AI is a branch of computer science aimed at enabling machines to autonomously learn, reason, and make decisions by simulating human intelligence. Machine learning is at the core of AI, using algorithms to allow computers to learn from data and generate predictions or decisions. The development of AI can be traced back to the 1950s, when Alan Turing proposed the "Turing Test" to assess whether machines possess intelligence^[16]. In 1956, at the Dartmouth Conference, John McCarthy first introduced the term "artificial intelligence," marking the official beginning of AI research^[17].

The development of AI in materials design can be divided into several key stages. In the early exploration phase (1980s–1990s), the focus was primarily on expert systems and rule-based reasoning methods. Researchers attempted to use knowledge bases and reasoning mechanisms to assist in materials design; however, due to limitations in computing power and data volume, these efforts largely relied on the experience and knowledge of experts. In the initial data-driven application phase (early 21st century), with the advancement of computer and automation technologies, the field of materials science gradually began accumulating large amounts of experimental data. Machine learning methods started to be introduced for analyzing materials data. In 2012, the rapid development of deep learning brought new opportunities for materials design^[18]. The emergence of technologies such as convolutional neural networks (CNN) and recurrent neural networks (RNN) made it possible to accurately quantify complex structure-property relationships in materials from large-scale datasets. At this stage, AI began demonstrating strong capabilities in materials discovery, optimization, and performance prediction.

MOFs new material design tasks typically involve two stages: (1) MOFs property prediction. End-to-end deep learning models can directly map molecular structures (such as molecular graphs, SMILES representations, and 3D structures) to target properties, making them suitable for handling large-scale data and complex nonlinear relationships. Unlike traditional machine learning, end-to-end deep learning automatically extracts hierarchical features from data through multi-layer neural networks, significantly reducing reliance on domain knowledge^[19]. This allows for the rapid screening of candidate structures with potentially excellent properties and provides initial guidance for subsequent optimization. (2) MOFs structure optimization and screening. At this stage, the incorporation of domain knowledge (such as chemical rules and structural constraints) is crucial. The outputs of end-to-end models may not be directly applicable due to issues such as chemical plausibility (e.g., bond length/angle constraints), synthetic feasibility (e.g., ligand accessibility), or stability concerns^[20]. Therefore, domain-knowledge-driven optimization must be introduced at this stage. End-to-end models efficiently narrow down the search space, while domain knowledge ensures that the final designed structures possess both high performance and practical operability.

The establishment of the MOFs prediction model can be further refined into several steps, including the construction of a standard dataset, model training and evaluation, and finally, model application (Figure 1). First, MOFs-related data are collected from public databases, literature, and experimental results. Next comes the feature engineering, which is the main focus of this paper, involving steps such as descriptor (feature) calculation and feature selection. The entire dataset is then divided into a training set and a test set; the training set is used for parameter tuning and learning the patterns inherent in the data, while the test set is employed to evaluate the model's performance on unseen data, thereby testing its generalization capability. Finally, the accuracy of the model is assessed. By validating the model using the test set, if the evaluation results indicate that the model's predictive accuracy meets the expected standards, it can be adopted as the final model for subsequent practical applications, such as new material screening and assisted design. However, if the model performs poorly on the test set—meaning insufficient predictive accuracy or excessive error—it is necessary to re-examine aspects such as the model structure, feature selection, hyperparameter settings, and make necessary adjustments or rebuild the model.

Deep learning is an important branch of machine learning, based on the concept of artificial neural networks, aiming to simulate the structure and function of the human brain. Deep learning models contain no fewer than 3 hidden layers (typically 5 to 20 layers)^[21], achieving automatic feature extraction through hierarchical nonlinear transformations, whereas traditional machine learning models (also known as shallow learning) usually have 0 to 2 layers or no layer structure at all^[22]. Unlike traditional machine learning methods, deep learning automatically learns features and patterns from large amounts of data through multi-layer neural networks. It excels at handling large-scale and complex datasets, such as images, audio, and natural language processing. As shown in Figure 1, the core advantage of deep learning lies in its powerful feature extraction capability, which allows it to automatically learn high-level feature representations from raw data, thereby reducing reliance on manual feature engineering^[23]. This automated feature learning capability has enabled deep learning to achieve remarkable success in many fields, such as image recognition in computer vision, language translation in natural language processing, and speech recognition.

