The motivation stems from a practical problem in climate modeling. Modern emissions models draw on national statistics, energy consumption, socio-economic indicators and environmental variables, producing high-dimensional feature spaces in which noise, nonlinear dependencies and multicollinearity can destabilize predictions. Even small percentage changes in emission trends can alter interpretations of global carbon budgets, so uncertainty in model estimates propagates directly into long-term policy projections. Dimensionality reduction, the process of trimming or transforming a dataset’s features down to its most informative core, is therefore central to building reliable models. The field divides such methods into feature selection, which keeps a subset of the original variables, and feature extraction, which transforms them into new composite dimensions. Both approaches carry trade-offs: selection preserves interpretability, while extraction can compress variance but obscure what the new components actually mean.

Classical tools each have well-documented weaknesses in this setting. Recursive Feature Elimination (RFE) trains models repeatedly and prunes the least important variables, but becomes slow and unstable as dimensionality grows. LASSO and Elastic Net apply linear penalties that shrink irrelevant coefficients to zero, yet they can miss the complex nonlinear relationships common in socio-economic and energy data. PCA compresses variance effectively but does not necessarily preserve what is predictive. Metaheuristics such as genetic algorithms and particle swarm optimization explore the search space randomly but risk premature convergence, while Bayesian optimization offers improvements at a computational cost the authors describe as prohibitively high for large-scale feature selection. Because the number of possible feature subsets grows exponentially with dimension, exhaustive evaluation is infeasible, and classical methods typically explore only restricted regions of this combinatorial landscape.

Quantum annealing offers a different route. The technique maps optimization problems onto a Quadratic Unconstrained Binary Optimization (QUBO) formulation, where each feature is a binary variable, set to 1 if selected and 0 if excluded, and the objective is encoded as an energy landscape to be minimized. Feature selection is inherently NP-hard, making it a natural fit for this formulation. The annealer starts in a simple quantum state and slowly evolves toward the problem’s ground state, with superposition and entanglement facilitating broad exploration that helps avoid local minima. Prior studies cited in the paper suggest the approach has promise: hybrid quantum annealing solvers were nearly ten times faster than classical simulated annealing on a large benchmark dataset, and on MNIST, selecting only 25 to 30 of 784 pixel features preserved image quality and even improved classification accuracy over using all features.

The novelty of QMICO lies in how it scores features inside the QUBO matrix. Relevance to the target is measured with mutual information (MI), which quantifies how much knowing a feature reduces uncertainty about the outcome and captures nonlinear dependencies without parametric assumptions. Redundancy between feature pairs is scored as a multiplicative combination of MI and absolute Pearson correlation (CO), so a pair is penalized only when it is redundant in both the linear and nonlinear senses. The authors argue this multiplicative design is stricter than additive or max-based alternatives, which can leave strongly overlapping features insufficiently penalized. Diagonal entries of the QUBO matrix reward relevance, off-diagonal entries punish redundancy, and a cardinality penalty with coefficient lambda keeps the selected subset close to a target size k. A trade-off parameter alpha balances the two forces.

The experimental setup was unusually thorough. The team used the Our World in Data carbon dataset with 13,706 records and 79 features, filtered to 60 predictors after applying a 40 percent missing-value threshold, with mean imputation and z-score scaling applied within each training fold to prevent information leakage. Exploratory analysis showed the dataset’s dependency structure is dominated by nonlinearity: 52 of 60 features, or 86.7 percent, exceeded the study’s threshold for predominantly nonlinear association with CO2 emissions. The framework was evaluated with three regression models representing different learning paradigms, Random Forest Regression, Artificial Neural Networks and Gradient Boosting Regression, across five target dimensionalities from 50 down to 10 features, three discretization settings, and two annealing solvers: Fixstars Amplify AE, a GPU-based digital annealer, and D-Wave Leap Hybrid BQM, which combines classical optimization with access to D-Wave’s superconducting quantum annealer. In total, 270 distinct experiments were run.

