A method and system for optimizing chemical reaction conditions based on deterministic cost constraints
By optimizing chemical reaction conditions through structured storage and a multi-task Gaussian process model, the difficulties of Bayesian optimization methods in handling heterogeneous information are solved, achieving efficient and low-cost chemical reaction optimization and providing reliable yield and cost decision support.
Patent Information
- Application Number
- CN202611564840.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-10
- Publication Date
- 2026-11-17
AI Technical Summary
Existing Bayesian optimization methods face difficulties in processing heterogeneous information in chemical experiments, resulting in ineffective use of experimental data, high costs and low efficiency, inability to accurately handle qualitative phenomena and censored observations, and inability to provide reliable experimental decision-making basis.
A chemical reaction condition optimization method based on deterministic cost constraints is adopted. By storing experimental data in a structured manner, a multi-task Gaussian process model is trained, a joint model is constructed for yield prediction and cost analysis, and experimental conditions are optimized by combining the acquisition function to achieve data-driven iterative optimization.
It significantly improves the utilization efficiency of scarce experimental data, provides reliable yield prediction and cost decision-making basis, reduces experimental costs, avoids invalid experiments, and outputs multiple sets of trade-off solutions to adapt to actual process requirements.
Smart Images

Figure CN123120086A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of computer-aided chemical experiment design technology, and in particular to a method and system for optimizing chemical reaction conditions based on deterministic cost constraints. Background Technology
[0002] In recent years, with the deepening integration of artificial intelligence and experimental science, Bayesian optimization, as a sequential experimental design method based on probabilistic surrogate models and acquisition functions, has been widely applied in the field of chemical reaction condition optimization.
[0003] Most existing Bayesian optimization methods for chemical reaction optimization implicitly assume a high-throughput automated experimental platform, namely, that the amount of data is relatively abundant, each experimental point can produce a regular yield value, and the observation noise follows a symmetrical distribution. However, in most ordinary artificial chemical laboratories, the experimental data, in addition to the precise yield value, also contain a large amount of qualitative phenomenon records and censored observations. These various heterogeneous information appear intertwined in the experimental records, and the yield observation values show a unidirectional low bias due to the inherent characteristics of chemical experiments such as separation loss and transfer loss. This complex and mixed data characteristic makes it difficult for existing Bayesian optimization methods to process this irregular information. Existing Bayesian optimization methods lack the ability to mathematically rigorously process information such as failure reasons, semi-quantitative records, visual observations, and separation losses. As a result, the experimental data obtained after the experiment is conducted is unusable, and the same experiment needs to be repeated, which is costly and inefficient. Summary of the Invention
[0004] In view of this, this application proposes a method, system, electronic device, and nontransient computer storage medium for optimizing chemical reaction conditions based on deterministic cost constraints.
[0005] In a first aspect, the present invention provides a method for optimizing chemical reaction conditions based on deterministic cost constraints, comprising: S1 stores the chemical reaction configuration information input by the user in a structured manner. The user configuration information includes the value range of each reaction condition, preset rules, deterministic cost model, termination conditions, and initial reaction condition set; experiments are conducted on the initial reaction condition set to obtain initial experimental data; S2 uses pre-defined rules to uniformly transform the initial experimental data to obtain structured observation data; S3 inputs structured observation data into a multi-task Gaussian process model, trains the multi-task Gaussian process model, and obtains the yield predictor parameter package. S4 constructs a joint model based on the yield predictor parameter package and the deterministic cost model. It samples the candidate reaction condition set within the range of reaction condition values and uses the joint model to conduct experimental simulations on the candidate reaction condition set to obtain simulation experimental data. S5 acquires historical experimental data and constructs the current Pareto front using this data. Based on the current Pareto front, it constructs a data acquisition function, collects simulation experimental data, and selects a recommended experimental list from the candidate response condition set with the goal of maximizing the data acquisition function. S6 performs experiments on the recommended experimental list to obtain the latest experimental data, replaces the initial experimental data with the latest experimental data, returns to step S2, and continues until the termination condition is met, stops the loop, and obtains the latest Pareto front. S7 selects representative Pareto optimal conditions from the latest Pareto frontier and extracts target response conditions from these representative Pareto optimal conditions.