Traditional machine learning and deep learning are key technologies behind today's data-driven decision-making and intelligent systems, each with its own unique advantages and disadvantages. Models built using traditional machine learning have strong interpretability, typically providing explainable results that facilitate understanding the basis of model decisions. They also offer high computational efficiency, with relatively fast training speeds when dealing with small to medium-sized datasets, making them suitable for resource-constrained environments. Traditional machine learning excels at handling datasets with clear features and patterns, such as tabular data. However, it places high demands on feature selection and preprocessing, requiring expertise from domain specialists; its ability to handle nonlinear, large-scale, and high-dimensional data is limited; and it heavily relies on features, with generalization performance affected by feature quality and selection. In contrast, end-to-end deep learning is well-suited for large-scale data and can effectively capture complex nonlinear relationships, such as those found in images, speech, and natural language^[24]. It also enables automated feature learning, extracting more abstract feature representations directly from raw data and reducing the need for manual feature engineering. Consequently, end-to-end deep learning often achieves higher predictive accuracy in big data and complex tasks. Training deep learning models, however, requires substantial computational resources and time, such as graphics processing unit (GPU) acceleration. The complexity and large number of parameters in deep learning models make their decision-making processes difficult to interpret, resulting in a lack of transparency. Moreover, deep learning relies heavily on large amounts of labeled data, and insufficient data can lead to degraded model performance^[25]. Choosing between traditional machine learning and deep learning requires a comprehensive assessment of factors such as data size, task complexity, and resource availability (Table 1).

Over the past several decades, both experimental and computational studies on MOFs have made remarkable progress, generating a wealth of experimental and simulation data. Table 2reveals the developmental trajectory of this process, starting from early experimental and molecular simulation studies and gradually evolving into cutting-edge methods that employ machine learning techniques to process and analyze these data. In the early 20th century, research on MOF materials was primarily conducted through experimental approaches, including synthesis, characterization, and measurement of adsorption and diffusion properties. During this period, the development of computational tools was relatively slow, and in most cases, experimental data were not systematically collected or organized, thus no corresponding databases were established. With the improvement of computational power and the advancement of computational methods, detailed simulation studies of MOFs gradually emerged, enabling researchers to predict and analyze MOF properties. Through meticulous comparison and calibration of simulation results, simulation methodologies were refined. The first computational datasets and databases for MOFs began to appear, such as the hMOF and CoRE MOF databases^[26-27]. From the beginning of the 21st century to the present, this stage is characterized by the production of large amounts of high-quality data and the widespread application of machine learning methods. Vast quantities of computational and experimental data have been systematically collected and organized, leading to the establishment of several large-scale MOF databases, such as MOF-DB and CSD-MOF databases^[28-29]. The evolution of MOF structural and performance data and databases—from non-existence to abundance—is a result of the combined advancements in automated material synthesis and characterization, enhanced computational capabilities, and improved data processing technologies.

Structural feature descriptors are parameters used to characterize molecular structural features, reflecting geometric, topological, and compositional characteristics of molecules. They are commonly applied in fields such as cheminformatics, drug design, and materials science. Among these, topological descriptors are widely used structural features, primarily focusing on the cage-like structure of MOFs, i.e., the connectivity between metal nodes and organic ligands. Topological descriptors can help researchers understand and predict the physicochemical properties of MOFs, such as stability, porosity, adsorption capacity, and catalytic activity. Some common topological descriptors include: (1) describing the types and quantities of metal nodes and organic ligands in MOFs; (2) characterizing the bonding states of each metal node with organic ligands; and (3) describing the network structure types of MOFs, such as cubic, tetrahedral, and octahedral structures. For example, Batra et al.^[37]developed a comprehensive chemical feature extraction program to obtain information on MOF metal nodes, organic ligand connectivity, and their molar ratios, and employed various machine learning models to predict the water stability of MOFs. Through feature importance analysis, it was found that the number of cyclic divalent nodes or six-membered rings, as well as the number of hydrogen bond acceptor sites, significantly influence water stability. In addition, structural information such as MOF surface area, pore size, and metal type is also frequently used as input features for machine learning models^[38].Structural feature descriptors directly represent the geometric and topological intrinsic properties of MOFs (such as pore size, specific surface area, and topological type), typically obtained directly from crystal structure data or theoretical simulations, offering high data reliability. They are suitable as core features for model input, but their static nature may limit their ability to predict dynamic processes (such as adsorption kinetics).

