Intelligent catalyst design method based on multi-modal fusion and causal inference

Abstract

The invention discloses an intelligent catalyst design method based on multi-modal fusion and causal inference. Comprising the following steps: constructing a multi-modal pre-training Catal-PTM model, generating a preliminary catalyst structure, obtaining constraint conditions of catalyst generation, generating a candidate catalyst structure, generating reaction parameters and a synthesis path of a candidate catalyst, generating the candidate catalyst, testing the candidate catalyst, and arranging and storing data. And carrying out closed-loop optimization to obtain an optimal formula of the catalyst. The method has the beneficial effects that unified characterization is realized through a multi-modal pre-training model, it is ensured that the generated catalyst conforms to physical and chemical laws in combination with a causal discovery and intervention mechanism, and a complete autonomous closed-loop system is constructed through deep integration of a generative design engine and an automatic synthesis platform; the core goal of improving the research and development efficiency and the success rate of the catalyst by several times is finally achieved from'trial and error 'to'rational design' and from'manual operation 'to'machine independence'.

Description

Technical Field

[0001] This invention relates to a method for developing catalysts, and more particularly to an intelligent design and synthesis method for catalysts based on multimodal data fusion, causal inference, and automated closed-loop feedback, belonging to the field of catalyst material development technology. Background Technology

[0002] Catalysts are indispensable key materials in chemical engineering, energy, and environmental protection. Their traditional development relies primarily on an experience-driven "trial and error" approach. This method typically involves designing the catalyst's composition and structure, experimenting with and optimizing synthesis conditions, characterizing reaction performance and structural properties, and using this feedback for improvement. However, with the increasing complexity of catalytic systems (involving high-dimensional parameters such as multiple components, multiple structures, and multiple reaction conditions), traditional methods face inherent limitations such as lengthy development cycles (usually requiring several years), high costs (a single experiment can cost tens to hundreds of thousands of yuan), and poor data reproducibility.

[0003] Current AI-assisted catalyst design largely relies on supervised learning (such as predicting adsorption energy and activity descriptors), which accelerates the screening process but has the following bottlenecks.

[0004] 1. Data heterogeneity: Catalytic design involves computational data (such as DFT), experimental data (such as XRD, TPR) and textual data (literature), which are difficult to effectively integrate using traditional ML models.

[0005] 2. Lack of causal relationship: Most models are based on correlation prediction (such as descriptor-activity) and lack modeling of the causal mechanism between "structure-synthesis-performance", resulting in poor extrapolation.

[0006] 3. Closed-loop failure: Many systems are still in the "offline prediction" stage and have failed to form a real-time closed-loop optimization with automated synthesis and characterization hardware.

[0007] 4. In complex reaction systems, the physicochemical feasibility of AI-generated schemes and the reliability of synthetic routes still need to be improved.

[0008] 5. Incomplete automation loop: Although automation platforms such as AI-EDISON have emerged, data gaps still exist in their "design-synthesis-testing-learning" loop.

[0009] 6. Insufficient real-time optimization and adaptability: Existing systems mostly adopt batch learning mode, which makes it difficult to learn and quickly adjust the real-time data stream generated by automated experiments online.

[0010] 7. Lack of physicochemical constraints in the generation results: Existing generation models (such as Catal-GPT) often neglect the actual synthetic feasibility (such as thermodynamic stability, kinetic barriers, etc.), resulting in a low success rate of laboratory synthesis (usually <30%). Summary of the Invention

[0011] Objective: This invention addresses the problems existing in current technologies by proposing a catalyst intelligent design method based on multimodal fusion and causal inference. It achieves unified characterization through a multimodal pre-trained model (Catal-PTM), combines causal discovery and intervention mechanisms to ensure the generated catalyst conforms to physicochemical laws, and constructs a complete autonomous closed-loop system through deep integration of a generative design engine and an automated synthesis platform, ultimately achieving a breakthrough in improving synthesis feasibility by more than three times. The technical effect of this application is to achieve a paradigm shift in R&D from "trial and error" to "rational design," and from "manual operation" to "machine autonomy," ultimately achieving the core goal of increasing catalyst R&D efficiency and success rate several times over.

[0012] Technical solution: A catalyst intelligent design method based on multimodal fusion and causal inference, comprising the following steps:

[0013] Step 1: Construct a multimodal pre-trained Catal-PTM model. The multimodal data includes heterogeneous data from multiple sources, including fused text, images, spectral data, and experimental parameters, to form a digital characterization of the catalyst's properties and performance.

[0014] The priority attributes of each property are ranked by the QSAR model and a linear function is calculated. Based on the general direction of the linear function, a machine learning prediction model is used to calculate complex functions of activation energy, selectivity and adsorption energy, which provide structural and coordination requirements for catalyst generation.

[0015] Step 2: Generate a preliminary catalyst structure. Input the required data for the multimodal catalyst into the Catal-PTM model for digital characterization to generate a preliminary catalyst structure.

[0016] Step 3: Obtain the constraints for catalyst generation, input multimodal data into the causal discovery and intervention module, so that the causal discovery and intervention module can identify key causal paths and direct the causal constraint module to perform directional constraint training, and output the causal constraint information structure.

[0017] Step 4: The generated candidate catalyst structures are input into the generative design engine using the causal constraint information output in Step 3 and the processed catalyst information output by the multimodal pre-training module trained in Step 2. The generative design engine consists of a GAN and a diffusion model working in series. The GAN generates diverse and reasonable candidate catalyst structures based on upstream information. The candidate catalyst structures are used as the "draft" or conditional input for the diffusion model. The diffusion model is then refined and optimized, and the final design result is output.

[0018] Step 5: Generate reaction parameters and synthesis pathways for candidate catalysts. Import the candidate catalyst structures output in Step 4 into the AI ​​decision engine, which will generate the module for optimizing reaction parameters and the optimal synthesis pathway. While receiving the upstream structure, the reaction parameters and synthesis pathways will be designed.

[0019] Step 6: Generate candidate catalysts. Input the generated candidate catalyst structure, optimal reaction route and parameters into the automated robotic synthesis workstation. The automated workstation will generate a small number of candidate catalysts.

[0020] Step 7: Candidate catalyst testing and data processing and storage. The generated candidate catalysts are tested using high-throughput standard equipment and the data is collected and fed back into the database.

[0021] Step 8: Closed-loop optimization to obtain the optimal catalyst formulation. The optimal catalyst formulation is explored through the expectation improvement EI intelligent exploration. For the optimal solution catalyst that meets the requirements, its formulation and synthesis route are output. For the formulation data that does not meet the requirements, it is re-introduced into the original multimodal preprocessing model in a closed loop to continue closed-loop optimization until the optimal catalyst formulation is generated.

[0022] Preferably, the method for constructing the multimodal pre-trained Catal-PTM model in step one is as follows:

[0023] Graphical-textual comparison learning: enables the model to learn to associate representational maps with textual descriptions or crystal structures;

[0024] Masked language / graph modeling: Randomly masking parts of text words or graph data segments, allowing the model to make predictions and learn deep features;

[0025] The pre-trained multimodal Catal-PTM model can output a unified representation vector h for any catalytic data, which can be used for downstream prediction tasks; its loss function is:

[0026]

[0027] It is a masked language model loss, used for text modality, to enable the model to learn to reason and deduce based on context, to enable the model to deeply understand text semantics, and to learn the language rules in the field of chemistry;

[0028] It is a contrastive learning loss, used to bring the representations of different modes of the same catalyst closer together in the vector space, and vice versa.

[0029] It is a cross-modal reconstruction loss, which ensures that the representation retains complete information, requiring the model to be able to reconstruct the representation from one modality;

[0030] These are the balancing hyperparameters, consisting of three balancing coefficients used to adjust the importance of different loss terms in the total loss, and can be adjusted according to the actual task requirements.

