A partially quantum‑informed crystal graph network for direct adsorption energy prediction in catalysis Authors: Ericsson Tetteh Chenebuah, Michel Nganbe, Linhao Liu, Alain Beaudelaire Tchagang Published: 2026 ABSTRACT The rational design of heterogeneous catalysts requires predictive tools that balance quantum-level accuracy with computational efficiency for estimating adsorbate energetics. Although recent machine learning approaches have shown strong performance, they mostly depend on total energy decomposition and/or force supervision, where forces are obtained as the gradient of the potential energy with respect to atomic positions. As a result, they tend to be computationally complex, often requiring the prediction of several components, including the surface, the adsorbate, and the combined system. To make a contribution in addressing this challenge, we introduce Q-CatNet, a partially quantum-informed crystal graph catalyst network designed for direct prediction of initial structure-to-adsorption energy correlation for both single-atom and multi-atom (molecular) adsorbates on bulk catalyst surfaces. The model extends the Crystal Graph Convolutional Neural Network (CGCNN) framework by incorporating electrostatic Hamiltonian-inspired edge descriptors together with global features derived from density of states (DOS) and X ray diffraction (XRD) signals, enabling the capture of both local bonding interactions and system-wide physical correlations. Benchmarking on a curated dataset confirms its robustness: Q-CatNet outperforms the image-based Fourier-Transformed Crystal Property (FTCP) representation by 46% and exceeds several invariant graph-based architectures by margins ranging from 8% to 38%. These results highlight Q-CatNet as a physically grounded framework that bypasses the need for total energy decomposition, enabling a faster and practical adsorption‑energy prediction towards accelerated catalyst discovery in materials science and engineering.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1038/s41598-026-56391-y. Introduction Catalysis lies at the heart of many critical processes for a sustainable energy future, including water splitting, fuel cells, and ammonia synthesis. Central to these technologies are complex chemical reactions such as the hydrogen evolution reaction (HER), oxygen reduction reaction (ORR), oxygen evolution reaction (OER), and nitrogen reduction reaction (NRR). Historically, designing effective catalysts has been a difficult task, often complicated by the intricate interplay of several factors that influence catalytic activity, selectivity, and stability. Traditional methods, including high-throughput density functional theory (DFT)1 calculations and Edisonian-based empirical screening2, while effective, are often constrained by computational expense and the immense chemical and structural space. To overcome such a challenge and accelerate the discovery and design of novel catalysts, the field of materials science is increasingly turning to advanced approaches at the expense of traditional techniques. In recent years, machine learning (ML) techniques have become transformative data-driven tools in catalysts design and informatics, enabling unprecedented efficiency in predictive modeling and optimization. Among ML techniques, Graph Neural Networks (GNNs) have proven to be particularly effective, especially for modeling and understanding the complex structures of materials3,4. In principle, GNNs are a class of deep learning models that operate on graph-structured data (containing nodes and edges) by automatically extracting hierarchical representations within a data structure, capturing both local and global environments. They have garnered significant attention for their ability to learn directly from the crystal structures of materials due to their adept functionality at identifying the connections between atoms by representing them as nodes and bonds as edges in a graph. This ability enables more precise property predictions and material design compared to other ML methods that use simplified models. Numerous studies have demonstrated the excellent ability of GNNs to predict a wide range of material properties – such as formation energy, band gap, and elastic moduli – with accuracies comparable to DFT calculations, but at a fraction of the computational cost5–8. Classic GNN frameworks like Crystal Graph Convolutional Neural Network (CGCNN)5, MatErials Graph Network (MEGNet)6, and Atomistic Line Graph Neural Network (ALIGNN)7, have been validated on both inorganic and organic materials databases such as the Materials Project (MP)9, JARVIS-DFT10, and Open Quantum Materials Database (OQMD)11. For instance, the CGCNN is widely recognized as the pioneering model that showcased the use of convolutional neural networks to extract significant features from a crystal graph for property prediction. The CGCNN encodes edge features using the discretized distances between contributing atoms in a unit cell crystal, achieving remarkable Mean Absolute Error (MAE) results at 0.039 eV/atom and 0.388 eV on the formation energy and bandgap targets, respectively5. Similarly, the universal MEGNet expanded the capabilities of GNNs by integrating global state attributes along with conventional atom and edge features, and was seen to improve CGCNN capabilities by about 28.2% and 14.9% on the same target properties, respectively6. Likewise, ALIGNN has been shown to improve traditional GNNs for property prediction by integrating line graphs that capture angular relationships between atoms, thereby enabling explicit account for three-body interactions, and effectively modeling bonding geometries. The ALIGNN model achieved a MAE of 0.218 eV in bandgap prediction7. The aforementioned models, among many others in the field, highlight the potential of GNNs as key tools for forward predictive modeling and autonomous discovery. In the context of catalysis, GNNs have been successfully deployed for rapid prediction of catalytic activity and selectivity, enabling efficient screening of novel materials. Applications span diverse catalytic systems or descriptors, including homogeneous, heterogeneous, electro-, and photocatalysis, further highlighting their versatility and broad applicability in advancing catalytic science12–15. Moreover, among the myriad of descriptors that serve as predictive proxies in modeling catalytic performance with GNNs, the adsorption energy (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{E}_{ads}$$\end{document}) stands out as arguably the most critical and universally applicable parameter for explaining and predicting catalytic activity and microkinetics16. By definition, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{E}_{ads}$$\end{document} is the strength of interaction between a reactant or intermediate species and the catalyst surface, and is a measure of the prominent energetic landscape for surface reactions to occur. For heterogeneous systems involving a single adsorbate on an idealized bulk surface, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{E}_{ads}$$\end{document} can be mathematically expressed as:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{E}_{ads}=\:{E}_{system}-{E}_{bulk}-{E}_{gas}$$\end{document} Here, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{system}\:$$\end{document}represents the overall DFT-computed energy of the combined adsorbate and bulk material system, modeled as a slab. The terms \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{bulk}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{gas}$$\end{document} correspond to the DFT-computed energies of the clean bulk surface and the isolated gas-phase adsorbate molecule, respectively. According to the Sabatier principle, optimal heterogenous catalytic activity often lies at the peak of a volcano plot relationship where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{E}_{ads}\approx\:0$$\end{document}. Therefore, too weak (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{E}_{ads}>0$$\end{document}) or too strong (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{E}_{ads}<0$$\end{document}) adsorptions would lead to suboptimal catalytic performance17. As such, accurately predicting \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{E}_{ads}\:$$\end{document}through machine learning techniques, particularly in the context of the Sabatier principle, has become a significant focus of research and has been pursued by numerous investigators. For example, it was demonstrated that straightforward tabular-dataset models like the Adaboost regression model can achieve moderate reliability in predicting \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{E}_{ads}$$\end{document} with a MAE score at 0.398 eV18 when trained with about 24,000 heterogenous crystal systems from the Open Catalyst 2020 (OC20)19 dataset. Considering GNNs in particular, state of the art models, whether trained to directly predict adsorption energies or to estimate them indirectly through total energy decomposition, have been widely employed for baseline \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{E}_{ads}$$\end{document} prediction in prior studies, thereby establishing a strong foundation for subsequent advancements. For example, the original OC20 study utilized three established GNN architectures, namely CGCNN5, SchNet8, and DimeNet++20, to map the initial structure to the relaxed energy. These models achieved best in domain mean absolute error values of 0.615 eV, 0.987 eV, and 0.562 eV, respectively, on the test set19. More recently, modern equivariant GNNs, including Geometric Message Passing Neural Network (GemNet)21,22, EquiformerV223,24, Higher-Order Equivariant Message Passing Neural Network (i.e., MACE)25,26, and Universal Models for Atoms (UMA)27, have also been applied to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{E}_{ads}\:$$\end{document}prediction tasks. Among these, UMA has demonstrated the strongest performance, achieving a MAE of 31.8 meV on in-domain relaxed structure to energy and force tasks (S2EF). However, the UMA approach relies on indirect total energy decomposition, which, although more accurate, is computationally more demanding than direct prediction methods. It further requires energy evaluation on fully relaxed structures and depends on force supervision to shape the underlying energy landscape, increasing both training and inference cost compared to methods that bypass these additional components. This raises an important question of whether such computationally intensive approaches can be bypassed by employing direct models that learn from the initial structure and map adsorption energies without requiring force supervision. However, even for direct approaches such as those demonstrated in Ref19., a substantial performance gap still remains between current GNN predictions and the level of experimental chemical accuracy typically desired for thermochemical properties, which is approximately 0.0434 eV per atom, or 1 kcal/mol28. This discrepancy could likely be a function of the intrinsic complexities of adsorbate–bulk surface interactions, including quantum mechanical effects such as charge transfer, binding site specificity, and orbital hybridization, which remain challenging to capture with current ML descriptors and architectures. These limitations highlight the ongoing need for innovative approaches in descriptor design and model development, aimed at bridging this accuracy gap and advancing the predictive capabilities of ML in direct \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{E}_{ads}$$\end{document} prediction. Perhaps, one promising direction could involve the explicit incorporation of electrostatic periodic effects that capture the many-body interactions characteristic of crystal environments. This includes modeling Coulombic forces, long-range electrostatics, and Fourier-transformed periodic features. To