Cement makes 5 to 8 percent of global CO₂ emissions. Much of the cut the industry needs has to come from replacing clinker, the kiln-fired ingredient, with industrial byproducts such as slag and fly ash. The machine-learning models used to design these low-clinker mixes are trained on historical records of mixes that someone chose to cast and report. That record is not a random sample. Mixes with very high replacement that crumbled were rarely written down. High replacement also tends to arrive together with high doses of chemical admixtures and low water content. A model fit to such data learns that high replacement "works," and it is most confident exactly where the data are most selectively thinned.
This error is costly. Before a new mix can be poured, it has to pass physical strength tests that take 28 days of curing and real money. Those tests are the transaction cost that makes contractors reluctant to adopt low-carbon binders. This proposal builds a mix-design procedure that treats each qualification test as the scarce resource. It uses a bound on strength that stays valid when failed mixes are missing from the data, so it proposes candidates that pass on the first or second try. The claim to be tested is concrete: under outcome-dependent selection in the training records, this procedure reaches a certified low-carbon mix in fewer tests than standard gradient-boosting or generative pipelines.
Core Claim
The file's three directions meet at one point. The friction that blocks low-clinker adoption is the cost and delay of verifying a mix. Selection bias in mix databases inflates predicted strength in the high-replacement region, so it directly raises that friction by producing failed verification tests. Existing concrete models do not model the missing failures. Research on prediction of compressive strength of fly ash and slag mixed concrete based on machine learning fits XGBoost to ingredient quantities. Prediction of Concrete Compressive Strength Based on ISSA-BPNN-AdaBoost fits a boosted neural network. Accelerated Design and Deployment of Low-Carbon Concrete for Data Centers trains a conditional variational autoencoder, a generator that samples mixes conditioned on target strength and global warming potential (GWP). All three treat the observed mixes as the population. Existing causal-inference work does handle selection, but never on a physical design loop where every query has a price.
We are careful about which bias we mean. In a lab dataset where every ingredient is recorded, ordinary confounding is limited: a flexible regression on all ingredients already estimates the effect of changing one of them, provided the data cover the region. The two failures that matter are different. The first is weak overlap: high replacement almost never appears with low admixture dose, so predictions there are extrapolations. The second is outcome-dependent selection: whether a mix enters the record depends partly on its strength. The first is detectable from the data. The second is not, and so it needs an explicit sensitivity assumption. The contribution is to handle both inside a test-budgeted certification loop and to show that the second is the one that costs tests.
Research Program
We would build three pieces.
The first piece is a strength model with a lower bound that accounts for selection. We fit an outcome model for 28-day strength as a function of the full mix. This uses the method of Automatic debiased machine learning and sensitivity analysis for sample selection models, which corrects estimates when outcomes are observed only for a non-random subset, using an automatically learned correction weight. Its sensitivity analysis bounds how far estimates move if selection depends on unobserved strength. We reuse that sensitivity machinery. We do not claim the data identify the selection mechanism; they cannot. Instead, a single parameter $\Gamma \ge 1$ caps how much more likely a strong mix is to be recorded than an equally composed weak one. For each candidate mix $x$, the method returns a lower bound $\underline{f}_\Gamma(x)$ on its expected strength. At $\Gamma = 1$ this reduces to an ordinary debiased estimate, and the bound widens as $\Gamma$ grows.
The second piece handles overlap. Representation-balancing methods such as Learning Representations for Counterfactual Inference and Estimating individual treatment effect: generalization bounds and algorithms penalize the distance between treated and control groups in a learned feature space. The second of these proves that effect error is bounded by fitting error plus that distance. Representation Learning for Treatment Effect Estimation from Observational Data shows that aggressive balancing can erase useful local structure.
We take the constraint from this line of work, not the architecture. Replacement level is a continuous dose. Estimating Continuous Treatment Effects with Two-Stage Kernel Ridge Regression estimates dose-response curves without estimating a treatment density, and it reports that density-based double machine learning degrades under weak overlap. We use it as the dose-response estimator for replacement fraction. We also add an explicit support penalty that shrinks $\underline{f}_\Gamma$ toward the specification threshold wherever kernel-estimated covariate density is low. Generalization Bounds for Estimating Causal Effects of Continuous Treatments instead bins the dose. It serves as a second continuous-dose comparison.
