Function-driven material structure
Through functional characterization methods and density functional theory combined with machine learning algorithms, the catalyst surface structure is predicted, which solves the problem of predicting complex catalyst surface structure, and realizes accurate prediction of catalyst active sites and performance optimization, reducing experimental costs.
Patent Information
- Application Number
- CN202380032321.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2022-04-08
- Filing Date
- 2023-04-10
- Publication Date
- 2024-12-06
AI Technical Summary
The prior art is difficult to effectively predict the surface structure and active sites of complex multi-element catalysts, which leads to difficulty in predicting catalytic performance and high cost of experimental structure analysis. The dynamic surface structure evolves during the reaction process, making it difficult to accurately predict catalytic performance.
By using functional characterization method, the functional characteristics of the unit cell are calculated based on density functional theory and machine learning algorithms, the accumulated functional characteristics match the experimental measurement results, and iteratively optimize the surface structure configuration to generate High-quality training data sets to predict the surface configuration of the catalyst.
It realizes efficient prediction of the catalyst surface structure, reduces the cost and complexity of experimental structure analysis, and can accurately predict the active sites and catalytic properties of the catalyst, which is suitable for the prediction of electrocatalytic activity of polymetallic nanoparticles.
Smart Images

Figure CN119096253A_ABST
Abstract
Description
[0001] CROSS-REFERENCE TO RELATED APPLICATIONS
[0002] This utility patent application claims priority to U.S. Provisional Application No. 63 / 329,326, filed on April 8, 2022, entitled “FUNCTION-INFORMED MATERIALS STRUCTURES,” the contents of which are incorporated herein by reference in their entirety.
[0003] background
[0004] The present disclosure relates to apparatus, systems and methods for material structures. More specifically, the present disclosure relates to apparatus, systems and methods for function-informed material structures.
[0005] Exploring the ultra-large-scale and multi-dimensional design space of complex multi-element materials to find disruptive catalysts that were previously out of reach can be extremely challenging, for example, given the combinatorial explosion of possibilities when considering combinations of constituent elements such as stoichiometry, size, shape, and structure. Smart strategies leveraging high-throughput experiments, computation, and machine learning (ML) can help overcome serendipity and / or improve the success rate of materials discovery. Summary of the invention
[0006] The present application discloses one or more of the features recited in the appended claims and / or the following features which alone or in any combination may constitute patentable subject matter.
[0007] According to one aspect of the present disclosure, a method for predicting material structure information based on functional characterization may include: (a) representing the surface of a material as an ensemble of unit cells; (b) determining a pool of possible unit cells based on one or more material input properties; (c) calculating the functional properties of the unit cells in the pool; (d) determining a combination of unit cells in the pool to represent the potential surface structure of the material, and calculating the corresponding cumulative functional properties of the material based on the previously calculated functional properties of the individual unit cells in the combination; and (f) verifying whether the calculated cumulative functional properties match at least one experimental measurement result of the same functional property of the material.
[0008] According to another aspect of the present disclosure, a method for predicting material structure information based on functional characterization may include: (a) representing the surface of a material as an ensemble of unit cells; (b) determining a pool of possible unit cells based on one or more material input properties; (c) calculating the functional properties of the unit cells in the pool; (d) determining a combination of unit cells from the pool to represent the potential surface structure of the material, and calculating the corresponding cumulative functional properties of the material based on the previously calculated functional properties of each unit cell in the combination; and (f) verifying whether the calculated cumulative functional properties match at least one experimental measurement result of the same functional property of the material. In some embodiments, the method may include (e) repeating steps (a)-(f) with respect to the material surface of at least one other material to generate a data set containing the cumulative functional properties of each verified material. The data set can be configured to train a global machine learning algorithm for predicting the surface configuration of the unit cells of another material based on one or more material input properties of another material.
[0009] In some embodiments, calculating the functional properties may include at least one of: determining reactant adsorption energies and associated current densities associated with electrocatalysis, and determining electronic properties associated with optical or magnetic properties of the material. In some embodiments, determining the functional properties of the unit cell may be based solely on density functional theory (DFT) calculations, or on a combination of DFT and machine learning.