make a contribution in this regard, the present study proposes a partially Quantum-informed Catalysis Network (Q-CatNet) model for improving direct prediction capabilities of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{E}_{ads}$$\end{document} as a primary descriptor of catalytic activity, alongside other target properties relevant to simulating adsorbate–bulk surface interactions. Unlike other models that computationally tackle \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{E}_{ads}\:$$\end{document}prediction from a total energy decomposition approach and induce force supervision into the machine learning loss function, Q-CatNet directly maps the unrelaxed or initial adsorbate-bulk material structure to the resolved \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{E}_{ads}$$\end{document} value without the need for force supervision. The model is developed and evaluated within a controlled setting focused on gas–intermetallic surface systems derived from a curated subset of the OC20 dataset, allowing systematic analysis of feature design and model behavior. Model generalizability investigations are deferred to future studies. The main contributions of this work are summarized as follows: Development of the Q-CatNet architecture: A physics-informed graph neural network that integrates Coulombic interactions and periodic crystalline descriptors to directly model adsorbate–bulk surface interactions across a wide range of unary, binary, and ternary bulk crystal surfaces.Validation on single-atom adsorbates: The Q-CatNet is systematically benchmarked on four representative single-atom adsorbate systems (C*, H*, N*, and O*) extracted from the OC20 dataset. These adsorbates were chosen for their central role in catalytic and energy-conversion processes. The resulting dataset enables robust ablation studies and baseline comparisons, ensuring consistent calibration of model efficiency.Extension to complex molecular systems: Beyond single-atom species, the model’s predictive capacity is further assessed using eight multi-atomic (molecular) adsorbates, enabling comprehensive evaluation of Q-CatNet’s capability to capture the intricate dynamics of molecular adsorption on metallic surfaces. Development of the Q-CatNet architecture: A physics-informed graph neural network that integrates Coulombic interactions and periodic crystalline descriptors to directly model adsorbate–bulk surface interactions across a wide range of unary, binary, and ternary bulk crystal surfaces. Development of the Q-CatNet architecture: A physics-informed graph neural network that integrates Coulombic interactions and periodic crystalline descriptors to directly model adsorbate–bulk surface interactions across a wide range of unary, binary, and ternary bulk crystal surfaces. Validation on single-atom adsorbates: The Q-CatNet is systematically benchmarked on four representative single-atom adsorbate systems (C*, H*, N*, and O*) extracted from the OC20 dataset. These adsorbates were chosen for their central role in catalytic and energy-conversion processes. The resulting dataset enables robust ablation studies and baseline comparisons, ensuring consistent calibration of model efficiency. Validation on single-atom adsorbates: The Q-CatNet is systematically benchmarked on four representative single-atom adsorbate systems (C*, H*, N*, and O*) extracted from the OC20 dataset. These adsorbates were chosen for their central role in catalytic and energy-conversion processes. The resulting dataset enables robust ablation studies and baseline comparisons, ensuring consistent calibration of model efficiency. Extension to complex molecular systems: Beyond single-atom species, the model’s predictive capacity is further assessed using eight multi-atomic (molecular) adsorbates, enabling comprehensive evaluation of Q-CatNet’s capability to capture the intricate dynamics of molecular adsorption on metallic surfaces. Extension to complex molecular systems: Beyond single-atom species, the model’s predictive capacity is further assessed using eight multi-atomic (molecular) adsorbates, enabling comprehensive evaluation of Q-CatNet’s capability to capture the intricate dynamics of molecular adsorption on metallic surfaces. Methods Statistical distribution and preprocessing of the training dataset The Q-CatNet model was trained using an integrated dataset derived from two complementary sources: the Open Catalyst 2020 (OC20)19 dataset and the Materials Project (MP)9 database. The OC20 dataset, generated from large-scale first-principles calculations, provides detailed adsorption energies and surface configurations for a diverse set of stable crystalline materials, many of which are cross-referenced to entries in the MP repository. These parent structures were drawn from a pool of 11,451 thermodynamically stable compounds, each characterized by negative formation energies and energy-above-hull values below 0.1 eV/atom. Collectively, the dataset spans 55 distinct chemical elements, encompassing both elemental and intermetallic systems of catalytic interest. To ensure coherence between the two databases, a systematic data mapping strategy was employed, linking OC20 and MP entries via their unique MP identifiers (MP-IDs). For each matched structure, three categories of information were extracted: (i) adsorption energies of adsorbate–surface systems, (ii) site-specific structural and coordination descriptors, and (iii) unit cell representations of the bulk parent crystals (limited to a maximum of 80 atoms per unit cell to preserve computational efficiency). Rigorous data preprocessing and cleaning were performed to maintain data quality. Duplicate entries sharing identical surface features but inconsistent adsorption energies were removed. Additionally, MP-derived samples lacking band structure data were excluded to ensure the completeness of Density of States (DOS) features. The curated dataset was divided into two principal subsets. The first subset comprises 37,433 unique single-atom adsorbate radicals across four species: carbon (C*), hydrogen (H*), nitrogen (N*), and oxygen (O*), with nearly uniform distributions (24.63%, 26.04%, 24.65%, and 24.68%, respectively). This subset was primarily utilized for model calibration, ablation studies, and benchmarking against baseline deep learning architectures. It also provides a strong foundation for inverse design workflows aimed at accelerating catalyst discovery. The second data subset contains 50,456 unique molecular adsorbate configurations spanning eight multi-atomic species: hydroxyl radical (OH, 15.73%), methyl radical (CH₃, 14.6%), methane (CH₄, 12.48%), methanol (CH₃OH, 11.53%), ethanol (C₂H₅OH, 10.71%), nitric oxide (NO, 14.05%), ammonia (NH₃, 12.83%), and methyl nitroxyl (CH₃NNOH, 7.9%). This subset was used to evaluate Q-CatNet’s extrapolative capability in modeling the complex physics of molecular adsorption on bulk crystal surfaces. To visualize the data structure and diversity, Fig. 1 summarizes the statistical and chemical characteristics of both subsets. Figure 1(A) shows a t-distributed Stochastic Neighbor Embedding (t-SNE) projection of the learned representations for the single-atom adsorbate dataset. Distinct clustering patterns corresponding to different adsorbate species are evident, indicating chemically meaningful separation. The transparent circular overlays denote adsorption energy magnitudes, where larger radii indicate more optimal (thermoneutral) binding energies (i.e., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{E}_{ads}\approx\:0$$\end{document} eV). Complementarily, Fig. 1(B) presents box-and-violin plots illustrating the statistical distributions of adsorption energies across the four single-atom species. Hydrogen–surface systems exhibit the narrowest and most symmetric distribution centered near thermoneutral binding, whereas nitrogen and oxygen systems show broader, skewed distributions corresponding to positive (endothermic) and negative (exothermic) regimes, respectively. Figure 1(C) depicts ball-and-stick models of the eight molecular adsorbates considered in this study, while Fig. 1(D) presents their corresponding adsorption energy distributions, revealing higher variability and broader extrema compared to single-atom systems. These observed trends reflect intrinsic chemical variability in adsorbate binding strengths and reinforce the importance of a diverse dataset for enabling Q-CatNet’s generalizability across different adsorption regimes. Fig. 1Dataset visualization and statistical distribution: (A) Two-dimensional t-SNE projection of the learned representations for single-atom adsorbates (C*, H*, N*, and O*), revealing well-separated clusters. Transparent circles indicate adsorption energy magnitudes, with larger radii corresponding to more optimal binding (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{E}_{ads}$$\end{document} ≈ 0 eV); (B) Box-and-violin plots showing adsorption energy distributions for single-atom adsorbates. Hydrogen systems display narrower, centered distributions, while nitrogen and oxygen exhibit broader and asymmetric trends; (C) Ball-and-stick representations of the eight molecular adsorbates included in this study: OH, CH₃, CH₄, CH₃OH, C₂H₅OH, NO, NH₃, and CH₃NNOH; and (D) Adsorption energy distributions for molecular adsorbates, highlighting broader energy ranges and variability across species, indicative of the complex chemisorption behavior within the dataset. Dataset visualization and statistical distribution: (A) Two-dimensional t-SNE projection of the learned representations for single-atom adsorbates (C*, H*, N*, and O*), revealing well-separated clusters. Transparent circles indicate adsorption energy magnitudes, with larger radii corresponding to more optimal binding (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{E}_{ads}$$\end{document} ≈ 0 eV); (B) Box-and-violin plots showing adsorption energy distributions for single-atom adsorbates. Hydrogen systems display narrower, centered distributions, while nitrogen and oxygen exhibit broader and asymmetric trends; (C) Ball-and-stick representations of the eight molecular adsorbates included in this study: OH, CH₃, CH₄, CH₃OH, C₂H₅OH, NO, NH₃, and CH₃NNOH; and (D) Adsorption energy distributions for molecular adsorbates, highlighting broader energy ranges and variability across species, indicative of the complex chemisorption behavior within the dataset. Q-CatNet input descriptor design The Q-CatNet model incorporates key partially quantum-informed descriptors to capture both local bonding interactions and system-level physical correlations. In this context, quantum-informed denotes the use of descriptors derived from operators rooted in the quantum mechanical Hamiltonian, along with features obtained from electronic structure calculations, rather than implying an explicit quantum solver. These physics-guided attributes are integrated within a graph-based framework across node ( v), edge ( u), and global state (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{G}$$\end{document}) levels. The node features comprise 14 carefully selected physicochemical properties that are transformed into one-hot encoding forms of size \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{v}_{i}\in\:{\left\{\mathrm{0,1}\right\}}^{\left(1\times\:230\right)}$$\end{document}, thereby capturing essential thermo- and electronic characteristics relevant to catalytic systems. A detailed description of the selected physicochemical attributes is provided in the Supplementary Information, where their effectiveness as input representations for structure–property learning has been well established in prior materials informatics studies29–32. Meanwhile, the edge features are expanded beyond the conventional bond-length attribute used in the traditional CGCNN5 to include bond angles (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\theta\:$$\end{document}), differences in electronegativity (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:\chi\:$$\end{document}), and three quantum-informed descriptors that capture the underlying physics of atomic environments. These descriptors are derived by approximating mathematical components of the computationally intensive Schrödinger equation, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{H}{\Psi\:}=E{\Psi\:}$$\end{document}, whereby E is the total electronic energy of the system and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\Psi\:}$$\end{document} is the wave function31. The current study applies such quantum-informed descriptors, as represented in their encoded Coulomb matrix33, Ewald sum matrix34, and Sine matrix35,36 forms, for efficiently simulating adsorbate interaction on bulk metallic surfaces. Specifically, the Coulomb matrix ( CM) serves as an approximation of the nucleus-to-nucleus repulsion and electrostatic interactions present in the Hamiltonian (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{H}$$\end{document}) of the Schrödinger equation, and can be mathematically expressed as:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{M}_{ij}^{coulomb}=\left\{\begin{array}{c}0.5{Z}_{i}^{2.4}\:\:\:\:\:\:\:\:\:\:\:\:\:\:if\:i=j\\\:\frac{{Z}_{i}{Z}_{j}}{\left|{R}_{i}-{R}_{j}\right|}\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:if\:i\ne\:j\end{array}\right.