-
Molecular potential energy is a physical quantity that describes the potential energy of interactions among atoms in a molecular system. Molecular potential captures various interactions among atoms within the molecular system—such as covalent bonds, electrostatic interactions, and van der Waals forces—thereby accurately depicting the energy and dynamic behavior of molecules. It serves as a critical foundation for molecular dynamics simulations, molecular structure optimization, and molecular property predictions. Its applications span molecular dynamics simulations, chemical reaction kinetics, drug design, and materials design. Constructing high-precision, high-efficiency molecular potentials is of significant theoretical importance and practical application value for accelerating scientific research and technological innovation.
Traditional molecular potential calculations typically rely on classical force fields [1] or quantum chemistry methods [2]. Classical force fields are based on chemical knowledge, predefining atomic types and the positional relationships among atoms, such as bond angles, bond lengths, and dihedral angles, in order to reduce the degrees of freedom of the potential energy surface (PES) of the system [3]. However, this approach also limits the accuracy of the simulations. Quantum chemistry methods, such as density functional theory (DFT), calculate the distribution of each electron and perform computations at the molecular level to determine the total energy of the molecular system. Compared to classical force fields, quantum chemistry methods are more accurate but come with extremely high computational costs, making them difficult to be applied to large-scale systems or long-time-scale molecular dynamics simulations.
In recent years, with the rapid development of machine learning technologies, Behler and Parrinello pioneered a high-dimensional potential energy surface representation method based on deep neural networks [4], achieving large-scale molecular simulations driven by multi-layer perceptrons (MLP) for the first time. GPUMD was proposed by Fan et al. [5], which is a toolkit for building precise machine-learned potentials and conducting highly efficient atomistic simulations. However, this method relies on predefined descriptors that are often designed for specific systems, which limits its generalizability. The rise of graph neural networks (GNNs) [6] has provided new insights for constructing molecular potential functions. GNN approaches can be further divided into invariant GNNs and equivariant GNNs. The development of invariant GNN methods has seen contributions such as the deep tensor neural network proposed by Schütt et al. [7], which introduced continuous filter convolutions and led to the development of SchNet [8], achieving rotation-invariant energy predictions and smooth, differentiable potential energy surfaces. Additionally, Gilmer et al. [9] proposed message passing neural networks (MPNN), which abstract current research on GNNs into a general framework for processing graph data. Its flexibility and scalability have made it one of the foundational models in the GNN field, with message-passing design being critical for learning graph representations [10]. Since its inception, various GNNs with different aggregation and update functions have been developed for expressive graph representation learning [11−13]. Initially, models such as PhysNet [14] proposed by Unke et al. only utilized distance information from geometric data, while Gasteiger et al. [15] introduced angle information into GNN models with DimeNet, marking a significant advancement. It is evident that the development trajectory of invariant GNN methods primarily revolves around progressively increasing the geometric information of the systems to enhance prediction accuracy.
