The catalysts that clean cold exhaust need two properties. They must work below about 200 °C, and they must keep working after months of heat and humidity. They also use platinum-group metals, and the aim is to use as little of those as possible. In practice the low-temperature number is cheap to measure: one light-off run, a few hours, records the temperature at which the catalyst converts 90% of carbon monoxide and hydrocarbons. The durability number is expensive. It requires aging the sample at high temperature for days, as in the 100 h hydrothermal aging at 700 °C of Activating low-temperature diesel oxidation by single-atom Pt on TiO2 nanowire array. Every automated discovery loop therefore runs as a funnel. The optimizer decides which candidates earn the expensive aging test, and those aged results become its only training data for durability. This is the same feedback loop studied in algorithmic fairness, where a lender only observes repayment for applicants it approved. In both settings the model's own choices decide which labels exist, and its errors stay hidden in the regions it never tests.
We propose to treat the aging stage of a closed-loop catalyst search as a selective-labels problem. The selection policy is ours, so we can randomize it and know the exact probability that each candidate was sent to aging. With those probabilities, a doubly robust estimator from causal inference can give unbiased estimates of how the claimed platinum-lean, low-temperature Pareto front actually behaves after aging. A Pareto front is the set of candidates where no objective can be improved without worsening another. The fairness literature cannot use this tool cleanly because it rarely knows the selection probabilities. A self-driving lab can know them by construction. The practical deliverable is a search loop whose final report includes a confidence interval on aged performance, and an empirical answer to whether standard funnel optimizers systematically overstate how durable their fronts are.
Core Claim
In low-temperature oxidation there is a physical reason to expect the cheap and expensive objectives to be coupled in a misleading way. The features that make a catalyst light off early are the ones that make it fragile. These include highly dispersed or single-atom metal, weakened metal–support interaction (which lowered toluene T50 from 210 °C to 170 °C for CuO/TiO2 in the work summarized in this direction), and reducible-oxide promoters whose oxygen chemistry depends on surface hydroxyl and water desorbing above about 150 °C (Exploring the roles of oxygen species in H2 oxidation at β-MnO2 surfaces using operando DRIFTS-MS).
A loop that ages only the fresh-light-off winners is therefore biased in two ways. It learns durability only where activity is high, and it never sees the moderately active but robust compositions that may dominate the true aged front.
Our claim has two parts. First, standard multi-objective Bayesian optimization run as a funnel produces aged-front estimates that are biased in a measurable direction. Second, a randomized aging policy with a floor on each candidate's aging probability, combined with doubly robust correction, removes most of that bias while giving up little of the achieved front.
What Exists and What Differs
Two papers supply the optimization machinery, and both assume that the data the optimizer collects is an unbiased view of a fixed landscape. Navigating the unknown with AI: multiobjective Bayesian optimization of non-noble acidic OER catalysts ran q-noisy expected hypervolume improvement (qNEHVI) over 1,011 Co–Mn–Sb–Sn–Ti oxides. qNEHVI is the acquisition rule of Parallel Bayesian Optimization of Multiple Noisy Objectives with Expected Hypervolume Improvement, which picks batches of points expected to enlarge the area dominated by the Pareto front. The study traded activity against dissolution and reached its front after 61 compositions, a 17x speedup over grid search. Evolution-guided Bayesian optimization for constrained multi-objective optimization in self-driving labs added an evolutionary search and constraint pre-repair (EGBO) and reported 14% higher hypervolume than qNEHVI in silver nanoparticle synthesis. Neither study asks whether its expensive objective was measured on a biased subset, and neither addresses low-temperature thermal catalysis.
On the catalysis side, Hoang et al. motivate the problem. They achieved T90 of about 160 °C with 5x less platinum-group metal, but as a single hand-designed point, with no front relating loading, light-off and aging.
The feedback-loop papers supply the statistics. Causal Modeling for Fairness In Dynamical Systems used doubly robust estimators to evaluate lending policies from logs with missing outcomes. Long-term Fairness with Selective Labels studies exactly the missingness we face. Achieving Long-Term Fairness in Sequential Decision Making shows how badly static models drift when decisions shape later data: unfairness reaches 0.481 at step 5 for a demographic-parity model.