[0010] In some embodiments, a combination of unit cells representing a potential surface structure can be determined based on one or more of a Monte Carlo simulation and a machine learning algorithm characterized as one or more of a deep learning model, a generative adversarial network (GAN), a transformer model, a reinforcement learning model, and an integrated model. The machine learning model can include one of a random forest and a genetic algorithm.
[0011] In some embodiments, the global machine learning algorithm for predicting surface structure can be based on one or more of the following: a deep learning model, a generative adversarial network (GAN), a transformer model, a reinforcement learning model, and an integrated model. The global machine learning model can include one of a random forest and a genetic algorithm.
[0012] In some embodiments, verifying whether the calculated functional property matches at least one experimental measurement can include determining whether a prediction threshold is met. Determining whether the prediction threshold is met can include determining whether a difference between the calculated functional property and the functional property measured experimentally is within a predetermined value range. Verification can include determining that the calculated functional property does not match an experimental measurement of a material surface of the material, and repeating steps (a)-(f) until the underlying surface structure that produces the calculated functional property matches the at least one experimental measurement.
[0013] In some embodiments, the configuration for training the machine learning algorithm may not require performing steps (i)-(e) on a further material. One or more material input properties of a material may be defined only as a composition of the material. One or more material input properties of a further material may be defined only as a composition of the further material.
[0014] In some embodiments, determining a combination of unit cells may include predicting potential surface structures as a deterministic set of unit cells. In some embodiments, determining a combination of unit cells may include predicting potential surface structures as a probabilistic set of unit cells.
[0015] According to another aspect of the present disclosure, a system may include at least one processor that executes instructions stored in a memory to perform the method described above or below.
[0016] According to another aspect of the present disclosure, a method for predicting material structure information related to functional characterization may include: determining the functional properties of unit cells in a unit cell pool on a material surface; determining a predicted structure of the material surface based on the unit cell pool; and determining a predicted material activity based on the determined functional properties and the determined predicted structure, and outputting a characterization of the material structure information on the material surface based on the predicted material activity.
[0017] In some embodiments, determining the predicted material activity may include determining whether a threshold prediction of the material activity is achieved, and redetermining the predicted structure in response to determining that the threshold prediction is not achieved. The threshold prediction of the material activity may be determined by comparing the predicted material activity to experimental results. In response to determining that the threshold prediction is achieved, an output may be performed. The output representation may include a data set for training a machine learning model to predict the surface configuration of a unit cell of a new material based on material input properties.
[0018] The additional features, alone or in combination with any other features, including those described above and those listed in the claims, may constitute patentable subject matter and will become apparent to those skilled in the art upon consideration of the following detailed description of illustrative embodiments that illustrate what is presently believed to be the best way to practice the invention. The emerging paradigm of heterogeneous catalysis suggests that the surfaces of complex catalysts have a myriad of different active sites that define the adsorption energies of reactants and intermediates, and thus the activity and selectivity of the catalyst. The adsorption energy of an active site depends on its local electronic and geometric configuration and can be tuned. For example, such tuning can be achieved through strategies such as mixing new elements, changing size and shape, controlling crystal structure and / or oxidation state, and / or modifying the catalyst support. BRIEF DESCRIPTION OF THE DRAWINGS
[0019] The detailed description makes particular reference to the accompanying drawings, in which:
[0020] Figure 1 Comparison of the predicted and experimental ORR activity trends of Ag-Ir-Pd-Pt-Ru ML with that of Pt.
[0021] Figure 2 is an illustration of crystal structure phase mapping for a separation task, where a phase diagram is inferred from a set of XRD patterns in material composition space (a), requiring identification of pure phase prototypes and their composition-dependent modifications. Pattern #73 (d) shows the input (a, b) and output (cf), with each separation pattern shown in (c).
[0022] Figure 3 is an example of a workflow diagram for determining the right combination of unit cells with known composition, size, and experimental results. Unit cells can be generated from the composition, and DFT simulations can be run for each unit cell. The model can iteratively explore combinations of unit cells, and the predicted activity (based on DFT simulations) can be compared to experimental results. Once a suitable threshold is reached, the predicted structure can be paired with the input data. DETAILED DESCRIPTION
[0023] For the purposes of promoting an understanding of the principles of the present disclosure, reference will now be made to a number of illustrative embodiments shown in the drawings and specific language will be used to describe the same.