$$\end{document} Here, Z is the atomic number of atom i, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left|{R}_{i}-{R}_{j}\right|=\left|{R}_{ij}\right|$$\end{document} is the interatomic cartesian distance between atoms i and j in atomic units31,33. The Ewald sum matrix extends the functionality of the Coulomb matrix by providing a more accurate representation of electrostatic interactions. It achieves this by incorporating long-range Coulombic effects through both real and reciprocal space components, effectively capturing the periodic boundary conditions characteristic of solid-state crystal systems. Moreover, the Sine matrix offers a computationally efficient alternative by heuristically modeling the periodic Coulombic interactions using a sine function, thereby approximating the influence of a periodic potential in a simplified yet physically meaningful manner. In the present research, the quantum-informed descriptors are calculated by importing their respective Python modules from the dscribe36. It shall be noted that the aforementioned edge features are only computed for neighboring atoms that are located within a natural spherical cut-off, which is based on the covalent radii of neighboring atoms, and identified via the adjacency matrix. This research identifies such neighboring atoms to build the adjacency matrix by enabling the Atomic Simulation Environment (ASE) NeighborList37 module with non-self interaction settings. To enable efficient representation within the model, the overall computed edge features are subsequently transformed into one-hot encoding vectors of size \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{u}_{ij}\in\:{\left\{\mathrm{0,1}\right\}}^{\left(1\times\:60\right)}$$\end{document}. Furthermore, this study incorporates global state attributes to adequately simulate the electronic and periodic dynamics that are associated with the catalyst system. Such idea has been implemented in the past using the MEGNet model6, ensuring that message-passing operations within the graph network are continuously informed by system-wide context, facilitating the propagation of relevant information across all nodes and edges. Specifically, two primary global attributes are encoded as one-dimensional vectors with a combined dimension of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{G}\in\:{\mathbb{R}}^{\left(1\times\:1204\right)}$$\end{document}. These include the density of states (DOS) signal \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:g\left(E\right)\in\:{\mathbb{R}}^{\left(1\times\:1024\right)}$$\end{document} of the bulk system, sampled within ± 10 eV of the Fermi energy, and the X-ray diffraction (XRD) intensity profile \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{I}_{XRD}\in\:{\mathbb{R}}^{\left(1\times\:180\right)}$$\end{document} of the adsorbate-bulk system, evaluated over a Bragg angle range up to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:2\theta\:=180^\circ\:$$\end{document}. In general, the DOS signal provides vital insights into the catalyst’s electronic structure, especially in the context of d-band theory, which plays a key role in determining how the catalyst interacts with the adsorbate16. On the other hand, Fourier-transformed X-ray Diffraction (XRD) reveals information about the combined adsorbate-bulk crystal structure, along with other critical details such as phase composition and defect chemistry, which are essential factors for understanding and predicting catalytic performance38–40. Although lattice parameters, atomic positions, and symmetry are in principle encoded within graph representations through node coordinates and periodic connectivity, these representations operate in real space and rely on finite-radius message passing. In contrast, XRD patterns offer the advantage of encoding the crystal structure in reciprocal space, capturing long-range periodicity and global structural correlations that are not trivially accessible through local neighborhood aggregation. Finally, as further explored in ablation studies, the global feature space can be extended to include explicit surface-sensitive descriptors, namely the Miller indices ( hkl) and a surface shift parameter, resulting in an expanded representation of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{G}\in\:{\mathbb{R}}^{\left(1\times\:1208\right)}$$\end{document}. These additional features provide direct information on surface orientation and termination, enabling improved modeling of the facet-dependent adsorption behavior while reinforcing the core role of DOS and XRD as the foundational global descriptors in Q-CatNet. Figure 2 presents the Q-CatNet descriptor framework alongside three representative adsorption sites where a hydrogen atom may bind to the intermetallic surface of the TiPt binary alloy. These sites correspond to specific crystallographic planes typically used in modeling heterogeneous catalysis, as they reflect the atomic arrangement at the surface. The considered adsorption geometries include: (1) Hollow – where the adsorbate resides at the center of multiple neighboring atoms, enabling multi-center coordination; (2) Bridge – where the adsorbate is positioned between two adjacent surface atoms; and (3) On-top – where the adsorbate binds directly above a single surface atom, resulting in lower coordination. To enhance binding stability and coordination, this study prioritizes adsorption configurations in the following order: hollow, followed by bridge, and lastly on-top. These preferred sites are determined using the AdsorbateSiteFinder module from the Pymatgen adsorption library41, applied to a conventional unit cell. The use of the conventional cell, obtained via the SpacegroupAnalyzer module, rather than the primitive cell, offers improved alignment with experimental observations and better supports the construction of periodic slab models for DFT, which is essential for simulating infinite surfaces42. Fig. 2The Q-CatNet graphical framework illustrating quantum-informed node, edge, and global attributes for a TiPt bulk crystal, obtained from the Materials Project (MP) database. The hydrogen adsorption energy, calculated from the OC20 dataset, is 4.5 × 10⁻³ eV. Three adsorption sites are modeled to simulate adsorbate-surface interactions: hollow, bridge, and on-top, prioritized in this order. The primary global attribute vector, of length 1204, corresponds to the Density of States (DOS) and the single crystal X ray powder diffraction pattern (XRD). In its extended variant configuration, Q‑CatNet further incorporates surface‑specific features related to the Miller indices ( hkl) and a surface shift parameter, yielding an expanded global representation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{G}\in\:{\mathbb{R}}^{\left(1\times\:1208\right)}$$\end{document}. The Q-CatNet graphical framework illustrating quantum-informed node, edge, and global attributes for a TiPt bulk crystal, obtained from the Materials Project (MP) database. The hydrogen adsorption energy, calculated from the OC20 dataset, is 4.5 × 10⁻³ eV. Three adsorption sites are modeled to simulate adsorbate-surface interactions: hollow, bridge, and on-top, prioritized in this order. The primary global attribute vector, of length 1204, corresponds to the Density of States (DOS) and the single crystal X ray powder diffraction pattern (XRD). In its extended variant configuration, Q‑CatNet further incorporates surface‑specific features related to the Miller indices ( hkl) and a surface shift parameter, yielding an expanded global representation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{G}\in\:{\mathbb{R}}^{\left(1\times\:1208\right)}$$\end{document}. Q-CatNet Graph-based learning model Given a catalyst system represented in its conventional unit cell, the Q-CatNet model encodes it as an input graph G defined by node, edge, and global-state attributes. Using supervised learning, the Q-CatNet learns a forward mapping function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:f\left( G \right) \mapsto \Delta E_{{ads}}$$\end{document}, which regressively predicts the target adsorbate energy. As illustrated in Fig. 3, the mapping model in this study is constructed by integrating deep neural networks, specifically by combining a CGCNN layer with fully connected linear layers. Fundamentally, the CGCNN layer performs nonlinear convolution and pooling operations to iteratively update atomic (node) feature vectors through message-passing interactions with neighboring atoms, as specified by the adjacency matrix of the adsorbate-bulk system. Accordingly, the graph-based update of an atom \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{v}_{i}$$\end{document} at a given convolution step is governed by the following Eq.5:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\ddot{v}_{i} = \dot{v}_{i} + \sum\limits_{{j,k}} {\sigma \left( {\dot{z}_{{\left( {i,j} \right)_{k} }} \dot{W}_{f} + \dot{b}_{f} } \right) \odot g\left( {\dot{z}_{{\left( {i,j} \right)_{k} }} \dot{W}_{s} + \dot{b}_{s} } \right)}$$\end{document} Here, the concatenated vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\dot{z}}_{{\left(i,j\right)}_{k}}=\left[{\dot{v}}_{i}\:\left|\right|{\:\dot{v}}_{j}\:\left|\right|{\:\dot{u}}_{ij}\right]$$\end{document} includes the central atom feature \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\dot{v}}_{i}$$\end{document}, neighboring atom feature \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\dot{v}}_{j}$$\end{document}, and the quantum-informed edge feature\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\:\dot{u}}_{ij}$$\end{document}. The functions \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:$$\end{document} and g represent the sigmoid and softplus activation functions, respectively, which introduce nonlinearity relative to weight matrices \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\dot{W}}_{f}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\dot{W}}_{s}$$\end{document}, and their associated biases \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\dot{b}}_{f}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\dot{b}}_{s}$$\end{document}. The mathematical operation ⊙ applies element-wise multiplication, functioning as a feature-level attention mechanism that filters out irrelevant interatomic interactions while highlighting significant bonding characteristics within the catalyst system. To satisfy permutation invariance with respect to atomic indexing and the crystalline unit cell choice, the convolutional operations are average-pooled