Density functional theory, or DFT, has been the workhorse of computational materials science for decades. Resting on the foundational theorems of Hohenberg and Kohn, which show that all ground-state properties of a many-electron system are uniquely determined by the electron density, DFT reduces the intractable many-body Schrödinger equation to a manageable single-electron problem through the Kohn–Sham formalism. In practice, the approach yields exactly the quantities that matter for energy applications: electronic band structures, densities of states, formation energies, adsorption energies, and the Gibbs free energies of hydrogen adsorption that underpin catalyst design. The catch is cost. A single DFT calculation for a complex oxide or perovskite can demand days of high-performance computing time, and exhaustively screening even ten thousand candidates can consume tens of thousands of CPU-hours. For the vast combinatorial space of possible energy materials, standalone DFT simply cannot keep pace.

Machine learning offers a way out. By training predictive models on databases of DFT-computed properties, researchers can estimate bandgaps, formation energies, and adsorption energies for new candidate structures orders of magnitude faster than direct simulation. Classical algorithms still hold their own in certain regimes: random forests and gradient boosting methods such as XGBoost and LightGBM perform robustly across benchmarks, while kernel-based approaches like Gaussian process regression are notably data-efficient and provide native uncertainty estimates, a valuable property when deciding which expensive calculations to run next. But the current accuracy frontier belongs to deep learning architectures that read crystal structures directly. Graph neural networks such as the Crystal Graph Convolutional Neural Network, MEGNet, SchNet, DimeNet++, CHGNet, and the equivariant MACE model represent atoms and bonds as nodes and edges in a graph, learning end to end from raw structure without hand-crafted descriptors and, in their equivariant variants, respecting the rotational and translational symmetries of physics itself.

The review is particularly clear-eyed about the practical trade-offs among these models. The three most accurate architectures, DimeNet++, CHGNet, and MACE, also carry the highest computational cost per prediction, and none of the leading graph models provides native, well-calibrated uncertainty quantification. That matters because uncertainty estimates are precisely what active learning pipelines use to decide which candidates deserve expensive DFT follow-up, so teams using graph networks typically bolt on ensembling or dropout-based proxies that are themselves imperfectly calibrated. Simpler crystal graph models scale most readily to campaigns screening hundreds of thousands to millions of candidates, while the classical low-data methods remain, somewhat counter-intuitively, the most portable across unfamiliar materials classes. Model choice, the author stresses, depends on the target property, training set size, and downstream needs, not on accuracy figures alone.

At the heart of the modern workflow sits the active learning loop, an iterative dance between statistics and quantum mechanics. A trained machine learning model screens a large candidate library, flags the predictions where its uncertainty is highest, and prioritises those structures for targeted DFT calculation. The new results are appended to the training set and the model is retrained, and the cycle repeats. Reported efficiency gains vary by domain, but case studies reviewed in the article describe reductions in the required number of experiments or calculations by roughly an order of magnitude. In one landmark campaign, Tran and Ulissi trained an active-learning surrogate across intermetallic surfaces spanning 31 elements, reaching adsorption energy mean absolute errors of just 0.17 to 0.18 electronvolts and guiding the identification of 258 candidate surfaces across 102 alloys for hydrogen evolution, along with 131 surfaces across 54 alloys for carbon dioxide reduction.

None of this would be possible without the open-data infrastructure that has grown up over the past decade. The Materials Project contains more than 154,000 inorganic compounds computed within a standardised PBE framework, the Open Quantum Materials Database offers thermodynamic data for over a million compositions, and AFLOW provides electronic and thermomechanical data for more than 3.5 million compounds, uniquely including phonon dispersions for lattice stability assessment. Workflow automation frameworks such as pymatgen, FireWorks, AiiDA, and Atomate orchestrate thousands of calculations across computing clusters while tracking provenance and recovering from errors. The review identifies this shared data commons as arguably the single most important enabler of data-driven discovery, transforming DFT from a bespoke per-study activity into reusable community infrastructure, while also cautioning that shared infrastructure concentrates shared data-quality problems.