[0024] The optimal adsorption energy of a (perfect) catalytic site for a key reaction intermediate must be neither "too strong" nor "too weak". If the intermediate molecules bind too strongly to the surface, they can poison the catalyst. If they bind too weakly, the intermediate may be released from the active site prematurely, e.g. before undergoing a subsequent transformation, typically by electron or proton transfer. By optimizing the combination and number of optimal active sites for key reaction intermediates, the activity and / or selectivity of the catalyst for the desired chemical transformation can be increased.
[0025] The active site framework can provide a way to computationally describe catalytic processes and guide the discovery of new catalysts. However, ab initio simulations generally require knowledge of the surface structure and can therefore only be applied to a very limited subset of materials that have been synthesized and characterized. Traditional experimental characterization of the surface structure of complex catalysts can be very slow and / or may require state-of-the-art expensive instrumentation. Extending ab initio efforts to new and as-yet uncharacterized materials could benefit from first predicting the likely structure of each material, which poses a computationally expensive global optimization problem.
[0026] These problems may be further complicated by the fact that the surface structure of the catalyst is often dynamic and evolves during the reaction, e.g., to a new quasi-stable state that is different from the initial configuration. Therefore, knowing only the surface structure of the original catalyst is not sufficient to accurately predict its catalytic performance using ab initio simulations, but rather requires understanding the structure in the operando state. Unfortunately, the operando structure not only accounts for only a small fraction of the already very limited characterized structures, but its characterization may be technically more challenging. The devices, systems, and methods disclosed herein relate to methods for revealing the operando configuration of active sites on the surface of the catalyst based on high-throughput experimental functional characterization of the catalyst, computational modeling, and / or machine learning (ML), thereby avoiding extremely slow and / or expensive experimental structural analysis.
[0027] First principles density functional theory (DFT) calculations have been widely used to calculate the adsorption energies of reactants on catalyst surfaces. However, performing a large calculation for all exposed atoms and nearest neighbor atoms in a catalyst can be computationally expensive and may not be feasible for high-throughput workflows. Alternatively, the catalyst surface can be represented by a combination of smaller (e.g., 4x4x2) atomic unit cells. Possible unit cells for a given catalyst composition can be constructed by assigning a random selection of constituent elements to each lattice site.
[0028] Once a possible set of unit cells for the catalyst has been determined, DFT can be used to calculate the corresponding reactant adsorption energies for the entire set of unit cells or for a random subset of the unit cells, and then ML can be used to predict the remaining adsorption energies, either directly or through a combination of direct and ML measurements. The adsorption energy, in turn, is related to the current density of the unit cell and, therefore, to the electrocatalytic activity of the unit cell.
[0029] For example, Pedersen et al. previously showed that the current ji at surface site i can be modeled using the Koutecky-Levich equation (1), where jD represents the diffusion-limited current and jk,i represents the kinetically-limited current (2).
[0030]
[0031] Here ΔE is the adsorption free energy of the reactant, ΔEopt is the optimal adsorption free energy of the reactant given by the Sabatier principle, e is the elementary charge, U is the potential applied relative to the reversible hydrogen electrode (RHE), and k B is the Boltzmann constant and T is the absolute temperature. The total current density of the unit cell is then the sum of the current densities of all active sites N on the unit cell surface (3), and the total current density of the catalyst is the sum of the current densities of all units weighted by the probability of finding each unit cell type.
[0032]
[0033] This framework has been successfully used to predict the electrocatalytic activity of multi-metallic thin films in the oxygen reduction reaction (ORR) Figure 1 ) 1 . However, for thin films, surface segregation effects are negligible and the surface composition at each pixel of the substrate can be assumed to be the same as the bulk composition at the time of deposition and is therefore known. For discrete 3D complex nanomaterials, the situation is quite different. Multimetallic nanoparticles can form heterogeneous structures in which some constituent elements are alloyed with other constituent elements and phase-separated. Without a robust and accurate surface structure prediction strategy, it will be challenging to use such a framework for off-the-shelf multimetallic nanoparticle electrocatalysts.