after R iterations to produce a feature representation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\ddot{v}}_{i}^{R}$$\end{document} that contains the vectorized information of the overall adsorbate-bulk system. This representation is supplemented with additional quantum context by concatenating with the global-state attributes. Fully connected linear layers are then employed to model the complex relationships between the combined crystal features and their corresponding target properties. The training is performed by minimizing the difference between the predicted property \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{\widehat{E}}_{ads}$$\end{document} and the DFT calculated property \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{E}_{ads}$$\end{document}, defined by a mean squared error cost function. Further details on the architecture and optimized hyperparameters for training the Q-CatNet model are provided in the Supplementary Information. Fig. 3Illustration of the Q-CatNet model’s learning process, which maps a crystal graph (comprising nodes, edges and global attributes) to the adsorption energy (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{E}_{ads}$$\end{document}) of a catalyst system. Message-passing operations are conducted among atoms within the conventional unit cell by applying R convolutional layers, followed by fully connected linear layers. Illustration of the Q-CatNet model’s learning process, which maps a crystal graph (comprising nodes, edges and global attributes) to the adsorption energy (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{E}_{ads}$$\end{document}) of a catalyst system. Message-passing operations are conducted among atoms within the conventional unit cell by applying R convolutional layers, followed by fully connected linear layers. Results Single-atom adsorbate energy prediction performance A key future application of the Q-CatNet model lies in its ability to estimate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{E}_{ads}$$\end{document} in previously unexplored physical regimes, using limited knowledge from first-principles calculations and/or experimental data on heterogeneous catalysis systems. In this study, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{E}_{ads}$$\end{document} is predicted for four distinct single-atom adsorbates (i.e., C*, H*, N*, and O*) interacting with bulk binary and ternary metallic surfaces. In the prediction exercise, the model’s performance is assessed using three standard regression metrics: Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and coefficient of determination (R²), as summarized in Table 1. The prediction errors are based on a 90/10 train-test split exercise, with each single-atom adsorbate contributing 10% of its total dataset to maintain a balanced and homogeneous data distribution across the evaluation set. Additional analyses are provided to evaluate the model’s robustness on unseen catalyst systems, with specific attention to the impact of training sample size and ablation studies. Overall, the nominal Q-CatNet (XRD + DOS) model achieved a MAE of 0.294 eV on the test set across all adsorbate–surface combinations, indicating strong potential for generalization in predicting \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{E}_{ads}$$\end{document} for unknown catalysts. Among the individual adsorbates, Hydrogen yielded the lowest MAE of 0.186 eV, highlighting its relatively high accuracy and model reliability in this subset. In contrast, Oxygen exhibited the highest coefficient of R² at 84.46%, despite a slightly higher MAE of 0.318 eV, compared to Hydrogen’s 75.07% R². This discrepancy arises from the narrower variance observed in the training distribution of Hydrogen-surface systems, which can limit the model’s ability to explain variability in predictions. Conversely, the broader spread in data for Oxygen, Nitrogen, and Carbon systems enables the model to better capture diverse adsorbate behaviors, as reflected by their consistently high R² values (all exceeding 80%). In this regard, Fig. 4 presents parity plots illustrating the regression performance of the model and highlights the adsorbate-specific prediction distributions. These plots provide visual confirmation of the model’s predictive accuracy and consistency, with most data points closely aligned along the diagonal, indicating strong agreement between predicted and true values. Table 1Prediction performance of the Q-CatNet (XRD + DOS) model for single-atom and multi-atom adsorbate systems. The table summarizes the mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²) on the test set for Q-CatNet predictions across both atomic and molecular adsorbates. Among the single-atom systems, hydrogen–surface catalysts exhibit the lowest MAE values, reflecting highly stable and reliable predictions, while oxygen–surface systems achieve the highest R² scores, indicating superior generalization potential across diverse surface environments. For multi-atom adsorbates, the nitric oxide (NO) and hydroxyl (OH) systems demonstrate the best overall predictive accuracy, balancing low prediction errors with strong parity correlations to DFT-computed adsorption energies. These results collectively highlight Q-CatNet’s robustness and transferability across varying adsorbate complexities, from simple atomic interactions to multi-center molecular adsorption dynamics.Adsorbate species\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{E}_{ads}\:$$\end{document}MAE (eV)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{E}_{ads}\:$$\end{document}RMSE (eV)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{E}_{ads}\:$$\end{document}R2 (%)Single-atom adsorbate Carbon (C*)0.3500.79180.30 Hydrogen (H*)0.1860.42575.07 Nitrogen (N*)0.3280.74683.18 Oxygen (O*)0.3180.70684.46 Overall0.2940.67982.56Multi-atom (molecular) adsorbate Hydroxyl (OH)0.2650.65472.71 Methyl (CH3)0.3080.65838.79 Methane (CH4)0.2240.62614.62 Methanol (CH3OH)0.2580.58143.36 Ethanol (C2H5OH)0.3060.71340.36 Nitric oxide (NO)0.3440.80783.21 Ammonia (NH3)0.2430.57053.57 Methyl nitroxyl (CH3NNOH)0.6791.25166.07 Overall0.3110.72885.72Combined (single-atom + multi-atom) dataset Overall0.3060.72486.77 Table 1Prediction performance of the Q-CatNet (XRD + DOS) model for single-atom and multi-atom adsorbate systems. The table summarizes the mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²) on the test set for Q-CatNet predictions across both atomic and molecular adsorbates. Among the single-atom systems, hydrogen–surface catalysts exhibit the lowest MAE values, reflecting highly stable and reliable predictions, while oxygen–surface systems achieve the highest R² scores, indicating superior generalization potential across diverse surface environments. For multi-atom adsorbates, the nitric oxide (NO) and hydroxyl (OH) systems demonstrate the best overall predictive accuracy, balancing low prediction errors with strong parity correlations to DFT-computed adsorption energies. These results collectively highlight Q-CatNet’s robustness and transferability across varying adsorbate complexities, from simple atomic interactions to multi-center molecular adsorption dynamics.Adsorbate species\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{E}_{ads}\:$$\end{document}MAE (eV)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{E}_{ads}\:$$\end{document}RMSE (eV)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{E}_{ads}\:$$\end{document}R2 (%)Single-atom adsorbate Carbon (C*)0.3500.79180.30 Hydrogen (H*)0.1860.42575.07 Nitrogen (N*)0.3280.74683.18 Oxygen (O*)0.3180.70684.46 Overall0.2940.67982.56Multi-atom (molecular) adsorbate Hydroxyl (OH)0.2650.65472.71 Methyl (CH3)0.3080.65838.79 Methane (CH4)0.2240.62614.62 Methanol (CH3OH)0.2580.58143.36 Ethanol (C2H5OH)0.3060.71340.36 Nitric oxide (NO)0.3440.80783.21 Ammonia (NH3)0.2430.57053.57 Methyl nitroxyl (CH3NNOH)0.6791.25166.07 Overall0.3110.72885.72Combined (single-atom + multi-atom) dataset Overall0.3060.72486.77 Prediction performance of the Q-CatNet (XRD + DOS) model for single-atom and multi-atom adsorbate systems. The table summarizes the mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²) on the test set for Q-CatNet predictions across both atomic and molecular adsorbates. Among the single-atom systems, hydrogen–surface catalysts exhibit the lowest MAE values, reflecting highly stable and reliable predictions, while oxygen–surface systems achieve the highest R² scores, indicating superior generalization potential across diverse surface environments. For multi-atom adsorbates, the nitric oxide (NO) and hydroxyl (OH) systems demonstrate the best overall predictive accuracy, balancing low prediction errors with strong parity correlations to DFT-computed adsorption energies. These results collectively highlight Q-CatNet’s robustness and transferability across varying adsorbate complexities, from simple atomic interactions to multi-center molecular adsorption dynamics. Fig. 4Regressive parity plots illustrating the deviation between the Q-CatNet (XRD + DOS) model’s predicted adsorption energies and the actual-DFT values for the four studied single-adsorbate-surface catalyst systems. The model’s predictive performance on the test set, quantified by R² values, is estimated at 80.30%, 75.07%, 83.18%, and 84.46% for the Carbon, Hydrogen, Nitrogen, and Oxygen adsorbate systems, respectively. Regressive parity plots illustrating the deviation between the Q-CatNet (XRD + DOS) model’s predicted adsorption energies and the actual-DFT values for the four studied single-adsorbate-surface catalyst systems. The model’s predictive performance on the test set, quantified by R² values, is estimated at 80.30%, 75.07%, 83.18%, and 84.46% for the Carbon, Hydrogen, Nitrogen, and Oxygen adsorbate systems, respectively. Multi-atom (molecular) adsorbate energy prediction performance The predictive capability of the Q-CatNet (XRD + DOS) framework was further extended to multi-atom (molecular) adsorbate surface systems to evaluate its robustness in capturing the energetics of chemically complex surface interactions. As shown in Fig. 5A, the aggregate regressive parity distribution across all molecular adsorbates exhibits a strong linear correlation, with an overall coefficient of determination (R²) of 85.72%, confirming that the model predictions are in close agreement with DFT calculated adsorption energies. The individual parity plots for the eight distinct molecular adsorbates (Fig. 5B) further elucidate Q CatNet’s chemical specificity and its potential for generalization across structurally and electronically diverse species. A key observation from this analysis lies in the contrast between the Mean Absolute Error (MAE) and R² metrics for each adsorbate system. As visualized in the violin plots (Fig. 1D), these differences are closely linked to the intrinsic variance of the true \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{E}_{ads}$$\end{document} distributions. Methane (CH₄) provides the most illustrative example. Although it achieves the lowest MAE of 0.224 eV, which reflects high absolute predictive precision, it also produces the lowest R² value at 14.62%. This apparent inconsistency arises from the narrow and tightly clustered \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{E}_{ads}$$\end{document} distribution centered around − 3 eV, reflecting methane’s inherently weakly activated yet exothermic adsorption behavior. Besides, this observation aligns well with prior DFT and experimental findings43 that attribute such energy ranges to the dissociative chemisorption of methane on transition metal and oxide surfaces, leading to moderately exothermic adsorption energies and consequently low intrinsic variance in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{E}_{ads}$$\end{document}. Conversely, molecular species with broader \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{E}_{ads}$$\end{document} distributions, such