Schütt et al. [16] introduced the concept of vectors in their pioneering work on PaiNN, dividing message passing into routes carried by scalar geometric information (distances) and routes carried out by vector geometric information (directions). This approach further transformed the entire network architecture into an equivariant network, significantly improving the prediction accuracy for properties such as forces and thermal conductivity, marking the entry into the era of equivariant GNNs. Satorras et al. [17] proposed E-GNN, which eliminated the high computational costs associated with high-order tensors and introduced the E(n) group, designing an equivariant message passing mechanism that ensures the message passing and aggregation operations remain equivariant under geometric transformations. This significantly enhanced the model’s robustness and generalization capabilities. Building on this, Brandstetter et al. [18] designed an improved message passing mechanism that explicitly incorporates geometric quantities such as angles and interatomic distances, along with physical quantities like energy and force, thereby enhancing the model’s feature extraction capability and prediction accuracy. Wang et al. [19] proposed an equivariant message passing mechanism that allows for the interaction and updating of vector and scalar features, facilitating more efficient feature extraction while simplifying the computationally intensive Clebsch-Gordan product, enabling the network to efficiently extract geometric features at a lower computational cost. However, previous molecular potential prediction models still face challenges related to insufficient feature extraction and inadequate function fitting accuracy. Further improving the precision of molecular potential energy predictions remains a key focus in the research of these models. Meanwhile, Liu et al. [20] recently introduced the Kolmogorov−Arnold network (KAN), which has gained widespread recognition for its high accuracy in small-scale AI+Science tasks. This network serves as an alternative to MLP, noted for its improved performance and interpretability. Unlike MLP, KAN applies activation functions to the connections between nodes, making these activation functions learnable, which enhances KAN’s ability to approximate complex functions. Furthermore, the emergence of attention mechanisms [21] has profoundly influenced various fields within machine learning [22−26]. The self-attention mechanism in Transformer effectively captures global information and long-range dependencies. In 2017, Veličković et al. [27] proposed the graph attention network (GAT), a GNN based on attention mechanisms. GAT integrates the correlations between vertex features more effectively into the model, resulting in significant performance improvements. Many studies have demonstrated that applying graph attention mechanisms to molecular graphs can effectively extract interactions between atoms [28−31].
In this work, we propose a novel graph neural network for efficient prediction of molecular potential energy, based on attention mechanisms and Kolmogorov−Arnold (RAKAN). To address the issue of insufficient feature extraction, inspired by the graph attention mechanism [27], we introduce the edge-featured graph attention (EGAT) module, which incorporates edge features that significantly influence molecular potential energy. This allows the model to more accurately capture interactions between atoms and facilitates more efficient feature extraction. Moreover, since we extract features from the central atom within a truncated radius, we overlook the fact that interactions exist among all atoms, albeit these interaction might be weak for distant atoms. To extract features more accurately without increasing computational costs, inspired by the Transformer [21], we propose the Distance Aware Transformer (DTF) module. In this module, we use interatomic distances as an attention bias when calculating attention coefficients, enabling DTF to consider both atomic features and their spatial relationships simultaneously. This enhances the model’s ability to capture global information and obtain implicit dependencies between nodes. To tackle the issue of inadequate function fitting accuracy, we design a KAN-gated equivariant module that leverages the KAN [20] to better approximate complex functions, thereby yielding more precise molecular potentials. RAKAN surpasses the existing state-of-the-art approaches on the MD17 benchmark dataset and demonstrates outstanding performance on the QM9 dataset. This indicates that RAKAN not only predicts molecular potential energy with higher accuracy but also possesses exceptionally strong feature extraction capabilities. The contributions of RAKAN can be summarized as follows: (i) the introduction of the EGAT module, which utilizes both edge features and atomic features to capture the relationships between atoms, resulting in more accurate representations of atomic interactions; (ii) the proposal of the DTF module, which uses interatomic distances as biases to adjust attention scores, affecting all atoms rather than just those within a truncated distance, thus acquiring global information; (iii) the design of the KAN-gated equivariant module, which allows for more precise function fitting.
-
In this section, we describe the framework of RAKAN, which is an equivariant GNN based on the MPNN framework. RAKAN predicts potential energy, atomic forces, or other quantum chemical properties by using the system’s atoms and atomic coordinates as inputs. As shown in FIG. 1, the model consists of an embedding layer, multiple RAB layers, and an output layer.