When constructing machine learning models, the chemical properties of MOFs are crucial for understanding their behavior and performance. In recent years, researchers have designed sets of stoichiometric features to describe MOFs, aiming to predict and optimize MOF performance through machine learning and data-driven approaches. These include Stoichiometric-120 Features, Stoichiometric-45 Features, Sine Coulomb Matrix, and others. These stoichiometric descriptors primarily focus on the molecular composition and relative proportions of elements, typically used to quantitatively characterize the composition and chemical properties of compounds.

Stoichiometric-120 Features are stoichiometric descriptors calculated based on the chemical composition of MOFs, specifically the ratio and type of metal centers and organic ligands^[39]. Each feature in Stoichiometric-120 Features represents a chemical or structural property of MOFs. This descriptor set comprises 120 properties, primarily calculated according to the ratio and type of metal centers (such as Zn, Cu, etc.) and organic ligands (such as benzenedicarboxylic acid, etc.) in MOFs, as well as several other chemical properties: average atomic weight, average group number, average period number, maximum atomic number difference, and average atomic number. Stoichiometric-45 Features include 45 independent descriptors, each representing a chemical or structural property of MOFs^[40]. The Sine Coulomb matrix is a matrix descriptor used to characterize material structures, capturing geometric and electronic properties by considering interactions between atoms in the material, particularly Coulomb interactions^[41]. The Orbital Field Matrix (OFM) descriptor is based on information about atomic orbitals within the material, constructing a matrix to represent the local environment and electronic characteristics of the material, with each matrix element containing information related to specific atomic orbitals^[42]. The Average Smooth Overlap of Atomic Positions (SOAP) descriptor is an advanced machine learning feature used to capture local environmental information of materials and molecules, by calculating the three-dimensional density distribution of the atomic local environment and converting it into a mathematical representation that can be effectively processed and compared, such as vectors or similarity metrics. Using these descriptors and the kernel ridge regression machine learning algorithm, researchers predicted the quantum chemical properties (such as band gap) of over 14,000 experimentally synthesized MOFs from the QMOF database, with all model prediction accuracies exceeding 0.64. This provides an effective solution for avoiding time-consuming and labor-intensive density functional theory calculations in future research^[35]. Recently, Bai et al.^[43]developed a predictive model based on machine learning algorithms and applied it to high-throughput screening of MOF catalysts for carbon dioxide cycloaddition reactions. The descriptors used were easily obtainable structural and physicochemical properties of MOFs selected according to the reaction mechanism, for example, using OMS (Open metal sites) charge to replace the Lewis acidity of catalytic centers. With this model, 239 highly active catalysts were successfully screened, and the catalytic performance of the preferred material MOF-76(Y) was experimentally verified. However, the stoichiometric characteristics of MOFs (such as metal/ligand ratio and elemental composition) must be obtained through chemical analysis or theoretical calculations, and data integrity is closely related to experimental methods.

Thermodynamic property descriptors are primarily used to characterize the energy properties, stability, and reactivity of MOFs, including adsorption energy, binding energy, desorption energy, band gap, and others. These descriptors can quantitatively represent the performance of MOFs in fields such as gas adsorption, catalysis, batteries, and supercapacitors. Bucior et al.^[44]developed a novel energy descriptor based on the interaction between MOFs and guest molecules. Using this descriptor, they constructed a machine learning model capable of predicting the gas adsorption capacity of MOFs across multiple databases, achieving a prediction accuracy within 3 g/L. Furthermore, by applying the constructed model to virtually screen a database containing over 50,000 MOFs, they identified an outstanding candidate material. Under storage conditions of 77 K and 100 bar, and release conditions of 160 K and 5 bar, this material exhibited a hydrogen delivery capacity of 47 g/L (simulated value: 54 g/L), demonstrating the effectiveness of the developed descriptor and the broad application prospects of machine learning-driven new material design. Thermodynamic properties rely on experimental measurements or high-precision calculations (such as density functional theory), and data may contain significant errors or gaps. Data noise or missing values need to be addressed through interpolation, data augmentation, or robust models (such as adversarial training); otherwise, it may lead to a decline in the model's generalization performance.