as Nitric Oxide (NO) and Hydroxyl (OH), exhibit higher R² values at 83.21% and 72.71%, respectively, despite comparable or slightly higher MAE values at 0.344 eV and 0.265 eV. Their wide \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{E}_{ads}$$\end{document} distribution ranges (approximately 15 eV for NO and 10 eV for OH) arise from the diversity of possible binding motifs, spin states, and surface coordination environments, which contribute to a more complex energy landscape that the model successfully captures. Overall, these results underscore the importance of interpreting MAE and R² jointly in the context of data variance and chemical diversity. While MAE quantifies absolute energy deviation and serves as a critical indicator for chemical accuracy, R² more effectively reflects the model’s ability to reproduce energy trends across wide ranging adsorption configurations. Collectively, the findings demonstrate Q-CatNet’s adaptability to both low variance systems such as CH₄ and high variance systems such as NO and OH, highlighting its potential as an adept predictive framework for heterogeneous catalytic adsorption energetics. Fig. 5Regressive parity plots showing the Q-CatNet (XRD + DOS) prediction performance on the test set for multi-atom adsorbates. (A) Aggregate regression for all molecular adsorbates showing an overall R² = 85.72%. (B) Individual regression plots for eight representative adsorbates (OH, CH₃, CH₄, CH₃OH, C₂H₅OH, NO, NH₃, and CH₃NNOH) revealing distinct predictive characteristics across chemically diverse systems. Model validation was performed using a 90/10 train-test split with each adsorbate species contributing 10% of its total dataset to ensure a balanced and statistically representative evaluation subset. Regressive parity plots showing the Q-CatNet (XRD + DOS) prediction performance on the test set for multi-atom adsorbates. (A) Aggregate regression for all molecular adsorbates showing an overall R² = 85.72%. (B) Individual regression plots for eight representative adsorbates (OH, CH₃, CH₄, CH₃OH, C₂H₅OH, NO, NH₃, and CH₃NNOH) revealing distinct predictive characteristics across chemically diverse systems. Model validation was performed using a 90/10 train-test split with each adsorbate species contributing 10% of its total dataset to ensure a balanced and statistically representative evaluation subset. Impact of training sample size on model generalizability To evaluate the robustness and generalization capability of Q-CatNet (XRD + DOS), the present study systematically investigates how reducing the training sample size influences predictive performance, as it relates to the single-atom data subset. This analysis provides critical insights into the model’s stability under data-scarce regimes, which are often encountered in real-world applications where generating large volumes of DFT data is computationally prohibitive. For this purpose, the training data was incrementally reduced in steps of 10%, while maintaining a balanced representation of adsorbate species in the test set to preserve uniform distribution. Figure 6 summarizes the outcomes, presenting the predictive performance in terms of MAE (Fig. 6(A)) and R² (Fig. 6(B)) on test set. Across all adsorbates, the model demonstrated strong resilience to reductions in training data. On average, the overall MAE reduces by approximately 5.5 meV per 1% increase in train data size, highlighting the model’s inherent stability. Among individual adsorbates, hydrogen single-atom adsorbates exhibited the most robust performance, with MAE improving by 3.4 meV whereas carbon, oxygen, and nitrogen followed with slightly higher variations, recording changes of 5.4 meV, 6.1 meV, and 6.4 meV, respectively. In terms of R², the predictive performance improves at an overall average rate of 0.446% per 1% increase in train size. Notably, nitrogen and oxygen systems showed greater resilience, with relatively modest performance drops of 0.412% and 0.425%, respectively. By contrast, carbon and hydrogen systems exhibited steeper declines, at 0.433% and 0.829%, respectively. These performance trends can be directly correlated with the statistical distributions of adsorption energies in the training dataset (i.e., Table 1; Fig. 1(D)). Specifically, the narrower variance of hydrogen–surface interactions around 0 eV led to a more consistent and favorable MAE performance, explaining hydrogen’s superior stability across varying training sample sizes. Conversely, carbon, oxygen, and nitrogen systems, which are characterized by broader and more skewed \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{E}_{ads}$$\end{document} distributions, demonstrated higher baseline MAE values and comparatively lower stability. Taken together, these findings emphasize that Q-CatNet is not only capable of learning robust adsorption–surface interaction patterns but also maintains efficiency even under significant reductions in training data. This resilience underscores the model’s practical potential for accelerating catalyst discovery in data-limited domains, where transferability and reliability are essential. Fig. 6Influence of increasing train data proportion on adsorbate energy prediction, revealing a stable improvement in Q-CatNet’s performance, reflecting the model’s capacity to capture underlying patterns rather than memorize the training set. (A) MAE results with an overall slope of − 5.5 meV per 1% increase in train size; (B) R² results with an overall slope of 0.4462% per 1% increase in train size. Influence of increasing train data proportion on adsorbate energy prediction, revealing a stable improvement in Q-CatNet’s performance, reflecting the model’s capacity to capture underlying patterns rather than memorize the training set. (A) MAE results with an overall slope of − 5.5 meV per 1% increase in train size; (B) R² results with an overall slope of 0.4462% per 1% increase in train size. Q-CatNet predictive simulation of other DFT target properties The present research also experiments on the applicability of the Q-CatNet model beyond catalytic adsorption energy prediction to encompass other critical DFT-derived properties that quantify material stability and functionality. The results are summarized in Table 2 showing the model’s performance across a suite of scalar properties, namely: formation energy (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{f}$$\end{document}), band gap (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{g}$$\end{document}), energy above the convex hull (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{hull}$$\end{document}), and total DFT-computed energy (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{tot}$$\end{document}), as well as a vectorized representation of the one-dimensional interpolated Density of States (DOS) signal. Scalar targets were evaluated using standard regression error metrics (MAE, RMSE, and R²), while the DOS signals were assessed using similarity-based measures, reflecting the structural and directional agreement between predicted and reference vectors. This benchmarking exercise was carried out on ~ 52,000 newly extracted compounds from the Materials Project, and employs an ablated version of Q-CatNet that integrates all edge-computed features (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{u}_{ij}\in\:{\left\{\mathrm{0,1}\right\}}^{\left(1\times\:60\right)}$$\end{document}) alongside their respective XRD-derived global-state attributes (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{I}_{XRD}\in\:{\mathbb{R}}^{\left(1\times\:180\right)}$$\end{document}) of the bulk crystal lattice. All results were obtained under a consistent 90/10 train–test protocol. Figure 7 further provides parity plots for scalar targets, offering a visual confirmation of the model’s predictive accuracy. Overall, Q-CatNet demonstrates robust performance across all tested properties. Notably, formation energy predictions achieved MAE of 0.038 eV/atom with a R² of 97.19%, representing a substantial improvement with respect to our prior works29,30,38,39,44. For energy above the convex hull, the model delivered an MAE of 0.024 eV/atom and RMSE of 0.104 eV/atom, paired with a moderate R² of 61.3%, indicating reasonable yet less reliable predictions compared to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{f}$$\end{document}. For the electronic band gap, Q-CatNet attained a MAE of 0.097 eV, markedly outperforming the original CGCNN benchmark5, which reported 0.388 eV when trained on ~ 16,000 Materials Project structures, thereby demonstrating strong predictive capability for electronic properties. Similarly, predictions of total DFT energy exhibited a high R² of 94.31%, despite a larger MAE of 0.559 eV, highlighting the model’s ability to capture global energy trends even in the presence of larger absolute errors. For the vector property, Q-CatNet was evaluated on its ability to reproduce normalized DOS signals \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:g\left(E\right)\in\:{\mathbb{R}}^{\left(1\times\:1024\right)}$$\end{document}. This task is particularly important, as DOS signals serve as the foundational input for constructing the Q-CatNet descriptor used in inverse design of novel catalysts. Prediction quality was assessed using three complementary similarity metrics: Jaccard similarity (0.761), Cosine similarity (0.922), and Pearson correlation (0.825), with each metric emphasizing a distinct facet of agreement. The Jaccard score captures overlap in active spectral components; Cosine emphasizes directional similarity irrespective of magnitude; and Pearson quantifies linear correlation in amplitude variations. The higher Cosine score suggests that Q-CatNet excels at reproducing the overall structural and directional characteristics of the DOS, even if minor discrepancies remain in feature presence (Jaccard) or signal amplitude (Pearson). This is consistent with the model’s prioritization of capturing global DOS features rather than exact peak magnitudes, a choice motivated by the importance of DOS shape and orientation for downstream tasks such as electronic structure analysis, materials classification, and screening for inverse design. Collectively, these results highlight Q-CatNet’s versatility as a unified framework capable of predicting both scalar thermodynamic and electronic properties, as well as complex vectorized descriptors. Table 2Predictive performance of an ablated version of the Q-CatNet across various target properties. Scalar outputs from DFT are evaluated using regression metrics: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and coefficient of determination (R2). Vector-valued DFT properties, are assessed using similarity metrics including Jaccard index, Cosine similarity, and Pearson correlation coefficient.Scalar Target PropertyMAERMSER2 (%)Formation energy, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{f}$$\end{document} (eV/atom)0.0380.13197.19Energy above the convex hull, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{hull}$$\end{document} (eV/atom)0.0240.10461.30Energy band gap, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{g}$$\end{document} (eV)0.0970.39085.25DFT total energy, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{tot}$$\end{document} (eV)0.5592.94994.31Vector target propertyJaccardCosinePearsonNormalized Density of State (DOS) signal0.7610.9220.825 Table 2Predictive performance of an ablated version of the Q-CatNet across various target properties. Scalar outputs from DFT are evaluated using regression metrics: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and coefficient of determination (R2). Vector-valued DFT properties, are assessed using similarity metrics including Jaccard index, Cosine similarity, and Pearson correlation coefficient.Scalar Target PropertyMAERMSER2 (%)Formation energy, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{f}$$\end{document} (eV/atom)0.0380.13197.19Energy