[0036] One way to solve this bottleneck is to reconstruct the surface structure of the catalyst by mapping the corresponding predicted current density to the experimental measurement value. That is, the possible surface configuration of the unit cell of the catalyst with a limited composition and size can be generated by a computer, and the corresponding current density can be calculated as described above. In an illustrative embodiment, the calculated current density is compared with the value measured by the experiment, and if it is not within a predefined confidence interval, a new unit cell surface configuration is generated, and a new theoretical current density value is calculated. This iterative process can be repeated until the calculated value and the experimental value are consistent, at which time the mapping can be advanced to the next component.
[0037] A variety of computational and machine learning methods can be deployed at this stage, with varying degrees of efficiency and scalability. For example, the most likely surface configuration can be calculated using Monte Carlo simulations. Iterative optimization of the surface configuration to match the measurements can benefit from, for example, reinforcement learning or generative adversarial network modeling. The latent space of the ML model can be designed to incorporate prior scientific knowledge and fundamental constraints.
[0038] Regardless of the combination of computational and machine learning methods used, the goal is to train a model that can accurately predict the distribution of active sites on a catalyst surface based on composition, size, and experimental parameters without the need for tedious structural characterization. However, to achieve this ambitious goal, large, high-quality training functional datasets may be required.
[0039] Constrained ML for reconstructing composite data from data about individual components that combine in unknown ways is not necessarily new. For example, deep learning has previously been combined with constrained reasoning to enable the identification of crystalline phases or their mixtures in X-ray diffraction measurements of synthetic materials ( Figure 2 ) 2 Phase mapping of crystal structures is automated. In this case, the X-ray diffraction pattern of the crystal phase mixture resembles the surface structure of the catalyst, and the individual crystal phases resemble the unit cell.
[0042] Although examples of some fragmentary aspects of the proposed concept may already exist in the literature and sometimes in other fields; to verify the feasibility of the idea, combining these aspects can be done in a new and non-obvious way, namely using DFT to calculate the current density for a given unit cell and reconstruct the complex catalyst surface of the individual catalysts from the unit cell in a manner similar to solving the phase separation problem in crystal structures. Furthermore, extending the concept beyond just reconstructing the surface of a single catalyst can provide the ability to build a model that predicts the surface configuration of any catalyst within the design space. Figure 3 A schematic diagram of an example workflow is provided.
[0043] like Figure 3 As shown, an exemplary iterative workflow begins with a set of elements within a material space labeled "Material Space". The next step in the process utilizes the set of elements within "Material Space" to "Generate Cell Configurations" that contain the complete combinatorial set of elements in "Material Space". All generated unit cell configurations can then be filtered in a "Cell Filtering" process, which can reduce the number of potential cells used in the iterative process and create a "Cell Pool".
[0044] Density Functional Theory (DFT) simulations can be performed for each unit cell in the "Unit Cell Pool" in the process labeled "Run DFT by Unit Cell" to generate "DFT Results", which contains the activity predictions for each unit cell. Next, the "Model" can take the inputs of "Composition", "Nanoparticle Size", and "Unit Cell Pool" to generate an initial structure in the process block labeled "Generate Structure", which contains a set of deterministic / probabilistic structures of the unit cell labeled "Predicted Structure" for a given "Composition" and "Nanoparticle Size".
[0045] The "Calculate Estimated Catalytic Activity" process block can add the predicted activities from the "DFT Results" for each unit cell in the unit cell present in the "Predicted Structure" to calculate the total predicted activity (labeled "Predicted Activity") for all unit cells within the ensemble structure. The "Predicted Activity" can be compared to the "Electrochemical Experimental Results" in the "Error Calculation" process to score the "Predicted Activity". If this "Error Calculation" is above a specified threshold, another iteration can be performed to generate a new structure and repeat the evaluation. This workflow can be iterated until the "Error Calculation" is below the target threshold.
[0046] Once a satisfactory deterministic / probabilistic structure is predicted within this workflow, the “composition”, “nanoparticle size”, “electrochemical experimental results”, and “predicted structure” can be paired together as the “final data point”. At this point, the workflow can start again for the next nanoparticle within the “material space” and can be repeated until the desired training dataset for the global machine learning model is generated. The global machine learning model pre-trained on the generated dataset can minimize or even completely eliminate the need for additional computational modeling and / or workflows, as shown in Figure 3 As mentioned and suggested in , it is used to predict the (surface) structure of nanomaterials.