[0031] Preferably, the method for obtaining the catalyst performance function in step one is as follows:

[0032] S3.1 Constructing a QSAR quantitative structure-activity relationship model

[0033]

[0034] In the formula, y represents the performance of the catalyst, x1, x2, ... represent the factors affecting the catalyst activity, and w1, w2, ... represent the weights of each factor. ϵ is the intercept term, also known as the deviation, and ϵ is the error term;

[0035] S3.2, Quantitative Structure-Activity Relationship Model: Predictive Model After Machine Learning

[0036] Catalytic Performance )= + ϵ

[0037] In the formula, y represents the target property to be predicted, Descriptori describes the characteristics of the catalyst, f() is the complex nonlinear function learned by the machine learning model, and ϵ is the error term.

[0038] Preferably, the method for obtaining the constraints on catalyst formation in step three is as follows:

[0039] S4.1 Constructing a Causal Discovery and Intervention Model

[0040] (1). Structural Causal Model (SCM)

[0041]

[0042] in, yes The set of parent nodes, It is an unobserved exogenous variable;

[0043] (2). Causal discovery

[0044] The module runs the PC (Peter-Clark) algorithm, which examines whether each pair of variables is independent given a subset of other variables, thereby gradually removing edges and determining the causal direction. The goal of causal discovery is to automatically find the set of parent nodes from the observed data, reconstruct the complete causal graph, and infer the causal relationship between variables.

[0045] (3) Causal intervention

[0046] When a set of variables Z satisfies the backdoor criterion with respect to (X,Y), the intervention formula can be expressed as:

[0047]

[0048] in For the intervention distribution, it represents the probability that variable Y takes the value y when variable X is forcibly set to the value x;

[0049] This indicates an intervention operation, meaning that an external force forcibly sets the variable X to a specific value x, thereby severing all edges pointing to X and making X no longer affected by its original cause;

[0050] To satisfy the backdoor criterion, the set of hybrid variables can be represented as "precursor concentration" and "calcination atmosphere" in the catalyst.

[0051] It represents the conditional probability of outcome Y occurring given cause X and confounding factors Z;

[0052] To estimate the intervention effect, it is necessary to examine the influence of X on Y within each hybrid layer Z=z, and then take a weighted average of the results of all layers.

[0053] S4.2 Causal Constraints

[0054] When generating new catalyst structures, a causal regularization term is introduced into the generator loss function: constraining GAN.

[0055]

[0056] Where G is the generator, whose goal is to capture the data distribution of real data. The input is usually random noise z, and the output is the generated data. In catalyst design, G's goal is to generate novel and theoretically efficient catalyst atomic structures.

[0057] D is the discriminator, and its goal is to estimate the probability that a sample is real from the training data rather than generated by G. In catalyst design, the task of D is to determine whether a given atomic structure is a real, known, and highly efficient catalyst, or a "fake" structure generated by G.

[0058] E represents the expected value, which is the average of all possible outcomes.

[0059] These are samples collected from the real data distribution;

[0060] Z represents the noise sample first sampled from the prior noise distribution;

[0061] To correct the parameters, It is the standard loss function for Generative Adversarial Networks (GANs), enhancing the adversarial relationship between the generator G and the discriminator D; in machine learning loss functions, where... It is a penalty item that penalizes the generated results that violate the identified causal relationship.

[0062] Preferably, in step four, the generative design engine optimizes and generates designs in a direction that satisfies performance targets, guided by the loss function. Specifically, this includes:

[0063] (1) Standard GAN

[0064]

[0065] Where G is the generator. Given a sample of real data, input a random noise vector z and a performance condition c;

[0066] D is the discriminator, which determines whether a catalyst comes from the real dataset or is faked by the generator, and evaluates whether it meets condition c.

[0067] As a condition, Given a potential space vector, by adjusting z, the generator can produce different catalyst variants;

[0068] To limit losses, To constrain weights;

[0069] (2) Noise prediction loss module for diffusion model

[0070] The generated candidate catalysts need to be evaluated quickly. The diffusion model generates samples by learning the data distribution, and its reverse process is used to evaluate the "uncertainty" of catalyst performance.

[0071] The training objective of the diffusion model, i.e., the noise prediction loss function, is as follows:

[0072]

[0073] in The original data point represents a catalyst structure with ideal and well-defined performance; in catalyst design, it represents a real high-performance catalyst sequence or a stable catalytic active site structure.

[0074] The diffusion time step represents the degree of disturbance or uncertainty. The larger the catalyst is, the greater the "interference" to its structure and the higher the uncertainty of its performance.

[0075] The noise added is randomly sampled Gaussian noise, an external uncertainty that affects catalyst performance;

[0076] It is a noise prediction network, representing a performance prediction model. Its task is to predict the deviation noise of the performance of a catalyst under "uncertainty" noise interference from the ideal state. The trained model can quickly predict the performance based on the catalyst's descriptor.

[0077] An ideal catalyst is a descriptor of its molecular formula, crystal structure, protein amino acid sequence, or active site.

[0078] (3) The concatenation of GAN and diffusion model

[0079] First, leveraging the speed advantage of GANs, a large number of candidate catalyst structures with a certain degree of diversity and rationality are rapidly generated based on existing causal constraints and multimodal inputs. Then, these candidate structures are used as the "draft" or conditional input of the diffusion model, which is then refined and optimized to finally output high-quality design results.

[0080] Preferably, the reaction parameters and synthesis route for generating the candidate catalyst in step five are as follows:

[0081] The AI ​​decision engine generates optimized reaction parameters and the optimal synthesis route through constrained optimization and sequence decision-making, given the target catalyst structure. Find the optimal synthesis path It can be represented as:

[0082]

[0083] Represents the set of all possible composition paths. , , These represent the cost, time risk, and security risk of the path, respectively. This represents the catalyst structure actually synthesized via path S. It is a loss function representing the difference between the synthesized result and the target structure, ensuring that the synthesized product is the designed structure. These are weighting coefficients, used to balance the importance of different objectives;

[0084] Parameter layer optimization:

[0085] When optimizing chemical reaction conditions, it is necessary to consider multiple continuous variables; different combinations of variables will affect the response value.

[0086] Central Composite Design (CCD): CCD is an experimental design method used to construct second-order response surface models. It allows for accurate estimation of the influence of various variables and their interactions on the response value with a relatively small number of experiments. Its model formula is:

[0087]

[0088] Where Y is the target response. The encoded values ​​representing the i-th and j-th factors; For constant terms, The coefficient of the linear term represents the main effect of factor i. The coefficient of the squared term represents the curvature effect of factor i. This represents the interaction effect between factors i and j, where K is the number of factors. For example, when k=3, the combinations are: i=1,j=2; i=1,j=3; i=2,j=3) (machine error term). To iterate through all the different factors and combine them in pairs (e.g., when k=3, the combinations are: i=1, j=2; i=1, j=3; i=2, j=3);

[0089] The value of the fitted second-order model lies in its coefficients. A positive coefficient for the first-order term indicates that the response value increases as the factor increases, while a negative coefficient indicates the opposite. The coefficient for the squared term represents the curvature effect of the factor. A negative squared term indicates the existence of a maximum value, meaning that the factor is not necessarily better the higher or lower it is, but rather achieves the best effect at a certain coding level.

[0090] After calculating experimental points by designing CCD experiments and fitting the coefficients of the above model through regression analysis, the model is used to predict the optimal reaction conditions and analyze the influence of each factor and its interaction on the yield.

[0091] Preferably, the method for generating the candidate catalyst in step six is ​​as follows:

[0092] The output of the AI ​​decision engine is a standardized synthesis protocol that is machine-readable and contains structured information;

[0093] This protocol can be directly transmitted to automated robot workstations via API interface to drive robotic arms, liquid processors, and reactors for fully automated synthesis.

[0094] The bill of materials requires the precursor name, molecular formula, purity, and mass / volume to ensure accurate weighing and avoid side reactions;

[0095] The reaction sequence requires step numbers, operation types (heating, stirring, dropping), target parameters, and a clearly defined execution order;

[0096] The equipment control requirements include reactor type, temperature program, stirring rate, and pH value settings, which are directly controlled by the automated hardware.