The results favored the quantum-assisted pipeline across most settings. QMICO-based variants topped the R² rankings at nearly every dimensionality, and QMICO1-PCA, the Fixstars version combined with PCA, achieved the best overall average performance. The largest single improvement over the no-reduction baseline was 8.25 percent at 50 features. Notably, the QMICO variants remained consistently positive even at the most aggressive compression level of 10 features, where LASSO actually dipped slightly below baseline. The ten features selected in the highest-relevance quadrant included energy-related and cumulative emission indicators such as oil-related CO2, share of global CO2, and total greenhouse gas emissions excluding land-use change, with mutual information values approaching 3.0. On computational cost, classical feature reduction was faster, with RFE taking 2.6 seconds, but QMICO-based methods produced cleaner subsets that trained faster downstream: QMICO1 recorded the lowest total combined time at 36.08 seconds.

The head-to-head solver comparison revealed a striking behavioral difference. Under identical QUBO formulations and time limits, the Fixstars digital annealer converged smoothly and produced nearly identical feature subsets across moderate hyperparameter changes, with Jaccard similarity reaching 1.000 at some settings. D-Wave, by contrast, explored the energy landscape more stochastically: at 20 features, similarity between two hyperparameter settings dropped to 0.143, meaning roughly half or more of the selected features differed. This stochasticity broadened exploration but introduced instability, reflected in larger swings in downstream R² as the relevance-redundancy balance shifted. Despite these differences, post-hoc tests found no statistically significant gap between the two solvers’ overall performance, suggesting the framework is robust across heterogeneous annealing platforms even though Fixstars showed slightly stronger average results.

Validation on two small-scale, country-level datasets reinforced the pattern with an instructive twist. On USA emissions data from the EDGAR database, with 24 features, QMICO variants achieved an average R² of 0.9356, a 1.76 percent improvement over baseline, selecting 12 of 24 features; adding PCA there yielded a marginal 0.12 percent decrease, likely because excluded features carried mostly noise. On China’s CEADs dataset, with 48 features, the QMICO-PCA variants reached an exceptional R² of 0.9896, a 4.29 percent gain, with the selector identifying just 5 features from 48, an aggressive roughly 10-to-1 reduction. Here PCA added a modest 0.10 percent improvement, consistent with the authors’ observation that PCA’s benefit grows with the dimensionality and information density of the unselected feature pool.

The authors also translate the gains into policy terms. Global CO2 emissions currently stand at roughly 36 to 37 gigatonnes per year, so a 4 to 7 percent reduction in predictive error corresponds to an annual uncertainty reduction of about 1.4 to 2.6 gigatonnes, accumulating to roughly 14 to 26 gigatonnes over a ten-year projection horizon, a magnitude comparable to the total annual emissions of major industrialized economies such as Japan or Germany. The study’s limitations are acknowledged: evaluation was confined to carbon-emissions data, downstream models used default hyperparameters, and current quantum platforms impose problem-size and cost constraints. Still, the findings suggest that hybrid quantum-classical workflows, pairing annealer-driven subset search with classical variance compression, can offer a competitive and statistically robust alternative for high-dimensional environmental modeling, with future work pointing toward quantum-assisted hyperparameter tuning and adaptive relevance-redundancy weighting.

Subject of Research: Hybrid quantum annealing-based feature selection combined with PCA for improving carbon-emissions prediction models

Article Title: Improving dimensionality-reduction for carbon-emissions prediction using hybrid quantum annealing-based feature selection and PCA

Article References: Munawar, G., Maulidevi, N. U., & Surendro, K. (2026). Improving dimensionality-reduction for carbon-emissions prediction using hybrid quantum annealing-based feature selection and PCA. Discover Artificial Intelligence, 6(1), Article 1012. https://doi.org/10.1007/s44163-026-02089-z

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02089-z

Keywords: quantum annealing, feature selection, principal component analysis, carbon emissions prediction, QUBO, mutual information, dimensionality reduction, machine learning, climate policy, D-Wave, Fixstars Amplify, regression models