above the convex hull, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{hull}$$\end{document} (eV/atom)0.0240.10461.30Energy band gap, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{g}$$\end{document} (eV)0.0970.39085.25DFT total energy, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{tot}$$\end{document} (eV)0.5592.94994.31Vector target propertyJaccardCosinePearsonNormalized Density of State (DOS) signal0.7610.9220.825 Predictive performance of an ablated version of the Q-CatNet across various target properties. Scalar outputs from DFT are evaluated using regression metrics: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and coefficient of determination (R2). Vector-valued DFT properties, are assessed using similarity metrics including Jaccard index, Cosine similarity, and Pearson correlation coefficient. Fig. 7Parity plots from test set evaluation comparing Q-CatNet-predicted scalar properties: (A) formation energy (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{f}$$\end{document}), (B) band gap (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{g}$$\end{document}), (C) energy above the convex hull (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{hull}$$\end{document}), and (D) total energy (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{tot}$$\end{document}) against DFT reference values from the Materials Project. Parity plots from test set evaluation comparing Q-CatNet-predicted scalar properties: (A) formation energy (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{f}$$\end{document}), (B) band gap (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{g}$$\end{document}), (C) energy above the convex hull (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{hull}$$\end{document}), and (D) total energy (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{tot}$$\end{document}) against DFT reference values from the Materials Project. Discussion Ablation studies of the Q-CatNet and comparison to other invariant baseline models This study conducts an ablation analysis to evaluate the sensitivity of Q-CatNet’s performance to its integral edge and global feature components in adsorbate energy prediction. Additionally, predictive results are benchmarked against four baseline models spanning diverse graph-based and image-based deep learning approaches that are reliably applied in molecular and materials informatics. Table 3 summarizes the MAE (eV), RMSE (eV), and R² (%) scores for the six models evaluated, as it relates to the single-atom adsorbate dataset. A comparative visualization of the MAE scores for these models is shown in Fig. 8(A). The ablation begins with the CGCNN in its baseline form, incorporating only simple edge features derived from bond length. This is compared to an edge-enhanced CGCNN, which integrates all considered edge attributes, including bond length, bond angle, electronegativity difference, Coulomb matrix, Ewald-sum matrix, and Sine matrix, while excluding global state attributes. Incorporating enhanced edge features yields a modest improvement in overall accuracy, reducing MAE from 0.422 eV (plain CGCNN) to 0.409 eV (edge-enhanced CGCNN), an improvement of approximately 3%. However, these results suggest that edge-enhanced attributes alone offer limited gains unless combined with global state attributes. When both are present, as demonstrated in Q-CatNet, the predictive accuracy improves substantially, with MAE reductions of 30.3% and 28.1% compared to the plain CGCNN and edge-enhanced CGCNN, respectively. This underscores the synergistic role of edge and global features in capturing complex adsorption-surface interactions. To further probe the contribution of global attributes, a simplified Q-CatNet-XRD variant was developed in which the global features are derived solely from XRD patterns. This scenario mimics cases where DFT–derived DOS data are unavailable or computationally expensive to obtain, while XRD data can be analytically generated directly from Fourier transformations of the reciprocal crystal lattice. Interestingly, Q-CatNet-XRD maintains competitive accuracy, achieving an MAE of 0.322 eV, representing a 21.3% improvement over the edge-enhanced CGCNN and only a 9.5% reduction in accuracy relative to the full, non-ablated Q-CatNet model. This finding highlights that XRD-based features can serve as reliable surrogates for global quantum descriptors in scenarios where electronic-structure information is limited. As a result, the XRD‑derived features (structural, reciprocal‑space) provide complementary system‑level structural information that strengthens the learned structure–property relationships. Consequently, their integration with DOS (electronic) descriptors in Q‑CatNet represents a physically motivated fusion of distinct yet synergistic information domains. Beyond the ablation studies, Q-CatNet’s performance is benchmarked against one image-based and three graph-based deep learning architectures. The chosen image-based baseline is the Fourier Transformed Crystal Property (FTCP) descriptor45, which is an invertible representation that encodes atomic positions and types into the reciprocal (Fourier) space of a crystal lattice. This preserves all spatial information, making it particularly suited for predicting properties naturally expressed in reciprocal space, such as band structures, phonon dispersions, and diffraction patterns. When applied to adsorbate energy prediction, FTCP yields an MAE of 0.544 eV, indicating that the Q-CatNet achieves a relative improvement of 46%. Besides, even the plain CGCNN surpasses FTCP by 22.4%, reinforcing the advantage of graph-based over image-based crystal representations. In theory, this performance gap is attributed to the inherent invariance of graph neural networks to rotation, translation, and permutation, properties that ensure physical consistency. In contrast, image-based approaches require discretizing atomic structures into voxel or grid formats, which can degrade precision and significantly increase input dimensionality. Finally, Q-CatNet is benchmarked against three reputable invariant deep learning architectures: Graph Isomorphism Network Convolution (GINConv)46, Neural Network Convolution (NNConv)47, and the Continuous-filter Convolutional Neural Network for Molecules and Materials (SchNet)8. For a fair comparison, GINConv and NNConv were augmented with the same global attributes used in Q-CatNet and trained on the same dataset split for evaluating the Q-CatNet. The results show that Q-CatNet outperforms GINConv by 8.7%, NNConv by 26.3%, and SchNet by 38.2% in MAE accuracy, underscoring its enhanced representational capacity and suitability for accelerating forward-design catalysis simulation and materials discovery. Furthermore, to investigate the role of explicit surface information, an extended variant denoted as Q-CatNet (XRD + DOS + Surface) was developed by augmenting the global descriptor space with surface-specific features extracted from the OC20 dataset. These include the Miller indices ( hkl), which define the crystallographic orientation of the exposed surface facet, and a surface shift parameter, which accounts for variations in slab termination and atomic registry along the surface normal. Physically, the Miller indices govern the coordination environment and atomic packing density of surface atoms, which are known to strongly influence adsorption energetics, while the surface shift captures subtle geometric variations that can alter adsorption sites and local bonding configurations. From the investigation, incorporating these features results in a notable improvement in predictive performance, reducing the MAE from 0.294 eV for the full Q-CatNet (XRD + DOS) model to 0.237 eV. This corresponds to an additional ~ 19.4% improvement, highlighting the importance of explicitly encoding surface-sensitive information in adsorption energy prediction tasks. These findings are consistent with established surface science principles, where adsorption energies vary significantly across different crystallographic facets and terminations, demonstrating that surface geometry descriptors, such as generalized coordination numbers, are essential for capturing facet-dependent reactivity48. Importantly, while the inclusion of surface descriptors enhances accuracy, the good performance of Q-CatNet (XRD + DOS) and even Q-CatNet (XRD) demonstrates that the model retains meaningful predictive capability in the absence of explicit surface features. This supports the central premise of this work: that global electronic (DOS) and reciprocal-space structural (XRD) descriptors provide a robust and transferable foundation for learning adsorption energetics. In this context, surface features can be interpreted as auxiliary refinements that improve fidelity when available, rather than mandatory inputs for model functionality. This flexibility is particularly valuable for high-throughput screening scenarios, where detailed surface annotations may not always be readily accessible. In the Supplementary Information, we provide additional details on runtime performance and computational scalability, along with extended comparisons to more advanced equivariant models that were additionally evaluated as baselines. Table 3Comparative performance evaluation of Q-CatNet, its ablated CGCNN5 variants, and baseline models for adsorbate energy prediction. Results are reported as MAE (eV), RMSE (eV), and R² (%). Baseline comparisons include both image-based (FTCP45 and graph-based (GINConv46, NNConv47, and SchNet8 architectures. Q-CatNet consistently achieves the highest overall accuracy across all evaluated models, with further improvement observed upon incorporating explicit surface descriptors (Miller indices and surface shift).Adsorbate speciesMetricPlain CGCNNEdge-Enhanced CGCNNFTCPGINConvNNConvSchNetQ-CatNet (XRD)Q-CatNet (XRD + DOS)Q-CatNet (XRD + DOS + Surface)CarbonMAE0.4830.4850.5850.3500.4600.5680.3610.350 0.243 RMSE0.9250.9500.9510.7150.9461.0350.8180.791 0.612 R273.0571.5769.4183.5371.1466.4178.4280.30 87.03 HydrogenMAE0.2900.2580.4060.2440.2760.3220.2090.186 0.141 RMSE0.6040.5320.7000.5010.5530.6310.4990.425 0.385 R249.6060.9134.1560.7152.2733.6861.14 75.07 72.25 NitrogenMAE0.4600.4630.6230.3610.4460.4830.3980.328 0.307 RMSE0.8670.9151.0380.8110.8760.8700.914 0.746 0.847 R277.3174.7169.0379.7676.4177.2774.28 83.18 81.32 OxygenMAE0.4630.4380.5700.3380.4200.5420.3240.318 0.263 RMSE0.8820.8350.9230.6990.8210.9940.7180.706 0.678 R275.7178.2770.1883.4277.1565.3282.5384.46 85.05 OverallMAE0.4220.4090.5440.3220.3990.4760.3220.294 0.237 RMSE0.8260.8210.9090.6890.8100.8930.7500.679 0.649 R274.1874.5167.7681.1473.9368.6977.6182.56 83.84 Significant values are in bold. Table 3Comparative performance evaluation of Q-CatNet, its ablated CGCNN5 variants, and baseline models for adsorbate energy prediction. Results are reported as MAE (eV), RMSE (eV), and R² (%). Baseline comparisons include both image-based (FTCP45 and graph-based (GINConv46, NNConv47, and SchNet8 architectures. Q-CatNet consistently achieves the highest overall accuracy across all evaluated models, with further improvement observed upon incorporating explicit surface descriptors (Miller indices and surface shift).Adsorbate speciesMetricPlain CGCNNEdge-Enhanced CGCNNFTCPGINConvNNConvSchNetQ-CatNet (XRD)Q-CatNet (XRD + DOS)Q-CatNet (XRD + DOS + Surface)CarbonMAE0.4830.4850.5850.3500.4600.5680.3610.350 0.243 RMSE0.9250.9500.9510.7150.9461.0350.8180.791 0.612 R273.0571.5769.4183.5371.1466.4178.4280.30 87.03 HydrogenMAE0.2900.2580.4060.2440.2760.3220.2090.186 0.141 RMSE0.6040.5320.7000.5010.5530.6310.4990.425 0.385 R249.6060.9134.1560.7152.2733.6861.14 75.07 72.25 NitrogenMAE0.4600.4630.6230.3610.4460.4830.3980.328 0.307 RMSE0.8670.9151.0380.8110.8760.8700.914 0.746 0.847 R277.3174.7169.0379.7676.4177.2774.28 83.18 81.32 OxygenMAE0.4630.4380.5700.3380.4200.5420.3240.318 0.263 RMSE0.8820.8350.9230.6990.8210.9940.7180.706 