[0097] Using online sensors (pH, temperature, spectrum) and sampling points, it provides real-time feedback and quality control; it also detects maximum temperature / pressure thresholds and issues warnings for hazardous operations.

[0098] Preferably, the method for obtaining the optimal catalyst formulation through closed-loop optimization in step eight is as follows:

[0099] The results synthesized by the automated workstation (yield, purity, characterization data) form a feedback loop to continuously optimize the AI ​​decision engine model.

[0100]

[0101] New data This is used to update previous models, forming an autonomous closed loop of "design-synthesis-testing-learning" to continuously iterate and improve the success rate of the synthesis strategy;

[0102] Automated closed-loop feedback system:

[0103] The system employs an expected improvement EI intelligent exploration method to discover new catalyst formulations, and Bayesian optimization guides experiments through EI functions to iteratively find the optimal catalyst formulation.

[0104]

[0105] In the formula, m represents a set of candidate catalysts. Let m be the unknown true performance of candidate point m. Given the best performance achieved so far, and E as the expected value, the AI ​​agent model predicts the probability distribution of f(m) and calculates the average of all possible improvements.

[0106] Given a small number of known catalyst formulations and performances in the initial dataset, a surrogate model is constructed to optimize the acquisition function EI. The candidate formulation m with the largest EI is calculated. The candidate catalyst m is experimentally verified, synthesized, and tested. The performance f(m) of the new data points is obtained. The new data is added to the dataset, and the performance is tested to see if it meets the requirements. If it does, the optimal catalyst formulation is output. If it does not, the process is repeated to return to the multimodal pre-trained model to construct an algebraic model. This process is repeated until the optimal catalyst formulation can be output, thus maximizing the acquisition function.

[0107] Beneficial effects: Multimodal pre-trained Catal-PTM model: integrates text, numerical, spectral and crystal structure information, extracts unified representation through self-supervised learning, and solves the problem of data heterogeneity.

[0108] Causal discovery and intervention module: Based on the structural causal model (SCM), it analyzes the causal chain between synthesis parameters, structural features and catalytic performance, and can simulate the intervention effect (such as "the effect of increasing calcination temperature on selectivity").

[0109] Causal constraints in generative design: Introducing causal regularization terms into generative adversarial networks (GANs) or diffusion models ensures that the generated new catalyst candidate structures not only have superior performance but also conform to physicochemical laws.

[0110] Fully automated closed-loop feedback system: integrates AI decision-making core with automated robot synthesis workstation and high-throughput characterization equipment to achieve unmanned intervention and continuous optimization throughout the entire process of "design-synthesis-testing-analysis".

[0111] The invention inputs catalyst requirements through multimodal data, performs unified characterization via a multimodal pre-trained Catal-PTM model, and uses a QSAR quantitative structure-activity model to prioritize attributes and calculate linear functions. Based on the general direction of the linear functions, a machine learning prediction model calculates complex functions related to activation energy, selectivity, adsorption energy, etc., providing reasonable structural and coordination requirements for catalyst generation. The proprietary multimodal pre-trained Catal-PTM model aims to solve the problem of existing systems' inability to effectively integrate multi-source heterogeneous data such as text (research literature), images (microstructure), spectral data (XPS, XRD), and experimental parameters, forming a comprehensive and high-quality digital characterization of catalyst properties and performance.

[0112] Subsequently, causal discovery, intervention, and causal constraints are used to regulate the properties of the catalyst to maintain its maximum effect while satisfying feasibility. This ensures that the generated catalyst not only has superior performance but also conforms to physicochemical laws, improving the synthetic feasibility by more than three times. By introducing causal discovery, intervention, and constraint mechanisms, this approach overcomes the limitations of traditional machine learning that relies solely on correlation prediction. It aims to clearly identify the key causal factors affecting catalyst performance (such as activation energy and selectivity) and their intrinsic mechanisms of action, thereby guiding the generation of design schemes that conform to physicochemical laws and improving the interpretability and reliability of the model.

[0113] The synthesis of catalysts is costly. To achieve greater accuracy and standardization, the obtained data is imported into a generative design engine (GAN / diffusion model) for further standardization. The physicochemical laws derived from causal inference are used as constraints and embedded into machine learning prediction and generative design engines (such as GAN / diffusion models) to ensure that the generated candidate catalyst structures not only have superior performance but also have high synthetic feasibility and stability, aiming to increase the success rate of laboratory synthesis by more than 3 times.

[0114] Candidate catalysts are generated, and an AI decision engine produces optimized formulations and synthetic routes. The essence of AI-generated synthetic routes is a constrained optimization and sequence decision problem. Given a target catalyst structure C* (the pre-designed optimal structure), the optimal synthetic route S* is sought.

[0115] The data is then directly transmitted to an automated robotic workstation (such as the AI-EDISON platform) via an API interface, driving robotic arms, liquid processors, reactors, and other equipment to perform fully automated synthesis of the target catalyst. Simultaneously, high-throughput characterization equipment monitors the process, recording performance and data structure in real time. The synthesis results (yield, purity, characterization data) from the automated workstation form a feedback loop for continuous optimization of the AI ​​model. This establishes a complete chain from "intelligent design" to "automated synthesis" and then to "high-throughput testing." By driving the automated robotic workstation (such as the AI-EDISON platform) to execute synthesis tasks via the API interface and using characterization data to optimize the AI ​​model in real time, a self-iterating "design-synthesis-testing-learning" closed-loop system is formed, greatly reducing manual intervention and accelerating the R&D process.

[0116] Simultaneously, the expected improvement (EI) intelligent exploration catalyst formulation is adopted to test the output catalyst. If it is the optimal solution catalyst, the formulation and synthesis route are output. For the poor performance, closed-loop optimization is performed, and the data is returned to the upstream to continue closed-loop optimization until the optimal solution is found. This forms an autonomous closed loop of "design-synthesis-testing-learning", which continuously iterates to improve the success rate of synthesis strategies and gathers collective wisdom to continuously improve catalyst design capabilities. Attached Figure Description

[0117] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0118] Figure 1 This is a flowchart illustrating the framework of the present invention;

[0119] Figure 2 This is a diagram showing the relationship between the causal discovery and intervention modules of this invention.

[0120] Figure 3 This is a graph showing the factors affecting the activity of the present invention. Detailed Implementation

[0121] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0122] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0123] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can include direct contact between the first and second features, or contact between the first and second features through another feature between them. Furthermore, "above," "over," and "on top" of the second feature includes the first feature directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature includes the first feature directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature.

[0124] like Figure 1 As shown, a catalyst intelligent design method based on multimodal fusion and causal inference includes the following steps:

[0125] Step 1: Design a complete Catal-PTM model by building a pre-trained Catal-PTM model to prepare for subsequent steps;

[0126] Step 2: Input the multimodal data into the Catal-PTM model. By inputting simple catalyst requirements and some information about the required catalyst, the Catal-PTM model can generate a preliminary catalyst structure. Simultaneously, input the multimodal data into the causal discovery and intervention module, enabling the module to identify key causal pathways.

[0127] Step 3: The causal discovery and intervention module identifies key causal paths and directs the causal constraint module to perform directional constraint training, outputting the constraint information structure to the generative design engine. At the same time, the trained multimodal pre-trained model also outputs processed catalyst information to the generative design engine.

[0128] Step 4: Upstream information is input into the generative design engine, where GAN and diffusion model work in tandem. GAN takes the lead, leveraging its speed advantage to rapidly generate a large number of candidate catalyst structures with a certain degree of diversity and rationality based on existing causal constraints and multimodal inputs. These candidate structures are then used as "drafts" or conditional inputs for the diffusion model. The diffusion model then refines and optimizes these structures, such as improving structural stability, correcting unreasonable bond lengths and angles, and enhancing specific functional groups, ultimately outputting high-quality design results.

[0129] Step 5: The high-quality design structure results are imported into the AI ​​decision engine, which generates the module for optimizing reaction parameters and the optimal synthesis path. While receiving the upstream structure, the module designs the reaction parameters and path to better complete the design.