-
Behler et al. [4, 32] proposed the localization approximation, which posits that the interactions between atoms are primarily determined by the localized chemical environment surrounding the central atom, rather than by globally considering the influence of all other atoms in the system on that central atom. This localization approximation reduces computational costs on one hand and also prepares the necessary groundwork to ensure the model’s transferability on the other hand. The localized chemical environment refers to the total interaction of an atom with its neighboring atoms within a given cutoff radius
$ r_{\text{cut}} $ . In this work, the cutoff is set to 5 Å, and a cosine cutoff is typically used to avoid discontinuities. The function$f_\text{c}(r_{ij})$ is defined as follows:where
$ r_{ij} $ represents the distance between atoms$ i $ and$ j $ , and$ r_{\text{cut}} $ is the cutoff radius.where
$\beta_k$ and$\mu_k$ are the parameters for the center and width of the radial basis function$k$ . The value of$\mu$ is the equidistant value between$\exp(-r_{ij})$ and 1, while$\beta$ is defined as$\left [2k^{-1}(1 - \exp(-r_{\text{cut}}))\right]^{-2}$ [14].The embedding of
$Z$ is given by:The features of atom
$i$ are updated as follows:Then we obtain the output atomic features
$X^0$ of the embedding layer, where$N(i)$ denotes the set of neighbors of atom$i$ , and$\odot$ represents the Hadamard product. This means we first perform embedding on the atomic nodes, and then aggregate the features of the neighboring nodes of atom$i$ .The edge features
$e_{ij}^0$ are computed as follows:Since the features of the edge are related to the atoms, the symmetric message is constructed using
$x_i + x_j$ . -
We first employed an improved graph attention mechanism, EGAT, which incorporates edge features into the calculation of attention weights for an initial update of atomic features. Then, we performed feature updates through both vector and scalar message passing, incorporating a residual mechanism [33] to prevent the degradation problem that can occur during the training process as the depth of the neural network increases.
Previous graph attention mechanisms only focused on node features and neglected edge features. However, it is well known that in molecular potential energy prediction, information such as the length of edges has a significant impact on the molecular potential energy. If there are no edge features, the GAT only focuses on the relationships between atoms. From the perspective of a central atom, different surrounding atoms appear to be at the same distance and connected in the same way to the central atom, which is clearly unreasonable. Edge features include inter-atomic distances and bond types, and may even encompass learned interaction features. These features not only capture the spatial structural information of atoms but also provide important information for calculating molecular potential energy. Therefore, considering the interactions between atoms and the special characteristics of molecular structures, we incorporate edge features into the GAT. Furthermore, to learn richer representations, we employed a multi-head attention mechanism and set the number of heads to 8.
We now describe the implementation details of EGAT. First, the attention coefficients are computed. For a vertex
$ i $ , the similarity coefficients between$ i $ and each of its neighbors$ j \in N_i $ , as well as the edge connecting them and the vertex itself, are calculated individually. Specifically, a linear transformation is applied to both atom features and edge features to increase their dimensionality, which is a common feature augmentation technique, defined aswhere
$ [\cdot \parallel \cdot] $ denotes the concatenation of the transformed features of vertices$ i $ and$ j $ , and$a[\cdot]$ is a function that maps the concatenated high-dimensional features to a scalar value. This mapping is implemented by a fully connected layer.$W_x^k$ is a learnable weight matrix, where$k$ denotes the$k$ -th head. The subscript$x$ in$W_x^k$ indicates that it is the weight matrix for node features, while$W_e^k$ represents the weight matrix for edge features.After obtaining the raw similarity coefficients
$ \alpha_{ij}^k $ , they are normalized to produce the attention coefficients$ \beta_{ij}^k $ using the softmax function. It is given byFinally, with the attention coefficients computed, a weighted aggregation is performed over the neighbor nodes’ features, and the outputs from all attention heads are concatenated to form the updated node features
$ x_i' $ . It is given bywhere