In machine learning-assisted materials design, feature selection is an important task aimed at identifying the most representative features from a large number of descriptors to build efficient and accurate predictive models^[45]. Using high-dimensional and correlated features in machine learning can lead to increased model complexity, longer training times, and a higher risk of overfitting. There are various methods for feature reduction, including manually examining the Pearson correlation coefficient matrix, as well as automated approaches such as recursive feature addition, recursive feature elimination, univariate feature filtering, and wrapper feature selection. Different methods have their own advantages and disadvantages; selecting an appropriate method can effectively reduce feature dimensionality and improve model performance. For example, Batra et al.^[37]demonstrated that recursive feature elimination can significantly reduce feature dimensionality from 149 to approximately 30 while improving model accuracy. Additionally, dimensionality reduction techniques are widely used to address the complexity of feature spaces, including principal component analysis, t-distributed stochastic neighbor embedding, and uniform manifold approximation and projection. These methods aim to reduce the ratio between feature dimensionality and the number of training samples, making models more efficient. They also effectively identify and remove irrelevant or redundant features, helping to mitigate overfitting^[46].

The missing features in MOF descriptors and the noise in experimental data constitute key bottlenecks that limit the predictive accuracy of machine learning models. Specifically, difficult-to-determine structural features (such as pore size and topological type), noisy thermodynamic properties (such as adsorption heat and band gap), and missing chemical compositions (such as metal/ligand ratios) collectively pose core challenges for model optimization. Machine learning addresses these feature-related challenges through various strategies, including data noise suppression, missing data reconstruction, and multi-source data fusion.

Traditional machine learning methods (such as linear regression and random forests) have achieved certain success in low-dimensional data preprocessing through statistical imputation, feature selection, and regularization techniques. Deep learning also efficiently preprocesses high-dimensional and complex features. First, robust loss functions and uncertainty modeling can reduce the interference of noisy characteristics caused by experimental and computational errors in physicochemical properties on the model^[47]. Second, for the high missing rate of chemical composition descriptors in public databases, deep learning models such as Graph Attention Networks (GATs) demonstrate significant advantages, effectively leveraging existing data information to infer missing values^[48]. Additionally, generative models like Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) can generate samples similar to the real data distribution, which are used to fill in missing values for MOFs' chemical compositions and physicochemical properties. For instance, Choudhary et al.^[49]used ALIGNN (Atomistic Line Graph Neural Network) to predict 29 properties of unlabeled materials. Combining traditional imputation techniques with deep generative features provides a solution for prediction tasks involving high noise and high missing data.

CNN is a deep learning model primarily used for image and video processing. CNN extracts features through convolutional layers, pooling layers, and fully connected layers, and can effectively capture local features and spatial structures in images (Figure 3a). The convolutional layer is the core of CNN, processing image data through convolution operations. A convolutional layer contains multiple learnable filters, each of which identifies different patterns and generates feature maps^[50]. CNN can extract features hierarchically, with the level of feature abstraction increasing as the network deepens. Researchers have proposed a general framework for predicting MOF gas adsorption performance, using the potential energy surface as the sole descriptor and employing CNN to process three-dimensional energy images, enabling efficient prediction of MOFs' CO₂adsorption performance^[51]. Compared to traditional geometric descriptor models, this model exhibits superior performance and requires less training data. Hung et al.^[52]developed a gas adsorption property prediction method based on chemically encoded CNN. This method uses atomic positions and corresponding chemical information to represent MOF structural features. Trained on a dataset of Henry constants for CO₂and CH₄in nearly ten thousand MOF structures obtained from molecular simulations, this CNN model demonstrates excellent accuracy in predicting the adsorption performance of CH₄and CO₂.