[0130] Step 6: Input the generated high-quality structure design and optimal reaction route and parameters into the automated robotic synthesis workstation. The automated workstation will generate a small amount of catalyst as a candidate catalyst.

[0131] Step 7: The generated candidate catalysts are tested using high-throughput standard equipment and data is collected and fed back into the database.

[0132] Step 8: Explore new catalyst formulations through Expectation Improvement (EI) intelligent exploration. For the optimal solution catalyst that meets the requirements, output its formulation and synthesis route. For formulation data that does not meet the requirements, re-close the loop and import it into the original multimodal preprocessing model to continue closed-loop optimization until the optimal catalyst formulation is generated.

[0133] 1. Catal—PTM Model Module

[0134] To address the data silo problem, we designed the Catal-PTM model, whose pre-training tasks include:

[0135] 1.1 Graphical and textual comparison learning: Enables the model to associate representational spectra (such as XRD, Raman) with textual descriptions (such as "spinel structure") or crystal structures.

[0136] 1.2 Masked Language / Graph Modeling: Randomly mask parts of text words or graph data segments, allowing the model to make predictions and learn deep features.

[0137] The pre-trained Catal-PTM model can output a unified representation vector h for any catalytic data, which can be used for downstream prediction tasks. Its loss function is:

[0138]

[0139] It is a masked language model loss, used for text modality, which enables the model to learn to reason and deduce based on context, allowing the model to deeply understand text semantics (such as literature descriptions and synthesis steps), and learn the language rules in the field of chemistry.

[0140] It is a contrastive learning loss used to bridge the gap between different modal representations of the same catalyst, bringing the representations of different modalities describing the same catalyst (such as the text "high dispersion" and TEM images) closer in the vector space, and vice versa. Solving the problem of data heterogeneity and establishing the correlation between different modal information is the key to achieving unified representation.

[0141] It is a cross-modal reconstruction loss that ensures the representation retains complete information. It requires the model to be able to reconstruct data from another modality, such as a crystal structure diagram, from a representation of one modality, for example, a textual description. It ensures that the learned unified representation contains sufficiently complete information and that no key features are lost during the fusion process.

[0142] These are the balancing hyperparameters, consisting of three balancing coefficients used to adjust the importance of different loss terms in the total loss. They can be adjusted according to the actual task requirements; for example, they can be increased when more structural information is being considered.

[0143] This model addresses the data heterogeneity problem in traditional methods. The loss function, through three collaborations, ensures that the model output has a unified representation with rich semantic information, strong cross-modal correlation, and high information integrity, providing high-quality feature input for downstream tasks. It can achieve the prediction accuracy of traditional methods with thousands of data sets with only 50 sets of samples, reducing the error from 20% to below 5%, and providing reasonable structural and compatibility requirements for catalyst generation.

[0144] 2. QSAR model module under multimodal training

[0145]

[0146] This model is a weighted summation model. The catalyst's performance (y, activity) is considered as a linear combination of its various characteristics (x1, x2, ..., such as specific surface area, metal loading). Each characteristic is preceded by a weight (w1, w2, ...), representing the degree to which that characteristic contributes to the performance. The intercept term, also known as the bias, is used in all descriptors. When both are zero, it represents a baseline value for catalyst performance. ϵ is the error term.

[0147] The factors affecting catalyst activity are as follows: Figure 3 As shown, a series of factors, such as composition and structure of active components, additives, supports, physical parameters, temperature, pressure, reactant concentration, and chemical deactivation, all have a certain influence on catalyst activity and can be used as the value of x.

[0148] For a large amount of known catalyst data, linear regression can quickly identify which structural features have the greatest impact on the target reaction (such as OER and HER). For example, the model may find that w2 (corresponding to platinum particle size) has a large and positive weight, meaning that reducing the platinum particle size may be an effective strategy to improve catalytic activity. The weights provide a clear direction for optimization. When designing a new catalyst, features with high weights can be prioritized for optimization, and then physical parameters, composition and structure, electronic properties, reaction conditions, etc., can be adjusted according to their contribution to activity to form a catalyst with an optimized structure. However, complex structures cannot be represented by a simple QSAR model. The simple QSAR model can serve as a guide for downstream processing, i.e., the starting point for constructing relationships.

[0149] 3. Machine learning prediction model module following the QSAR model

[0150] Catalytic Performance )= + ϵ

[0151] Where y represents the target property to be predicted (e.g., activity, selectivity, adsorption energy), Descriptori is the characteristic (descriptor) describing the catalyst, such as d-band center, coordination number, specific surface area, metal loading, etc., f() is a complex nonlinear function learned by the machine learning model, and ϵ is the error term. The complex function f() is obtained through neural networks, random forests, and Gaussian processes. This approach addresses the problem of data silos by unifying heterogeneous data such as text, images, and spectra into machine-understandable representations. It creates a high-quality, in-depth digital profile of the catalyst, laying the foundation for subsequent accurate prediction and generation, and is expected to improve data utilization efficiency by more than 50%.

[0152] 4. Causal Discovery and Intervention Module

[0153] Traditional machine learning models learn This system reveals, through structural causal modeling (SCM), that... , where y and x represent a set of variables that affect the catalyst effect.

[0154] (1). Structural Causal Module (SCM)

[0155]

[0156] in yes The set of parent nodes (i.e., direct causes). These are unobserved exogenous variables (such as unknown factors or random noise). The following example is used to facilitate understanding, as shown in Table 1.

[0157] Table 1. Parent Node Set Display

[0158]

[0159] The parent node set PA(Y) is {X3}. This means that if we want to predict or calculate the activity (Y) of the catalyst, it is sufficient to know the number of active sites (X3). The parent node set PA(X3) is {X1, X2}. The number of active sites (X3) is determined by the particle size (X1) and dispersion (X2) of the active component.

[0160] (2) Causal discovery

[0161] The core is the conditional independence test. The module runs the Peter-Clark (PC) algorithm, which tests whether each pair of variables is independent given a subset of other variables, thereby progressively removing edges and determining the causal direction. The goal of causal discovery is to automatically find these parent node sets from the observed data, thus reconstructing the complete causal graph. Causal discovery aims to infer causal relationships between variables from observed data.

[0162] (3) Causal intervention

[0163] The goal is to quantify how much impact changing a variable (intervention) will have on the target variable.

[0164] One of its core formulas is the backdoor adjustment formula. When a set of variables Z satisfies the backdoor criterion with respect to (X,Y) (i.e., Z is not a descendant of X and can block all promiscuous paths between X and Y), the intervention formula can be expressed as:

[0165]

[0166] in The intervention distribution represents the probability that variable Y takes the value y when we forcibly set variable X to the value x;

[0167] This indicates an intervention operation, meaning that an external force forcibly sets the variable X to a specific value x, thereby severing all edges pointing to X and making X no longer affected by its original cause;

[0168] To satisfy the backdoor criterion, the set of mixed variables in a catalyst can be represented as "precursor concentration", "calcination atmosphere", etc.

[0169] It represents the conditional probability of outcome Y occurring given cause X and confounding factors Z;

[0170] To estimate the intervention effect, it is necessary to examine the influence of X on Y within each hybrid layer Z=z, and then perform a weighted average of the results for all layers. Estimate the causal effect of "calcination temperature (X)" on "catalyst activity (Y)". When it is known that "precursor type (Z)" simultaneously affects both X and Y (as a confounding factor), group the data according to "precursor type (Z)", and within each group, calculate the effect of calcination temperature on activity (mathematically expressed as...). The results from each group were weighted and averaged according to the distribution ratio of precursor type in the overall population (P(Z)), ultimately yielding a pure causal effect after eliminating the interference of precursor type. All parameters affecting activity can participate in the intervention of the causal module, optimizing and simplifying the optimal parameters of the catalyst.