$ \sigma $ represents the activation function, and the ELU (exponential linear unit) function is used in this work,$\parallel$ denotes concatenation.To obtain a more expressive representation of the local environment, neurons need not be scalars; they can also be geometric objects such as vectors [34−37]. While previous studies [27] primarily focused on updating node features, inspired by the work of Wang et al. [19], we adopt a vector-scalar interaction message passing mechanism to propagate and update both vector and scalar features. The key operations are given as follows:
where
$x_i$ denotes the scalar features of atom$i$ , and$e_{ij}$ represents the edge features between atoms$i$ and$j$ .$\vec{v}_i$ denotes the vector features of atom$i$ , which are initialized as the zero vector$\vec{0}$ . The superscript$l \in \{0,1,2,\ldots,L\}$ indicates the index of the RAB layer.$\vec{r}_{ij}$ represents the vector from atom$i$ to atom$j$ .$\mathrm{Rej}_{\vec{r}_{ij}}(\vec{v}_i^l)$ denotes the component of$\vec{v}_i^l$ perpendicular to$\vec{r}_{ij}$ , referred to as vector rejection. The symbol$\langle \cdot, \cdot \rangle$ denotes the inner product.$M_s^l$ and$M_v^l$ are the scalar and vector message functions, respectively, while$U_s^l$ ,$U_v^l$ , and$U_e^l$ denote the update functions for scalar features, vector features, and edges, respectively. Eqs. (9) and (11) describe the message passing and aggregation processes, whereas Eqs. (10), (12), and (13) illustrate the update procedures. Through this equivariant message passing neural network (MPNN), the node features$ X $ , edge features$ E $ , and vector features$ \vec{V} $ are updated effectively. -
Transformer [21] is a deep learning model based on the self-attention mechanism and multi-head attention mechanism. It was originally proposed by Vaswani et al. in 2017 for natural language processing (NLP) tasks such as machine translation. The Transformer model abandons traditional recurrent neural networks (RNNs) and convolutional neural networks (CNNs), relying entirely on the attention mechanism to capture global dependencies within input sequences, thereby achieving significant improvements in both performance and efficiency. Transformer model has demonstrated remarkable success across multiple fields. There have also been efforts to adapt the Transformer architecture to graph data [38]. The standard Transformer is designed to learn from sequential data such as text [21]. A primary challenge in applying Transformers to graphs lies in how to encode structural information of the graph. To address this assue, Ying et al. proposed Graphormer [39], based on the standard Transformer, which can effectively encode structural information.
To balance computational speed and accuracy, previous works have defined the local chemical environment of a central atom by the neighboring atoms within a cutoff radius, i.e., local features. As shown in FIG. 2, the larger the cutoff radius is, the closer the local chemical environment of the central atom approaches its true environment. The computations between atoms thereby become more complete and accurate. However, a larger cutoff radius implies higher computational cost, especially because the fully connected calculations among atoms within the radius grow geometrically. In the self-attention mechanism of Transformer [21], each input element interacts with all other elements to compute attention weights, which represent the importance of each element to the others. This mechanism enables automatic contextual focus, capturing global information and long-range dependencies.
The key to applying the Transformer on graphs is the effective encoding of the graph’s structural information into the model, because the self-attention mechanism only computes the semantic similarity between nodes without considering the graph’s structural information. Inspired by Ying et al. [39], who were the first to propose incorporating the shortest-path spatial encoding and edge features along the shortest path into the attention scores; considering that in molecular systems, interactions between atoms are strongly influenced by distance; we incorporate inter-atomic distances as attention biases to dynamically adjust the weights of distant and nearby atoms, thereby encoding the graph’s structural information into the model. This allows the DTF model to simultaneously consider atomic features and their spatial relationships, thereby obtaining a more comprehensive global representation. This approach captures not only interactions within the cutoff radius but also long-range interactions.