[0047] Although sometimes described in the context of nanoparticles and electrochemical activity, Figure 3 The workflow mentioned and suggested is for reference, but in some embodiments, this approach can be translated to other classes of nanomaterials, including nanofilms, nanosheets, core-shell nanostructures, quantum dots, halide perovskites, metal-organic frameworks, and other electronic, catalytic, optical, and mechanical properties.
[0048] Within the present disclosure, the electronic band structure of a material can provide information about its electronic properties and can be used to predict properties such as conductivity, semiconductor behavior, and the presence of a band gap. The density of states can represent the number of electronic states that an electron can occupy at a given energy level. Therefore, it can provide insight into the electronic structure and help predict properties such as conductivity and / or light absorption.
[0049] Charge carrier mobility can be an important property of electronic materials because it affects the speed at which charge carriers (e.g., electrons or holes) move through the material under an applied electric field. For magnetic materials, properties such as magnetic moment, exchange interaction, and / or magnetic anisotropy can be calculated to understand and / or predict the magnetic behavior of the material. Dielectric properties such as dielectric constant and / or loss tangent are very important for materials used for capacitors, insulators, and / or other applications for storing and / or transmitting electrical energy. Providing a calculated evaluation of chemical stability can involve examining the reactivity of the material and / or resistance to chemical degradation, corrosion, and / or oxidation. The adsorption energy associated with the adsorption of reactants onto the catalyst surface is important for evaluating the activity of the catalyst. Strong adsorption can promote the breaking of reactant bonds, while weak adsorption can lead to poor catalytic performance. Computational models can be used to predict the adsorption energy of various reactants on the catalyst surface. The electronic properties of the catalyst, such as the density of states and / or the position of the Fermi level, can affect its catalytic performance. For example, in heterogeneous catalysis, electronic properties can affect the adsorption energy of reactants and / or the activation energy barrier of the reaction step.
[0050] In the present disclosure, a deterministic method for defining a unit cell may include precisely selecting which unit cells are present on the surface in what proportion, for example, each unit cell is present or absent. Additionally or alternatively, a probabilistic method for defining a unit cell may include a probability profile of a particular unit cell or cells that can be found on the surface within a certain probability, and in some embodiments, different unit cells may have different acceptable probabilities.
[0051] The concepts within this disclosure can invert the current paradigm from structure to function, and instead describe how to use large-scale functional data to gain structural insights about materials. Although the designs within this disclosure have been described using electrocatalysis as an example, this inverted paradigm can be applied to any structure-function material relationship.
[0052] Within the present disclosure, functional screening methods may include: scanning electrochemical methods (scanning droplet cell, scanning electrochemical cell microscopy, scanning electrochemical microscopy), optical detection methods (fluorescence / phosphorescence on / off, electrochromic detection), spatially isolated parallel experiments (microwell arrays, microelectrode arrays, parallel backing bed reactors), spectroscopic methods (scanning Raman, IR thermal imaging, UV-VIS, (AT)-FTIR, high-throughput NMR), parallel / scanning product collection (microfluidics, scanning droplets, capillary probes), and product analysis (mass spectrometry, NMR, gas chromatography, liquid chromatography, IR, UV-VIS).
[0053] Data generated by the functional screening method may include: current-potential trace, current density, onset potential, overpotential, optical image, fluorescence image, fluorescence intensity, transparency, color, conductivity, mass spectrum, NMR spectrum, Raman spectrum, IR spectrum, UV-VIS spectrum, IR thermography, product / reagent ratio, product conversion rate, turnover rate, by-product formation rate, temperature. Data generated by the functional screening method may be related to one or more materials. Data generated by the functional screening method may be related to a specific location or area on the sample. Data generated by the functional screening method can be measured over time.
[0054] The following references are incorporated by reference in their entirety and include at least the specific portions mentioned herein: Batchelor, TAA et al, Complex-Solid-Solution Electrocatalyst Discovery by Computational Prediction and High-Throughput Experimentation**. Angew. Chem., Int. Ed. 60, 6932-6937 (2021); Chen, D. et al, Automating crystal-structure phase mapping by combining deep learning with constraint reasoning. Nature Machine Intelligence. 3, 812-822 (2021).
[0055] Although the present disclosure has been shown and described in detail in the foregoing drawings and description, it should be considered illustrative rather than restrictive, and it should be understood that only illustrative embodiments thereof have been shown and described and that all changes and modifications within the spirit of the disclosure are desired to be protected.