[0171] For example, the goal is to design a high-performance platinum-based hydrogen evolution reaction (HER) catalyst. The causal discovery and intervention module would then... Figure 2 Work:

[0172] Input: Historical catalyst dataset. Each sample contains features such as the particle size of platinum nanoparticles, coordination number of surface atoms, adsorption energy for hydrogen intermediates, and experimentally measured hydrogen evolution activity.

[0173] Causal discovery: The module runs algorithms such as PC to discover and construct causal graphs through conditional independence tests.

[0174] Causal Intervention: Based on the causal diagram above, the module uses a backdoor adjustment formula to calculate the intervention effect. By adjusting for confounding factors (such as "particle size"), it estimates that P(activity|do(coordination number=6)) is much higher than P(activity|do(coordination number=9)), thus concluding that reducing the surface atomic coordination number is a strong causal driver for improving HER activity.

[0175] The causal discovery and intervention module acts like the team's "chief scientist," analyzing data to tell us: "Based on causal analysis, we should focus on optimizing variable X, adjusting it to a value close to Y for optimal results!" The generated causal knowledge package is the core driving force behind the efficient and reliable innovation of the next module, the causal constraint module.

[0176] By exploring causal and interventional distributions, this approach breaks through the traditional "black box" model, revealing and harnessing the intrinsic causal mechanisms influencing catalyst performance (such as selectivity and activation energy). It shifts the focus of generation from "correlation" to "causality," ensuring that the design scheme conforms to physicochemical laws and improving the interpretability and reliability of generated candidate materials by 60%.

[0177] 5. Causal Constraint Module in Generative Design

[0178] When generating new catalyst structures, we introduce a causal regularization term into the generator loss function: constraining GAN.

[0179]

[0180] Where G is the generator, its goal is to capture the data distribution of real data, the input is usually random noise z, and the output is the generated data. In catalyst design, the goal of G is to generate novel and theoretically efficient catalyst atomic structures.

[0181] D is the discriminator, and its goal is to estimate the probability that a sample comes from the training data (real) rather than from G (generated). In catalyst design, D's task is to determine whether a given atomic structure is a real, known efficient catalyst or a "fake" structure generated by G.

[0182] E represents the expected value, which can be understood as the average of all possible outcomes.

[0183] xt represents a sample collected from the real data distribution.

[0184] Z represents the noise sample first sampled from the prior noise distribution.

[0185] To correct the parameters, It is the standard loss function for Generative Adversarial Networks (GANs), enhancing the adversarial relationship between the generator G and the discriminator D. In machine learning loss functions...

[0186] in It is a penalty term that penalizes generated results that violate the identified causal relationship, for example:

[0187] The contradiction between structural stability and synthesis temperature

[0188] • Trade-off between activity and selectivity

[0189] • Balancing cost and performance

[0190] This ensures that the generated catalyst not only has superior performance but also conforms to physicochemical laws, increasing the feasibility of synthesis by more than 3 times.

[0191] In this model, the discriminator D aims to maximize V(D,G). For a sample xt from the real data distribution, i.e., a known efficient catalyst structure, the value of D(Xt) should be as large as possible, approaching 1 (the probability of being classified as true), so logD(X) should be as large as possible (approaching 0). For a sample G(z) generated by the generator G (i.e., the generated candidate catalyst structure), the value of D(G(z)) (the probability of being misclassified as "true") should be as small as possible, approaching 0, so log(1-D(G(Z))) should be as large as possible. D needs to strive to accurately distinguish between real data and generated data, thereby improving its discriminative ability.

[0192] The goal of generator G is to minimize V(D,G) because it wants its generated sample G(Z) to be indistinguishable from real data, making the discriminator D think it comes from real data, thereby improving its generation ability. Therefore, the value of D(G(Z)) should be as large as possible, and the smaller log(1-D(G(Z))) will be, so G needs to minimize this term.

[0193] In this model, the generator G acts like a novice designer, randomly generating atomic structures that may not conform to chemical laws or have poor catalytic performance. The discriminator D, like an experienced expert, learns from a database of known high-performance catalysts, gaining the ability to identify the structures of good catalysts. Initially, D easily detects the inferior structures generated by G. As D continuously identifies the "fake" structures generated by G and the "real" structures in the database, its discriminative ability grows stronger. It can more precisely capture key features in the structure of highly efficient catalysts (such as specific active site coordination environments and electronic structure characteristics). G continuously adjusts its parameters based on feedback from D (i.e., the discriminative results). It learns which structural features are more likely to be considered "good" catalysts by D. Thus, it begins to generate increasingly realistic atomic structures that better match the characteristics of high-performance catalysts. When G progresses to the point where its generated catalyst structures are so realistic that D cannot distinguish between real and fake structures, the game reaches an equilibrium. At this point, the data distribution learned by G is infinitely close to the actual catalyst structure distribution. G has become a powerful "AI catalyst designer," capable of generating a large number of novel catalyst candidate structures that are likely to have high activity.

[0194] After training, inputting noise into the generator G will output novel catalyst candidate structures that are both realistic and reasonable. By introducing causal constraints, the generator is no longer a blind imitator, but an "AI catalyst designer" that understands the rules of chemistry. It can directly generate a large number of candidate catalysts with reasonable structures and satisfactory performance.

[0195] The core responsibility of this module is to impose causal relationships. The GAN component primarily serves as a tool or means to learn causal structures in the data (e.g., to ensure the generator's output samples conform to a certain causal relationship through adversarial training), construct causal graphs, or as a regularization term to guide the model to focus on causal features. It does not directly generate the final design output.

[0196] 6. Generative Design Engine Module

[0197] Generative models are optimized and generated in a direction that satisfies specific performance objectives, guided by a loss function.

[0198] (1) Standard GAN

[0199]

[0200] Where G is the generator, xt is a sample of real data, the input is a random noise vector z and the performance condition c (e.g., "TOF > 10 s⁻¹"), and the output is a novel catalyst structure or composition. D is the discriminator, which determines whether a catalyst comes from a real dataset or is faked by the generator, and evaluates whether it satisfies condition c. As a condition, Given a potential space vector, by adjusting z, the generator can produce an infinite number of different catalyst variants. To limit losses, To constrain the weights, all constraints must satisfy the following rules. See Table 2 for the rules.

[0201] Table 2 Catalyst Constraints

[0202]

[0203] As can be seen from the table and the catalyst, the formation energy of the generated catalyst structure must be negative to satisfy the monotonicity constraint, that is, increasing the specific surface area (SBET) should not lead to a decrease in activity.

[0204] The structure should meet the rules of catalysis. The charge distribution of active sites (e.g., Pt charge) can serve as a catalytic descriptor for quantitatively predicting their turnover frequency (TOFi). The system is designed with a first coordination layer (direct coordination environment), a second coordination layer (non-covalent interactions such as hydrogen bonding and electrostatics), and an outer coordination layer (overall reaction environment, such as confinement effects) to optimize catalytic performance. Nanoscale spatial confinement can stabilize transition states or destabilize ground states, thereby altering reaction pathways, lowering energy barriers, and even changing reaction mechanisms.

[0205] To satisfy the synergistic effect rule, for reactions requiring multiple steps in series, active centers with different functions (such as acidic and metal active sites) should be designed for synergistic catalysis to achieve efficient execution of the reaction sequence. For reactions involving proton transfer, efficient "proton highways" (such as those via ordered proton relay groups) should be designed to significantly improve the proton transfer rate.

[0206] Stability and lifetime requirements should be met. Migration and sintering of active components at high temperatures should be suppressed through strong interfacial interactions (such as Pt-OP anchoring), the use of stabilizing supports, or the addition of structural additives. Anti-poisoning capabilities can be enhanced by adjusting the electronic structure or surface properties to reduce the adsorption strength of poisons (such as carbon precursors generated during the reaction or impurities carried in the raw materials) on the catalyst surface.

[0207] Synthesis and preparation rules should be followed, prioritizing abundant Earth elements, low-cost raw materials, and low-energy-consumption preparation methods to ensure the economic feasibility of the catalyst. Environmentally friendly synthesis routes, precursors, and solvents should be selected to reduce waste generation, and recycling or harmless treatment at the end of the catalyst's lifespan should be considered.