The DTF module consists of a multi-head self-attention mechanism with an added distance bias (as illustrated in FIG. 3) and a feed-forward network (FFN), employing layer normalization [40] (LayerNorm, LN). The distance encoding shown in the FIG. 3 represents the distance matrix between atoms in the system. Distances between atoms within the cutoff radius are calculated based on their coordinates, while the distances between atoms outside the cutoff radius and the central atom are initialized to zero. The specific implementation details are as follows:
The queries
$ Q $ , keys$ K $ , and values$ V $ of the multi-head self-attention all come from the same input sequence$ X $ . They are respectively transformed by learnable linear projection matrices$ W_Q $ ,$ W_K $ , and$ W_V $ . By computing multiple attention heads in parallel, the model captures information from different subspaces of the input data, thereby enhancing both the expressiveness and generalization ability of the model.First,
$ Q $ ,$ K $ , and$ V $ are linearly projected to obtain the query, key, and value for each attention head, denoted as$ Q_i $ ,$ K_i $ , and$ V_i $ .where subscript
$ i $ indicates the$ i $ -th attention head.Then, the base attention scores
$A_{\text{base}_i}$ for each head are computed via scaled dot-product attention, defined aswhere
$d_k$ is the number of columns of the$Q_i$ and$K_i$ matrices, that is, the dimension of the vectors.Using an MLP (multilayer perceptron) to perform distance projection, then element-wise multiplied by the cutoff function, yielding the distance bias coefficient
$ W_{d_i} $ . It is defined aswhere
$\text{SiLU}(D W_i^1) W_i^2$ is a two-layer MLP, SiLU is the activation function, and$D$ is the distance matrix.$f_\text{c}$ is the cutoff function defined above, and$ \odot $ signifies the hadamard product.The final attention coefficients are obtained by adding the base attention scores and the distance bias, followed by a softmax normalization to produce the attention weights
$ \alpha_i $ :Finally, each attention head output is computed by weighting the value vectors by the attention weights:
The outputs of all
$ h $ attention heads are concatenated and linearly transformed back to the original feature space using a learnable weight matrix$ W^O $ :where
$ \parallel $ denotes concatenation.In this work, the number of attention heads
$ h $ is set to 8, a common choice in Transformer-based models. -
The atomic features
$X^l$ , edge features$e_{ij}^l$ , and vectors$\vec{v}^l$ pass through the ViSA module, respectively yielding residual values$\Delta X^l$ ,$\Delta e_{ij}^l$ , and$\Delta \vec{v}^l$ . Then, the DTF operation is applied to all the atomic features, enabling the capture of implicit relationships among atoms. Additionally, by differentiating the importance of atoms, the training of the network is accelerated.Finally, the updates are performed using residual connections as follows:
where LayerNorm refers to layer normalization. By normalizing the inputs, it stabilizes the input distribution during training, reduces internal covariate shift, and thus contributes to enhanced training stability. After multiple layers of the RAB layer, the feature extraction is completed.
-
The KAN [20] is an emerging neural network architecture based on the Kolmogorov-Arnold theorem. These theorems prove that any continuous multivariate function can be represented as a finite superposition of univariate functions. This discovery provides a solid mathematical foundation for the design of KAN, enabling it to replace the fixed linear weights in traditional neural networks with learnable univariate functions, thereby enhancing the model’s flexibility and interpretability.
In recent years, KAN has attracted widespread attention in machine learning applications, primarily due to its flexibility and interpretability. Unlike traditional multilayer perceptrons (MLPs) which use fixed activation functions, KAN employs learnable univariate functions, the difference between MLP and KAN is illustrated in the FIG. 4). The characteristic of KAN allows it to perform excellently in data fitting and complex function learning tasks. For example, Liu et al. demonstrated the superior performance of KAN in data fitting and solving partial differential equations (PDEs), showing that even smaller KAN models can achieve or surpass the performance of MLPs in these tasks [20, 41, 42]. Inspired by Schütt et al. [16] who proposed gated equivariant modules, we have designed a KAN-gated equivariant module.
The implementation details of the KAN-gated equivariant module are as follows:
In this work, B-spline curves are used as the learnable activation functions in KAN for fitting [20]. After training through two layers, the final
$ X $ is used to predict molecular potential energy. The total potential energy of the molecule is the sum of the potential energies of all atoms in the molecule, and the forces are computed as the negative gradient of the predicted total potential energy with respect to the atomic coordinates:where
$ N $ denotes the total number of atoms within the molecule. -
We evaluated RAKAN on several mainstream benchmark datasets, namely MD17 [7, 43, 44] and QM9 [45], for the prediction of energies, forces, and other molecular properties.
MD17 consists of molecular dynamics (MD) trajectory data of seven small organic molecules, including aspirin, ethanol, malonaldehyde, naphthalene, salicylic acid, toluene, and uracil. Each molecule contains tens of thousands to hundreds of thousands of distinct conformations (snapshots from MD trajectories). This dataset is commonly used for training and evaluating machine learning models on molecular energy and force prediction tasks.
