Audited Desk Rejection for Hydrogen-Carrier Screening

Computational searches for hydrogen-storage molecules work like a journal that desk-rejects almost every submission. A cheap test, such as a rough energy estimate or a similarity check against known carriers, discards most candidates. A slower calculation looks at the survivors, and only a handful reach the expensive, accurate quantum-chemistry tier. Real editorial boards have a safeguard that these pipelines lack. The editor-in-chief does not fully trust the associate editors. The board occasionally re-examines what was rejected, so it can learn which gatekeepers are throwing away good work. We propose to build that safeguard into molecular screening. The highest-accuracy tier would spend a small, fixed share of its budget re-evaluating a random sample of each cheap tier's rejections. From that sample it estimates how many good molecules each tier is silently losing, and it adjusts that tier's cutoff while the search runs.

This matters because the target is narrow. A liquid organic hydrogen carrier (a liquid molecule that stores hydrogen by chemically bonding it and releases it on heating) needs at least 5.5 weight-percent hydrogen and a release energy between 40 and 70 kJ per mole of hydrogen. Higher capacity tends to push release energy out of that window. In the QM9-LOHC reference set, 54.1% of reactions fall in the unfavourable 70–120 kJ/mol band and only 3,040 of 10,373 sit in the target window (Accurate Dehydrogenation Enthalpies Dataset for Liquid Organic Hydrogen Carriers). When good candidates lie near a sharp boundary, a cheap filter with a slightly biased cutoff can remove a large share of them, and nobody measures the loss.

Core Claim

A multi-fidelity screening cascade can monitor itself. Random audits of rejected candidates at the top fidelity give an unbiased estimate of each tier's miss rate, and that estimate can be used to re-set cutoffs online. Under a fixed compute budget, we expect this to recover more in-window carriers than a fixed-threshold cascade. The advantage should be largest when the search moves into chemistry the cheap filters were not tuned on, such as nitrogen heterocycles.

The governance literature motivates both the idea and its vocabulary. Principal-agent theory describes how delegated agents drift from the principal's goals when monitoring is weak. Its remedies are pre-vetting agents and independent supervisory checks (Dilemmas and Governance Strategies in the Implementation of the Modern Apprenticeship System in Chinese Vocational Colleges: A Perspective from Principal-Agent Theory). The journal structures documented for IEEE transactions describe a cascade in which an editor-in-chief sets standards and associate editors gate their own areas (IEEE Transactions on Cybernetics publication information; IEEE Transactions on Information Theory publication information). We do not borrow the metaphor loosely. We borrow one specific mechanism, the independent audit of rejections, and make it a statistical estimator.

The obvious alternative pairing would be a board of language-model "associate editors" debating molecules. We treat that as a baseline, not the proposal. Its evidence base comes entirely from question-answering and household-task benchmarks. It would also need to beat numerical methods that never talk.

Research Program

The base system copies the four-stage pipeline of Uncovering novel liquid organic hydrogen carriers. That pipeline screened more than 160 billion molecules from ZINC15 and GDB-17 with a structure filter, a heuristic dehydrogenation index, density functional theory (DFT, a mid-cost quantum calculation), and capacity and melting/boiling-point models. It produced 41 candidates. Its rejections are final, and that is the property we change.

Each cheap tier $k$ keeps its cutoff but sends a random fraction $q_k$ of its rejections to the top tier, which uses G4MP2 composite thermochemistry. If $m_k$ rejections are audited and $h_k$ of them turn out to be in-window hits, the estimated number of hits that tier $k$ discarded is

$$\hat M_k = \frac{N_k}{m_k}\, h_k,$$

where $N_k$ is the total number of molecules tier $k$ rejected. This is an unbiased estimate, with binomial standard error shrinking as $1/\sqrt{m_k}$, because the audited rejections are a uniform sample. Each tier's cutoff is then moved to equalise the marginal hits per unit of compute across tiers. In plain terms, the next unit of compute goes wherever the audits say good molecules are being lost fastest.

A second ingredient decides what moves a molecule up the cascade. Voting or Consensus? Decision-Making in Multi-Agent Debate found that consensus works as a strict filter and voting works as an explorer. Consensus beat voting by 2.8% on average on knowledge tasks, and voting beat consensus by 13.2% on reasoning tasks. We turn this into a testable hypothesis about screening rules. Promotion would be by "any-vote": a molecule moves up if any single property checker scores it strongly. Final acceptance would be by consensus: every checker must pass. We test this against all-pass and majority rules.

A third design choice follows from Peacemaker or Troublemaker. There, an independent judge cut disagreement collapse on CommonsenseQA from about 81% to 41% (disagreement collapse is the rate at which agents abandon correct disagreeing answers to agree with peers). Our analogous requirement is that the auditor share no training data or fitted parameters with the tiers it audits. We ablate this by auditing with a surrogate fitted on the same data as the tier.