[0208] The catalyst must meet the mechanical performance requirements and have sufficient mechanical strength to resist wear, impact and pressure drop in the reactor, which is especially important in industrial reactors such as fluidized beds.

[0209] Characterization and verification rules should be met, and in-situ spectroscopy (such as in-situ Raman spectroscopy) and microscopy techniques should be used to explore the structural evolution and reaction mechanism of catalysts under real reaction conditions, providing a direct basis for design.

[0210] GANs in generative design engines are typically more complex and powerful, focusing on generating high-fidelity, diverse outputs. The training of GANs in generative design engines, however, focuses more on the quality of the final output.

[0211] Guided by causal constraints, thousands of candidate catalyst molecular structures that meet the requirements can be automatically and efficiently created. The catalyst design cycle is shortened from "years" to "weeks" or even "days", and the chemical space explored is expanded by more than 10^4 times compared with traditional methods.

[0212] (2) Noise prediction loss module for diffusion model

[0213] The generated candidate catalysts need to be evaluated rapidly. The diffusion model generates samples by learning the data distribution; its inverse process is profoundly analogous to the "uncertainty" assessment of catalyst performance. The training objective of the diffusion model, namely the noise prediction loss function, is as follows:

[0214]

[0215] in The original data point represents a catalyst structure with ideal and well-defined performance. In catalyst design, it represents a real high-performance catalyst sequence or a stable catalytic active site structure. The diffusion time step represents the degree of perturbation or uncertainty. The larger the value of T, the greater the "interference" received by the catalyst structure, and the higher the uncertainty of its performance. The noise added is randomly sampled Gaussian noise, representing external uncertainties that affect catalyst performance, such as variations during synthesis and uncontrollable fluctuations under reaction conditions. This is a noise prediction network, representing a performance prediction model. Its task is to predict the deviation (noise) from the ideal state of a catalyst under "uncertainty" (noise) interference. A trained model can quickly predict the performance (such as adsorption energy, TOF) of a catalyst based on its descriptor, replacing some expensive DFT calculations or experiments.

[0216] An ideal catalyst can be a descriptor of a molecular formula, crystal structure, protein amino acid sequence, or active site (such as d-band center, coordination number). Changing... This involves randomly perturbing the catalyst structure, randomly replacing active metal atoms, changing the pore size of the support material, and disrupting the bond lengths of adsorption sites to create random catalyst states. A smaller t-value indicates a smaller perturbation; a larger t-value indicates a larger perturbation. Through AI adaptive modifications, the model can learn the entire evolution spectrum of catalysts, from minor modifications to complete reconstruction. The model needs to know the extent to which the structure has been perturbed in order to accurately predict how to "repair" it. If the AI's predictions are corrected... With respect to the actual direction of destruction The complete consistency demonstrates that AI has a profound understanding of "how to create order from chaos," and a deep understanding of... In essence, the diffusion model further reinforces the selectivity of catalyst formation.

[0217] (3) The concatenation of GAN and diffusion model

[0218] GAN models are characterized by fast generation speed, single forward propagation, low training stability, and the need for fine-tuning. Diffusion models, on the other hand, have slower generation speed, require multi-step iterative denoising, and exhibit higher training stability and a smoother training process. Based on this analysis, a collaborative approach is adopted, leveraging their respective strengths. GANs take the lead, utilizing their speed advantage to rapidly generate a large number of diverse and reasonable candidate catalyst structures based on existing causal constraints and multimodal inputs. These candidate structures are then used as the "draft" or conditional input for the diffusion model. The diffusion model is then refined and optimized based on this, such as improving structural stability, correcting unreasonable bond lengths and angles, and enhancing specific functional groups, ultimately outputting high-quality design results.

[0219] 7. AI Decision Engine and Module for Generating Optimization Parameters and Synthesis Path

[0220] The essence of AI-generated synthetic pathways is a constrained optimization and sequence decision problem. Given a target catalyst structure... (For a pre-designed structure), find the optimal synthesis path. It can be represented as:

[0221]

[0222] This represents the set of all possible composition paths. , , These represent the cost, time risk, and safety risk (such as toxicity, high temperature and high pressure) of the route, respectively. This indicates the structure of the catalyst actually synthesized via path S. It is the difference loss function between the synthesis result and the target structure, ensuring that the synthesized product is the designed structure. This is crucial for heterogeneous catalysis (such as supported nanocatalysts and single-atom catalysts) because synthesis deviations can significantly affect active sites.

[0223] It is a weighting coefficient that balances the importance of different objectives.

[0224] This module transforms molecular structures from the digital world into real-world physical objects, resolving the disconnect between design and synthesis. It enables 24 / 7 unmanned, precise synthesis, reducing experimental labor costs by 90% and controlling batch-to-batch variation to below 5% through standardized operations.

[0225] This formula makes a preliminary determination of the global path optimization, but simple linear determination has certain errors. However, it points out the general direction. Based on this, the path is continuously improved through theoretical calculations and algorithmic predictions. Performance.

[0226] Parameter layer optimization:

[0227] When optimizing chemical reaction conditions, we typically need to consider multiple continuous variables (also known as factors, such as temperature, pressure, concentration, and time). Different combinations of these variables will affect the response value we are interested in. This layer involves fine-tuning continuous parameters such as temperature, pressure, concentration, and time, which is a multi-parameter collaborative optimization problem.

[0228] Central Composite Design (CCD): CCD is an experimental design method used to efficiently construct second-order response surface models, thereby accurately estimating the influence of individual variables and their interactions on the response value with a relatively small number of experiments. Its model formula is:

[0229]

[0230] Where Y is the target response (e.g., yield, selectivity). These represent the encoded values ​​of the i-th and j-th factors. (For temperature T, the encoded value is...) = (T - T0) / ΔT, where T0 is the center point temperature and ΔT is the step size. This is a constant term (an estimate of the global average response). The coefficient for the linear term represents the main effect (linear influence) of factor i. The coefficient of the squared term represents the curvature effect (nonlinear effect) of factor i. This represents the interaction effect (synergistic or antagonistic effect) between factors i and j, where K is the number of factors (e.g., if temperature, pressure, and concentration are optimized simultaneously, then K=3). For example, when k=3, the combinations are: i=1,j=2; i=1,j=3; i=2,j=3) (machine error term). To iterate through all the different factors and combine them in pairs (e.g., when k=3, the combinations are: i=1,j=2; i=1,j=3; i=2,j=3).

[0231] The value of the fitted second-order model lies in its coefficients. Positive coefficients in the first-order terms indicate that an increase in this factor leads to an increase in the response value (e.g., yield); negative coefficients indicate the opposite. The quadratic coefficients represent the curvature effect of this factor. Negative quadratic coefficients indicate the existence of a maximum value, meaning that the factor is not necessarily better the higher or lower it is, but rather achieves its optimal effect at a certain coding level (e.g., near 0). The interaction coefficients are the essence of CCD, quantifying the synergistic or antagonistic effects between factors. For example, a positive β... 12 (Corresponding to X1X2) indicates that when both factors are at a high or low level at the same time, they have a positive synergistic effect on the response value; a negative value indicates that a high level of one factor requires a low level of the other factor to cooperate.

[0232] By analyzing these coefficients and the generated model, the software can generate response surface plots (3D) and contour plots (2D). From the contour plots, it can be seen that if the contour lines are elliptical, it usually indicates a significant interaction between the two factors; if they are circular, it indicates a weak interaction. By taking the partial derivative of the second-order model and setting it to zero, the optimal parameter combination (coded values) can be mathematically solved, and then the actual process conditions (e.g., temperature 215°C, pressure 12.5 MPa) can be calculated using the coded formula. Any set of process conditions can be substituted into the model to predict its theoretical yield or performance, greatly reducing the cost of experimental trial and error.

[0233] By designing and calculating experimental points using CCD experiments, and fitting the coefficients of the aforementioned model through regression analysis, the model can be used to predict optimal reaction conditions and analyze the influence of various factors and their interactions on the yield. Optimization can then be used to improve the feasibility of the synthetic pathway and better guide downstream processes.