0.678 R275.7178.2770.1883.4277.1565.3282.5384.46 85.05 OverallMAE0.4220.4090.5440.3220.3990.4760.3220.294 0.237 RMSE0.8260.8210.9090.6890.8100.8930.7500.679 0.649 R274.1874.5167.7681.1473.9368.6977.6182.56 83.84 Significant values are in bold. Comparative performance evaluation of Q-CatNet, its ablated CGCNN5 variants, and baseline models for adsorbate energy prediction. Results are reported as MAE (eV), RMSE (eV), and R² (%). Baseline comparisons include both image-based (FTCP45 and graph-based (GINConv46, NNConv47, and SchNet8 architectures. Q-CatNet consistently achieves the highest overall accuracy across all evaluated models, with further improvement observed upon incorporating explicit surface descriptors (Miller indices and surface shift). Significant values are in bold. Fig. 8Overall performance evaluation and reliability assessment of the Q-CatNet model: (A) Comparative prediction analysis (MAE) across ablation variants and baseline architectures, highlighting the impact of the developed partially quantum-informed feature integration; (B) Five-fold cross-validation results based on an 80/20 train–test split, demonstrating low inter-fold variability and confirming the model’s robustness, stability, and consistency across different catalyst–adsorbate data partitions. Overall performance evaluation and reliability assessment of the Q-CatNet model: (A) Comparative prediction analysis (MAE) across ablation variants and baseline architectures, highlighting the impact of the developed partially quantum-informed feature integration; (B) Five-fold cross-validation results based on an 80/20 train–test split, demonstrating low inter-fold variability and confirming the model’s robustness, stability, and consistency across different catalyst–adsorbate data partitions. Standard uncertainty quantification and residual error analysis The present study conducts standard uncertainty quantification and residual error analysis to assess the predictive reliability and systematic behavior of the Q-CatNet model. As illustrated in Fig. 8(B), results from a five-fold cross-validation demonstrate the model’s statistical robustness against variational uncertainties in data partitioning. Across all folds, Q-CatNet exhibited low prediction variability, indicating strong resilience to sampling bias. The MAE, RMSE, and R² showed minimal inter-fold fluctuation, with standard deviations of ± 0.010 eV, ± 0.025 eV, and ± 1.42%, respectively. These narrow deviations confirm that the model’s performance remains stable across adsorbate–catalyst compositions, reflecting reliable feature generalization and consistent learning behavior. Considering the residual analysis, Fig. 9 displays residual error plots for single-atoms adsorbate energy predictions on bulk crystal surfaces. As shown in Fig. 9(A), the overall residuals, defined as the difference between predicted and true adsorption energies (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{\widehat{E}}_{ads}$$\end{document} − \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{E}_{ads}$$\end{document}), are largely centered around zero, suggesting that the model achieves good global calibration across the test set. The approximately symmetric spread of residuals along the horizontal axis indicates the absence of severe systematic bias. That is to say, neither consistent overestimation nor underestimation dominates the predictions. However, the presence of moderate scatter at the extremes of the energy range implies that prediction uncertainty increases for strongly bound or weakly bound species, a behavior commonly observed when the training data distribution is uneven across adsorption strength regimes. To better understand element-specific model behavior, the residuals were decomposed by adsorbate identity, as shown in Fig. 9(B). Each elemental subset (i.e., C*, H*, N*, and O*) exhibits distinct dispersion patterns that reflect both the chemical diversity of bonding environments and the model’s representational sensitivity to different adsorbate–substrate interactions. The residuals for carbon, nitrogen, and oxygen adsorbates display slightly broader scatter and mild skews. By contrast, hydrogen adsorbates show narrower distributions with more symmetric error profiles, indicative of higher predictive stability and better local calibration. This element-resolved analysis highlights that the model’s performance is not uniform across chemical classes and that localized biases may arise from imbalances in chemical environments or surface coordination motifs present in the training data. The statistical distributions of the residuals, presented in Fig. 9(C), further confirm these observations. The histograms for all four adsorbate species are approximately Gaussian but exhibit heavy tails compared to a perfect normal distribution, reflecting the presence of a limited number of large-error cases. These tails correspond to rare adsorption configurations that are either underrepresented in the training dataset or exhibit complex surface relaxations not fully captured by the graph representation. The central peak around zero residual supports the reliability of the model’s mean predictions, while the broadening of the wings quantifies the epistemic uncertainty associated with out-of-domain configurations. Overall, the residual and uncertainty analyses demonstrate that the Q-CatNet framework sufficiently captures the essential physics of adsorbate binding across diverse chemical species, while also identifying interpretable regions of higher predictive uncertainty. These diagnostics are crucial for deploying such models in high-throughput catalyst screening, where understanding both predictive accuracy and reliability directly influences the confidence of materials discovery outcomes. Experimental relevance and learning capacity of Q-CatNet for novel surface chemistries The development of Q-CatNet represents a significant advancement in the integration of quantum-informed implicit representations within graph neural network (GNN) architectures. By embedding physically interpretable edge and global state features, Q-CatNet demonstrates consistent superiority over traditional and invariant graph-based models, including CGCNN, SchNet, GINConv, and NNConv, as well as the image-based FTCP framework. Its success highlights not only the importance of improved featurization but also the necessity of physics-driven representations for achieving transferable, data-efficient, and chemically meaningful predictions in heterogeneous catalysis. Although Q-CatNet’s best-performing configuration achieves an overall MAE of 0.237 eV from the Q-CatNet-XRD-DOS-surface evaluation, which is less accurate when compared to the targeted threshold for general chemical accuracy estimated at 0.043 eV/atom28, it still represents a marked improvement over comparable machine learning surrogate for direct adsorbate energy prediction. Importantly, this improvement in performance is suggested to be interpreted within the inherent complexity of heterogeneous catalysis, where even DFT itself can vary by 0.1–0.3 eV depending on the pseudopotential functional choice or surface termination criteria49. Thus, Q-CatNet’s results fall well within the practical uncertainty bounds of first-principles computation, positioning it as an efficient surrogate for DFT-level accuracy. Besides, the model’s ability to efficiently perform across chemically diverse adsorbates and surfaces indicates substantial progress toward prior data-driven approximations of quantum behavior. Moreover, unlike existing GNNs that employ generic or statistical global state embeddings, such as MEGNet6, Q-CatNet integrates quantum-informed implicit descriptors that are definitively tied to the underlying physical phenomena governing adsorption processes. The inclusion of Coulombic-derived models, density-of-states (DOS) and X-ray diffraction (XRD) signals provides a tripartite representation of the catalyst’s electronic and crystallographic environments, effectively linking local chemical bonding with long-range structural periodicity. This approach allows Q-CatNet to capture both the site-specific reactivity (through enhanced edge features like bond length, bond angle, electronegativity difference, Coulombic, Ewald-sum and Sine matrices) and the global system behavior (through DOS/XRD/surface features). This synergistic coupling of local and global physics contrasts with MEGNet’s more abstract global state vector, which, while effective for cohesive energy or formation energy prediction, often lacks the physical specificity required for complex surface-adsorbate interactions. Q-CatNet’s feature construction bridges this gap, yielding consistently lower MAEs across both single-atom and multi-atom adsorption systems. Besides, a notable strength of Q-CatNet lies in its architectural flexibility when global features are unavailable or partially missing. As demonstrated in the ablation study, a variant of Q-CatNet relying solely on XRD-derived global attributes retained an MAE of 0.322 eV, only ~ 9.5% less accurate than the primary model incorporating both DOS and XRD, and ~ 35.8% to the full configuraiton that includes surface features. This demonstrates that the model can operate under practical conditions where DOS data, typically obtained from DFT calculations, may be computationally expensive or noisy. XRD patterns, by contrast, can be easily derived analytically from reciprocal lattice transformations or experimentally measured with high fidelity. Thus, Q-CatNet can adaptively degrade to a “simplified physics mode,” maintaining predictive reliability while significantly reducing input data requirements. Fig. 9Residual error for single-atoms adsorbate energy predictions on bulk crystal surfaces: (A) Overall residual error distribution for the Q-CatNet model on the test set. The scatter plot shows the residuals, defined as the difference between predicted (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{\widehat{E}}_{ads}$$\end{document}) and true adsorption energies (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{E}_{ads}$$\end{document}), plotted against the true binding energies for all adsorbate–surface systems; (B) Element-resolved residual error analysis for individual adsorbate species (C, H*, N*, O*). Carbon and nitrogen species exhibit slightly broader and negatively skewed distributions, suggesting mild underprediction for strongly bound configurations, while hydrogen and oxygen adsorbates show narrower and more symmetric residual distributions, reflecting improved local calibration and stability; (C) Histograms of residual errors for each adsorbate species with Gaussian kernel fits. The residual distributions for C*, H*, N*, and O* are approximately centered around zero but exhibit noticeable heavy tails, indicating a small number of large-error cases. Residual error for single-atoms adsorbate energy predictions on bulk crystal surfaces: (A) Overall residual error distribution for the Q-CatNet model on the test set. The scatter plot shows the residuals, defined as the difference between predicted (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{\widehat{E}}_{ads}$$\end{document}) and true adsorption energies (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varDelta\:{E}_{ads}$$\end{document}), plotted against the true binding energies for all adsorbate–surface systems; (B) Element-resolved residual error analysis for individual adsorbate species (C, H*, N*, O*). Carbon and nitrogen species exhibit slightly broader and negatively skewed distributions, suggesting mild underprediction for strongly bound configurations, while hydrogen and oxygen adsorbates show narrower and more symmetric residual distributions, reflecting improved local calibration and stability; (C) Histograms of residual errors for each adsorbate species with Gaussian kernel fits. The residual distributions for C*, H*, N*, and O* are