G4MP2 is not ground truth everywhere. It matches 14 known carriers to a mean absolute error of 6.3 kJ/mol, but it breaks down systematically on aniline-like nitrogen species (Accurate Dehydrogenation Enthalpies Dataset for Liquid Organic Hydrogen Carriers). We therefore flag those classes and report results with and without them.

As a secondary, solid-state setting, we use the HyMARC metal-hydride data with a learned capacity surrogate as the cheap tier. There the audit question becomes how much a surrogate's errors cost. The hybrid model of A Hybrid Machine Learning Framework for Predicting Hydrogen Storage Capacities in Metal Hydrides overpredicted Al₁₁O₁₈ at 5.98% against a DFT value of 4.61%, which is exactly the kind of error an audit should catch.

Baselines

The naive reference point is the fixed-threshold cascade with cutoffs set as in Harb et al. and no audits. It runs on the same candidate pool, with the same surrogates and the same total compute budget measured in G4MP2-equivalent units. The audit system spends part of its budget on audits, so it must earn that spending back. The headline count from that work is 41 candidates out of more than 160 billion molecules. We found no reported recall on QM9-LOHC for this pipeline on Lacuna, so we compute that number ourselves.

The strongest competitor is the same cascade with cutoffs widened by a safety margin tuned offline on a labelled calibration split of QM9-LOHC. It uses an identical tuning budget, and its margin is chosen to maximise recall at the fixed cost. This is the honest version of "just be less strict." It should match auditing in-distribution, so the audit system beats it only if it recovers more in-window reactions on a held-out chemical family at equal cost.

The third baseline is multi-fidelity active learning. A surrogate model of release energy is retrained on every top-tier result, and the next molecules are chosen by expected improvement, under the same budget. This baseline has no published number in the file; no reported result found on Lacuna for multi-fidelity active learning on QM9-LOHC.

The fourth baseline is a language-model editorial board. Specialist "associate editor" agents for thermodynamics, phase behaviour and synthesizability use drafting-in-isolation and a judge, following Embodied LLM Agents Learn to Cooperate in Organized Teams and the debate papers above. It gets the same tool access and budget.

Two ablations isolate the key ingredients. The first keeps the audit's compute but spends it on extra top-tier evaluation of promoted molecules instead of rejected ones. This tests whether looking at rejections specifically is what helps. The second replaces any-vote promotion with all-pass promotion.

Why not just…?

Why not just loosen the thresholds? This is the most serious threat. In-distribution, a well-tuned margin may do everything audits do. The difference shows up only when the search moves to chemistry where the cheap filter's bias changes, and an offline margin cannot detect that change. Our held-out-family experiment settles it. If the tuned margin matches audits on held-out nitrogen heterocycles at equal cost, we will report that audits add only monitoring value, not hits.

Why not just use multi-fidelity Bayesian optimisation? Active learning picks the most promising molecules. It never revisits molecules the cheap filter discarded, so it cannot estimate the cascade's miss rate. The two methods may combine: audits could supply unbiased negatives to the surrogate. We test active learning alone, audits alone, and both together. If active learning alone wins, the simpler method is the recommendation.

Why not just let a language-model editorial board deliberate? The multi-agent evidence shows that structure helps language agents. A designated leader improved completion time by 9.76% in VirtualHome-Social, and a judge halved collapse. But none of these results involve numerical property boundaries, where a calculation is cheaper and more reliable than an argument. We include the board as a baseline. We expect it to lose on hits per unit of compute, and we will report it fairly if it does not.

Proposed Experiments

The main experiment simulates the cascade on QM9-LOHC, using its G4MP2 enthalpies as the top tier. Cheap tiers are the dehydrogenation index, B3LYP DFT (recomputed where it is not released, which is feasible for molecules of up to nine heavy atoms), and an ML surrogate. Because every label is known, recall of the 3,040 in-window reactions can be measured exactly at budgets from 1% to 20% of full G4MP2 cost. We also check that $\hat M_k$ falls within its confidence interval of the true miss count.

The decisive check is a family-shift split. Cutoffs and margins are tuned on carbocycles and oxygen heterocycles, then deployed on nitrogen heterocycles. That is the family where Harb et al. found release energies are systematically lowered and where G4MP2 has known failures. Audits should beat the tuned margin there or the idea fails. A secondary run on HyMARC measures recovered high-capacity hydrides against the gradient-boosting surrogate ($R^2 = 0.794$) and the hybrid surrogate ($R^2 = 0.85$) as cheap tiers. A small out-of-sample screen of GDB-17 fragments then checks whether audit-recovered candidates survive top-tier evaluation at the same rate as candidates the cascade promoted normally.