[0234] 8. Connect to the automated robot workstation module.

[0235] The output of an AI decision engine is a standardized synthesis protocol that is readable by machines. It typically contains the following structured information, detailed in Table 3.

[0236] Table 3 Structured Information

[0237]

[0238] This protocol allows direct API communication to automated robotic workstations (such as the AI-EDISON platform), driving robotic arms, liquid processors, reactors, and other equipment for fully automated synthesis. The bill of materials requires precursor names, molecular formulas, purity, and mass / volume to ensure accurate weighing and avoid side reactions. The reaction sequence requires step numbers, operation types (heating, stirring, dropping), and target parameters, clearly defining the execution order. Equipment control specifies reactor type, temperature program, stirring rate, and pH settings, directly controlling the automated hardware. Online sensors (pH, temperature, spectrum) and sampling points provide real-time feedback and quality control. Simultaneously, it monitors maximum temperature / pressure thresholds and issues warnings for hazardous operations to ensure experimental safety.

[0239] 9. Closed-loop optimization and continuous learning module

[0240] The system is not a one-time operation. The results synthesized by the automated workstations (yield, purity, characterization data) form a feedback loop to continuously optimize the AI ​​model.

[0241]

[0242] New data This is used to update previous models, thus forming an autonomous closed loop of "design-synthesis-testing-learning" to continuously iterate and improve the success rate of the synthesis strategy.

[0243] Automated closed-loop feedback system

[0244] The system employs Expectation Improvement (EI) intelligent exploration of new catalyst formulations, with Bayesian optimization guiding experiments through EI functions, significantly accelerating the catalyst development process. Its workflow clearly demonstrates how to iteratively search for the optimal catalyst formulation:

[0245]

[0246] In the formula, m represents a set of candidate catalysts. The unknown true performance (e.g., activity) of candidate point m. E represents the best performance currently tested, and E is the expected value. The AI ​​agent model (such as a Gaussian process) predicts the probability distribution of f(m) and calculates the average of all possible improvements.

[0247] Given a small number of known catalyst formulations and performances in the initial dataset, a surrogate model (such as a Gaussian process) is constructed to optimize the acquisition function (EI). The candidate formulation m with the largest EI is calculated. The candidate catalyst m is experimentally verified, synthesized, and tested. The performance f(m) of the new data points is obtained. The new data is added to the dataset, and the performance is tested to see if it meets the requirements. If it does, the optimal catalyst formulation is output. If it does not, the process returns to continue constructing an algebraic model from the multimodal pre-trained model. This process is repeated until the optimal catalyst formulation is output, thus maximizing the acquisition function.

[0248] By collecting experimental results, we continuously feed back and optimize upstream AI models, forming a self-improving flywheel. We build a cross-institutional collaborative network to share knowledge without data leaving its domain, increasing model iteration speed by 3 times and continuously raising the synthesis success rate.

[0249] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0250] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A catalyst intelligent design method based on multimodal fusion and causal inference, characterized in that, Includes the following steps: Step 1: Construct a multimodal pre-trained Catal-PTM model. The multimodal data includes heterogeneous data from multiple sources, including fused text, images, spectral data, and experimental parameters, to form a digital characterization of the catalyst's properties and performance. The priority attributes of each property are ranked by the QSAR model and a linear function is calculated. Based on the general direction of the linear function, a machine learning prediction model is used to calculate complex functions of activation energy, selectivity and adsorption energy, which provide structural and coordination requirements for catalyst generation. Step 2: Generate a preliminary catalyst structure. Input the required data for the multimodal catalyst into the Catal-PTM model for digital characterization to generate a preliminary catalyst structure. Step 3: Obtain the constraints for catalyst generation, input multimodal data into the causal discovery and intervention module, so that the causal discovery and intervention module can identify key causal paths and direct the causal constraint module to perform directional constraint training, and output the causal constraint information structure. Step 4: The generated candidate catalyst structure is input into the generative design engine by taking the causal constraint information output in Step 3 and the processed catalyst information output by the multimodal pre-training module trained in Step 2. The generative design engine works in series with GAN and diffusion model. GAN generates candidate catalyst structures with diversity and rationality based on upstream information. The candidate catalyst structure is used as a "draft" or conditional input for the diffusion model; The diffusion model is refined and optimized to produce the final design results. Step 5: Generate reaction parameters and synthesis pathways for candidate catalysts. Import the candidate catalyst structures output in Step 4 into the AI ​​decision engine, which will generate the module for optimizing reaction parameters and the optimal synthesis pathway. While receiving the upstream structure, the reaction parameters and synthesis pathways will be designed. Step 6: Generate candidate catalysts. Input the generated candidate catalyst structure, optimal reaction route and parameters into the automated robotic synthesis workstation. The automated workstation will generate a small number of candidate catalysts. Step 7: Candidate catalyst testing and data processing and storage. The generated candidate catalysts are tested using high-throughput standard equipment and the data is collected and fed back into the database. Step 8: Closed-loop optimization to obtain the optimal catalyst formulation. The optimal catalyst formulation is explored through the expectation improvement EI intelligent exploration. For the optimal solution catalyst that meets the requirements, its formulation and synthesis route are output. For the formulation data that does not meet the requirements, it is re-introduced into the original multimodal preprocessing model in a closed loop to continue closed-loop optimization until the optimal catalyst formulation is generated.

2. The catalyst intelligent design method based on multimodal fusion and causal inference according to claim 1, characterized in that, The method for constructing the multimodal pre-trained Catal-PTM model in step one is as follows: Graphical-textual comparison learning: enables the model to learn to associate representational maps with textual descriptions or crystal structures; Masked language / graph modeling: Randomly masking parts of text words or graph data segments, allowing the model to make predictions and learn deep features; The pre-trained multimodal Catal-PTM model can output a unified representation vector h for any catalytic data, which can be used for downstream prediction tasks; its loss function is: ； It is a masked language model loss, used for text modality, to enable the model to learn to reason and deduce based on context, to enable the model to deeply understand text semantics, and to learn the language rules in the field of chemistry; It is a contrastive learning loss, used to bring the representations of different modes of the same catalyst closer together in the vector space, and vice versa. It is a cross-modal reconstruction loss, which ensures that the representation retains complete information, requiring the model to be able to reconstruct the representation from one modality; These are the balancing hyperparameters, consisting of three balancing coefficients used to adjust the importance of different loss terms in the total loss, and can be adjusted according to the actual task requirements.

3. The catalyst intelligent design method based on multimodal fusion and causal inference according to claim 1, characterized in that, The method for obtaining the catalyst performance function in step one is as follows: S3.1 Constructing a QSAR quantitative structure-activity relationship model ； In the formula, y represents the performance of the catalyst, x1, x2, ... represent the factors affecting the catalyst activity, and w1, w2, ... represent the weights of each factor. ϵ is the intercept term, also known as the deviation, and ϵ is the error term; S3.2, Quantitative Structure-Activity Relationship Model: Predictive Model After Machine Learning Catalytic Performance( )= +ϵ; In the formula, y represents the target property to be predicted, Descriptori describes the characteristics of the catalyst, f() is the complex nonlinear function learned by the machine learning model, and ϵ is the error term.