These papers also constrain the design. They warn that identifiability fails under unobserved confounding and a misspecified causal graph. That warning is why we insist on known, randomized propensities rather than estimated ones.
Adaptive Combinatorial Experimental Design: Pareto Optimality for Decision-Making and Inference proves that exploiting the best arms and estimating differences between arms are conflicting goals in bandits. We use this result to frame the propensity floor as a deliberate point on that trade-off, not a free lunch.
Research Program
Each candidate $x$ is a composition and synthesis recipe, for example a Pt loading on TiO2 with a Cu, Mn, Fe or Co oxide promoter. Each receives a cheap fresh light-off measurement $y_c$ (T90 under a water-containing simulated exhaust) and a known cost (platinum-group metal mass). The loop then sends it to aging with probability $\pi(x, y_c) \ge \pi_{\min}$. That probability is a tempered softmax of the acquisition value, floored at $\pi_{\min}$. If aged, the candidate yields $y_a$, its aged T90.
For any quantity $g(y_a)$ we care about, we estimate the mean over a candidate set of size $n$ as
$$\hat\mu_{DR} = \frac{1}{n}\sum_{i=1}^{n}\Big[\hat m(x_i, y_{c,i}) + \frac{S_i}{\pi_i}\big(g(y_{a,i}) - \hat m(x_i, y_{c,i})\big)\Big].$$
Here $S_i$ is 1 if candidate $i$ was aged, $\pi_i$ is its logged aging probability, and $\hat m$ is the Gaussian-process surrogate's prediction of $g$. Two choices of $g$ matter most. The first is the indicator that the aged T90 stays below a target, which gives the "aging survival rate" of the claimed front. The second is the aged T90 itself, averaged within regions of composition space. The estimator is unbiased whenever the propensities are correct, whether or not the surrogate is. Because we log the propensities ourselves, that condition holds by design.
The corrected regional estimates feed back into the surrogate as a calibration offset. The acquisition then stops trusting durability predictions in regions it has under-sampled.
The program has three parts. The first builds the randomized funnel on top of qNEHVI and EGBO. The second characterizes how $\pi_{\min}$ trades achieved hypervolume against estimation error, an empirical version of the Xie et al. frontier. The third tests whether the bias exists in real low-temperature chemistry, not only in simulation.
Baselines
The naive reference is a uniform-random funnel: candidates are chosen at random, and a fixed random fraction is aged. It is unbiased but wasteful, and it sets the floor for achieved front quality.
The strongest existing method is qNEHVI run as a standard funnel, in which the top candidates by fresh light-off and cost are aged. Jenewein et al. report reaching their activity–stability front after evaluating 6% of the space (61 of 1,011 compositions, 17x faster than grid search). EGBO is the stronger constrained variant, reported at 14% higher hypervolume and 24% better front uniformity than qNEHVI by Low et al. We run both from public BoTorch-style implementations, with the same batch size (15, matching Jenewein), the same aging budget, the same surrogate kernel family, and the same hyperparameter tuning budget, all on identical candidate pools.
A third baseline is a multi-task Gaussian process that conditions aged predictions on the observed fresh value. This is the principled "the model already handles missing data" answer.
The ablation that isolates our ingredient keeps the randomized policy with its propensity floor but uses the plain surrogate without doubly robust correction. A second ablation applies doubly robust correction with propensities estimated after the fact from a deterministic funnel. Comparing these separates the value of randomization from the value of the estimator.
Beating the baselines means two things. Our method must cut the absolute error of the estimated aging survival rate of the reported front by at least half relative to the qNEHVI funnel, with nominal 90% confidence-interval coverage. It must also lose no more than 5% of the true aged hypervolume relative to EGBO at equal budget.
Why not just…?
Why not just age every candidate? Then there is no selection bias, but there is also almost no search. Hoang et al.'s aging protocol takes 100 h per run at 700 °C. Aging all 61 candidates that Jenewein-scale campaigns evaluate would cost roughly 6,000 furnace-hours against a few hours of light-off each. The funnel exists because of this cost ratio, and our question is how to keep the funnel honest.