Claims
1. A method for predicting material structure information based on functional characterization, the method include: (a) Representing the surface of a material as a collection of unit cells; (b) determining a pool of possible unit cells based on one or more material input properties; (c) calculating functional properties of the unit cells within the cell; (d) determining a combination of unit cells from the pool that represent the underlying surface structure of the material, and calculating a corresponding cumulative functional property of the material based on the previously calculated functional properties of the individual unit cells in the combination; (f) verifying whether the calculated cumulative functional property matches at least one experimental measurement of the same functional property of the material; Repeat steps (a)-(f) with respect to the material surface of at least one other material to generate a data set comprising the cumulative functional properties of each verified material, wherein the data set is configured to train a global machine learning algorithm for predicting the surface configuration of a unit cell of another material based on one or more material input properties of the other material.
2. The method according to claim 1, in, Calculating the functional properties includes at least one of determining reactant adsorption energies and associated current densities associated with electrocatalysis, and determining electronic properties associated with optical or magnetic properties of the material.
3. The method according to claim 1, in, The functional properties of the unit cell are determined based on density functional theory (DFT) calculations alone, or on a combination of DFT and machine learning.
4. The method according to claim 1, in, A combination of unit cells representing the underlying surface structure is determined based on one or more of a Monte Carlo simulation and a machine learning algorithm characterized as one or more of a deep learning model, a generative adversarial network (GAN), a transformer model, a reinforcement learning model, and an integrated model.
5. The method according to claim 4, in, The machine learning model includes one of a random forest and a genetic algorithm.
6. The method according to claim 1, in, The global machine learning algorithm for predicting the surface structure is based on one or more of the following: a deep learning model, a generative adversarial network (GAN), a transformer model, a reinforcement learning model, and an integrated model.
7. The method according to claim 6, in, The global machine learning model includes one of a random forest and a genetic algorithm.
8. The method according to claim 1, in, Verifying whether the calculated functional property matches at least one experimental measurement includes determining whether a prediction threshold is met.
9. The method according to claim 8, in, Determining whether the prediction threshold is met includes determining whether a difference between the calculated functional property and the experimentally measured functional property is within a predetermined range of values.
10. The method according to claim 1, in, Verification includes determining that the calculated functional property does not match an experimental measurement of a material surface of the material, and repeating steps (a)-(f) until the underlying surface structure that produces the calculated functional property matches at least one experimental measurement.
11. The method according to claim 1, in, The configuration for training the machine learning algorithm does not require performing steps (i)-(e) on the further material.
12. The method according to claim 1, in, One or more material input properties of the material are defined only as a composition of the material.
13. The method according to claim 1, in, One or more material input properties of the further material are defined only as a component of the further material.
14. The method according to claim 1, in, Determining a combination of unit cells includes predicting the potential surface structure as a deterministic set of unit cells.
15. The method according to claim 1, in, Determining a combination of unit cells includes predicting the potential surface structure as a probabilistic set of unit cells.
16. A system, wherein the system include: At least one processor, the at least one processor executing instructions stored in the memory to perform the method of claim 1.
17. A method for predicting material structure information related to functional characterization, the method include: Determining the functional characteristics of the unit cells in the unit cell pool on the surface of the material; Determining a predicted structure of the surface of the material based on the unit cell pool; as well as A predicted material activity is determined based on the determined functional properties and the determined predicted structure, and a representation of material structural information of the material surface is output based on the predicted material activity.
18. The method according to claim 17, in, Determining the predicted material activity includes determining whether a threshold prediction of the material activity is achieved, and redetermining the prediction structure in response to determining that the threshold prediction is not achieved.
19. The method according to claim 18, in, The threshold prediction for material activity is determined by comparing the predicted material activity with experimental results.
20. The method according to claim 18, in, In response to determining that the threshold prediction is achieved, outputting is performed.
21. The method according to claim 17, in, The output representation includes a data set used to train a machine learning model to predict the surface configuration of a unit cell of a new material based on the material input properties.
Citation Information
Cited By
Alloying layer-by-layer control method for multilayer metal films of DFB laser
CN122428114A
Alloying layer-by-layer control method for multilayer metal thin film of dbr laser
CN122428114B