4. The catalyst intelligent design method based on multimodal fusion and causal inference according to claim 1, characterized in that, The method for obtaining the constraints on catalyst generation in step three is as follows: S4.1 Constructing a Causal Discovery and Intervention Model (1). Structural Causal Model (SCM) ； in, yes The set of parent nodes, It is an unobserved exogenous variable; (2). Causal discovery The module runs the PC (Peter-Clark) algorithm, which examines whether each pair of variables is independent given a subset of other variables, thereby gradually removing edges and determining the causal direction. The goal of causal discovery is to automatically find the set of parent nodes from the observed data, reconstruct the complete causal graph, and infer the causal relationship between variables. (3) Causal intervention When a set of variables Z satisfies the backdoor criterion with respect to (X,Y), the intervention formula can be expressed as: ； in For the intervention distribution, it represents the probability that variable Y takes the value y when variable X is forcibly set to the value x; This indicates an intervention operation, meaning that an external force forcibly sets the variable X to a specific value x, thereby severing all edges pointing to X and making X no longer affected by its original cause; To satisfy the backdoor criterion, the set of mixed variables in the catalyst can be represented as "precursor concentration" and "calcination atmosphere". It represents the conditional probability of outcome Y occurring given cause X and confounding factors Z; To estimate the intervention effect, it is necessary to examine the influence of X on Y within each hybrid layer Z=z, and then take a weighted average of the results of all layers. S4.2 Causal Constraints When generating new catalyst structures, a causal regularization term is introduced into the generator loss function: constraining GAN. ； Where G is the generator, whose goal is to capture the data distribution of real data. The input is usually random noise z, and the output is the generated data. In catalyst design, G's goal is to generate novel and theoretically efficient catalyst atomic structures. D is the discriminator, and its goal is to estimate the probability that a sample is real from the training data rather than generated by G. In catalyst design, the task of D is to determine whether a given atomic structure is a real, known, and highly efficient catalyst, or a "fake" structure generated by G. E represents the expected value, which is the average of all possible outcomes. These are samples collected from the real data distribution; Z represents the noise sample first sampled from the prior noise distribution; To correct the parameters, It is the standard loss function for Generative Adversarial Networks (GANs), enhancing the adversarial relationship between the generator G and the discriminator D; in machine learning loss functions, where... It is a penalty item that penalizes the generated results that violate the identified causal relationship.

5. The intelligent catalyst design method based on multimodal fusion and causal inference according to claim 1, characterized in that, In step four, the generative design engine optimizes and generates designs in a way that satisfies performance targets, guided by the loss function. Specifically, this includes: (1) Standard GAN ； Where G is the generator. Given a sample of real data, input a random noise vector z and a performance condition c; D is the discriminator, which determines whether a catalyst comes from the real dataset or is faked by the generator, and evaluates whether it meets condition c. As a condition, Given a potential space vector, by adjusting z, the generator can produce different catalyst variants; To limit losses, To constrain weights; (2) Noise prediction loss module for diffusion model The generated candidate catalysts need to be evaluated quickly. The diffusion model generates samples by learning the data distribution, and its reverse process is used to evaluate the "uncertainty" of catalyst performance. The training objective of the diffusion model, i.e., the noise prediction loss function, is as follows: ； in The original data point represents a catalyst structure with ideal and well-defined performance; in catalyst design, it represents a real high-performance catalyst sequence or a stable catalytic active site structure. The diffusion time step represents the degree of disturbance or uncertainty. The larger the catalyst is, the greater the "interference" to its structure, and the higher the uncertainty of its performance. The noise added is randomly sampled Gaussian noise, an external uncertainty that affects catalyst performance; It is a noise prediction network, representing a performance prediction model. Its task is to predict the deviation noise of the performance of a catalyst under "uncertainty" noise interference from the ideal state. The trained model can quickly predict the performance based on the catalyst's descriptor. An ideal catalyst is a descriptor of its molecular formula, crystal structure, protein amino acid sequence, or active site. (3) The concatenation of GAN and diffusion model First, leveraging the speed advantage of GANs, a large number of candidate catalyst structures with a certain degree of diversity and rationality are rapidly generated based on existing causal constraints and multimodal inputs. Then, these candidate structures are used as the "draft" or conditional input of the diffusion model, which is then refined and optimized to finally output high-quality design results.

6. The intelligent catalyst design method based on multimodal fusion and causal inference according to claim 1, characterized in that, The reaction parameters and synthesis route for generating the candidate catalyst in step five are as follows: The AI ​​decision engine generates optimized reaction parameters and the optimal synthesis route through constrained optimization and sequence decision-making, given the target catalyst structure. Find the optimal synthesis path It can be represented as: ； Represents the set of all possible composition paths. , , These represent the cost, time risk, and security risk of the path, respectively. This represents the catalyst structure actually synthesized via path S. It is a loss function representing the difference between the synthesized result and the target structure, ensuring that the synthesized product is the designed structure. These are weighting coefficients, used to balance the importance of different objectives; Parameter layer optimization: When optimizing chemical reaction conditions, it is necessary to consider multiple continuous variables; different combinations of variables will affect the response value. Central Composite Design (CCD): CCD is an experimental design method used to construct second-order response surface models. It allows for accurate estimation of the influence of various variables and their interactions on the response value with a relatively small number of experiments. Its model formula is: ； Where Y is the target response. The encoded values ​​representing the i-th and j-th factors; For constant terms, The coefficient of the linear term represents the main effect of factor i. The coefficient of the squared term represents the curvature effect of factor i. This represents the interaction effect between factors i and j, where K is the number of factors. For example, when k=3, the combinations are: i=1,j=2; i=1,j=3; i=2,j=3) (machine error term). To iterate through all the different factors and combine them in pairs (e.g., when k=3, the combinations are: i=1, j=2; i=1, j=3; i=2, j=3); The value of the fitted second-order model lies in its coefficients. A positive coefficient for the first-order term indicates that the response value increases as the factor increases, while a negative coefficient indicates the opposite. The coefficient for the squared term represents the curvature effect of the factor. A negative squared term indicates the existence of a maximum value, meaning that the factor is not necessarily better the higher or lower it is, but rather achieves the best effect at a certain coding level. After calculating experimental points by designing CCD experiments and fitting the coefficients of the above model through regression analysis, the model is used to predict the optimal reaction conditions and analyze the influence of each factor and its interaction on the yield.

7. The catalyst intelligent design method based on multimodal fusion and causal inference according to claim 6, characterized in that, The method for generating the candidate catalyst in step six is ​​as follows: The output of the AI ​​decision engine is a standardized synthesis protocol that is machine-readable and contains structured information; This protocol can be directly transmitted to automated robot workstations via API interface to drive robotic arms, liquid processors, and reactors for fully automated synthesis. The bill of materials requires the precursor name, molecular formula, purity, and mass / volume to ensure accurate weighing and avoid side reactions; The reaction sequence requires step numbers, operation types (heating, stirring, dropping), target parameters, and a clearly defined execution order; The equipment control requirements include reactor type, temperature program, stirring rate, and pH value settings, which are directly controlled by the automated hardware. Using online sensors (pH, temperature, spectrum) and sampling points, it provides real-time feedback and quality control; it also detects maximum temperature / pressure thresholds and issues warnings for hazardous operations.

8. The intelligent catalyst design method based on multimodal fusion and causal inference according to claim 7, characterized in that, The method for obtaining the optimal catalyst formulation through closed-loop optimization in step eight is as follows: The results synthesized by the automated workstation (yield, purity, characterization data) form a feedback loop to continuously optimize the AI ​​decision engine model. ； New data This is used to update previous models, forming an autonomous closed loop of "design-synthesis-testing-learning" to continuously iterate and improve the success rate of the synthesis strategy; Automated closed-loop feedback system: The system employs an expected improvement EI intelligent exploration method to discover new catalyst formulations, and Bayesian optimization guides experiments through EI functions to iteratively find the optimal catalyst formulation. ； In the formula, m represents a set of candidate catalysts. Let m be the unknown true performance of candidate point m. Given the best performance achieved so far, and E as the expected value, the AI ​​agent model predicts the probability distribution of f(m) and calculates the average of all possible improvements. Given a small number of known catalyst formulations and performances in the initial dataset, a surrogate model is constructed to optimize the acquisition function EI. The candidate formulation m with the largest EI is calculated. The candidate catalyst m is experimentally verified, synthesized, and tested. The performance f(m) of the new data points is obtained. The new data is added to the dataset, and the performance is tested to see if it meets the requirements. If it does, the optimal catalyst formulation is output. If it does not, the process is repeated to return to the multimodal pre-trained model to construct an algebraic model. This process is repeated until the optimal catalyst formulation can be output, thus maximizing the acquisition function.

Images