[0006] In one embodiment, the preset rules include a phenomenon vocabulary and a materialized descriptor table. The experimental data is uniformly transformed using these preset rules to obtain structured observation data, including: The initial experimental data includes experimental conditions, precise yield, censored records with only interval information, and qualitative descriptions. The experimental conditions are transformed into descriptor vectors based on the materialized descriptor table; The exact yield is transformed using a logit numerical transformation to obtain the measured logit transformed value. For censored records with only interval information, the censored records with only interval information are encoded as left censored or right censored according to the interval constraint direction given by the censored records with only interval information. The boundary value of the interval information is used as the censoring boundary, and the censoring boundary is transformed by logit to obtain the censoring direction mark and the boundary logit transformation value. The terminology of the queried phenomena maps qualitative descriptions to numerical levels, thus obtaining the numerical levels of qualitative phenomenon labels. The descriptor vector, measured logit transform value, censored direction marker and boundary logit transform value, and qualitative phenomenon label numerical level of the same experiment are output as structured observation data.
[0007] In one embodiment, structured observation data is input into a multi-task Gaussian process model. The training of the multi-task Gaussian process model to obtain a yield predictor parameter package includes: Construct a latent yield function, input the descriptor vector into the latent yield function to obtain the logit transformation value of the true yield, construct a latent phenomenon function, input the descriptor vector into the latent phenomenon function to obtain the potential severity of each qualitative phenomenon; Construct a one-sided biased likelihood function by inputting the logit transformation values of the true yield and the measured logit transformation values into the one-sided biased likelihood function to obtain the probability of observing the measured logit transformation value at the true yield. Construct a censoring likelihood function by inputting the logit transformation values of the true yield, the censoring direction marker, and the boundary logit transformation values into the censoring likelihood function to obtain the probability of the censored record appearing at the true yield. Combine the one-sided biased likelihood function and the censoring likelihood function to obtain the composite likelihood function.
[0008] Construct an ordinal likelihood function and set a segmentation threshold for each level of a qualitative phenomenon. The segmentation threshold is a boundary value used to divide adjacent qualitative phenomenon level intervals. Input the potential severity of the qualitative phenomenon, the numerical level of the qualitative phenomenon label, and the segmentation threshold into the ordinal likelihood function to obtain the probability of observing the level label under the potential severity of the phenomenon. A task-related matrix is constructed, and the latent functions of yield and phenomenon are correlated through the task-related matrix. The total likelihood is constructed by the composite likelihood function and the ordinal likelihood function. Joint inference is performed by maximizing the total likelihood. The task-related matrix and model parameters are jointly learned to obtain the predictor parameter package.
[0009] In one embodiment, the step of selecting a recommended experimental list from the candidate response condition set with the goal of maximizing the acquisition function specifically includes: Each candidate reaction condition in the set of candidate reaction conditions is input into the joint model to obtain the mean yield, uncertainty, and deterministic cost. Monte Carlo sampling was performed on the mean yield and uncertainty, and combined with deterministic costs to form a set of simulation data; Based on the current Pareto front, the expected hypervolume improvement of the simulation dataset to the current Pareto front is calculated and used as the acquisition function value; The candidate reaction conditions with the largest acquisition function values will be output as a list of recommended experiments.
[0010] In one embodiment, the user configuration information further includes a single operation cost, the termination condition includes a termination confirmation round and a termination threshold, and the determination of whether the termination condition is met includes: Input the cost of a single operation and the deterministic cost, calculate the total batch budget and unit cost return rate required to execute the recommended experiment list, and determine that the termination condition is met when the unit cost return rate of the consecutive termination confirmation rounds is lower than the termination threshold.
[0011] In one embodiment, the step of screening representative Pareto optimal conditions from the latest Pareto frontier and extracting target response conditions from the representative Pareto optimal conditions includes: Input the latest Pareto frontier, extract the reaction conditions, the mean predicted yield corresponding to the reaction conditions, the predicted uncertainty interval, and the deterministic material cost from the latest Pareto frontier, and sort them from high to low yield; and select representative Pareto optimal conditions based on at least one of the following: uncertainty threshold, experimental support strength, condition discrimination degree, and qualitative phenomenon label, and extract the target reaction conditions from the representative Pareto optimal conditions.