The third piece is certification with a test budget. Global warming potential is linear in the recipe: $\mathrm{GWP}(x) = \sum_i e_i x_i$, where $x_i$ is the mass of ingredient $i$ per cubic meter and $e_i$ is its emission factor. We would take the $e_i$ from the CSI Environmental Product Declaration tool used by Accelerated Design and Deployment of Low-Carbon Concrete for Data Centers. At each round the procedure proposes
$$ x_t = \arg\min_{x \in \mathcal{C}} \mathrm{GWP}(x) \quad \text{subject to} \quad \underline{f}_\Gamma(x) \ge f_{\text{spec}}, $$
where $\mathcal{C}$ is the candidate set and $f_{\text{spec}}$ is the required strength. The proposed mix is tested, the result is added to the data, and the procedure repeats until a mix passes. Each test has a unit cost plus a fixed 28-day delay. The quantity we care about is the total number of tests to the first certified mix below a regional GWP benchmark.
This structure borrows from Bi-Level Decision-Focused Causal Learning for Large-Scale Marketing Optimization: Bridging Observational and Experimental Data, which couples large biased logs with a small clean experiment and trains the estimator for the downstream decision. The difference is in where the clean data come from and what they cost. Bi-DFCL assumes a randomized trial is already available. Here the clean data are the tests themselves, each one expensive, so choosing which clean point to buy is the central decision. Tests also correct the selection bias: every failed test is a record that the historical data were missing.
Transaction-cost theory frames the setting but does not supply the method. Coase vs. Pigou in the Petroleum Market argues qualitatively that monitoring and verification costs erode market remedies for emissions. Ma et al. finds that perceived administrative complexity deters firms even when the economics are clear. Feasible supply of steel and cement within a carbon budget is likely to fall short of expected global demand finds that feasible Paris-compliant cement supply covers only 22–56% of projected 2050 demand, with no recycling loop to fall back on, so material substitution has to carry the load. These papers motivate treating verification cost as the binding constraint. We make that cost measurable in tests. We do not address the bargaining questions in Mitigating Externalities while Learning: an Online Version of the Coase Theorem. Its assumption of frictionless transfers is a separate problem.
Baselines
The naive reference point is to do nothing new: propose the regional benchmark mix with no added replacement, which passes but saves no carbon. A second naive point greedily picks the lowest-GWP mix in the training data that meets specification. Together these bound the cost-versus-savings tradeoff from both ends.
The strongest correlational competitor is XGBoost predict-then-optimize, following Research on prediction of compressive strength of fly ash and slag mixed concrete based on machine learning. That paper reports $R^2 = 0.95$ on the UCI Concrete test split and a 3.5% average error on 10 physical validation mixes. Prediction of Concrete Compressive Strength Based on ISSA-BPNN-AdaBoost reports $R^2 = 0.964$ on UCI and $0.970$ on the Chopra et al. set. We will run both inside the same certification loop, with a point prediction replacing $\underline{f}_\Gamma$. They get the same candidate set, the same test budget, and the same hyperparameter search budget, retuned after each round.
The generative competitor is the conditional variational autoencoder of Accelerated Design and Deployment of Low-Carbon Concrete for Data Centers. It reports up to 42% GWP reduction against regional benchmarks, and field slabs at 152–157 kg CO₂ eq./m³ against 281–334. Its samples are filtered by predicted strength, and it gets the same budget.
The strongest uncertainty-aware competitor is a Gaussian-process or conformal lower confidence bound on the same features, run as standard Bayesian optimization. This is the baseline most likely to match us, because it too is cautious outside the data.
Causal comparisons are two-stage kernel ridge regression and ADMIT, each without our selection bound. For scale only, two-stage kernel ridge regression reports a mean MISE of 1.2466 on Job Corps, against 2.4230 for random-forest double machine learning and 2.1065 for neural-network double machine learning. That is a different benchmark and not directly comparable.
Two ablations isolate the key ingredient. The first sets $\Gamma = 1$, which keeps overlap handling but assumes no outcome-dependent selection. The second drops the support penalty.
Beating the baselines means certifying a mix at or below a fixed GWP target in fewer expected tests than the best baseline, with non-overlapping bootstrap intervals over repeated selection draws. It also means doing so without a higher final GWP. No reported result was found on Lacuna for tests-to-certification on any concrete dataset, so every baseline number on this metric will be produced by us.
Why not just…?
