: rank candidates by GNN-predicted barrier and take the top $K$, or all below $\gamma$, with no error control. It sets the power ceiling and shows the realized false discovery proportion when nothing is guaranteed. CatTSunami reports an average 91% success rate in finding transition states within 0.1 eV of DFT on OC20NEB. That figure is a point accuracy, not a selection error rate, and it is computed only over reactions where the comparison exists.
The strongest existing competitor is cfBH calibrated on converged reactions only. This is exactly what a careful practitioner would do today with published tools. It uses the same GNN, the same calibration budget in DFT calls, the same $q$, and the same score family as our method. We found no reported result on Lacuna for cfBH or any conformal selection method on reaction barriers. The nearest reported numbers are from other domains: mCS-learn reaches a selection power of 0.431 at $q=0.3$ on the ADMET benchmark (22,805 molecules), where the Bonferroni intersection baseline reaches 0.000 (Multivariate Conformal Selection). Beating converged-only cfBH means two things. First, holding realized FDR at or below $q$ in regimes where converged-only cfBH exceeds it. Second, retaining at least roughly 80% of its power where both are valid.
The covariate-correction competitor is cfBH with inverse-probability weighting. Each converged calibration point is reweighted by one over its predicted probability of converging. That probability comes from a classifier on structural and elemental features plus GNN outputs, tuned by cross-validation on the same calibration reactions. The ensemble competitor thresholds a deep-ensemble variance at a level chosen to hit $q$ on a held-out converged split.
Two ablations isolate the key ingredient. The first uses only the lower bound on censored labels, which is the anti-conservative direction. It should break validity, showing that the direction of substitution matters. The second replaces certified bounds with GNN-predicted barriers for the failed reactions, testing whether physics, rather than simply having more labels, drives the result. A further ablation swaps symmetric for asymmetric scores and measures interval width at equal coverage.
Why not just…?
Why not just use deep ensembles or dropout variance? These estimates carry no finite-sample guarantee. They are known to be overconfident under distribution shift (Conformal Prediction for Rigorous Uncertainty Quantification), and oxide surfaces show large out-of-domain error spikes (OC22). More to the point, an ensemble tuned on converged reactions inherits the same dropout bias as converged-only cfBH. The ensemble baseline reports realized FDR on the full-label ground truth, so this is measured rather than assumed.
Why not just reweight by predicted convergence probability? This is the real threat. If failure is well predicted by observable features, weighting is valid and loses less power than worst-case substitution. The evidence against assuming so comes from Joyal-Desmarais et al.. In their study, covariate adjustment of self-selected survey samples increased the discrepancy from census-matched ground truth in 45% of about 17,000 specifications, because selection acted on unobserved factors. Whether NEB failure behaves that way is empirical. Our semi-synthetic experiment has true labels for every dropped reaction, so it settles the question directly. If weighting matches our method, we will report that and recommend it where convergence is predictable.
Why not just rerun failed NEBs until they converge? Reruns are expensive. Some searches never converge under a fixed budget, and those residual failures are precisely the hard tail. Reruns shrink the problem but do not remove the dependence of the reference set on the outcome. We report how coverage error scales with the number of rerun attempts.
Why not just use JOMI? JOMI corrects for the bias of reporting intervals only on the selected candidates. It assumes clean, exchangeable calibration labels. The two corrections are complementary, and our censored labels can feed JOMI's swap construction as a secondary deliverable.
Proposed Experiments
The decisive experiment is semi-synthetic, on OC20NEB: the 932 RPBE DFT NEBs from CatTSunami, restricted to dissociation reactions and emphasizing O–H cleavage wherever the dataset contains it. The design has four steps. First, run the ML NEB workflow on every reaction. Second, use the realistic failure flags it produces to decide which reactions lose their labels. Third, construct certified intervals for those reactions from the DFT single points along their failed paths. Fourth, repeat random calibration and test splits many times. Because the true DFT barrier is known for every reaction, the realized false discovery proportion and power of each method are exact.
We add a stress condition that makes dropout depend more strongly on the barrier. This maps how far converged-only cfBH drifts past $q$ as selection strengthens, and whether our method stays below it.
The second experiment is prospective. The target is water dissociation on oxide-supported metal and bifunctional surfaces drawn from OC22 compositions, where the GNN is least reliable. We calibrate on a random draw of attempted DFT NEBs, keeping every one, and select with $q=0.2$ jointly on barrier and $\Delta G_{H^*}$. We then verify the selected set with full DFT. Success means the realized false discovery proportion is at or below $q$ and the selection is non-empty, with fewer false picks than converged-only cfBH at comparable selection size.
The audit reports three quantities. The first is the fraction of calibration reactions censored. The second is the median width of the certified intervals. The third is the gap between barriers of converged and failed reactions, which directly measures self-selection in barrier data.
References
- Selection by Prediction with Conformal p-values
- Multivariate Conformal Selection
- Confidence on the focal: Conformal prediction with selection-conditional coverage
- A Strongly Coupled Cluster Heterostructure with Pt–N-Mo Bonding for Durable and Efficient H2 Evolution in Anion-Exchange Membrane Water Electrolyzers
- Oxophilic Ce single atoms-triggered active sites reverse for superior alkaline hydrogen evolution
- CatTSunami: Accelerating Transition State Energy Calculations with Pretrained Graph Neural Networks
- Bias-Aware Conformal Prediction for Metric-Based Imaging Pipelines
- The Open Catalyst 2022 (OC22) Dataset and Challenges for Oxide Electrocatalysis
- How well do covariates perform when adjusting for sampling bias in online COVID-19 research? Insights from multiverse analyses