A new class of artificial intelligence models is cutting the time needed to identify promising catalytic materials from weeks to hours, and unlike the opaque neural networks that dominate materials science, these systems can explain which atomic features matter most. The approach, broadly called gray-box AI, sits between fully transparent “glass-box” models such as symbolic regression and inscrutable “black-box” deep learning. By pairing high-accuracy predictions with post-hoc explanation tools, researchers are gaining both speed and scientific insight, a combination that could reshape how chemists design materials for clean energy, fuel cells, and industrial chemistry.
What Gray-Box AI Actually Means for Chemistry
The term “gray-box” has circulated in machine learning for years, but its specific meaning in heterogeneous catalysis was sharpened in a recent perspective on AI for chemistry that distinguishes three tiers of model transparency. Black-box systems, including most deep neural networks, deliver predictions without revealing why a given surface or adsorbate combination performs well. Glass-box models, such as symbolic regression or decision trees, are fully readable but often sacrifice accuracy on complex datasets. Gray-box methods split the difference: they train a high-capacity model first, then apply interpretability techniques like SHAP values or feature-importance rankings to attribute predictions to specific physical variables.
Esterhuizen and co-workers formalized this distinction by proposing that SHAP-based post-hoc attribution from complex models counts as grey-box machine learning, separate from glass-box alternatives like symbolic regression. The practical payoff is that scientists can check whether a model’s reasoning aligns with known chemistry before trusting its screening recommendations. If a model flags a surface alloy as promising but attributes the prediction to an irrelevant descriptor, researchers know to investigate further rather than synthesize a dead-end material.
Massive Datasets That Make Speed Possible
Gray-box explanations would mean little without accurate underlying predictions, and accuracy at scale depends on training data. The Open Catalyst 2020 dataset, known as OC20, remains the largest public benchmark for heterogeneous catalysis. It contains approximately 1.28 million density functional theory relaxations and roughly 264.9 million single-point evaluations, all drawn from DFT calculations that simulate how molecules interact with catalyst surfaces, as detailed in the original OC20 benchmark description. OC20 defines three benchmark tasks, labeled S2EF, IS2RS, and IS2RE, each targeting a different stage of the computational pipeline that traditionally requires expensive quantum-mechanical simulations.
The Open Catalyst Project did not stop there. A follow-up dataset, OC22, extended the benchmark to oxide electrocatalysts, adding a second major collection of structures and evaluation settings. These oxide-focused benchmarks are central to water splitting and carbon dioxide reduction, two reactions at the heart of green hydrogen and carbon-capture technology. By covering both metallic alloys and oxides, the combined datasets let models generalize across a wider chemical space than any single research group could explore with DFT alone.
Graph Neural Networks as the Prediction Engine
Most top-performing models on these benchmarks use graph neural networks, or GNNs, which represent atoms as nodes and bonds or spatial proximity as edges. GemNet-OC, a GNN architecture developed for the Open Catalyst Challenge, demonstrated significant performance improvements on OC20 tasks while quantifying the tradeoff between model accuracy and computational cost. The GemNet-OC study showed that carefully designed message-passing schemes and geometric features can push prediction errors down while keeping inference times practical. That tradeoff matters for real screening campaigns: a model that is 2 percent more accurate but ten times slower may not be worth deploying when a lab needs to rank thousands of candidate surfaces in a single afternoon.
Yet GNNs are, by default, black boxes. A graph convolution can capture subtle many-body interactions between atoms, but the learned weights do not map neatly onto textbook descriptors like d-band center or electronegativity. This is precisely where gray-box tools enter the pipeline. GNNExplainer, a widely cited interpretability method for graph neural networks, identifies which subgraphs, edges, and node features contribute most to a given output. Applied to a catalyst-screening GNN, it can highlight that a prediction depends heavily on the coordination environment of a particular metal atom, giving chemists a testable physical hypothesis rather than a bare number.