RNN is a deep learning model well-suited for handling sequential data, particularly time series and text data. By using hidden states to remember previous information, RNN captures temporal dependencies within sequences. The basic structure of an RNN (Figure 3b) is as follows: (1) Input layer: receives sequential data; (2) Hidden layer: saves past information through recurrent connections and updates the current state; (3) Output layer: generates outputs related to the input sequence. RNN can be extended into various forms, including Long Short-Term Memory networks (LSTM) and Gated Recurrent Units (GRU), both of which can alleviate the vanishing gradient problem in standard RNNs and are suitable for capturing longer-term dependencies. For example, MOF structures can be encoded into specific SMILES representations, where each SMILES is defined by three components: inorganic nodes (metals), organic ligands, and topological structure. When learning to generate new SMILES, the distribution of the next symbol is determined based on predictions from the RNN model, and properties such as gas adsorption capacity and thermodynamic stability of the newly generated MOF materials are evaluated, helping researchers identify high-performance MOF materials^[53]. Recently, Zhang et al.^[54] combined Monte Carlo tree search with an RNN algorithm to build a deep learning model for designing novel MOFs with high adsorption performance for methane and carbon dioxide.

Graph neural networks (GNNs) are deep learning models specifically designed to handle graph-structured data. Unlike CNNs, which can only process conventional Euclidean data, GNNs can operate on non-Euclidean data composed of nodes and edges, modeling local and global relationships between nodes through a message-passing mechanism (Figure 3c). The core idea is to iteratively update node features, so that each node's representation not only includes its own attributes but also aggregates information from neighboring nodes. This feature enables GNNs to effectively capture the complex topological structures and dynamic interactions within graph data. Traditional neural networks perform poorly when dealing with graph data (network structures composed of nodes and edges), as graph data exhibits irregularity and complex connectivity. MOFs have complex and diverse structures, where atoms can be regarded as nodes in a graph, chemical bonds as edges connecting nodes, and weak interactions such as coordination bonds and hydrogen bonds can be encoded through edge weights or types. Conventional molecular modeling methods (such as those based on geometric descriptors) struggle to effectively capture this complexity, whereas GNNs can naturally handle the graph-structured data of MOFs and extract their topological features (e.g., node connection patterns, subgraph structures). In terms of model complexity, GNNs are generally more complex than RNNs. RNNs process sequential data, and their parameter count grows linearly with sequence length; in contrast, GNNs must simultaneously model both node features and edge relationships, resulting in a parameter count that grows polynomially with the number of nodes and edges^[55]. Xie et al.^[56]proposed the Crystal Graph Convolutional Neural Network (CGCNN) model, which represents the structure of MOFs using a crystal graph that encodes atomic information and interatomic bonding interactions, and then constructs a CNN to extract features and accurately predict properties such as formation energy, absolute energy, band gap, and Fermi energy. Lu et al.^[57]used CGCNN to develop a deep learning model with an accuracy greater than 0.81, applying it to screen the CSD database for high-performance hydrogen storage materials.

As shown in Figure 3d, GAN is a deep learning framework consisting of two neural networks: the generator and the discriminator. The generator is responsible for producing samples that resemble real data, while the discriminator evaluates whether the input samples are genuine or generated by the generator. Through this adversarial process, the two networks continuously optimize each other, ultimately achieving the goal of generating high-quality samples^[58].GAN can learn the characteristics of existing MOF datasets and generate novel MOFs with potentially superior performance. For example, researchers have proposed an automated nanoporous material discovery platform driven by supramolecular VAE, aiming at the generative design of reticular materials. Taking MOF structures as an example, the platform's design process targets the separation of carbon dioxide from natural gas or flue gas. By jointly training with various MOFs identified as superior for gas separation, the autoencoder demonstrates strong optimization capabilities. The newly discovered MOFs also exhibit superior adsorption performance compared to some of the best reported MOF/zeolite materials. Furthermore, GAN shows great application potential in areas such as data augmentation and reverse design^[58-59]..

In addition to GANs, VAEs and diffusion models have also made significant progress in the generative design of MOFs. VAEs map MOF structures into a latent space using an encoder and then generate new structures through a decoder. For example, Yao et al.^[60]developed a VAE-based MOF generation framework that uses latent space interpolation to create MOFs with continuously varying porosity. Diffusion models generate high-quality MOF structures by gradually removing noise. For instance, Park et al.^[61]proposed a diffusion model-based MOF generation method that simulates the crystal growth process to produce MOFs with specific topological structures, whose structural stability is superior to that of MOFs generated by traditional methods. These generative models provide diverse tools for MOF design, and in the future, integrating multiple models could further enhance the efficiency and quality of new material generation.