The applications span the full clean-energy portfolio. For the hydrogen evolution reaction, machine learning models trained on DFT-computed hydrogen adsorption free energies exploit the volcano-plot framework rooted in the Sabatier principle, which locates the optimal catalyst where hydrogen binds neither too strongly nor too weakly. For photocatalytic water splitting, which requires a bandgap of roughly 1.8 to 3.1 electronvolts and band edges straddling the water redox potentials, machine learning-guided screening of 5,158 unexplored hybrid organic-inorganic perovskites identified six lead-free candidates with suitable bandgaps and room-temperature stability. In carbon dioxide reduction, active machine learning guided the experimental discovery of copper-aluminium electrocatalysts that convert CO2 to ethylene with the highest reported Faradaic efficiency for that reaction, validated by in situ X-ray absorption measurements. In batteries, a classifier trained on just 40 labelled materials screened 12,831 lithium-containing structures from the Materials Project and flagged 21 candidates for superionic conductivity, while machine learning interatomic potentials now simulate ionic transport at near-DFT accuracy.

The review also confronts the field’s growing pains with unusual candour. Machine learning models inherit every weakness of their training data: major databases overrepresent stable, well-known compounds while underrepresenting metastable phases, disordered alloys, and amorphous structures; inconsistent calculation parameters across aggregated sources introduce systematic offsets; and poorly converged structures inject label noise. The chronic underestimation of bandgaps by the standard PBE functional undermines photocatalyst predictions, with the more accurate hybrid and GW methods too expensive to generate large datasets, though transfer learning is emerging as a partial remedy. Graph networks remain black boxes, and uncertainty quantification is not a solved problem, with no single method uniformly well calibrated, especially in the out-of-distribution regime where genuine discovery happens. Most strikingly, the review highlights reproducibility concerns that have been raised about two of the field’s most famous demonstrations, the GNoME deep learning model that scaled discovery to 2.2 million predicted stable crystals and the A-Lab autonomous synthesis campaign, arguing that the community should release full structural and provenance data alongside every large-scale discovery claim as a norm rather than an exception.

Looking forward, the review maps out four directions that follow directly from these limitations. Autonomous materials-discovery laboratories will couple the active learning loop to robotic synthesis and characterisation platforms, extending the vision demonstrated by the A-Lab. Explainable AI, including physics-constrained neural networks and symbolic regression that extracts human-readable equations from trained models, will address the interpretability gap. Quantum computing, through variational quantum eigensolvers and quantum phase estimation, may eventually supply near-exact training labels for strongly correlated systems, although practical quantum advantage remains a multiyear to decade prospect. Multi-fidelity modelling will blend cheap, low-accuracy DFT data with sparse high-accuracy calculations to reconcile the accuracy-cost trade-off. Together with a fifth priority of independently re-executable benchmarking, these directions define the agenda for the coming decade.

The verdict of the review is ultimately optimistic but disciplined. Density functional theory provides the rigorous quantum mechanical foundation that gives machine learning models physical grounding, while machine learning supplies the statistical power to generalise those quantum insights across millions of unexplored candidates at negligible marginal cost. The two are most powerful not in isolation but as a closed, active learning loop, increasingly connected to robotic laboratories that can synthesise what the algorithms propose. Landmark studies, from the intermetallic electrocatalyst campaigns to the experimentally validated copper-aluminium CO2 catalysts, show that the framework already delivers real materials. Whether the clean energy revolution arrives on schedule, the author suggests, will depend less on the raw scale of these computational engines than on the discipline and verifiability with which they are extended to new chemistries.

Subject of Research: Integration of density functional theory and machine learning for accelerated discovery of energy materials

Article Title: Integration of density functional theory and machine learning for materials discovery in energy applications

Article References: Integration of density functional theory and machine learning for materials discovery in energy applications. (n.d.). https://doi.org/10.1007/s44371-026-00985-y

Image Credits: AI Generated

DOI: 10.1007/s44371-026-00985-y

Keywords: density functional theory, machine learning, materials discovery, energy materials, hydrogen evolution reaction, photocatalysis, CO2 reduction, battery materials, graph neural networks, active learning, high-throughput screening, Materials Project