QM9 comprises 133385 small organic molecules with up to nine heavy atoms of types C, N, O, and F, annotated with 12 quantum chemical properties. The dataset reports computed geometries optimized locally, along with corresponding thermodynamic, energetic, and electronic properties.
-
The hardware environment used in the experiments consisted of an Intel(R) Core(TM) i7-14700KF processor running at 3.40 GHz and an NVIDIA GeForce RTX 4080 GPU. The algorithm was developed using Python 3.9 and PyTorch 1.10, with PyTorch-Geometric version 2.1.0. Training was performed using the ADAM optimizer [46].
-
In the experiments, to evaluate the effectiveness of RAKAN on simulated data, we trained RAKAN on the MD17 dataset under a limited data setting. For each small molecule dataset, only 950 uniformly sampled conformations were used for training, 50 conformations for validation, and the remaining conformations were used as the test set. We employed a combined mean squared error (MSE) loss for predicting both energy and forces, with a weight of 0.05 assigned to the energy loss and 0.95 to the force loss. Learning rate decay was applied, with a decay factor set to 0.8. The RAB network was configured with 9 layers, and the embedding dimension was assigned a value of 256.
We compared RAKAN with state-of-the-art algorithms including SchNet [8], DimeNet [15], SpookyNet [47], GemNet [48], NequIP [49], SO3KRATES [50], and ViSNet [19], as shown in Table I. RAKAN achieves the smallest mean absolute error (MAE) in predicting both energies and forces across all molecules, with particularly notable performance in energy prediction. There are 5 molecules with cyclic rigid planar structures and 2 molecules containing only linear flexible structures in the MD17 dataset (see FIG. S1 in supplementary materials, SM). By comparing the prediction results of RAKAN and ViSNet, we find that the improvement in prediction accuracy is greater for molecules with cyclic structures. For example, RAKAN achieves a 56.8% lower MAE than ViSNet in energy prediction for the uracil molecule. In contrast, the improvement in prediction accuracy is relatively smaller for molecules with purely linear structures; for instance, RAKAN’s MAE for energy prediction of the malonaldehyde molecule is 40.0% lower than that of ViSNet. Additionally, Table SI in SM presents the correlation between the energies and forces predicted by RAKAN or ViSNet [19] and the reference values from DFT calculations in the MD17 dataset, reflecting the fitting performance of the models. RAKAN achieves the highest
$R^2$ score.To more intuitively demonstrate the correlation between our model’s predictions and the data from the MD17 dataset, we randomly selected the molecules aspirin and salicylic acid. We compared the forces and energies predicted by our model, RAKAN, and the baseline model ViSNet with the corresponding values from the MD17 dataset to validate the effectiveness of our algorithm, as shown in the FIG. 5. It can be observed that the predicted values from our model have a strong correlation with the reference values in the MD17 dataset. Furthermore, compared with the predictions of ViSNet, our model’s predicted values show smaller deviations from the reference values, indicating superior predictive capability. This also confirms that our model possesses stronger feature extraction and function fitting abilities.
To evaluate the computational efficiency of RAKAN, we measured the inference time on the MD17 test set using an NVIDIA GeForce RTX 4080 GPU. We report the inference time per molecule and compare it with the baseline model (see Table II).
-
To comprehensively evaluate our model’s predictive performance and generalization ability on limited datasets, we selected from the MD17 dataset the molecule with the greatest number of atoms, aspirin (C9H8O4), and the molecule with the fewest atoms, ethanol (C2H6O). We further reduced their training set sizes to 350, 550, and 750 samples, while keeping the validation set size fixed at 50. The rest of the samples were utilized as the test set. These results were compared against our baseline model, ViSNet.