approximately centered around zero but exhibit noticeable heavy tails, indicating a small number of large-error cases. Figure 10 provides a comprehensive visualization of the chemical diversity and surface composition distribution within the test set, highlighting potential novel systems that the Q-CatNet model can reliably predict with near-zero adsorption energy error. Specifically, among the 3,745 evaluated single-atom adsorbate-catalyst configurations, approximately 46% of binary and 11% of ternary compositions are predicted within an absolute error threshold of 0.1 eV relative to DFT-calculated values. Notably, about 8% of binary and 2% of ternary systems achieve exceptional precision, with prediction errors below 0.01 eV, demonstrating the model’s strong performance across varied chemistries. This observation further reinforces Q-CatNet’s capacity for accurately learning the underlying quantum–chemical relationships governing adsorption phenomena, enabling confident extrapolation to unseen materials and composition spaces. Supplementary data provide representative examples of surfaces for which Q-CatNet achieved perfect agreement with DFT values across multiple adsorbate species. In summary, Q-CatNet establishes a new paradigm in graph-based catalysis modeling by embedding physically meaningful descriptors into a scalable deep learning framework. It effectively bridges the gap between interpretable physics and data-driven performance, demonstrating robustness, adaptability, and extensibility to inverse design tasks. Beyond its current role as a predictive tool, Q-CatNet holds transformative potential for guiding autonomous discovery systems capable of learning, screening, and designing catalytic materials with reduced computational overhead and enhanced interpretability, advancing the broader goal of physics-informed artificial intelligence for materials innovation. Fig. 10Elemental distribution of catalyst surfaces in single-atom adsorbate evaluation set accurately predicted by Q-CatNet within 0.01 eV in absolute error relative to DFT-calculated values. The figure distinguishes (left) binary compositions (302 unique surfaces) and (right) ternary compositions (67 unique surfaces), showing the counts of unique chemical elements involved in each class. Elemental distribution of catalyst surfaces in single-atom adsorbate evaluation set accurately predicted by Q-CatNet within 0.01 eV in absolute error relative to DFT-calculated values. The figure distinguishes (left) binary compositions (302 unique surfaces) and (right) ternary compositions (67 unique surfaces), showing the counts of unique chemical elements involved in each class. Limitations and opportunities for improvement of the Q-CatNet framework Despite its strong performance in direct adsorption energy prediction, Q-CatNet exhibits several structural and methodological limitations that warrant discussion. First, the architecture is inherently based on the invariant CGCNN framework and, therefore, lacks explicit E(3), SE(3), or SO(3) equivariance. In contrast, recent state-of-the-art models such as GemNet21,22, EquiformerV223,24, MACE25,26, and UMA27 incorporate rotational and translational equivariance directly into their message-passing schemes, enabling improved modeling of geometric transformations and force consistency. This explains why the present study initially considered benchmarking the performance of the Q-CatNet only against similar invariant models (i.e., GINConv46, NNConv47 and SchNet8 in the Table 3) for establishing a fair comparison. Besides, the aforementioned equivariant models have been extensively evaluated on the OC20 benchmark dataset using standardized splits, particularly the initial structure-to-relaxed energy (IS2RE) and structure-to-energy-and-forces (S2EF) tasks. For example, GemNet-OC22 reports a best MAE of 0.348 eV on the IS2RE task, while EquiformerV2 achieves an MAE of 0.309 eV. Notably, both models are trained primarily on S2EF splits that include force supervision, whereas Q-CatNet is trained using a simplified adsorption construction procedure based on direct attachment of adsorbates to bulk conventional unit cells without explicit force learning. Despite these methodological differences, Q-CatNet still achieves an MAE of 0.237 eV on the evaluated OC20 subset, suggesting that its performance is considerably competitive with advanced equivariant approaches. However, a rigorous comparison on the full OC20 dataset would be required to conclusively establish relative robustness and scalability. Furthermore, UMA, currently one of the most advanced models for adsorption prediction, follows a fundamentally different paradigm based on total energy decomposition. Unlike direct adsorption predictors, UMA estimates total energies of the combined slab–adsorbate system, the clean slab, and the isolated adsorbate, and subsequently computes adsorption energies through energy differences according to Eq. 1. This strategy has achieved remarkable reported performance, including MAE values as low as 0.0324 eV for in-domain adsorption energy prediction on the S2EF task. However, because UMA optimizes total energies rather than directly regressing adsorption thermodynamics, and relies on force-supervised training with surface reconstructions, it cannot be directly compared to Q-CatNet in this context. On the one hand, total‑energy models inherently describe surface reconstruction effects more accurately, and as such, they are expected to outperform models that regress adsorption energies directly22,50. On the other hand, Q-CatNet is explicitly designed as a direct, low-cost, and high-throughput adsorption energy predictor without the need for intermediate total energy decomposition nor force supervision. To further assess architectural portability, the current research implemented simplified versions of GemNet-T, EquiformerV2, and MACE using the same processed dataset developed for Q-CatNet, and trained under identical data preparation conditions and hardware resources. Upon training using the bulk-adsorbate configuration dataset, as preprocessed in this study and without S2EF supervision, these models yielded reduced performance, with R² values in the range of 50–60%, suggesting that equivariant architectures optimized for force-aware S2EF tasks may not directly transfer to simplified adsorption-only datasets without tailored training strategies. This observation is an important subject of our future studies, as the absence of equivariance in Q-CatNet also presents a great opportunity for improvement. The electrostatic descriptors and global-state features introduced in this work, particularly Coulombic interactions, periodic Ewald representations, XRD patterns, and DOS embeddings, encode long-range periodicity and electronic structure information that are not explicitly leveraged in many existing equivariant architectures. A promising direction for future research would involve integrating these physics-informed global descriptors into an SE(3)-equivariant backbone, thereby combining geometric tensor consistency with electronic and periodic structure awareness. Such a hybrid framework could potentially enhance long-range interaction modeling while preserving rotational equivariance, ultimately improving predictive accuracy and generalization across chemically diverse catalyst systems. In the Supplementary, implementation details of the equivariant models and architectural configurations under study are provided for reproducibility. Finally, as demonstrated in the present study, the performance of Q-CatNet is evaluated on a curated subset of the OC20 dataset, which is tailored to direct adsorption energy prediction for gas–intermetallic surface systems. While this enables controlled and consistent comparison across models trained under identical conditions, it does not constitute a full benchmarking against other community-wide tasks such as the complete OC20, OC2251, or CatBench52 evaluations. Therefore, the reported improvements should be interpreted within the context of this specific modeling objective, rather than as a universal advancement over all existing approaches. In particular, Q-CatNet demonstrates improved performance relative to baseline models for direct structure-to-adsorption-energy prediction without force supervision, but does not aim to surpass state-of-the-art equivariant models optimized for force-aware or total-energy-based benchmarks in its present form. Future work will focus on extending the evaluation to broader standardized datasets to further assess generalization and comparative performance. Conclusion In this work, we introduced Q-CatNet, a partially Quantum-informed Catalysis Network model designed to accelerate heterogeneous catalyst simulation and discovery by enhancing direct adsorbate energy prediction for single-atom and molecular adsorbate-bulk systems. Q-CatNet extends the traditional Crystal Graph Convolutional Neural Network (CGCNN) by incorporating: enriched edge descriptors including key Coulomb, Ewald-sum and Sine matrix representations; global state features derived from density of states (DOS) and X-ray diffraction (XRD) data; as well as auxiliary surface specific attributes related to Miller indices ( hkl planes) and surface shifting parameters. This combination enables Q-CatNet to capture both local bonding interactions and broader physical correlations across the system. Through systematic ablation studies, the study demonstrates the distinct contributions of integral edge and global feature components, highlighting that while edge-enhanced attributes alone provide marginal improvements, their synergistic integration with global state attributes significantly boosts predictive performance. Comparative benchmarking against baseline models further validated the robustness of Q-CatNet. Notably, Q-CatNet (XRD + DOS) outperformed the image-based Fourier Transformed Crystal Property (FTCP) representation by 46%, and surpassed well established invariant graph-based models with improvements spanning 8% to 38% in terms of prediction accuracy. These results reinforce the strength of graph-based architectures over image-based counterparts for materials informatics tasks, especially due to their inherent invariance to rotation, translation, and permutation as well as their ability to efficiently capture long-range periodic boundary conditions. The reported performance gains should therefore be interpreted within the specific scope of this study, which focuses on direct adsorption energy prediction for gas–intermetallic surface systems using a curated subset of the Open Catalyst (OC20) dataset. This controlled setup enables fair and consistent comparisons across models trained under identical data and preprocessing conditions. However, it does not represent a comprehensive evaluation against standardized community benchmarks or the full diversity of catalytic systems, and thus the conclusions should be viewed within this defined modeling context. Overall, Q-CatNet establishes itself as a powerful forward-design tool, offering improved accuracy, physical consistency, and scalability for catalysis research. By providing a reliable and low-cost framework for directly predicting adsorbate energies, it paves the way for more efficient screening, simulation, and rational discovery of heterogeneous catalysts. In future work, new strategies will be explored to extend Q-CatNet beyond the current controlled setting by integrating SE(3)-equivariant backbone architectures with the long-range electrostatic and electronic descriptors introduced in this study. This direction is intended not only to unite geometric symmetry preservation with improved periodic and electronic structure awareness, but also to enable systematic evaluation and improved generalization across broader, standardized community benchmarks that lie beyond the present study’s scope. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1