Explanation Tools Tailored to Scarce Lab Data
Large computational datasets like OC20 and OC22 are one side of the problem. The other side is experimental data, which is almost always smaller, noisier, and unevenly distributed across chemical families. A recent study in The Journal of Physical Chemistry C proposed a machine learning and explainable AI workflow designed specifically for scarce and imbalanced experimental catalyst datasets. In that work, the authors describe a combined modeling and attribution framework that applies both feature-importance rankings and instance-level explanations to verify that the model is learning genuine chemical trends rather than artifacts of data imbalance.
This matters because a model trained on 50 examples of platinum-group catalysts and five examples of earth-abundant alternatives will naturally skew toward the overrepresented class. Without explanation tools, a researcher might conclude that platinum alloys are universally superior when the model has simply never seen enough non-platinum data to learn otherwise. Gray-box diagnostics expose that bias before it wastes lab resources, for example by revealing that the model’s confidence is tightly correlated with data density rather than with physically meaningful descriptors.
Why Black-Box Speed Alone Falls Short
Much of the popular coverage around AI in catalysis focuses on raw acceleration: replacing a week of DFT calculations with a millisecond neural-network inference. That framing, while accurate, misses a deeper issue. Speed without understanding can send experimental teams chasing predictions that exploit dataset quirks rather than real physics. A screening model might latch onto a spurious correlation between surface area and activity in a narrow training set, then confidently recommend a family of materials that fail in the lab because the underlying mechanism was never captured.
Gray-box AI is a response to that risk. By forcing models to “show their work” through SHAP values, GNN explanations, or counterfactual examples, chemists can filter recommendations using domain knowledge. If the highest-ranked candidates are all justified by descriptors known to correlate with stability, selectivity, or adsorption strength, teams can move forward with more confidence. If not, they can refine the dataset, adjust the feature set, or switch to a different model architecture before committing to expensive synthesis and testing.
From Insight to Closed-Loop Discovery
The long-term vision goes beyond one-off screening campaigns. As automated laboratories and robotic platforms become more common, gray-box catalysts models could sit at the core of closed-loop discovery systems. In such a loop, a model proposes new compositions or surface structures, a robotic setup synthesizes and tests them, and the resulting data are fed back into the model for retraining. Interpretability is crucial here: it helps decide which experiments are most informative, not just which materials are most promising.
For example, attribution scores can highlight regions of chemical space where the model is both uncertain and highly sensitive to particular descriptors. Targeting those regions for new experiments accelerates learning, much like active learning strategies in other fields. Over time, the loop does more than optimize performance; it refines mechanistic understanding, revealing which structural motifs consistently drive activity across different reaction conditions and material classes.
Challenges and the Road Ahead
Despite the progress, gray-box AI for catalysis faces several challenges. Many interpretability methods were developed for images or tabular data and do not directly account for the symmetries and constraints of atomic systems. Explanations can also be unstable: small changes in input or model initialization sometimes yield different attribution maps, complicating their use as a basis for theory building. Moreover, there is a risk of over-trusting clean-looking explanations that are mathematically correct but chemically misleading.
Addressing these issues will likely require closer collaboration between method developers and domain experts. Chemists can help benchmark explanation quality against known mechanistic trends, while machine-learning researchers can adapt attribution techniques to respect physical invariances and conservation laws. As datasets like OC20 and OC22 continue to grow, and as experimental workflows catch up in scale, the combination of fast GNN predictors and robust gray-box diagnostics could turn AI from a black-box oracle into a genuine partner in catalyst design.
If that happens, the impact will extend well beyond a few percentage points of efficiency. By revealing why certain atomic arrangements work while others fail, gray-box AI can help transform isolated screening results into general design rules, rules that chemists can apply to entirely new reactions and material classes. In the race to develop catalysts for cleaner fuels, carbon-neutral chemicals, and more efficient industrial processes, that blend of speed and understanding may prove to be the decisive advantage.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