Why not just use XGBoost, whose $R^2$ is already 0.95? An $R^2$ computed on a random split of selected data measures fit to the region the data already cover. It says nothing about the thinned high-replacement region where low-carbon mixes live. The CVAE authors themselves had to override superplasticizer doses by hand because the model missed admixture interactions (Accelerated Design and Deployment of Low-Carbon Concrete for Data Centers). Our induced-selection experiment tests this directly: if XGBoost's failed-test rate does not rise when failures are deleted from the training data, the claim is wrong.
Why not just use an uncertainty-aware Bayesian optimizer? This is the real threat. If the historical data contain no outcome-dependent selection, a Gaussian-process lower bound should do about as well as our method, because the only problem left is overlap, which it already handles. Where it should fail is the case where failed mixes were removed. The surviving data are then internally consistent, so predictive uncertainty stays small while the mean is biased upward. The experiment varies selection strength from none to severe. We expect parity at zero selection and a growing gap as selection increases. If the gap does not appear, the selection bound adds nothing, and we would report that.
Why not just run lab experiments from scratch and ignore historical data? Pure active search wastes early tests on mixes that the historical data already rule out. The cold-start variant of the loop, with an empty initial dataset, is included to measure how many tests the historical prior saves.
Why not just use econometric policy evaluation of adoption? Methods such as the PSM-DID study in Achieving carbon neutrality: How does the construction of national high-tech zones affect the green innovation of enterprises? Based on quasi-natural experiments in pilot areas in China estimate a regional policy effect of 0.112 on green patents. They answer whether an institution lowered frictions on average. They cannot say which recipe to test next, which is the friction studied here.
Proposed Experiments
The main experiment uses the UCI Concrete Compressive Strength data: 1,030 mixes, seven ingredients plus curing age, and measured strength. Following the IHDP recipe of removing a non-random subpopulation, we induce outcome-dependent selection. Mixes with high slag-plus-fly-ash fraction are deleted from the training data with probability that rises as their measured strength falls, at several severities. The deleted mixes, together with a held-out split, form the pool of "lab" mixes. A proposal is scored by the nearest pool mix within a tolerance on every ingredient, and that mix's real measured strength serves as the test outcome. Because the covariates and outcomes are real measurements, we know the truth for every test. We report four quantities: tests to first certification, GWP of the certified mix, the failed-test rate, and the empirical coverage of $\underline{f}_\Gamma$ on pool mixes across values of $\Gamma$. The last one checks whether the bound is calibrated when the true selection strength is known.
A second, external check uses the Chopra et al. dataset (76 fly-ash mixes from a different source). Models trained on UCI propose fly-ash mixes, which are scored against Chopra measurements. This tests whether the conservative bound transfers across laboratories better than point predictors. With only 76 samples, we treat this as a sign check, not a precise estimate.
The decisive result is the plot of tests-to-certification against selection severity for our method, the Gaussian-process lower-bound baseline, and the $\Gamma = 1$ ablation. The proposal succeeds if the three curves coincide at zero selection and ours separates as selection grows. The GWP of certified mixes must stay within noise of the XGBoost pipeline's GWP on the runs where that pipeline eventually succeeds. If the lower bound only makes the method so conservative that it certifies high-clinker mixes, then the saved tests were bought with lost carbon savings, and the claim fails.
References
- Research on prediction of compressive strength of fly ash and slag mixed concrete based on machine learning
- Prediction of Concrete Compressive Strength Based on ISSA-BPNN-AdaBoost
- Accelerated Design and Deployment of Low-Carbon Concrete for Data Centers
- Automatic debiased machine learning and sensitivity analysis for sample selection models
- Learning Representations for Counterfactual Inference
- Estimating individual treatment effect: generalization bounds and algorithms
- Representation Learning for Treatment Effect Estimation from Observational Data.
- Estimating Continuous Treatment Effects with Two-Stage Kernel Ridge Regression
- Generalization Bounds for Estimating Causal Effects of Continuous Treatments
- Bi-Level Decision-Focused Causal Learning for Large-Scale Marketing Optimization: Bridging Observational and Experimental Data
- Coase vs. Pigou in the Petroleum Market
- Would information acquisition increase enterprises’ willingness to participate in the carbon market? An empirical investigation of equipment manufacturing enterprises in Henan Province, China
- Feasible supply of steel and cement within a carbon budget is likely to fall short of expected global demand
- Learning to Mitigate Externalities: the Coase Theorem with Hindsight Rationality
- Achieving carbon neutrality: How does the construction of national high-tech zones affect the green innovation of enterprises? Based on quasi-natural experiments in pilot areas in China