Why not just let a multi-task Gaussian process model the correlation between fresh and aged performance? This is the real threat. If aging decisions depend only on observed fresh values and the joint model is correctly specified, the missingness is ignorable and the Gaussian process is already consistent. Our argument is that sintering and water-blocking effects make the fresh-to-aged relationship strongly non-stationary across composition, so the model will be misspecified exactly where it has no aged data. The multi-task baseline settles this directly. In simulations where we control misspecification, and in the wet-lab campaign, if it matches doubly robust error, the contribution reduces to the propensity floor's exploration value and we will report it that way.
Why not just screen in silico and skip the loop? Static adsorption-energy predictors are built for zero-temperature clean surfaces. Their errors (about 0.4 eV energy MAE for DimeNet++ on OC20, 0.829 eV out-of-domain for GemNet-OC on OC22) are far larger than the 0.2 eV activation energies that matter here. OCx24 reached only R² = 0.59 even for the simpler hydrogen-evolution reaction and failed on CO2 reduction, where operating conditions dominate. Hydroxyl blocking below 150 °C and hydrothermal sintering are not visible to these models. We still use OC20-derived descriptors as surrogate inputs.
Why not just weight the objectives into one score? Linear weights cannot reach concave parts of a front (Cheng et al.). More importantly here, scalarization does nothing about which candidates get aged, so it inherits the same bias.
Proposed Experiments
The first stage is semi-synthetic, where ground truth is known. We fit a generative simulator to the Jenewein quinary-oxide data, with activity as the cheap objective and dissolution as the expensive one. We then add a tunable "fragility coupling" between them. We also run the MW7 and ZDT suites used by Low et al., with one objective made expensive and selectively observed. Across coupling strengths and values of $\pi_{\min}$, we measure the bias and RMSE of the aged survival-rate estimate, interval coverage, true aged hypervolume, and the false-Pareto rate, meaning the fraction of reported front members that are dominated after aging.
The decisive check is a sign test. If the qNEHVI and EGBO funnels do not overstate aged survival as coupling grows, the premise fails and the project stops there.
The second stage is a wet-lab campaign on Pt-lean TiO2-supported diesel oxidation catalysts with first-row transition-metal promoters. The three objectives are platinum-group metal loading, fresh water-containing T90 for CO and hydrocarbons, and T90 after hydrothermal aging. We run matched-budget campaigns with the qNEHVI funnel and with our randomized loop, using the Hoang et al. single-atom Pt result (T90 of about 160 °C at 5x lower metal loading) as the reference point on the front.
The decisive check comes after both campaigns finish. We age a held-out random audit set drawn from both campaigns' claimed fronts and compare each method's predicted survival rate with the audited one. No reported result was found on Lacuna for aged light-off Pareto fronts of this kind, so this audit is itself a new measurement. The proposal succeeds if our estimate falls inside its interval and the funnel's does not.
References
- Activating low-temperature diesel oxidation by single-atom Pt on TiO2 nanowire array
- Exploring the roles of oxygen species in H2 oxidation at β-MnO2 surfaces using operando DRIFTS-MS
- Navigating the unknown with AI: multiobjective Bayesian optimization of non-noble acidic OER catalysts
- Parallel Bayesian Optimization of Multiple Noisy Objectives with Expected Hypervolume Improvement
- Evolution-guided Bayesian optimization for constrained multi-objective optimization in self-driving labs
- Causal Modeling for Fairness In Dynamical Systems
- Long-term Fairness with Selective Labels
- Achieving Long-Term Fairness in Sequential Decision Making
- Adaptive Combinatorial Experimental Design: Pareto Optimality for Decision-Making and Inference
- The Open Catalyst 2020 (OC20) Dataset and Community Challenges
- The Open Catalyst 2022 (OC22) Dataset and Challenges for Oxide Electrocatalysts
- Open Catalyst Experiments 2024 (OCx24): Bridging Experiments and Computational Models
- Preference-Based Policy Iteration: Leveraging Preference Learning for Reinforcement Learning