[0012] Secondly, this application also provides a chemical reaction condition optimization system based on deterministic cost constraints, comprising: The data management and human-computer interaction module is used to structure and store the chemical reaction configuration information input by the user. The user configuration information includes the value range of each reaction condition, preset rules, deterministic cost model and initial reaction condition set; experiments are conducted on the initial reaction condition set to obtain experimental data; The observation coding module is used to uniformly transform experimental data through preset rules to obtain structured observation data. The multi-task Gaussian process model training module is used to input structured observation data into the multi-task Gaussian process model, train the multi-task Gaussian process model, and obtain the yield predictor parameter package. The simulation experiment module is used to construct a joint model based on the yield predictor parameter package and the deterministic cost model, sample within the range of reaction conditions to obtain a set of candidate reaction conditions, and conduct experimental simulations on the set of candidate reaction conditions through the joint model to obtain simulation experimental data. The experiment list recommendation module is used to acquire historical experimental data, construct the current Pareto front based on the historical experimental data, construct the acquisition function based on the current Pareto front, collect simulation experimental data, and select a recommended experiment list from the candidate reaction condition set with the goal of maximizing the acquisition function. The termination decision module is used to conduct experiments on the recommended experiment list to obtain the latest experimental data. The latest experimental data replaces the initial experimental data, and the results are returned to the observation coding module for unified transformation of the latest experimental data observations. This process continues until the termination condition is met, at which point the loop stops, and the latest Pareto front is obtained. The output filtering module is used to filter out representative Pareto optimal conditions from the latest Pareto frontier and extract target reaction conditions from the representative Pareto optimal conditions.
[0013] Thirdly, this application also provides an electronic device, including a processor and a memory, the memory having a stored computer program that, when executed by the processor, implements the chemical reaction condition optimization method based on deterministic cost constraints as described in any of the above embodiments.
[0014] Fourthly, this application also provides a non-transitory computer storage medium storing a computer program thereon, wherein the computer program, when executed by a processor, implements the chemical reaction condition optimization method based on deterministic cost constraints as described in any of the above embodiments.
[0015] The chemical reaction condition optimization method based on deterministic cost constraints proposed in this application has the following advantages over related technologies: 1. The chemical reaction condition optimization method based on deterministic cost constraints proposed in this application first completes the system initialization configuration, providing a standardized data foundation for subsequent full-process operation. This method performs unified structured coding processing on the original experimental data, allowing heterogeneous information such as qualitative phenomena and censored observations to participate in model training in a mathematically rigorous form, significantly improving the utilization efficiency of scarce experimental data. The method trains a multi-task Gaussian process model that can simultaneously handle the correlation between yield prediction and phenomenon labels, quantifying uncertainty while outputting the predicted mean, providing a reliable basis for experimental condition decisions. This method combines a probability model of yield and a deterministic cost model for joint recommendation, eliminating the need to repeatedly model cost objectives that can already be analytically solved, thereby reducing unnecessary computational overhead, improving experimental efficiency while lowering experimental costs.
[0016] 2. By combining a probabilistic yield prediction model with a deterministic cost analysis model, costs are directly calculated analytically, eliminating the need for redundant cost target modeling, thus avoiding model bias introduced by repetitive modeling and reducing computational overhead. The optimization process balances yield improvement with material cost constraints, outputting multiple Pareto frontier alternatives rather than a single fixed optimal condition. R&D personnel can flexibly select suitable reaction conditions based on project budgets and process requirements, better aligning with the decision-making needs of actual process development.
[0017] 3. Introducing unit cost return rate as a quantitative evaluation indicator, comprehensively considering material expenditures and labor operation costs, to measure the expected optimization benefits that continuing the experiment can bring. Relying on the return rate of consecutive rounds to determine whether to terminate the iteration, this breaks away from the model of stopping experiments entirely based on the subjective experience of researchers. This can avoid premature termination and missing out on better process conditions, and can also stop optimization in time when the marginal benefits are insufficient to cover the experimental investment, preventing ineffective experiments from causing unnecessary consumption of R&D resources. Attached Figure Description
[0018] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0019] Figure 1 A flowchart illustrating the chemical reaction condition optimization method based on deterministic cost constraints provided in this application embodiment.