As shown in Table III, on sufficiently small datasets, RAKAN achieved the smallest MAE in predicting both energy and forces. To more intuitively demonstrate RAKAN’s predictive performance, we visualized the prediction errors of ViSNet and RAKAN for aspirin and ethanol with a training set size of 550, relative to the reference values from MD17. As illustrated in the FIG. 6, ViSNet’s errors for aspirin mostly fall within the
$ \pm0.6 $ range, whereas RAKAN’s errors are predominantly confined within$ \pm0.5 $ . Similarly, for ethanol, ViSNet’s errors mainly lie within$ \pm0.3 $ , while RAKAN’s errors are largely controlled within$ \pm0.2 $ .RAKAN’s strong performance under limited data conditions demonstrates its ability to effectively leverage scarce data resources—especially considering the extremely high computational cost of DFT calculations—by capturing critical patterns and intrinsic relationships. This highlights RAKAN’s robust generalization capability and learning efficiency.
-
To further validate the effectiveness of the EGAT, DTF, and KAN-gated equivariant modules, we conducted ablation experiments by respectively removing EGAT, removing DTF, and replacing the KAN-gated equivariant module (KAN-G) with the gate equivariant module [16]. The experimental setup remained consistent, with 950 conformations uniformly sampled used for training, 50 conformations per molecule used for validation, and the remaining conformations designated as the test set. The experiments were performed on the molecule aspirin (C9H8O4) and ethanol (C2H6O). We report both the model’s performance and complexity.
The results, shown in Table IV, indicate that RAKAN achieves the best performance, while the variants exhibit varying degrees of decline. This demonstrates that the EGAT, DTF, and KAN-gated equivariant modules are all beneficial, each contributing to enhanced feature extraction and function fitting capabilities of the model, thereby enabling it to deliver superior performance.
-
To evaluate RAKAN’s molecular feature extraction and function fitting capabilities, we conducted experiments using the QM9 dataset to assess RAKAN’s performance in predicting other quantum chemical properties. Following previous studies [16, 51], we randomly split the dataset into a training set with 110,000 samples and a validation set with 10,000 samples, with the remaining samples used as the test set. We compared RAKAN with state-of-the-art algorithms including DimeNet++ [52], EGNN [17], PaiNN [16], ET [51], PaxNet [53], ComENet [54], SphereNet [34], and ViSNet [19]. As shown in Table V, RAKAN demonstrated the highest performance in 6 of the 12 chemical properties and delivered competitive outcomes for the other properties.
-
In this work, we proposed RAKAN, an efficient graph neural network for predicting molecular potential energy, which is based on attention mechanisms and the Kolmogorov–Arnold representation theorem. RAKAN demonstrated high accuracy and strong generalization ability on the MD17 dataset, as well as excellent performance in predicting other quantum chemical properties. RAKAN provides a more precise tool and a novel approach for molecular potential energy prediction and quantum chemical property estimation. In future research, we aim to investigate the combination of cutting-edge neural network architectures with molecular structures to develop neural networks specifically designed for molecular systems. Our work aims to accelerate scientific research and technological innovation.
RAKAN: An Efficient Residual-Connected Graph Neural Network for Predicting Molecular Potential Energy Based on Attention Mechanism and Kolmogorov−Arnold
- Received Date: 04/06/2025
- Available Online: 27/08/2026
-
Key words:
- Molecular potential energy /
- Geometric deep learning /
- Attention mechanism /
- Kolmogorov−Arnold
Abstract: Molecular potential energy prediction is fundamental to molecular dynamics simulations. Recent advances in geometric deep learning have yielded numerous models that balance speed and accuracy. However, existing approaches often struggle with limited feature extraction and suboptimal function fitting. Enhancing prediction accuracy remains a critical challenge. In this study, we introduce RAKAN, a novel residual-connected graph neural network that leverages attention mechanisms and Kolmogorov−Arnold representation to improve feature extraction and prediction precision. RAKAN surpasses the existing state-of-the-art approaches on the MD17 benchmark dataset and demonstrates superior performance in predicting chemical properties on the QM9 dataset. These results demonstrate RAKAN’s superior accuracy and robust feature learning capabilities, advancing the field of molecular potential energy prediction.
首页
登录
注册


DownLoad: