Computing device, method, computer storage medium and program product for optimizing chemical reaction conditions

Abstract

The present invention relates to a computer system using a computer model, and discloses a computing device, method, computer storage medium and program product for optimizing chemical reaction conditions. A computing device for optimizing chemical reaction conditions includes computing resources and multiple agents, and the multiple agents are executed by the computing resources. The multiple agents include a first agent, a second agent and a third agent. The first agent is configured to construct a hierarchical optimization tree to limit the search space. The second agent is configured to perform predictive reasoning on the search space. The third agent is configured to perform Bayesian optimization on the search space defined by the hierarchical optimization tree to determine the optimized chemical reaction conditions. The computing device according to the present invention overcomes the "cold start" and "curse of dimensionality" problems in existing chemical reaction condition optimization schemes, and can quickly and effectively obtain optimized chemical reaction conditions, improve the reaction performance of chemical reactions and reduce experimental costs.

Description

Technical Field

[0001] The present invention generally relates to computer systems utilizing computational models, and more particularly to computing devices, methods, computer storage media, and program products for optimizing chemical reaction conditions. Background Art

[0002] Chemical reactions are widely used in almost every field of modern society, including industrial production, pharmaceutical research and development, environmental protection, energy development and storage, food processing, agricultural cultivation, scientific research and teaching. Setting optimal chemical reaction conditions is crucial for improving chemical reaction performance, but it is also a very challenging task. Optimizing chemical reaction conditions often requires a repetitive cycle of chemical reaction experiment design, actual experimental execution, reaction condition testing, and reaction parameter analysis. It also requires the experimenter to accurately assess various reaction parameters (such as substrates, additives, solvents, concentrations, catalysts, and temperature), which is often time-consuming and costly, and places very high demands on the experimenter's expertise.

[0003] With the development of artificial intelligence technology, researchers have begun to focus on applying artificial intelligence technology to the optimization of chemical reaction conditions, and realize automatic optimization of chemical reaction conditions through methods such as machine learning modeling and prediction, generative artificial intelligence recommendations, etc., thereby improving the accuracy and efficiency of chemical reaction condition optimization and reducing experimental costs.

[0004] Some existing approaches use Bayesian optimization algorithms to optimize chemical reaction conditions, particularly for compound formulation screening, catalytic system design, and new material discovery (for example, see Malcolm Sim et al., "ChemOS 2.0: An orchestration architecture for chemical self-driving laboratories," DOI: 10.1016 / j.matt.2024.04.022). Bayesian optimization is a global optimization strategy consisting of two core components: a surrogate model, consisting of a prior distribution and an observation model, representing the unknown objective function. The acquisition function is used to obtain the next evaluation point after balancing exploration and exploitation. In this approach, researchers set the optimization objective, the surrogate model fits historical data using a Gaussian process, and the acquisition function recommends the next batch of variables for evaluation. These variables are then evaluated by a robot or human, and the evaluation results are returned. Bayesian optimization can rapidly predict experimental outputs and generate candidate solutions within a vast virtual experimental space. It then feeds limited real-world experimental results back to the model, enabling iterative self-learning and rapid convergence, significantly reducing the number and cost of actual experiments. However, in chemical reaction optimization scenarios, when the parameters of chemical reactions are multidimensional and the initial experimental data is extremely sparse, Gaussian processes are susceptible to cold starts and the "curse of dimensionality," resulting in slower convergence and lower quality recommendations. This limits the performance of Bayesian optimization in complex chemical scenarios.

[0005] Some existing approaches employ Monte Carlo tree search to recursively partition the search space through nonlinear partitioning and then employ Bayesian optimization to explore within these subdomains (see, for example, Linnan Wang et al., “Learning search space partition for black-box optimization using Monte Carlo tree search,” Advances in Neural Information Processing Systems, 2020). Monte Carlo tree search reduces the dimensionality of the search space, enabling Bayesian optimization to be applied to localized, high-potential subdomains, alleviating the “curse of dimensionality” problem caused by high-dimensional search spaces. However, in the context of chemical reaction optimization, the high-potential subdomains selected by Monte Carlo tree search may be ineffective for optimization due to the lack of chemical mechanism constraints. Iteratively performing Bayesian optimization in such ineffective subdomains wastes computational resources and time, increasing experimental costs. Furthermore, the “cold start” problem of Bayesian optimization remains unresolved.

[0006] Some existing approaches combine large language models (LLMs) with Bayesian optimization to address the cold-start problem in traditional Bayesian optimization (see, for example, Tennison Liu et al., "Large Language Models to Enhance Bayesian Optimization," arXiv preprint arXiv:2402.03921, 2024). Large language models generally use deep learning methods to train large neural network models to process and generate natural language text. Large language models typically use a knowledge base as background knowledge, enabling them to leverage the information in the knowledge base to perform natural language processing tasks such as text generation, machine translation, and intelligent question answering. In this approach, the Bayesian optimization problem is formulated in natural language, and the large language model provides highly accurate predictions based on real prior data to enable a warm-start of the Bayesian optimization to explore new candidates. Large language models can alleviate the cold-start problem of traditional Bayesian optimization to a certain extent, achieving superior performance in the early stages of experiments. However, large language models inherently suffer from hallucinations. In the context of chemical reaction optimization, especially in the absence of effective chemical mechanism constraints, recommendations based on large language models are subject to chemical hallucinations and quantization errors. Furthermore, chemical reaction optimization still requires Bayesian optimization over the high-dimensional search space of chemical reaction parameters, and the "curse of dimensionality" issue remains a serious problem.

[0007] Some existing approaches build hierarchical multi-agent system architectures, called ChemAgents, to autonomously complete the entire process from literature retrieval to experimental execution (see, for example, Tao Song et al., “A multi-agent-driven robotic AI chemist enabling autonomous chemical research on demand,” DOI: 10.26434 / chemrxiv-2024-w953h). A multi-agent system is a collection of agents that work together to solve complex problems, each playing different roles in the task. In this scheme, the multi-agent system ChemAgents includes four agents with specific roles: (1) a literature reader, which mines the literature database and provides knowledge related to the experimental objectives; (2) an experimental designer, which generates step-by-step experimental procedures based on the predetermined experiment and using the protocol library; (3) a robotic operator, which converts the step-by-step experimental procedures generated by the experimental designer into codes and commands for operating automated laboratory robots and experimental stations, facilitating and completing the robotic execution of chemical experiments; and (4) a computational executor, which searches for a suitable pre-trained machine learning model in the model library and uses the model to optimize the experiment. Through task decomposition and role collaboration, the scheme cycles between experimental planning, mechanism inference, and robotic execution, and has end-to-end experimental capabilities. However, in this scheme, only the computational executor is involved in the optimization of the experiment. The computational executor only involves searching for a suitable pre-trained machine learning model in the model library and combining the model with a Bayesian optimizer to determine the optimal MO-HEC composition. There is a lack of deep coupling between the agent and Bayesian optimization, and Bayesian optimization is still required to explore the huge chemical space.

[0008] In summary, existing solutions struggle with accurate optimization due to a lack of early experimental data. This results in an inability to make recommendations based on the fitted relationships, leading to a "cold start" problem and inefficient optimization. In the context of chemical reaction optimization, since reaction parameters often involve multiple dimensions (e.g., more than five), existing solutions require a large number of samples to converge in this high-dimensional search space. This "curse of dimensionality" leads to prohibitively high experimental costs and limited application. Furthermore, due to the lack of constraints on chemical mechanisms, the optimization process relies solely on historical data, resulting in wasted exploration of invalid or risky parameter search spaces and wasted experiments. Furthermore, while some solutions have made efforts to improve the performance of traditional Bayesian optimization, these approaches remain solely data-driven or knowledge-driven, with candidate recommendations from large language models and the spatial decomposition of Monte Carlo tree search operating independently, failing to achieve a closed-loop synergy between knowledge, data, and experimentation.

[0009] Therefore, there is a need in the art to solve at least one or some of the above problems to improve the efficiency and reliability of chemical reaction condition optimization. Summary of the Invention

[0010] The present invention is provided in order to provide further improved techniques for optimizing chemical reaction conditions.

[0011] One aspect of the present invention provides a computing device for optimizing chemical reaction conditions, comprising: computing resources; and multiple intelligent agents, the multiple intelligent agents being executed by the computing resources, the multiple intelligent agents comprising: a first intelligent agent configured to call a first model to: receive a first user input, the first user input comprising a target chemical reaction, a chemical reaction condition parameter dimension to be analyzed, a candidate parameter in each chemical reaction condition parameter dimension to be analyzed, and provided literature; determine, based on the provided literature, the importance of the chemical reaction condition parameter dimension to be analyzed to the reaction performance of the target chemical reaction; sort the chemical reaction condition parameter dimension to be analyzed based on the importance; and classify, for each candidate parameter in the chemical reaction condition parameter dimension to be analyzed, the candidate parameter according to the physicochemical properties of the candidate parameter. wherein a hierarchical optimization tree is constructed layer by layer based on the results of the sorting and the categories of the candidate parameters, and the hierarchical optimization tree is used to limit the search space for chemical reaction condition parameters; a second intelligent agent is configured to call a second model to: receive a second user input, the second user input includes a combination of candidate parameters associated with the target chemical reaction; and based on the second user input, perform predictive reasoning on the search space to determine reaction performance as a pseudo-label of the combination of candidate parameters, the reaction performance is associated with the combination of candidate parameters indicated by the second user input and the target chemical reaction; and a third intelligent agent is configured to call a third model to: receive a third user input, the third user input includes real prior data associated with the target chemical reaction, the hierarchical optimization tree and the pseudo-label; and based on the real prior data and the pseudo-label, iteratively perform a Bayesian optimization algorithm on the search space defined by the hierarchical optimization tree to determine an optimized combination of chemical reaction condition parameters.

[0012] In the computing device as described above, the chemical reaction condition parameter dimension to be analyzed is selected from a chemical reaction condition parameter dimension group, and the chemical reaction condition parameter dimension group includes at least one of a reactant dimension, a solvent dimension, a base dimension, a catalyst dimension, and a ligand dimension.

[0013] As a computing device as described in any of the above items, the first intelligent agent is configured to call the first model to perform at least one of the following for each candidate parameter in the chemical reaction condition parameter dimension to be analyzed: analyzing the provided literature to identify the key physicochemical properties of the candidate parameters, and classifying the candidate parameters according to the key physicochemical properties, wherein the key physicochemical properties indicate the physicochemical properties that affect the reaction performance of the target chemical reaction; searching for the quantitative values ​​of the physicochemical properties of the candidate parameters in a physicochemical property database, and performing similarity classification on the candidate parameters based on the quantitative values ​​of the physicochemical properties.

[0014] A computing device as described in any of the above items, wherein the key physicochemical properties include at least one of acidity and alkalinity, polarity, valence, steric effect, electronic effect, volatility, and functional group, and wherein the quantitative value of the physicochemical property includes quantitative data or structural indication describing the physicochemical property.

[0015] A computing device as described in any of the above items, the hierarchical optimization tree includes a root node and leaf nodes organized in a hierarchy, the root node indicates a search space including all chemical reaction condition parameter combinations associated with the target chemical reaction, each level of the leaf nodes indicates a chemical reaction condition parameter dimension, each leaf node indicates a category of candidate parameters, and the importance of the chemical reaction condition parameter dimensions associated with the levels of the leaf nodes decreases from the root node to the lowest level of the hierarchical optimization tree, wherein each category of candidate parameters includes a group of candidate parameters with similar physicochemical properties.

[0016] In a computing device as described in any of the above items, the hierarchical optimization tree is further used to define a sub-search space, which includes a subset of chemical reaction condition parameter combinations associated with the target chemical reaction, and wherein the sub-search space is determined by traversing the hierarchical optimization tree layer by layer.

[0017] A computing device as described in any of the above items, wherein the sub-search space is determined based on an upper confidence bound algorithm, from the root node of the hierarchical optimization tree to the lowest level, and the leaf node with the highest upper confidence bound is iteratively selected at each level of the hierarchical optimization tree.

[0018] A computing device as described in any of the above items, in the hierarchical optimization tree, the leaf nodes in the current level serve as child nodes, and the leaf nodes with the highest upper confidence bound in the previous level of the current level serve as parent nodes, and wherein the upper confidence bound of the child node is determined based on the reaction performance associated with the candidate parameter indicated by the child node, the number of visits to the child node, and the number of visits to the parent node.

[0019] In a computing device as described in any of the above items, the upper confidence bound of the child node is further updated by the following steps: using a Bayesian optimization algorithm to sample chemical reaction condition parameter combinations in the sub-search space; determining the reaction performance associated with the chemical reaction condition parameter combination based on the sampling results; updating the reaction performance associated with the candidate parameters indicated by the child node with the reaction performance associated with the chemical reaction condition parameter combination; updating the number of visits to the child node; and determining the updated upper confidence bound of the child node based on the updated reaction performance, the updated number of visits to the child node, and the number of visits to the parent node.

[0020] A computing device as described in any of the above items, wherein the second agent is configured to call a second model to: perform predictive reasoning on the sub-search space based on the second user input to determine reaction performance as a pseudo-label of the combination of candidate parameters, and wherein the third agent is configured to call a third model to: iteratively perform a Bayesian optimization algorithm on the sub-search space based on the real prior data and the pseudo-label to determine an optimized combination of chemical reaction condition parameters.

[0021] In a computing device as described in any of the above items, the second model is obtained through pre-training based on a basic large model, wherein the basic large model adopts causal language modeling loss and is pre-trained by performing a chemical reaction condition prediction task on a training set, wherein the chemical reaction condition prediction task is used to predict chemical reaction condition parameters for the chemical reaction given the reactants and products of the chemical reaction.

[0022] A computing device as described in any of the above items, wherein the second model is obtained through fine-tuning based on a pre-trained basic large model, wherein the pre-trained basic large model is further adjusted to generate predicted values ​​of reaction performance associated with the predicted chemical reaction condition parameters and the chemical reaction, and wherein the pre-trained basic large model is fine-tuned by a low-rank adaptation method based on the loss between the predicted value of the reaction performance and the true value of the reaction performance.

[0023] As a computing device as described in any of the above items, the third agent is configured to call a third model to: initialize the third model based on the pseudo-label; execute a Bayesian optimization algorithm on the search space via the initialized third model to determine an initial chemical reaction condition parameter combination; update the third model based on the actual reaction performance associated with the chemical reaction condition parameter combination; and execute a Bayesian optimization algorithm on the search space via the updated third model to determine the optimized chemical reaction condition parameter combination.

[0024] As a computing device as described in any of the above items, the third agent is configured to call a third model to: calculate the similarity between the pseudo label and the true reaction performance, wherein the pseudo label includes a first pseudo label; determine that the similarity between the first pseudo label and the true reaction performance is less than a similarity threshold; remove the first pseudo label from the pseudo label to update the third model; and execute a Bayesian optimization algorithm on the search space via the updated third model to determine the optimized chemical reaction condition parameter combination.

[0025] In a computing device as described in any of the above items, the third agent is configured to call a third model to: randomly remove one or more of the pseudo labels from the pseudo labels to update the third model; and execute a Bayesian optimization algorithm on the search space via the updated third model to determine the optimized chemical reaction condition parameter combination.

[0026] Another aspect of the present invention provides a method for optimizing chemical reaction conditions, comprising the following steps: S1: a first intelligent agent constructs a hierarchical optimization tree, the hierarchical optimization tree is used to limit the search space for chemical reaction condition parameters, the step S1 comprising: S11: the first intelligent agent calls a first model to perform the following steps: S111: receiving a first user input, the first user input comprising a target chemical reaction, a chemical reaction condition parameter dimension to be analyzed, a candidate parameter in each chemical reaction condition parameter dimension to be analyzed, and provided literature; S112: determining the importance of the chemical reaction condition parameter dimension to be analyzed to the reaction performance of the target chemical reaction based on the provided literature; S113: sorting the chemical reaction condition parameter dimension to be analyzed based on the importance; and S114: for each candidate parameter in the chemical reaction condition parameter dimension to be analyzed, classifying the candidate parameter according to the physicochemical properties of the candidate parameter; and S12: the first intelligent agent constructs the hierarchical optimization tree layer by layer based on the sorting result and the category of the candidate parameter; S2: the second intelligent agent performs prediction of the search space Reasoning to generate pseudo labels for combinations of chemical reaction condition parameters, said step S2 comprising calling a second model by the second agent to perform the following steps: S21: receiving a second user input, said second user input comprising a combination of candidate parameters associated with the target chemical reaction; and S22: performing predictive reasoning on the search space based on the second user input to determine a reaction performance as a pseudo label for the combination of the candidate parameters, said reaction performance being associated with the combination of the candidate parameters indicated by the second user input and the target chemical reaction; and S3: performing Bayesian optimization on the search space defined by the hierarchical optimization tree by a third agent to determine an optimized combination of chemical reaction condition parameters, said step S3 comprising calling a third model by the third agent to perform the following steps: S31: receiving a third user input, said third user input comprising true prior data associated with the target chemical reaction, the hierarchical optimization tree and the pseudo labels; and S32: iteratively performing a Bayesian optimization algorithm on the search space defined by the hierarchical optimization tree based on the true prior data and the pseudo labels to determine an optimized combination of chemical reaction condition parameters.

[0027] According to the method described above, the hierarchical optimization tree includes a root node and leaf nodes organized hierarchically, the root node indicates a search space including all chemical reaction condition parameter combinations associated with the target chemical reaction, each level of the leaf nodes indicates a chemical reaction condition parameter dimension, each leaf node indicates a category of candidate parameters, and from the root node to the lowest level of the hierarchical optimization tree, the importance of the chemical reaction condition parameter dimensions associated with the level of the leaf nodes decreases, each category of candidate parameters includes a group of candidate parameters with similar physical and chemical properties, and wherein the hierarchical optimization tree is also used to define a sub-search space, the sub-search space includes a subset of the chemical reaction condition parameter combinations associated with the target chemical reaction, and is determined by traversing the hierarchical optimization tree layer by layer.

[0028] The method as described in any of the above items, wherein the sub-search space is determined based on an upper confidence bound algorithm, from the root node of the hierarchical optimization tree to the lowest level, and the leaf node with the highest upper confidence bound is iteratively selected at each level of the hierarchical optimization tree.

[0029] As described in any of the above methods, in the hierarchical optimization tree, the leaf nodes in the current level serve as child nodes, and the leaf nodes with the highest upper confidence bound in the previous level of the current level serve as parent nodes, wherein the upper confidence bound of the child node is determined based on the reaction performance associated with the candidate parameters indicated by the child node, the number of visits to the child node, and the number of visits to the parent node, and wherein the upper confidence bound of the child node is further updated by the following steps: using a Bayesian optimization algorithm to sample chemical reaction condition parameter combinations in the sub-search space; determining the reaction performance associated with the chemical reaction condition parameter combination based on the sampling results; updating the reaction performance associated with the candidate parameters indicated by the child node with the reaction performance associated with the chemical reaction condition parameter combination; updating the number of visits to the child node; and determining the updated upper confidence bound of the child node based on the updated reaction performance, the updated number of visits to the child node, and the number of visits to the parent node.

[0030] As described in any of the above methods, step S22 includes: performing predictive reasoning on the sub-search space based on the second user input to determine reaction performance as a pseudo-label of the combination of the candidate parameters, and wherein step S32 includes: based on the real prior data and the pseudo-label, iteratively performing a Bayesian optimization algorithm on the sub-search space to determine an optimized chemical reaction condition parameter combination.

[0031] As described in any of the methods above, step S32 includes: S321: initializing the third model based on the pseudo-label; S322: executing a Bayesian optimization algorithm on the search space via the initialized third model to determine the initial chemical reaction condition parameter combination; S323: updating the third model based on the actual reaction performance associated with the chemical reaction condition parameter combination; and S324: executing a Bayesian optimization algorithm on the search space via the updated third model to determine the optimized chemical reaction condition parameter combination.

[0032] As described in any of the above methods, step S323 includes: S3231: calculating the similarity between the pseudo label and the true reaction performance, wherein the pseudo label includes a first pseudo label; S3232: determining that the similarity between the first pseudo label and the true reaction performance is less than a similarity threshold; S3233: removing the first pseudo label from the pseudo label to update the third model.

[0033] As in any of the above methods, the step S323 includes: S3234: randomly removing one or more of the pseudo labels from the pseudo labels to update the third model.

[0034] Another aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program implements the steps of any of the above methods when executed by a processor.

[0035] Another aspect of the present invention provides a computer program product, comprising a computer program, which implements the steps of any one of the above methods when executed by a processor.

[0036] The technical solution proposed in the present invention constructs a hierarchical optimization tree to limit the search space. On this basis, Bayesian optimization is performed to explore optimized chemical reaction condition combinations in the "high potential / low risk" search space that is conducive to the target chemical reaction. The predictive reasoning of the search space provides high-quality prior experimental data for Bayesian optimization, which can effectively solve the "cold start" problem caused by the lack of prior data in the early stages of the experiment, which can significantly improve the optimization efficiency. In addition, in the technical solution proposed in the present invention, the search space decomposition and the generation of pseudo-data are carried out under the constraints of chemical mechanisms. The synergistic effect of knowledge and data can determine the optimal chemical reaction conditions more quickly and reliably within limited experimental resources, significantly reducing the number and cost of experiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Various embodiments of the present invention are described with reference to the accompanying drawings.

[0038] Figure 1is a schematic block diagram of a computing device for optimizing chemical reaction conditions according to some embodiments of the present invention.

[0039] Figure 2 An exemplary process of optimizing chemical reaction conditions by a computing device according to some embodiments of the present invention is shown.

[0040] Figure 3-Figure 5 is a screenshot of an exemplary user interface for a computing device performing optimization of chemical reaction conditions according to some embodiments of the present invention.

[0041] Figure 6 is a schematic flow chart of a method for optimizing chemical reaction conditions according to some embodiments of the present invention.

[0042] Figure 7 is a flow chart of a first process associated with a method for optimizing chemical reaction conditions according to some embodiments of the present invention.

[0043] Figure 8 is a flow chart of a second process associated with a method for optimizing chemical reaction conditions according to some embodiments of the present invention.

[0044] Figure 9 is a flow chart of a third process associated with a method for optimizing chemical reaction conditions according to some embodiments of the present invention.

[0045] Figure 10 is a flow chart of exemplary operations associated with a third process in a method for optimizing chemical reaction conditions, according to some embodiments of the present invention.

[0046] Figure 11 is a block diagram of a computer-readable storage medium according to some embodiments of the present invention.

[0047] Figure 12 is a block diagram of a computer program product according to some embodiments of the present invention.

[0048] Figures 13-18 The diagram shows a comparison of the operating results of optimizing chemical reaction conditions using the technical solutions for optimizing chemical reaction conditions according to some embodiments of the present invention and existing optimization solutions. DETAILED DESCRIPTION

[0049] Next, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings, wherein, unless otherwise expressly stated, the same or similar reference numerals in different drawings represent the same or similar elements. In addition, it should be noted that exemplary embodiments of the present invention may perform the steps of the corresponding method in a different order than shown and described in the specification. In some embodiments of the present invention, the method may include more or fewer steps than those described in the specification and shown in the accompanying drawings. In addition, a single step adopted in some embodiments of this specification may be decomposed into multiple steps in other embodiments, or a plurality of steps adopted in some embodiments of this specification may be combined into a single step in other embodiments.

[0050] In the following description, numerous specific details are set forth. However, it should be understood that embodiments of the present invention may be practiced without these specific details. In other instances, well-known circuits, structures, and techniques are not shown in detail to avoid obscuring the understanding of this description.

[0051] References in the specification to "one embodiment," "an embodiment," "an example embodiment," etc., indicate that the described embodiment may include a particular feature, structure, or characteristic, but not every embodiment will necessarily include that particular feature, structure, or characteristic. Furthermore, such phrases do not necessarily refer to the same embodiment. Furthermore, when a particular feature, structure, or characteristic is described in conjunction with an embodiment, it is understood that it is within the knowledge of those skilled in the art to be able to affect such feature, structure, or characteristic in conjunction with other embodiments, whether or not explicitly described.

[0052] For purposes of the present invention, the phrase "A and / or B" means (A), (B), or (A and B). For purposes of the present invention, the phrase "A, B, and / or C" means (A), (B), (C), (A and B), (A and C), (B and C), or (A, B and C).

[0053] In this disclosure, the term "agent" refers to an agent capable of perceiving its environment and taking actions to achieve specific goals. An agent primarily refers to software code. An agent can be executed by the computing resources of a computing device. An agent can access corresponding models through an application programming interface (API) and invoke corresponding tools (e.g., document readers, code interpreters, calculators, etc.) to interact with various forms of input or implement corresponding functions.

[0054] In the present invention, the term "chemical reaction conditions" refers to a collection of parameters that promote the occurrence of a chemical reaction or affect the reaction performance of a chemical reaction (such as the reaction rate, product selectivity, yield, safety, etc. of a chemical reaction). Chemical reaction conditions are generally composed of various chemical reaction condition parameters, and are actually a combination of various chemical reaction condition parameters. Chemical reaction condition parameters can be divided into different dimensions, which are referred to as chemical reaction condition parameter dimensions in this article. As an example, the chemical reaction condition parameter dimensions may include, for example, reactant dimensions, catalyst dimensions, solvent dimensions, ligand dimensions, base dimensions, and the like. As another example, the chemical reaction condition parameter dimensions may also include reaction environment parameter dimensions to indicate environmental conditions that promote the occurrence of a chemical reaction or enhance the reaction performance of a chemical reaction, such as a specific temperature.

[0055] In the present invention, ordinal numbers such as "first," "second," and "third" are used to distinguish different instances of an object with the same name. The ordinal numbers such as "first," "second," and "third" do not indicate the relative order of the objects in time, space, ranking, or other aspects.

[0056] According to one aspect of the present invention, a computing device for optimizing chemical reaction conditions is provided.

[0057] Figure 1 is a block diagram of a computing device 100 for optimizing chemical reaction conditions according to some embodiments of the present invention.

[0058] The computing device 100 may be a local or remote computer, server, etc. The computing device 100 may include a computing resource 110 and multiple agents. The multiple agents may be executed by the computing resource 110 to enable the computing resource 110 to perform corresponding operations.

[0059] In some embodiments, the computing resources 110 may include a central processing unit (CPU), a graphics processing unit (GPU), and various other processing units or cores (e.g., arithmetic logic unit, integer unit, floating point unit, tensor unit, ray tracing core, etc.).

[0060] In some embodiments, each of the multiple agents can call a corresponding model through an API. As an example, the model may include a large language model, a multimodal model, a multimodal language model, etc. In some embodiments, the multiple models that can be called by the multiple agents can be deployed locally on the computing device 100. In some embodiments, the multiple models that can be called by the multiple agents can be deployed remotely from the computing device 100, for example, deployed in the cloud. In some embodiments, some of the multiple models that can be called by the multiple agents can be deployed locally on the computing device 100, and other models can be deployed remotely from the computing device 100. In some embodiments, each of the multiple agents can call various tools to interact with various forms of input or implement corresponding functions. For example, the agent can call a PDF reader to read a PDF document. The agent can call a python interpreter to interact with python code. The agent can call a calculator to implement calculation functions.

[0061] like Figure 1 As shown, the plurality of agents may include at least a first agent 122, a second agent 124, and a third agent 126. According to an embodiment of the present invention, each of the first agent 122, the second agent 124, and the third agent 126 may call appropriate models and / or tools to perform various interactions and / or implement various functions.

[0062] According to an embodiment of the present invention, the first agent 122 may be configured to call a first model to perform a decomposition of the search space. In some embodiments, the first model may include a large language model. In some embodiments, the first model may include one or more models.

[0063] First agent 122 may be configured to receive a first user input. In embodiments of the present invention, the first user input may include a target chemical reaction. The target chemical reaction may indicate the chemical reaction whose reaction conditions are to be optimized, and may include one or a class of chemical reactions. As an exemplary implementation, the target chemical reaction may be specified by describing the reactants and target products, or alternatively, the target chemical reaction may be specified directly by the name of the chemical reaction (e.g., Suzuki-Miyaura coupling reaction, Buchwald-Hartwig coupling reaction, tandem arylation reaction, enantioselective reaction, etc.). In embodiments of the present invention, the first user input may also include the chemical reaction condition parameter dimensions to be analyzed. The chemical reaction condition parameter dimensions to be analyzed may indicate the dimensions for which the chemical reaction conditions of the target chemical reaction are to be optimized. Taking the Suzuki-Miyaura coupling reaction as an example, the presence of a base, the choice of solvent, the activity of the substrate, the efficiency of the catalyst, and so on all affect its reaction performance. Therefore, the first user input may indicate any of the base, solvent, substrate, and catalyst dimensions as the chemical reaction condition parameter dimensions to be analyzed. In embodiments of the present invention, the first user input may also include candidate parameters for each chemical reaction condition parameter dimension to be analyzed. Within each chemical reaction condition parameter dimension, multiple chemical reaction condition parameters may be available for selection. In addition to indicating the chemical reaction condition parameter dimension to be analyzed via the first user input, the first user input may further indicate candidate parameters within each dimension. Still using the Suzuki-Miyaura coupling reaction as an example, in addition to indicating the base and solvent dimensions as the chemical reaction condition parameter dimensions to be analyzed via the first user input, the first user input may further indicate that candidate parameters within the base dimension include potassium hydroxide, sodium methoxide, potassium methoxide, potassium tert-butoxide, sodium tert-butoxide, and so on, and candidate parameters within the solvent dimension include tetrahydrofuran, dimethylformamide, acetonitrile, toluene, water, or a mixture of multiple solvents (e.g., a mixture of toluene, ethanol, and water). In embodiments of the present invention, the first user input may also include provided literature. The provided literature may include professional scientific literature in the field of chemistry and may be documents uploaded to the first agent 122 by the user or provided to the first agent 122 through a search of a literature library.

[0064] First agent 122 may be configured to determine the importance of a chemical reaction condition parameter dimension to be analyzed on the reaction performance of a target chemical reaction based on the literature provided in the first user input. As an exemplary implementation, first agent 122 may determine the importance of a chemical reaction condition parameter dimension to be analyzed based solely on explicit or implicit descriptions of the impact of the chemical reaction condition parameter dimension on the reaction performance in the literature associated with the target chemical reaction provided in the first user input. The greater the impact of a chemical reaction condition parameter dimension to be analyzed on the reaction performance (e.g., yield, selectivity, etc.) of the target chemical reaction, the higher the importance of the chemical reaction condition parameter dimension.

[0065] The first agent 122 may be configured to sort the chemical reaction condition parameter dimensions to be analyzed based on their importance. As an exemplary implementation, the first agent 122 may sort the chemical reaction condition parameter dimensions to be analyzed in descending order of importance based on an analysis of the provided literature.

[0066] The first agent 122 may be configured to classify each candidate parameter in the chemical reaction condition parameter dimension to be analyzed according to the physicochemical properties of the candidate parameters. As an exemplary implementation, the first agent 122 may perform clustering on the physicochemical properties of the candidate parameters in each chemical reaction condition parameter dimension to be analyzed based on an analysis of the provided literature, so as to classify each candidate parameter in the chemical reaction condition parameter dimension to be analyzed into different categories according to the physicochemical properties.

[0067] The first agent 122 can decompose the search space into a tree structure, constructing the search space into a hierarchical optimization tree. The hierarchical optimization tree can be constructed layer by layer based on the ranking results of the chemical reaction condition parameter dimensions to be analyzed by the first agent 122 according to their importance and the classification of the candidate parameters in each chemical reaction condition parameter dimension to be analyzed according to their physicochemical properties. The constructed hierarchical optimization tree can define the search space for the chemical reaction condition parameters.

[0068] The following will be combined Figure 2 The first box 210 in the figure further describes the operation of the first agent 122 in detail.

[0069] In an embodiment of the present invention, a hierarchical optimization tree is used to limit the search space for chemical reaction condition parameters, which decomposes the high-dimensional search space into a hierarchical tree structure, can effectively reduce the number of samples that need to be traversed to optimize the chemical reaction conditions, thereby effectively alleviating the "curse of dimensionality" problem caused by the high-dimensional search space. In addition, the construction of the hierarchical optimization tree is based on the influence of the chemical reaction condition parameters on the chemical reaction performance and their physicochemical properties, which means that the construction of the hierarchical optimization tree takes into account the chemical mechanism in the optimization of chemical reaction conditions. The optimization of chemical reaction conditions based on the hierarchical optimization tree is subject to chemical mechanism constraints and can effectively exclude invalid search subdomains, which can further improve the optimization efficiency and reduce experimental costs.

[0070] According to an embodiment of the present invention, the second agent 124 may be configured to call a second model to perform predictive reasoning on the search space. In some embodiments, the second model may include a large language model. In some embodiments, the second model may include one or more models.

[0071] The second agent 124 may be configured to receive a second user input. In an embodiment of the present invention, the second user input may include a combination of candidate parameters associated with the target chemical reaction. As an exemplary implementation, on the basis of the first agent 122 decomposing the search space for chemical reaction condition parameters into a hierarchical optimization tree, a search may be performed for the hierarchical optimization tree to obtain a combination of candidate parameters for the target chemical reaction. Alternatively, a combination of candidate parameters for the target chemical reaction may be searched in the entire search space for chemical reaction condition parameters. Still taking the Suzuki-Miyaura coupling reaction as the target chemical reaction as an example, the second user input received by the second agent 124 may include various combinations of candidate parameters for the Suzuki-Miyaura coupling reaction, such as various combinations of solvents, catalysts, ligands and bases that can be used to perform the Suzuki-Miyaura coupling reaction, and the like.

[0072] The second agent 124 can be configured to perform predictive reasoning on the search space based on the second user input to determine the reaction performance as a pseudo label of the combination of candidate parameters, and the reaction performance is associated with the combination of candidate parameters indicated by the second user input and the target chemical reaction. As an exemplary implementation, based on the indication of the second user input, predictive reasoning can be performed on the entire search space or on one or more sub-search subdomains indicated by the hierarchical optimization tree. This predictive reasoning can generate predicted reaction performance for different candidate parameter combinations. This predicted reaction performance can indicate the reaction performance of a target chemical reaction using a candidate parameter combination, and the reaction performance can be a data value (e.g., yield, selectivity, etc.). The predicted reaction performance of the combination of candidate parameters can serve as a pseudo label for the combination of candidate parameters. By performing predictive reasoning on the search space by the second agent 124, pseudo labels can be generated for the combination of all chemical reaction condition parameters included in the search space.

[0073] In an embodiment of the present invention, the second intelligent agent 124 can generate reliable pseudo-labels by learning historical knowledge in chemical literature, chemical experiment databases, etc., and use the pseudo-labels to mark the search space, which can effectively solve the problem of lack of prior data in the early stage of the experiment.

[0074] The following will be combined Figure 2 The second box 220 in further details the operation of the second agent 124.

[0075] According to an embodiment of the present invention, the third agent 126 may be configured to call a third model to perform Bayesian optimization on the search space to determine optimized chemical reaction conditions.

[0076] The third agent 126 can be configured to receive a third user input. In an embodiment of the present invention, the third user input can include real prior data associated with the target chemical reaction. As an exemplary implementation, the user can manually perform the target chemical reaction and collect the chemical reaction condition parameters used to perform the target chemical reaction and the reaction performance of the target chemical reaction as real prior data associated with the target chemical reaction, or alternatively, the third agent 126 can automatically retrieve historical experimental data such as chemical reaction condition parameters and reaction performance associated with the target chemical reaction from the chemical experiment database as real prior data. In addition, in an embodiment of the present invention, the third user input can also include a pseudo-label of a combination of the hierarchical optimization tree constructed by the first agent 122 and the chemical reaction condition parameters generated by the second agent 124. As an exemplary implementation, the third agent 126 can communicate directly or indirectly with the first agent 122 and the second agent 124 to obtain the hierarchical optimization tree from the first agent 122 and the pseudo-label of the combination of the chemical reaction condition parameters from the second agent 124.

[0077] The third agent 126 can be configured to iteratively execute a Bayesian optimization algorithm on the search space defined by the hierarchical optimization tree based on the real prior data and the pseudo-labels to determine an optimized combination of chemical reaction condition parameters. As an exemplary implementation, the third agent 126 can execute the Bayesian optimization algorithm on the search space annotated with pseudo-labels in the hierarchical optimization tree and further filter based on the real prior data. Through this iterative optimization, an optimized combination of chemical reaction condition parameters is obtained.

[0078] The following will be combined Figure 2 The third box 230 in the figure further describes the Bayesian optimization process performed by the third agent 126 in detail.

[0079] In this embodiment of the present invention, due to the predictive reasoning of the search space by the second agent 124, the third agent 126 is able to perform Bayesian optimization based on the pseudo data generated by the predictive reasoning. This provides high-quality prior experimental data for Bayesian optimization, effectively resolving the "cold start" problem caused by the lack of prior data in the early stages of the experiment. In addition, the first agent 122 constructs a hierarchical optimization tree to limit the search space. Performing Bayesian optimization on this basis can explore optimized chemical reaction condition combinations in the "high potential / low risk" search space that are favorable to the target chemical reaction, significantly improving optimization efficiency.

[0080] Through the multi-agent collaborative computing device proposed in the present invention, by decomposing the search space under the constraints of chemical mechanisms and generating high-quality pseudo data, knowledge and data work together to determine the optimal chemical reaction conditions more quickly and reliably within limited experimental resources, significantly reducing the number and cost of experiments.

[0081] References below Figure 2 The exemplary process shown is used to further describe the operation of multiple agents in the computing device proposed by the present invention and their coordination with each other.

[0082] Figure 2 An exemplary process of optimizing chemical reaction conditions by a computing device according to some embodiments of the present invention is shown.

[0083] Figure 2 The exemplary process shown in FIG. 1 may include a first block 210, which may be similar to the process described above with reference to FIG. Figure 1 The described first agent 122 is associated. The operation shown in the first block 210 may be a knowledge-driven operation.

[0084] In step 1, the first agent can call the inference large language model 213 to perform semantic analysis on the relevant documents 211, so as to sort the chemical reaction condition parameter dimensions to be analyzed to obtain the sorting result 214. In some embodiments, the chemical reaction condition parameter dimensions to be analyzed are selected from the chemical reaction condition parameter dimension group, and the chemical reaction condition parameter dimension group may include at least one of the reactant dimension, solvent dimension, base dimension, catalyst dimension, and ligand dimension. As an exemplary implementation, the chemical reaction condition parameter dimensions to be analyzed can be selected by the first agent based on the analysis of the relevant documents 211, or can be selected by the user based on prior knowledge to indicate the chemical reaction condition parameter dimensions that can promote the occurrence of the target chemical reaction or have an impact on the reaction performance of the target chemical reaction. The sorting result 214 is shown as importance 1, importance 2, importance 3, etc., and these importances decrease from top to bottom, corresponding to different chemical reaction condition parameter dimensions, wherein different chemical reaction condition parameter dimensions are distinguished by different colors in the sorting result 214. In Figure 2 In the example shown in , importance level 1 corresponds to the catalyst dimension, shown in red; importance level 2 corresponds to the ligand dimension, shown in blue; and importance level 3 corresponds to the base dimension, shown in green.

[0085] In step 2, the first agent may call the large language model 216 to analyze the physical and chemical property data 212, and classify the chemical reaction condition parameters in each chemical reaction condition parameter dimension according to the key attribute 215 to obtain a classification result 218. According to an embodiment of the present invention, the large language model 216 may be the same as the inference large language model 213, or may be different. In some embodiments, the inference large language model 213 and the large language model 216 may together constitute the above-mentioned combined large language model. Figure 1 The first model called by the first agent 122 is described.

[0086] In some embodiments of the present invention, the first agent can call the inference large language model 213 to conduct a comprehensive analysis of the relevant documents 211 to identify the key physical and chemical properties 217 of the chemical reaction condition parameters, and the first agent can call the large language model 216 to classify these chemical reaction condition parameters according to the identified key physical and chemical properties 217 to obtain a classification result 218. Each category of chemical reaction condition parameters includes a group of candidate parameter chemical reaction condition parameters with similar physical and chemical properties. Similar physical and chemical properties can indicate that the physical and chemical properties differ, for example, within 20%, within 15%, within 10%, within 5%, within 1%, etc. As an exemplary implementation, the chemical reaction condition parameters can be clustered according to the similarity of the key physical and chemical properties 217 to classify them into chemical reaction condition parameter categories. According to an embodiment of the present invention, the key physical and chemical properties 217 can indicate those physical and chemical properties that can be found in the physical and chemical property data 212 and affect the reaction performance of the target chemical reaction. The key physical and chemical properties 217 may include, for example, at least one of acidity, polarity, valence, steric effect, electronic effect, volatility, functional group, etc. Figure 2 In the example, key physicochemical attributes 217 show key attribute 1, key attribute 2, key attribute 3, and so on. For example, key attribute 1 indicates valence, is associated with the catalyst dimension, and is shown in red; key attribute 2 indicates steric effect, is associated with the ligand dimension, and is shown in blue; and key attribute 3 indicates basicity, is associated with the base dimension, and is shown in green. As an exemplary implementation, various catalysts in the catalyst dimension can be classified based on their valence.

[0087] In some embodiments of the present invention, the first agent can invoke the inference large language model 213 to search for quantitative values ​​of the physicochemical properties of chemical reaction condition parameters in a physicochemical property database (e.g., physicochemical property data 212). The first agent can then invoke the large language model 216 to perform similarity classification on the chemical reaction condition parameters based on the quantitative values, to obtain classification results 218. According to embodiments of the present invention, the quantitative values ​​of the physicochemical properties may include quantitative data or structural indicators describing the physicochemical properties of the chemical reaction condition parameters. The quantitative values ​​of the physicochemical properties may include, for example, the pKa value for acidity or alkalinity, the dielectric constant for polarity, the boiling point for volatility, and the number of specific functional groups. As an exemplary implementation, the first agent can invoke the inference large language model 213 to cluster the chemical reaction condition parameters based on the quantitative values ​​of the physicochemical properties, so that chemical reaction condition parameters with similar quantitative values ​​of the physicochemical properties are classified into the same category. For example, for various bases in the base dimension, these bases can be similarly classified based on their pKa values.

[0088] In some embodiments of the present invention, the large language model 216 may provide an indication of chemical reaction condition parameters at the boundary of a category. For example, the large language model may classify a chemical reaction condition parameter at the boundary of a category into any category adjacent to the boundary and indicate that the chemical reaction condition parameter is at the boundary of the category for manual verification by scientists.

[0089] For example, the classification results 218 are shown as category a, category b, category c, category d, category e, category f, and so on, where category a and category b indicate the classification of various catalysts in the catalyst dimension, and are shown in red with different shades of color to distinguish the two categories; category c and category d indicate the classification of various ligands in the ligand dimension, and are shown in blue with different shades of color to distinguish the two categories; category e and category f indicate the classification of various bases in the base dimension, and are shown in green with different shades of color to distinguish the two categories. The classification results 218 further indicate the chemical reaction condition parameters included in each category.

[0090] An exemplary hierarchical optimization tree 219 is also shown. The first agent invokes the inference large language model 213 to analyze the literature information in the vector database to determine the importance of the chemical reaction condition parameter dimensions to be analyzed to the reaction performance of the chemical reaction, and generates a sorted result 214 (e.g., a sorted list) (e.g., as shown in sorted result 214, the importance is ranked as follows: catalyst dimension > ligand dimension > base dimension). The first agent invokes the inference large language model 213 to further identify key physicochemical properties within each chemical reaction condition parameter dimension that significantly affect the chemical reaction performance, as shown in key physicochemical attributes 217. These may include valence associated with the catalyst dimension, steric effects associated with the ligand dimension, and pKa associated with the base dimension. Subsequently, the first agent invokes the large language model 216 to obtain detailed information on the key physicochemical attributes of each chemical reaction condition parameter from the physicochemical attribute data 212, and clusters these chemical reaction condition parameters into several categories based on the similarity of their key physicochemical attributes. On this basis, the first agent can construct a hierarchical optimization tree based on the importance ranking results of the chemical reaction condition parameter dimensions and the categories of the chemical reaction condition parameters within each chemical reaction condition parameter dimension to define the search space for the chemical reaction condition parameters. A hierarchical optimization tree is a tree structure with a root node and hierarchically organized leaf nodes. In the hierarchical optimization tree 219, its root node corresponds to the complete search space, which includes all chemical reaction condition parameter combinations associated with the target chemical reaction. According to an embodiment of the present invention, the first agent can first determine the root node of the hierarchical optimization tree and then construct the leaf nodes of the hierarchical optimization tree layer by layer. The highest level of the leaf node corresponds to the chemical reaction condition parameter dimension with the highest importance ranking (i.e., the catalyst dimension, shown in red); the next highest level of the leaf node corresponds to the chemical reaction condition parameter dimension with the next highest importance ranking (i.e., the ligand dimension, shown in blue); and the next level of the leaf node corresponds to the chemical reaction condition parameter dimension with the next highest importance ranking (i.e., the base dimension, shown in green). Furthermore, each leaf node in the hierarchical optimization tree can indicate a category of chemical reaction condition parameters. For example, as shown in the hierarchical optimization tree 219, the highest level includes two leaf nodes, which are labeled a and b, respectively, corresponding to category a and category b of the catalyst shown in the classification result 218; the next highest level includes two leaf nodes, which are labeled c and d, respectively, corresponding to category c and category d of the ligand shown in the classification result 218; and so on.

[0091] Figure 2 The exemplary process shown in FIG. 1 may include a second block 220, which may be similar to the process described above with reference to FIG. Figure 1The described second agent 124 is associated. The operation shown in the second block 220 may be a data driven operation.

[0092] like Figure 2 As shown, the second block 220 may include the operation of pre-training the base large model 222 using the unlabeled reaction data 221 in step 3. The input of the base large model 222 is a simplified molecular linear input canonical string, which is a one-dimensional linear representation that encodes molecular structures in computer text encoding (ASCII) and is widely used in structure retrieval and molecular modeling research. In embodiments of the present invention, a causal language modeling loss can be used to perform pre-training by performing a chemical reaction condition prediction task on the unlabeled reaction data 221 as a training set. According to embodiments of the present invention, the chemical reaction condition prediction task is used to predict the chemical reaction condition parameters for a chemical reaction given the reactants and products of the chemical reaction. The output of the base large model 222 is a simplified molecular linear input canonical string representing these chemical reaction condition parameters. After pre-training in step 3, a chemical large model 224 can be obtained from the base large language model.

[0093] According to an embodiment of the present invention, by constructing a chemical reaction condition prediction task and pre-training the basic large model through the chemical reaction condition prediction task, problems such as missing training set data, unknown sources, and large noise in reaction performance indicators can be effectively solved.

[0094] like Figure 2 As shown, the second block 220 may include the operation of fine-tuning the chemical model 224 using the labeled reaction data 223 in step 4. In an embodiment of the present invention, through prompt word engineering, the chemical model is guided to input the hidden state of the last character (as shown in yellow in token 225) in the simplified molecular linear input canonical string representing chemical reaction condition parameters into an additional projection layer 226, thereby outputting a predicted value of the reaction performance shown in 227. As an exemplary implementation, the chemical model 224 may be fine-tuned using a low-rank adaptation method based on the loss between the predicted value of the reaction performance (e.g., the predicted value of the reaction performance shown in output 227) and the actual value of the reaction performance (e.g., the labeled reaction data 223). The second block 220 further shows a prompt word 228 for guiding the fine-tuning of the chemical model. Guided by the prompt word 228, the chemical model can output a predicted value of the reaction performance.

[0095] In an embodiment of the present invention, the second agent can call the second model obtained through pre-training in step 3 and fine-tuning in step 4 to perform predictive reasoning on the search space, thereby generating a predicted value of the reaction performance associated with the target chemical reaction and the combination of chemical reaction condition parameters as a pseudo label for the search space.

[0096] According to an embodiment of the present invention, model fine-tuning can improve the accuracy of model prediction, reduce computational overhead, and alleviate the potential "hallucination" problem of large language models.

[0097] Figure 2 The exemplary process shown in FIG. 1 may include a third block 230, which may be similar to the process described above with reference to FIG. Figure 1 The described third agent 126 is associated. The operation shown in the third box 230 may be a Bayesian optimization operation performed in the search space defined by the hierarchical optimization tree.

[0098] In an embodiment of the present invention, the third intelligent agent can perform Bayesian optimization on the search space based on the hierarchical optimization tree obtained through the operation shown in the first box 210 and the pseudo-label obtained through the operation shown in the second box 220 to obtain an optimized combination of chemical reaction condition parameters (for example, the recommended sampling points marked with five-pointed stars in the third box 230).

[0099] In some embodiments of the present invention, the third agent is configured to call a third model to initialize the third model based on the pseudo-labels, and perform a Bayesian optimization algorithm on the search space via the initialized third model to determine an initial chemical reaction condition parameter combination, on this basis, update the third model based on the actual reaction performance associated with the chemical reaction condition parameter combination, and perform a Bayesian optimization algorithm on the search space via the updated third model to determine an optimized chemical reaction condition parameter combination. As an exemplary implementation, the third agent can update the third model based on the similarity between the pseudo-labels and the actual reaction performance. The third agent can call the third model to calculate the similarity between the pseudo-labels and the actual reaction performance, and determine that the similarity between the first pseudo-labels among these pseudo-labels and the actual reaction performance is less than a similarity threshold. On this basis, the first pseudo-labels are removed from these pseudo-labels to update the third model. The third agent can thus perform a Bayesian optimization algorithm on the search space via the updated third model to determine an optimized chemical reaction condition parameter combination. As another example implementation, the third agent can call a third model to randomly remove one or more of the pseudo labels from the pseudo labels to update the third model, whereupon the third agent can execute a Bayesian optimization algorithm on the search space via the updated third model to determine an optimized combination of chemical reaction condition parameters.

[0100] According to an embodiment of the present invention, Bayesian optimization is performed by updating the third model and adopting the updated third model. As the Bayesian optimization is iterated and real experimental data accumulates, pseudo data points are gradually removed, which can increase the weight of real data in the model update, thereby alleviating the impact of pseudo data noise.

[0101] In some embodiments of the present invention, the hierarchical optimization tree 219 can be updated based on the actual reaction performance associated with the chemical reaction condition parameter combination. For example, the upper confidence bound of the reaction performance corresponding to the chemical reaction condition parameter combination indicated by each leaf node of the candidate hierarchical optimization tree can be updated and calculated, and the leaf node of the candidate hierarchical optimization tree with the largest upper confidence bound is selected as the parameter combination space subdomain to be explored in the next round.

[0102] In some embodiments of the present invention, the second model (e.g., Figure 2 The described chemical macromodel 224 can be updated based on the actual reaction performance associated with the chemical reaction condition parameter combination. As an exemplary implementation, the actual reaction performance can be used to update the labeled reaction data 223, thereby updating the chemical macromodel 224. The updated chemical macromodel 224 can more accurately perform predictive inference on the search space to generate pseudo-labels.

[0103] According to an embodiment of the present invention, the updating of the first model and / or the second model is conducive to further improving the accuracy and efficiency of the third agent performing Bayesian optimization on the search space defined by the hierarchical optimization tree based on the hierarchical optimization tree and pseudo-labels.

[0104] Third box 230 shows the fitting results of Bayesian optimization performed on the search space, both without the pseudo-labels provided by the operation shown in second box 220 (shown as a blue curve, where blue shading represents variance) and with the pseudo-labels provided by the operation shown in second box 220 (shown as a purple curve, where purple shading represents variance). Furthermore, a true objective function (shown as a black curve) containing actual experimental observations is shown. The abscissa of the curve represents the chemical reaction condition parameter combinations, and the ordinate represents the reaction performance. As can be seen from third box 230, the chemical reaction condition parameter combination determined by Bayesian optimization, with the pseudo-labels provided by the operation shown in second box 220, can correspond to the highest reaction performance in the true objective function.

[0105] Figure 2 The exemplary process shown in FIG. 2 may further include a fourth block 240 . The operation shown in the fourth block 240 may be a decomposition operation of the search space based on the hierarchical optimization tree 219 .

[0106] In an embodiment of the present invention, the hierarchical optimization tree 219 may further define a sub-search space for chemical reaction condition parameters, where the sub-search space includes a subset of chemical reaction condition parameter combinations associated with the target chemical reaction. According to an embodiment of the present invention, the sub-search space may be determined by traversing the hierarchical optimization tree 219 layer by layer through the operation shown in the fourth block 240.

[0107] As shown in the fourth block 240, an upper confidence bound (UCB) algorithm can be used to traverse the hierarchical optimization tree 219 layer by layer to identify and sample high-value reaction condition subspaces, thereby narrowing the search space and improving the efficiency of the refined optimization phase. According to an embodiment of the present invention, the traversal of the hierarchical optimization tree 219 begins at its root node and, at each level of the hierarchical optimization tree, iteratively selects the child node with the highest upper confidence bound until reaching the final leaf node. For example, at the highest level, a leaf node with an upper confidence bound of UCB = 72 is selected. Exploring downward from this leaf node, a leaf node with an upper confidence bound of UCB = 86 is selected at the next highest level. Further exploration from this leaf node results in a leaf node with an upper confidence bound of UCB = 93 at the next lower level being selected.

[0108] In an embodiment of the present invention, in a hierarchical optimization tree, a leaf node in the current level is used as a child node, and the leaf node with the highest upper confidence bound in the previous level of the current level is used as its parent node, where the upper confidence bound of each child node can be calculated based on the following formula (1):

[0109] (1)

[0110] in, is the average reaction performance (e.g., yield or selectivity) of child node i, representing the “utilization” term, It is obtained by dividing the cumulative response performance Qi of the node by the number of visits ni. is the total number of visits to the current parent node, ni is the number of visits to child node i, and Cp is the exploration constant, which is used to balance the importance between “exploitation” and “exploration”.

[0111] According to an embodiment of the present invention, in addition to selecting the leaf node with the highest upper confidence bound to enter the next step of exploration, the above formula (1) also ensures that while tending to select child nodes with good historical performance, it also gives child nodes with fewer visits a certain amount of exploration opportunities, thereby avoiding premature convergence to a suboptimal solution.

[0112] In some embodiments of the present invention, the upper confidence bound of the leaf node can be further updated. According to an embodiment of the present invention, the search space decomposition operation shown in the fourth box 240 can be used to determine a sub-search space for chemical reaction condition parameters. By sampling on the sub-search space using Bayesian optimization, a set of chemical reaction condition parameter combinations can be obtained, and reaction performance (for example, yield, selectivity, etc.) associated with these chemical reaction condition parameter combinations can be obtained. As an exemplary implementation, corresponding experiments can be carried out based on these chemical reaction condition parameter combinations, or alternatively, the data set results can be queried in the simulation to evaluate these chemical reaction condition parameter combinations to obtain reaction performance associated with these chemical reaction condition parameter combinations. On this basis, the determined reaction performance associated with these chemical reaction condition parameter combinations can be used to update the reaction performance associated with these chemical reaction condition parameter combinations indicated by the child nodes in the hierarchical optimization tree, and the number of visits to the child nodes can be updated. Based on the above formula (1), the upper confidence bound of the updated child node is determined using the updated reaction performance, the updated number of visits to the child node, and the updated number of visits to the parent node. According to an embodiment of the present invention, the reaction performance associated with the chemical reaction condition parameter combinations obtained by evaluating the chemical reaction condition parameter combinations can be recorded as , is used as a reward signal, which is back-propagated to update the statistics of all nodes on the path from the currently evaluated leaf node to the root node.

[0113] Specifically, the upper confidence bound of a child node can be updated based on the following rules:

[0114] Update the number of visits to each node i on the path to:

[0115] Update the cumulative reaction performance of each node i on the path to

[0116] Therefore, the average response performance of the node It is also updated accordingly, thus affecting the UCB value calculation and node selection in subsequent iterations.

[0117] According to an embodiment of the present invention, by continuously and iteratively updating the upper confidence bounds of child nodes, the evaluation of each leaf node of the hierarchical optimization tree will be gradually refined, thereby being able to more effectively guide the selection of optimal chemical reaction conditions.

[0118] In some embodiments of the present invention, as shown in third block 230, Bayesian optimization can be performed on the sub-search space based on the decomposition of the search space in fourth block 240 to determine an optimized combination of chemical reaction condition parameters. According to an embodiment of the present invention, the second agent can call the second model to perform predictive reasoning on the sub-search space to determine reaction performance as a pseudo-label for the chemical reaction condition parameter combination included in the sub-search space, and the third agent can call the third model to iteratively perform a Bayesian optimization algorithm on the sub-search space based on real prior data and the pseudo-label to determine the optimized combination of chemical reaction condition parameters. According to an embodiment of the present invention, the third agent explores the optimized combination of chemical reaction conditions in the sub-search space, which can significantly improve optimization efficiency.

[0119] like Figure 2 The process shown is a coarse-to-fine process, wherein the first stage (e.g., the first box 210, the fourth box 240) is a coarse optimization stage, in which the search space is knowledge-driven decomposed by using the chemical prior knowledge extracted from the literature by the large language model, and the upper confidence bound algorithm is used to identify potential high-quality subspaces; the second stage (e.g., the second box 220, the third box 230) is a fine optimization stage, in which data-driven Bayesian optimization is performed within the selected sub-search space, and pseudo samples generated by the fine-tuned large language model are used to accelerate the search, thereby determining the optimal reaction conditions.

[0120] Next reference Figure 3-Figure 5 The examples described below further describe the operation of the computing device according to the embodiment of the present invention.

[0121] In an embodiment of the present invention, the first agent can call the first model and, guided by the prompt word example 1 below, determine the importance of the chemical reaction condition parameter dimensions to be analyzed and rank the chemical reaction condition parameter dimensions based on the importance. The ranking result is shown in the following output example 1. As shown in output example 1, in addition to the ranking result of the chemical reaction condition parameter dimensions to be analyzed, the reasons for the ranking can also be output.

[0122] ===========;

[0123] Prompt word example 1:

[0124] Target:

[0125] Analyze the provided excerpt from the scientific literature on the Suzuki-Miyaura coupling reaction. Your task is to rank the following reaction parameters in descending order of importance based on their reported impact on yield or importance in the literature.

[0126] Key Notes:

[0127] Your analysis, conclusions, and final ranking must be based solely on the literature excerpts provided. No external knowledge, prior knowledge, or general chemical intuition should be used. You must strictly adhere to any explicit statements or implications of the given text regarding the effects of each parameter.

[0128] Reaction parameters to be sorted:

[0129] [Solvent type] [Base type] [Catalyst type] [Ligand type]…

[0130] Literature provided:

[0131] [LITERATURE_1] [LITERATURE_2]

[0132] ===========;

[0133] ===========;

[0134] Output example 1:

[0135] In the two retrieved literature excerpts, the most significant influence on the yield (or final conversion efficiency) of the Suzuki–Miyaura coupling was the type of base, followed by the type of catalyst, then the type of solvent, and finally the type of ligand. Combining all quantitative data with the authors' qualitative evaluations, we obtained the following ranking from highest to lowest importance:

[0136] [1] Type of base > [2] Type of catalyst > [3] Type of solvent > [4] Type of ligand

[0137] The following is a point-by-point analysis and reasons based on the literature.

[0138] 1 Type of base:

[0139] By changing the rate-controlling step and mechanism, in the Pd / MIL-101 system, the 6 h yield difference caused by different bases ranged from <3% (K2CO3) to 82% (NaOMe), a span of nearly 30 times.

[0140] 2. Type of catalyst:

[0141] The catalyst, usually a type of palladium, is the basis of the reaction. The nature, concentration, and activity of the catalyst are closely related to the reaction outcome.

[0142] ===========;

[0143] In an embodiment of the present invention, the first agent can invoke the first model and, guided by the prompt Example 2 below, classify candidate parameters in each chemical reaction condition parameter dimension to be analyzed according to their physicochemical properties. The classification results are shown in Output Example 2 below. As shown in Output Example 2, in addition to the candidate parameter category, the basis and reason for the classification can also be output.

[0144] ===========;

[0145] Prompt word example 2:

[0146] Target:

[0147] Identify the physicochemical properties that most significantly affect the yield (or final conversion efficiency) of the Suzuki–Miyaura coupling reaction for each variable dimension and categorize the candidate chemical substances in each variable dimension according to the identified specific physicochemical properties. This categorization must preferably be based on quantitative data available in a comprehensive database of physicochemical properties.

[0148] Key Notes:

[0149] 1. Prioritize the use of quantitative data:

[0150] For each substance and property, one should first attempt to classify it according to specific, measurable quantitative values ​​(e.g., pKa for acidity or basicity, dielectric constant for polarity, boiling point for volatility, number of specific functional groups).

[0151] 2. Minimize the use of general knowledge or intuition:

[0152] Avoid reliance on non-quantitative general chemical knowledge or intuition. If quantitative data directly supporting the classification exist in the database, this should be clearly stated. If a category generally does not have a direct quantitative value but has clear structural indications (e.g., number of hydrogen bond donors), the inference process should be explained based on similar data.

[0153] 3. Strictly adhere to the given classification categories:

[0154] Classify the substance strictly according to the category provided for each property. If a substance cannot be clearly classified, or the data put it on the edge of the category, this ambiguity should be pointed out.

[0155] Candidate substances for classification:

[0156] [Base: potassium hydroxide, sodium methoxide, potassium methoxide, potassium tert-butoxide, sodium tert-butoxide, potassium carbonate, cesium carbonate, potassium borate, sodium borate, potassium acetate, cesium acetate, potassium benzoate, potassium hydrogen borate] [Catalyst: ...] ...

[0157] Available tools:

[0158] [GoogleScholar Toolkit], [Arxiv Toolkit], [PubMed Toolkit], [PubChem Toolkit], [Wikipedia Toolkit]

[0159] ===========;

[0160] ===========;

[0161] Output example 2:

[0162] In the Suzuki–Miyaura coupling, the primary responsibility of the base is to activate the boronic acid / boronic ester into a more accessible boronate salt, while maintaining the palladium-ligand active site in a deprotonated state. This ability is directly dependent on the base's strength, defined as its conjugate acid pK-a. The pK-a physicochemical parameter is the most readily available and quantitatively comparable physicochemical parameter in authoritative databases and was therefore chosen as the most significant discriminant characteristic in this dimension.

[0163] Classification threshold and basis:

[0164] grade pK-a (conjugate acid, 25 °C) Typical behavior of the corresponding base in coupling strong base ≥ 15 Borates can be rapidly generated in weakly polar solvents at room temperature, often leading to "increased rate + increased side reactions" Medium base 9 –<15 It has both activation and mildness, and is the most common high yield window weak base 5 –<9 Acceptable only under highly polar or high temperature conditions; often requires a synergistic solvent or phase transfer system Very weak base <5 Usually insufficient to drive coupling, only acts as a buffer or adds synergy

[0165] Quantitative classification of candidate bases:

[0166] Candidate base Conjugate acid pK-a Classification Main data sources potassium hydroxide <![CDATA[H2O]]> 15.7 strong base Sodium methoxide / potassium methoxide <![CDATA[CH3OH]]> 15.5 strong base Potassium tert-butoxide / Sodium tert-butoxide <![CDATA[(CH3)3COH]]> 16.5 ± 0.1 strong base Potassium carbonate / cesium carbonate <![CDATA[HCO3⁻ → CO3²⁻ (pK-a2)]]> 10.25 Medium base Potassium borate / sodium borate <![CDATA[B(OH)3]]> 9.24 ± 0.1 Medium base Potassium acetate / cesium acetate <![CDATA[CH3COOH]]> 4.76 weak base Potassium benzoate <![CDATA[C6H5COOH]]> 4.19 weak base Potassium hydrogen borate <![CDATA[B(OH)3 / H2BO3⁻]]> 9.24 (estimated, same as above) Medium base†

[0167] Note: If you need to continue to perform the same physical property-quantitative classification for the "catalyst" or "ligand" dimension, please provide a list of corresponding candidate substances so that the same process can be called for retrieval and classification.

[0168] ===========;

[0169] In an embodiment of the present invention, the third agent can call the third model to receive a third user input, the third user input including Figure 3 The prior data shown, Figure 4 The search space shown in FIG (corresponding to the root node of the hierarchical optimization tree), the exemplary pseudo-labels shown below, and the exemplary sub-search space shown below (corresponding to the combination of sub-nodes of the hierarchical optimization tree). The third agent can call the third model to iteratively perform the Bayesian optimization algorithm on the search space defined by the hierarchical optimization tree based on the real prior data and the pseudo-labels to determine the optimized chemical reaction condition parameter combination. The output of these optimized chemical reaction condition parameter combinations is as follows: Figure 5 shown.

[0170] ===========;

[0171] Example pseudo-labels:

[0172] tensor([[3.7643],

[0173] [1.1229],

[0174] [2.4587],

[0175] [5.0572],

[0176] [2.2192],

[0177] [4.8189],

[0178] [0.5794],

[0179] [3.5773],

[0180] [2.4137],

[0181] [0.3265],

[0182] … ])

[0184] ===========;

[0185] ===========;

[0186] Example sub-search space: [

[0188] "Base": [ [

[0190] "[OH-].[K+]",

[0191] "C[O-].[Na+]",

[0192] "C[O-].[K+]",

[0193] "CC(C)(C)[O-].[K+]",

[0194] "CC(C)(C)[O-].[Na+]"

[0195] ], [

[0197] "C(=O)(O)[O-].[K+].[K+]",

[0198] "C(=O)([O-])[O-].[Cs+].[Cs+]",

[0199] "[K+].[K+].[K+].[O-]P([O-])([O-])=O",

[0200] "[Na+].[Na+].[Na+].[O-]P([O-])([O-])=O"

[0201] ], [

[0203] "CC(=O)[O-].[K+]",

[0204] "CC(=O)[O-].[Cs+]",

[0205] "C1=CC=C(C=C1)C(=O)[O-].[K+]",

[0206] "[K+].[K+].OP(O)([O-])=O",

[0207] "C(=O)(O)[O-].[K+]" ]

[0209] ],

[0210] "Solvent": [ [

[0212] "CC1=CC(=C(C=C1)N)C",

[0213] "CN(C)C=O",

[0214] "CC#N",

[0215] "CO",

[0216] "CN1CCCC1=O"

[0217] ], [

[0219] "C1CCOC1",

[0220] "CC1CCCO1" ,

[0221] "C1=CC=C(C=C1)Cl" ,

[0222] "COCCOC"

[0223] ], [

[0225] "CC1=CC=CC=C1",

[0226] "C1COCCO1" ]

[0228] ],

[0229] "Catalyst": [

[0230] …

[0231] ],

[0232] … ]

[0234] ===========;

[0235] According to some aspects of the present invention, a method for optimizing chemical reaction conditions is provided.

[0236] Figure 6 The method for optimizing chemical reaction conditions according to some embodiments of the present invention is shown. Figure 1 The computing device 100 is described as performing.

[0237] The method may include step S1: constructing a hierarchical optimization tree by a first agent, wherein the hierarchical optimization tree is used to define a search space for chemical reaction condition parameters. Figure 7 The first process associated with step S1 will be described in detail.

[0238] The method may include step S2: performing predictive reasoning on the search space by a second agent to generate pseudo labels for combinations of chemical reaction condition parameters.

[0239] The method may further include step S3: performing, by a third intelligent agent, Bayesian optimization on a search space defined by the hierarchical optimization tree to determine an optimized combination of chemical reaction condition parameters.

[0240] According to an embodiment of the present invention, the search space for chemical reaction condition parameters is constructed as a hierarchical optimization tree, and Bayesian optimization is performed on the search space defined by the hierarchical optimization tree. In addition, the second intelligent agent performs predictive reasoning on the search space to generate pseudo labels, which can alleviate the "cold start" problem caused by the lack of effective implementation data in the early stage of the experiment, and is conducive to solving the "curse of dimensionality" problem caused by the high dimensionality of the search space for chemical reaction condition parameters.

[0241] Figure 7 A first process associated with a method for optimizing chemical reaction conditions according to some embodiments of the present invention is shown. Figure 1 The first agent 122 in the computing device 100 described above performs Figure 6 The described step S1 corresponds.

[0242] The first process may include step S11: the first agent calls the first model to analyze chemical reaction condition parameters.

[0243] In some embodiments, step S11 may include step S111: receiving a first user input, where the first user input includes a target chemical reaction, chemical reaction condition parameter dimensions to be analyzed, candidate parameters in each chemical reaction condition parameter dimension to be analyzed, and provided literature.

[0244] In some embodiments, step S11 may include step S112: determining the importance of the chemical reaction condition parameter dimension to be analyzed to the reaction performance of the target chemical reaction based on the provided literature.

[0245] In some embodiments, step S11 may include step S113: sorting the chemical reaction condition parameter dimensions to be analyzed based on the importance levels determined in step S112.

[0246] In some embodiments, step S11 may include step S114: for each candidate parameter in the chemical reaction condition parameter dimension to be analyzed, classifying the candidate parameter according to the physicochemical properties of the candidate parameter.

[0247] The first process may include step S12: the first agent constructs a hierarchical optimization tree layer by layer based on the sorting result of step S113 and the categories of the candidate parameters classified in step S114.

[0248] Figure 8 A second process associated with a method for optimizing chemical reaction conditions according to some embodiments of the present invention is shown. Figure 1 The second agent 124 in the computing device 100 described above calls the second model to execute Figure 6 The described step S2 corresponds.

[0249] The second process may include step S21 of receiving a second user input, the second user input including a combination of candidate parameters associated with a target chemical reaction.

[0250] The second process may further include step S22: performing predictive reasoning on the search space based on the second user input to determine a reaction performance as a pseudo label of the combination of candidate parameters, the reaction performance being associated with the combination of candidate parameters indicated by the second user input and the target chemical reaction.

[0251] Figure 9 A third process associated with a method for optimizing chemical reaction conditions according to some embodiments of the present invention is shown. Figure 1 The third agent 126 in the computing device 100 described above calls the third model to execute Figure 6 The described step S3 corresponds.

[0252] The third process may include step S31: receiving a third user input, where the third user input includes true prior data associated with the target chemical reaction, the hierarchical optimization tree constructed in step S1, and the pseudo labels generated in step S2.

[0253] The third process may further include step S32: iteratively executing a Bayesian optimization algorithm on the search space defined by the hierarchical optimization tree based on the real prior data and the pseudo labels to determine an optimized combination of chemical reaction condition parameters.

[0254] In some embodiments of the present invention, the first agent can construct a hierarchical optimization tree based on the ranking results of the importance of the chemical reaction condition parameter dimensions and the categories of the chemical reaction condition parameters in each chemical reaction condition parameter dimension to limit the search space for the chemical reaction condition parameters. The hierarchical optimization tree is a tree structure having a root node and leaf nodes organized hierarchically, wherein the root node of the hierarchical optimization tree corresponds to a complete search space, which includes all chemical reaction condition parameter combinations associated with the target chemical reaction. According to an embodiment of the present invention, the first agent can first determine the root node of the hierarchical optimization tree, and then construct the leaf nodes of the hierarchical optimization tree layer by layer. Each level of the leaf nodes of the hierarchical optimization tree can indicate a chemical reaction condition parameter dimension, and each leaf node in each level indicates a category of chemical reaction condition parameters. From the root node of the hierarchical optimization tree to the lowest level, the importance of the chemical reaction condition parameter dimensions associated with the level of the leaf node decreases.