[0020] Figure 2 This is a schematic diagram of a chemical reaction condition optimization system based on deterministic cost constraints provided in an embodiment of this application.
[0021] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation
[0022] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the embodiments of this application. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.
[0023] In some embodiments, this application provides a method for optimizing chemical reaction conditions based on deterministic cost constraints, including S1-S7: S1 stores the chemical reaction configuration information input by the user in a structured manner. The user configuration information includes the value range of each reaction condition, preset rules, deterministic cost model, termination conditions, and initial reaction condition set; experiments are conducted on the initial reaction condition set to obtain initial experimental data.
[0024] System initialization is understandably the starting point of the entire optimization process. Its purpose is to structure and organize the various scattered configuration information provided by the experimenters, establish a unified index relationship, and ensure that all subsequent steps can efficiently and accurately call the required data. Chemical reaction optimization involves configuration information across multiple dimensions, including boundary constraints of optimization variables, coding rules for phenomenon labels, unit price data of reagent costs, and control parameters for termination decisions. These information have different formats and sources. If used directly without structured processing, subsequent modules will frequently perform repeated parsing and format conversion, severely impacting computational efficiency and easily introducing human error. Therefore, this step uses a unified structured storage mechanism to transform heterogeneous configuration information into a standardized system initialization configuration package, laying the data foundation for the entire optimization process.
[0025] The initial experimental design aims to generate a set of initial recipes that are distributed as evenly as possible within the variable space. This initial experimental data provides the surrogate model with its initial training samples. Bayesian optimization is a data-driven iterative method, and the quality of its surrogate model is highly dependent on the distribution of the training data in the variable space. If the initial experimental points are concentrated in a local area, the model will be unable to accurately capture the changing trend of the global yield terrain, and subsequent recommendations are prone to getting trapped in local optima. Therefore, this embodiment uses the Latin hypercube sampling algorithm to generate initial recipes that evenly cover the entire variable space, ensuring that the surrogate model has a coarse-grained understanding of the global terrain from the first iteration.
[0026] S2 uses pre-defined rules to uniformly transform the initial experimental data to obtain structured observation data.
[0027] Understandably, the purpose of observation coding is to transform the raw data recorded by experimenters in various formats into a standardized numerical format that can be processed by machine learning models. The raw records of artificial chemical experiments exhibit significant heterogeneity, with three completely different data formats: precise yield values, censored observations, and qualitative phenomenon descriptions. Precise yield values can be directly input into the model, but their range is limited to 0 to 1, and direct use will cause the model to predict out of bounds. Censored observations only have boundary values known, and the specific numerical values of the true yield are unknown, making them unsuitable as direct training points. Qualitative phenomenon descriptions are in natural language, which the model cannot directly understand. This step uses a differentiated coding strategy to transform these three types of heterogeneous data into numerical forms that the model can compute, ensuring that censored observations and qualitative phenomena that would otherwise be discarded are retained and used in subsequent modeling.
[0028] S3 inputs structured observation data into a multi-task Gaussian process model, trains the multi-task Gaussian process model, and obtains the yield predictor parameter package.
[0029] It is understandable that training a multi-task Gaussian process model is the core of the entire optimization method. Its purpose is to train a probabilistic model based on structured observation data that can predict yield for any untested candidate conditions. Traditional response surface models can only provide point estimates and cannot quantify the reliability of the prediction. The core advantage of Bayesian optimization lies precisely in providing both the prediction mean and the prediction variance through a probabilistic surrogate model. In addition, due to the inherent one-way bias between the observed yield and the actual yield in chemical experiments, and the inclusion of censored observations and rank labels in the data, ordinary regression models cannot correctly handle these special data formats. It is necessary to construct a composite probabilistic model that can accommodate the above-mentioned heterogeneous information and asymmetric error structure.
[0030] S4: Construct a joint model based on the yield predictor parameter package and the deterministic cost model. Sample the candidate reaction conditions within the range of reaction condition values to obtain a set of candidate reaction conditions. Perform experimental simulation on the candidate reaction condition set using the joint model to obtain simulation experimental data.
[0031] This step first combines the yield predictor parameter package obtained in the previous step with the deterministic cost model stored in the initialization phase to build a joint model. Then, sampling is performed within the range of reaction conditions to generate a set of candidate reaction conditions. Finally, simulation-level experiments are conducted on this set of candidate reaction conditions using the joint model to obtain simulated experimental data. This step does not perform actual chemical experiments; instead, it relies on model simulation to complete the performance evaluation of a large number of candidate conditions in a virtual space. This eliminates the need for actual reagents and manual experimental costs, providing simulation samples for subsequent Pareto front calculations and acquisition function calculations.
[0032] S5: Obtain historical experimental data and construct the current Pareto front using the historical experimental data; construct the acquisition function based on the current Pareto front, collect simulation experimental data, and select a recommended experimental list from the candidate response condition set with the goal of maximizing the acquisition function.
[0033] This step first retrieves all historical experimental data and constructs the current Pareto front based on this data. This front represents the set of non-dominated optimal solutions at the current iteration stage under the two objective dimensions of yield and cost. Then, based on the current Pareto front, a data acquisition function is constructed. The data acquisition function is used to process the simulated experimental data, and maximizing the value of the data acquisition function is taken as the optimization objective. A recommended list of experiments that can be performed by the experimenters is selected from the candidate reaction condition set, realizing intelligent site selection in multi-objective scenarios and determining the actual chemical experiments to be carried out in the next round.
[0034] Understandably, the purpose of this step is to find a set of experimental conditions in the candidate variable space that perform well in both maximizing yield and minimizing cost, serving as the execution list for the next round of experiments. The core contradiction in chemical reaction optimization lies in the trade-off between high yield and low cost. Increasing the amount of catalyst often results in higher yield but also higher cost, while expensive solvents may increase yield but reduce economic efficiency. If yield is considered a black-box objective that can only be obtained experimentally, cost can be directly calculated analytically from the unit price and dosage of reagents, belonging to a completely known white-box objective. Existing technologies model cost as a black-box objective as well, which wastes computational resources and causes the Pareto front to deviate from the true value due to model errors. This step combines the probabilistic model of yield with the deterministic analytical model of cost, enabling the recommended results to accurately control material costs while pursuing yield improvement.
[0035] S6: Perform experiments on the recommended experimental list to obtain the latest experimental data, replace the initial experimental data with the latest experimental data, return to the observation unified transformation step, and continue until the termination condition is met, stop the loop, and obtain the latest Pareto front.
[0036] This step defines the complete iterative closed-loop logic of this invention. First, real chemical experiments are conducted against the recommended experimental list to obtain the latest experimental data. This latest data replaces the original initial experimental data. Then, the process jumps back to the observation-unified transformation step, initiating a new round of iteration. Termination judgment uses a termination condition to configure information. When the termination condition is met, the iteration loop ends, and the latest Pareto front is output as the data source for subsequent extraction of target reaction conditions. This closed-loop model achieves data-driven iterative optimization. Each round of newly added experimental information continuously optimizes the model and Pareto front, while the termination condition avoids infinite iteration and prevents the conduct of ineffective experiments without any gain.
[0037] S7: Select representative Pareto optimal conditions from the latest Pareto frontier, and extract the target response conditions from the representative Pareto optimal conditions.
[0038] This step is the final output of the entire optimization method. It takes the latest Pareto front obtained after the iteration terminates as input. The Pareto front contains a large number of non-dominated solutions that trade off between yield and cost. Not all of them will be directly output as the final result. It is necessary to select representative Pareto optimal conditions from them, and then extract the target reaction conditions from the representative Pareto optimal conditions for users. This completes the simplification and screening of a large number of Pareto trade-off solutions and outputs target reaction conditions with practical process reference value.