[0255] In addition, in some embodiments of the present invention, the hierarchical optimization tree can also define a sub-search space for chemical reaction condition parameters, which includes a subset of chemical reaction condition parameter combinations associated with the target chemical reaction and is determined by traversing the hierarchical optimization tree layer by layer. According to an embodiment of the present invention, an upper confidence bound (UCB) algorithm can be used to traverse the hierarchical optimization tree layer by layer to identify and sample high-value reaction condition subspaces, thereby narrowing the search space and improving the efficiency of the fine optimization stage. According to an embodiment of the present invention, the traversal of the hierarchical optimization tree starts from its root node, and in each level of the hierarchical optimization tree, the child node with the highest upper confidence bound is iteratively selected until the final leaf node is reached.

[0256] In some embodiments of the present invention, in a hierarchical optimization tree, a leaf node in the current level is used as a child node, and a leaf node with the highest upper confidence bound in the previous level of the current level is used as its parent node, where the upper confidence bound of each child node can be calculated based on the above formula (1).

[0257] According to an embodiment of the present invention, in addition to selecting the leaf node with the highest upper confidence bound to enter the next step of exploration, the above formula (1) also ensures that while tending to select child nodes with good historical performance, it also gives child nodes with fewer visits a certain amount of exploration opportunities, thereby avoiding premature convergence to a suboptimal solution.

[0258] In some embodiments of the present invention, the upper confidence bound of the leaf node can be further updated. According to an embodiment of the present invention, a set of chemical reaction condition parameter combinations can be obtained by sampling on the sub-search space using Bayesian optimization, and reaction performance (for example, yield, selectivity, etc.) associated with these chemical reaction condition parameter combinations can be obtained. As an exemplary implementation method, corresponding experiments can be carried out based on these chemical reaction condition parameter combinations, or alternatively, the data set results can be queried in the simulation to evaluate these chemical reaction condition parameter combinations to obtain the reaction performance associated with these chemical reaction condition parameter combinations. On this basis, the determined reaction performance associated with these chemical reaction condition parameter combinations can be used to update the reaction performance associated with these chemical reaction condition parameter combinations indicated by the child nodes in the hierarchical optimization tree, and the number of visits to the child nodes can be updated. Based on the above formula (1), the upper confidence bound of the updated child node is determined using the updated reaction performance, the updated number of visits to the child node, and the number of visits to the parent node. According to an embodiment of the present invention, the reaction performance associated with these chemical reaction condition parameter combinations obtained by evaluating the chemical reaction condition parameter combinations can be recorded as , is used as a reward signal, which is back-propagated to update the statistics of all nodes on the path from the currently evaluated leaf node to the root node.

[0259] According to an embodiment of the present invention, by continuously and iteratively updating the upper confidence bounds of child nodes, the evaluation of each leaf node of the hierarchical optimization tree will be gradually refined, thereby being able to more effectively guide the selection of optimal chemical reaction conditions.

[0260] In some embodiments of the present invention, based on the decomposition of the search space, the second agent can invoke the second model to perform step S22 by performing predictive reasoning on the sub-search space based on the second user input to determine pseudo-labels for the reaction performance as candidate parameter combinations. Based on this, the third agent can invoke the third model to perform step S32 by iteratively executing a Bayesian optimization algorithm on the sub-search space based on real prior data and pseudo-labels to determine an optimized combination of chemical reaction condition parameters. According to embodiments of the present invention, the third agent explores optimized chemical reaction condition combinations in the sub-search space, which can significantly improve optimization efficiency.

[0261] Figure 10 Some embodiments of the present invention are shown in FIG. Figure 9 Exemplary operations associated with the third process are described.

[0262] The operation may include step S321: initializing a third model based on the pseudo labels.

[0263] The operation may include step S322: executing a Bayesian optimization algorithm on the search space via the initialized third model to determine an initial chemical reaction condition parameter combination;

[0264] The operation may include step S323 of updating the third model based on actual reaction performance associated with the chemical reaction condition parameter combination.

[0265] In some embodiments, step S323 may include step S3231: calculating the similarity between the pseudo label and the true reaction performance, wherein the pseudo label includes a first pseudo label.

[0266] In some embodiments, step S323 may include step S3232: determining that the similarity between the first pseudo label and the true reaction performance is less than a similarity threshold.

[0267] In some embodiments, step S323 may include step S3233: removing the first pseudo label from the pseudo labels to update the third model.

[0268] In some embodiments, step S323 may include step S3234: randomly removing one or more of the pseudo labels from the pseudo labels to update the third model.

[0269] In some embodiments, step S3234 may be performed instead of steps S3231 , S3232 , and S3233 .

[0270] The operation may include step S324: executing a Bayesian optimization algorithm on the search space via the updated third model to determine an optimized combination of chemical reaction condition parameters.

[0271] According to another aspect of the present invention, a computer-readable storage medium is provided.

[0272] Figure 11 is a block diagram of a computer-readable storage medium 1100 according to some embodiments of the present invention.

[0273] The computer readable storage medium 1100 stores a computer program 1150. When the computer program 1150 is executed by the processor, the computer program 1150 realizes the above combination Figures 6-10 The steps of each method or process are described.

[0274] According to another aspect of the present invention, a computer program product is provided.

[0275] Figure 12 is a block diagram of a computer program product 1200 according to some embodiments of the present invention.

[0276] The computer program product 1200 may include a computer program 1150. When the computer program 1150 is executed by a processor, the computer program 1150 implements the above-mentioned Figures 6-10 The steps of each method or process are described.

[0277] To validate the effectiveness of our proposed solution for optimizing chemical reaction conditions, we conducted a comparative experiment using our proposed solution and existing solutions. The experimental data were obtained from a publicly available dataset collected and collated using high-throughput equipment (Science, 359(6374):429–434, 2018).

[0278] The following will refer to Figures 13-18 The experimental results are explained.

[0279] The present invention uses the scheme for optimizing chemical reaction conditions proposed by the present invention and the traditional Bayesian optimization scheme to perform ten rounds of long-iteration Suzuki-Miyaura coupling chemical reaction condition optimization tasks.

[0280] Figure 13 The optimization trends, results and comprehensive evaluation of convergence efficiency of the two schemes are shown. Figure 13 As shown, the solution proposed by the present invention (in Figure 13 The proposed method (denoted as ChemBOMAS in the original text) achieved a high yield in the first iteration, exceeding the yield of traditional Bayesian optimization solutions by approximately 30%. The yield performance was consistently better than that of traditional Bayesian optimization solutions throughout the iterations, with the highest yield improvement reaching approximately 17%. Furthermore, it is noteworthy that the proposed method converged in the fourth iteration, while the traditional Bayesian optimization solution had not yet converged by the ninth iteration. This demonstrates an improvement in optimization efficiency of over 50% over traditional Bayesian optimization solutions.

[0281] In addition, the present invention also repeated the above ten rounds of long-iteration Suzuki-Miyaura coupling chemical reaction condition optimization task. Figure 14 A comprehensive evaluation of the optimization trends, results, and convergence efficiency of the two schemes is shown when the above ten-round long iterative Suzuki-Miyaura coupling chemical reaction condition optimization task is repeated ten times.

[0282] like Figure 14 As shown, after 10 independent repeated experiments, the solution proposed by the present invention (in Figure 14The yield performance of the proposed method (denoted as ChemBOMAS in the figure) is consistently better than that of the traditional Bayesian optimization scheme, and the standard deviation of its experimental process is also significantly smaller than that of the traditional Bayesian optimization scheme.

[0283] To evaluate and compare the robustness and superiority of the proposed method in mitigating the "cold start" effect of Bayesian optimization, we conducted comparative experiments using four optimization tasks and three other representative optimization schemes. Each optimization task was set to run five rounds (i.e., four iterations) to focus on performance in the early stages of optimization. The four optimization tasks were the Suzuki-Miyaura coupling reaction, the Buchwald-Hartwig coupling reaction, a tandem arylation reaction, and an enantioselective reaction. The three representative optimization schemes were traditional Bayesian optimization (implemented using the BoTorch framework), LA-MCTS (implemented using code from Wang et al., https: / / doi.org / 10.48550 / arXiv.2007.00708), and LLAMBO (implemented using the default configuration provided in the code repository of the original authors in the paper by Liu et al., https: / / doi.org / 10.48550 / arXiv.2402.03921).

[0284] The comparative experimental results for these four optimization tasks are respectively Figures 15-18 Shown, where Figure 15 Comparative experimental results are shown for the Buchwald-Hartwig coupling reaction, Figure 16 The comparative experimental results are shown for the Suzuki-Miyaura coupling reaction, Figure 17 Comparative experimental results are shown for a tandem arylation reaction, Figure 18 The comparative experimental results shown are for enantioselective reactions. Figures 15-18 In the figure, the experimental results of the scheme proposed in the present invention are marked as ChemBOMAS, which are shown as purple dots and purple curves, the experimental results of the traditional Bayesian Optimization scheme are shown as cyan dots and cyan curves, the experimental results of the LA-MCTS scheme are shown as brown dots and brown curves, and the experimental results of the LLAMBO scheme are shown as yellow dots and yellow curves. Figure 15-17 The "target value" in indicates the yield, Figure 18 The "Target Value" in indicates the selectivity.

[0285] like Figures 15-18As shown in Figure 2, in the four chemical reaction optimization tasks, the scheme proposed by the present invention always outperforms all other schemes in terms of initial performance, and in each task, the scheme proposed by the present invention can identify the chemical reaction condition parameter combination that achieves the optimal target value within only four rounds of iterations. In addition, as Figure 16 As shown in the Suzuki-Miyaura coupling reaction task, LLAMBO achieved an optimal target value of 96.3% in the first round of iteration, slightly higher than the 92.56% of the initial round of the proposed solution. However, at the end of the optimization process, the proposed solution also achieved an optimal target value of 96.3%, which was on par with LLAMBO.

[0286] In addition, the present invention further verified the superior optimization capabilities of the proposed multi-agent collaborative solution based on ablation experiments. The proposed solution significantly outperformed three different pruned versions: (i) removing the first agent, (ii) removing the second agent, and (iii) removing both the first and second agents. A key finding was that removing the second agent performed better in the early stages of the experiment than removing the first agent, indicating that the knowledge-driven first agent contributes more significantly to alleviating the cold start problem in Bayesian optimization. It was also noted that the standard Bayesian optimization method, which completely excludes the first and second agents, performed the worst among all tested configurations, achieving the lowest optimization effect.

[0287] To further verify the effectiveness and applicability of the proposed scheme, an algorithm-driven practical wet experiment was conducted. This experiment used the proposed scheme for optimizing chemical reaction conditions to maximize the yield of the palladium-catalyzed cross-coupling reaction of borate esters and aryl chlorides. This complex optimization task came from a pharmaceutical company and was subject to four strict practical constraints: (1) there was a complete lack of prior data for reference; (2) there was a six-dimensional parameter space, which was reported to be more than 70 times the size of similar public studies, greatly increasing the difficulty of exploration; (3) cost control required a tenfold reduction in catalyst dosage compared to the traditional level, a condition known to significantly inhibit product formation; and (4) the number of experiments was limited to approximately 50 to reduce labor costs.

[0288] In this wet lab task, the proposed scheme for optimizing chemical reaction conditions successfully identified a combination of reaction parameters with a yield of 96%, significantly exceeding the 15% yield achieved by manual optimization by chemists using traditional single-variable control methods. Furthermore, three notable observations were made: First, in the initial rounds, the proposed scheme identified a combination of reaction parameters that achieved the highest yield of 90%, exceeding the preset yield threshold of 75%. Second, the combination with a yield of 96% was discovered early in the optimization process, specifically in the second iteration. Third, as the optimization process progressed, the proportion of combinations with yields exceeding 75% among the combinations recommended by the proposed scheme continued to increase, indicating that the multiple agents in the proposed scheme were continuously optimizing. In the first five iterations, the number of combinations with yields ≥75% was 1, 2, 3, 5, and 5, respectively. The above-mentioned excellent initialization performance, fast convergence ability and continuous improvement of the proxy model together demonstrate the efficiency and effectiveness of the proposed solution in accelerating chemical reaction optimization.

Claims

1. A computing device for optimizing chemical reaction conditions, characterized in that: include: Computing resources; as well as A plurality of intelligent agents, the plurality of intelligent agents being executed by the computing resources, the plurality of intelligent agents comprising: A first agent is configured to invoke a first model to: receiving a first user input, the first user input including a target chemical reaction, a chemical reaction condition parameter dimension to be analyzed, a candidate parameter in each chemical reaction condition parameter dimension to be analyzed, and provided literature; Determining the importance of the chemical reaction condition parameter dimension to be analyzed to the reaction performance of the target chemical reaction based on the provided literature; sorting the chemical reaction condition parameter dimensions to be analyzed based on the importance; For each candidate parameter in the chemical reaction condition parameter dimension to be analyzed, the candidate parameters are classified according to their physicochemical properties. wherein a hierarchical optimization tree is constructed layer by layer based on the ranking results and the categories of the candidate parameters, and the hierarchical optimization tree is used to limit the search space for chemical reaction condition parameters; A second agent is configured to invoke a second model to: receiving a second user input comprising a combination of candidate parameters associated with the target chemical reaction; and performing predictive inference on the search space based on the second user input to determine a reaction performance as a pseudo label for the combination of the candidate parameters, the reaction performance being associated with the combination of the candidate parameters indicated by the second user input and the target chemical reaction; and A third agent is configured to invoke a third model to: receiving a third user input, the third user input comprising true prior data associated with the target chemical reaction, the hierarchical optimization tree, and the pseudo-label; and Based on the true prior data and the pseudo labels, a Bayesian optimization algorithm is iteratively performed on the search space defined by the hierarchical optimization tree to determine an optimized combination of chemical reaction condition parameters.

2. The computing device of claim 1, wherein: The chemical reaction condition parameter dimension to be analyzed is selected from a chemical reaction condition parameter dimension group, and the chemical reaction condition parameter dimension group includes at least one of a reactant dimension, a solvent dimension, a base dimension, a catalyst dimension, and a ligand dimension.

3. The computing device of claim 1, wherein: The first agent is configured to call the first model to perform at least one of the following for each candidate parameter in the chemical reaction condition parameter dimension to be analyzed: Analyzing the provided literature to identify key physicochemical properties of the candidate parameters, and classifying the candidate parameters according to the key physicochemical properties, wherein the key physicochemical properties indicate physicochemical properties that affect the reaction performance of the target chemical reaction; The quantitative values ​​of the physicochemical properties of the candidate parameters are searched in a physicochemical property database, and the candidate parameters are classified according to the similarity based on the quantitative values ​​of the physicochemical properties.

4. The computing device of claim 3, wherein: The key physicochemical properties include at least one of acidity and alkalinity, polarity, valence, steric effect, electronic effect, volatility, and functional group, and the quantitative value of the physicochemical property includes quantitative data or structural indication describing the physicochemical property.

5. The computing device of claim 1, wherein: The hierarchical optimization tree includes a root node and leaf nodes organized hierarchically, the root node indicates a search space including all chemical reaction condition parameter combinations associated with the target chemical reaction, each level of the leaf nodes indicates a chemical reaction condition parameter dimension, each leaf node indicates a category of candidate parameters, and from the root node to the lowest level of the hierarchical optimization tree, the importance of the chemical reaction condition parameter dimensions associated with the levels of the leaf nodes decreases, wherein each category of candidate parameters includes a group of candidate parameters with similar physical and chemical properties.

6. The computing device of claim 5, wherein: The hierarchical optimization tree is further used to define a sub-search space, which includes a subset of chemical reaction condition parameter combinations associated with the target chemical reaction, and wherein the sub-search space is determined by traversing the hierarchical optimization tree layer by layer.

7. The computing device of claim 6, wherein: The sub-search space is determined based on an upper confidence bound algorithm, from a root node of the hierarchical optimization tree to the lowest level, by iteratively selecting a leaf node with a highest upper confidence bound at each level of the hierarchical optimization tree.

8. The computing device of claim 7, wherein: In the hierarchical optimization tree, the leaf nodes in the current level serve as child nodes, and the leaf nodes with the highest upper confidence bound in the previous level of the current level serve as parent nodes, and wherein the upper confidence bounds of the child nodes are determined based on the reaction performance associated with the candidate parameters indicated by the child nodes, the number of visits to the child nodes, and the number of visits to the parent nodes.

9. The computing device of claim 8, wherein: The upper confidence bound of the child node is further updated by the following steps: A Bayesian optimization algorithm is used to sample chemical reaction condition parameter combinations in the sub-search space; determining a reaction performance associated with the chemical reaction condition parameter combination based on the sampling results; updating the reaction performance associated with the candidate parameter indicated by the child node using the reaction performance associated with the chemical reaction condition parameter combination; Update the number of visits to the child node; as well as An updated upper confidence bound of the child node is determined based on the updated reaction performance, the updated number of visits to the child node, and the number of visits to the parent node.

10. The computing device of claim 7, wherein: The second agent is configured to invoke a second model to: performing predictive inference on the sub-search space based on the second user input to determine reaction performance as a pseudo label for the combination of the candidate parameters, And wherein the third agent is configured to call a third model to: Based on the true prior data and the pseudo labels, a Bayesian optimization algorithm is iteratively performed on the sub-search space to determine an optimized combination of chemical reaction condition parameters.

11. The computing device of claim 1 , wherein: The second model is obtained through pre-training based on a basic large model, wherein the basic large model adopts causal language modeling loss and is pre-trained by performing a chemical reaction condition prediction task on a training set, wherein the chemical reaction condition prediction task is used to predict chemical reaction condition parameters for the chemical reaction given the reactants and products of the chemical reaction.

12. The computing device of claim 11, wherein: The second model is obtained through fine-tuning based on a pre-trained basic large model, wherein the pre-trained basic large model is further adjusted to generate predicted values ​​of reaction performance associated with the predicted chemical reaction condition parameters and the chemical reaction, and wherein the pre-trained basic large model is fine-tuned by a low-rank adaptation method based on the loss between the predicted value of the reaction performance and the true value of the reaction performance.

13. The computing device of claim 1, wherein: The third agent is configured to invoke the third model to: Initializing the third model based on the pseudo labels; executing a Bayesian optimization algorithm on the search space via the initialized third model to determine an initial chemical reaction condition parameter combination; updating the third model based on actual reaction performance associated with the chemical reaction condition parameter combination; as well as A Bayesian optimization algorithm is executed on the search space via the updated third model to determine the optimized chemical reaction condition parameter combination.

14. The computing device of claim 13, wherein: The third agent is configured to call a third model to: calculating a similarity between the pseudo labels and the true reaction performance, wherein the pseudo labels include a first pseudo label; determining that a similarity between the first pseudo-label and the true reaction performance is less than a similarity threshold; removing the first pseudo label from the pseudo labels to update the third model; as well as A Bayesian optimization algorithm is executed on the search space via the updated third model to determine the optimized chemical reaction condition parameter combination.

15. The computing device of claim 13, wherein: The third agent is configured to call a third model to: randomly removing one or more of the pseudo labels from the pseudo labels to update the third model; as well as A Bayesian optimization algorithm is executed on the search space via the updated third model to determine the optimized chemical reaction condition parameter combination.

16. A method for optimizing chemical reaction conditions, characterized in that: The following steps are involved: S1: A hierarchical optimization tree is constructed by a first agent, wherein the hierarchical optimization tree is used to define a search space for chemical reaction condition parameters. Step S1 includes: S11: The first agent calls the first model to perform the following steps: S111: receiving a first user input, where the first user input includes a target chemical reaction, a chemical reaction condition parameter dimension to be analyzed, candidate parameters in each chemical reaction condition parameter dimension to be analyzed, and provided literature; S112: Determining the importance of the chemical reaction condition parameter dimension to be analyzed to the reaction performance of the target chemical reaction based on the provided literature; S113: sorting the chemical reaction condition parameter dimensions to be analyzed based on the importance; and S114: for each candidate parameter in the chemical reaction condition parameter dimension to be analyzed, classify the candidate parameter according to the physicochemical properties of the candidate parameter; and S12: The first agent constructs the hierarchical optimization tree layer by layer based on the sorting result and the categories of the candidate parameters; S2: A second agent performs predictive reasoning on the search space to generate pseudo labels for combinations of chemical reaction condition parameters. Step S2 includes the second agent calling a second model to perform the following steps: S21: receiving a second user input, wherein the second user input includes a combination of candidate parameters associated with the target chemical reaction; and S22: performing predictive reasoning on the search space based on the second user input to determine a reaction performance as a pseudo label of the combination of the candidate parameters, the reaction performance being associated with the combination of the candidate parameters indicated by the second user input and the target chemical reaction; and S3: performing Bayesian optimization on the search space defined by the hierarchical optimization tree by a third agent to determine an optimized combination of chemical reaction condition parameters, wherein step S3 includes calling a third model by the third agent to perform the following steps: S31: receiving a third user input, wherein the third user input includes real prior data associated with the target chemical reaction, the hierarchical optimization tree, and the pseudo label; and S32: Based on the true prior data and the pseudo labels, iteratively perform a Bayesian optimization algorithm on the search space defined by the hierarchical optimization tree to determine an optimized combination of chemical reaction condition parameters.

17. The method according to claim 16, wherein The hierarchical optimization tree includes a root node and hierarchically organized leaf nodes, wherein the root node indicates a search space including all chemical reaction condition parameter combinations associated with the target chemical reaction, each level of the leaf nodes indicates a chemical reaction condition parameter dimension, each leaf node indicates a category of candidate parameters, and from the root node to the lowest level of the hierarchical optimization tree, the importance of the chemical reaction condition parameter dimensions associated with the levels of the leaf nodes decreases, and each category of candidate parameters includes a group of candidate parameters with similar physical and chemical properties. And wherein, the hierarchical optimization tree is further used to define a sub-search space, which includes a subset of chemical reaction condition parameter combinations associated with the target chemical reaction and is determined by traversing the hierarchical optimization tree layer by layer.

18. The method according to claim 17, wherein The sub-search space is determined based on an upper confidence bound algorithm, from a root node of the hierarchical optimization tree to the lowest level, by iteratively selecting a leaf node with a highest upper confidence bound at each level of the hierarchical optimization tree.

19. The method according to claim 18, wherein In the hierarchical optimization tree, a leaf node in a current level is used as a child node, and a leaf node with a highest upper confidence bound in a level above the current level is used as a parent node, wherein the upper confidence bound of the child node is determined based on the reaction performance associated with the candidate parameter indicated by the child node, the number of visits to the child node, and the number of visits to the parent node. And wherein the upper confidence bound of the child node is further updated by the following steps: A Bayesian optimization algorithm is used to sample chemical reaction condition parameter combinations in the sub-search space; determining a reaction performance associated with the chemical reaction condition parameter combination based on the sampling results; updating the reaction performance associated with the candidate parameter indicated by the child node using the reaction performance associated with the chemical reaction condition parameter combination; Updating the number of visits to the child node; and An updated upper confidence bound of the child node is determined based on the updated reaction performance, the updated number of visits to the child node, and the number of visits to the parent node.

20. The method according to claim 19, wherein The step S22 includes: performing predictive inference on the sub-search space based on the second user input to determine reaction performance as a pseudo label for the combination of the candidate parameters, And wherein, the step S32 includes: Based on the true prior data and the pseudo labels, a Bayesian optimization algorithm is iteratively performed on the sub-search space to determine an optimized combination of chemical reaction condition parameters.

21. The method according to claim 16, wherein The step S32 includes: S321: Initializing the third model based on the pseudo label; S322: executing a Bayesian optimization algorithm on the search space via the initialized third model to determine an initial chemical reaction condition parameter combination; S323: updating the third model based on the actual reaction performance associated with the chemical reaction condition parameter combination; and S324: Execute a Bayesian optimization algorithm on the search space via the updated third model to determine the optimized chemical reaction condition parameter combination.

22. The method according to claim 21, wherein The step S323 includes: S3231: Calculating the similarity between the pseudo label and the true reaction performance, wherein the pseudo label includes a first pseudo label; S3232: Determine whether the similarity between the first pseudo label and the true reaction performance is less than a similarity threshold; S3233: Remove the first pseudo label from the pseudo labels to update the third model.

23. The method according to claim 21, wherein The step S323 includes: S3234: Randomly remove one or more pseudo labels from the pseudo labels to update the third model.

24. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 16 to 23 are implemented.

25. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 16 to 23 are implemented.

Images