[0039] This embodiment provides a standardized data foundation for subsequent processes through system initialization configuration; by performing unified structured coding on the raw data, heterogeneous information such as qualitative phenomena and censored observations can be mathematically rigorously incorporated into model training, significantly improving the utilization efficiency of scarce experimental data; by training a multi-task Gaussian process model, it can simultaneously handle the correlation between yield prediction and phenomenon labels, and quantify uncertainty while predicting the mean, providing a reliable basis for decision-making; through joint cost calculation suite and acquisition function optimization, it achieves joint recommendation of yield probability model and cost deterministic model, avoiding the computational overhead and model error caused by repeatedly modeling known analytical objectives; and by calculating the unit cost return rate in iterative loops. RtBased on this, the system determines whether to terminate the optimization, providing experimenters with a quantitative, data-driven termination criterion that effectively avoids premature termination or excessive experimentation. The final Pareto front provides multiple options for different cost budget requirements, rather than a single optimal solution, which is more in line with actual process development scenarios. This application can achieve optimization conditions that balance high yield and low cost with fewer experiments and lower total cost, and terminate the experiment in a timely manner when the marginal benefits no longer cover the experimental costs.
[0040] In some embodiments, the preset rules include a phenomenon vocabulary and a materialized descriptor table. The preset rules are used to uniformly transform experimental data to obtain structured observation data, including: The initial experimental data includes experimental conditions, precise yield, censored records with only interval information, and qualitative descriptions. The experimental conditions are transformed into descriptor vectors based on the materialization descriptor table. The exact yield is transformed using a logit numerical transformation to obtain the measured logit transformed value. For censored records with only interval information, the censored records with only interval information are encoded as left censored or right censored according to the interval constraint direction given by the censored records with only interval information. The boundary value of the interval information is used as the censoring boundary, and the censoring boundary is transformed by logit to obtain the censoring direction mark and the boundary logit transformation value. The terminology of the queried phenomena maps qualitative descriptions to numerical levels, thus obtaining the numerical levels of qualitative phenomenon labels. The descriptor vector, measured logit transform value, censored direction marker and boundary logit transform value, and qualitative phenomenon label numerical level of the same experiment are output as structured observation data.
[0041] In this embodiment, the received experimental conditions include reaction temperature, catalyst equivalent, base equivalent, substrate concentration, and solvent type. For discrete conditions such as solvents, the physical and chemical descriptor table is queried to read the corresponding physical and chemical parameters such as dielectric constant, number of donors, boiling point, and logP of the solvent, and these parameters are then concatenated to form a descriptor vector.
[0042] It is understandable that the raw records from chemical experiments are highly heterogeneous in format, including precise numerical yields, censored records with yields <5% (only boundary values), and qualitative descriptions in natural language such as tarring and precipitation. These raw records cannot be directly input into machine learning models, and traditional processing methods often discard censored records and qualitative phenomena, resulting in a waste of valuable small-sample experimental information. This step utilizes two pre-defined rules—a physical descriptor table and a phenomenon vocabulary—to perform conditional vectorization, yield interval transformation, censoring marking, and qualitative text digitization, respectively. This unifies the raw records of different formats into structured observation data readable by the model, preserving all valid information carried by various observations without fabricating false experimental information, thus providing a high-quality input dataset for subsequent model training.
[0043] In this embodiment, the received experimental conditions include reaction temperature, catalyst equivalent, base equivalent, substrate concentration, and solvent type. For discrete conditions such as solvents, the physical and chemical descriptor table is queried to read the corresponding physical and chemical parameters such as dielectric constant, number of donors, boiling point, and logP of the solvent, and these parameters are then concatenated to form a descriptor vector.
[0044] Logit transformation formula:
[0045] Taking a yield of 72% as an example, the corresponding decimal yield value is 0.72. Substituting this into the formula, we get...
[0046] If the original record shows a yield of less than 5%, it is classified as left censored, with a boundary value of 0.05.
[0047] If the original record shows a conversion rate greater than 95%, it is classified as right censoring, with a boundary value of 0.95.
[0048] The phenomenon vocabulary completes the mapping of textual descriptions to numerical levels, with the specific correspondences as follows: Tarification degree: Clear experimental phenomenon recorded as 0, slight discoloration as 1, significant darkening or small amount of gelatinous substance as 2, and abundant black tar as 3. Palladium black precipitation: None recorded as 0, a small amount of black particles as 1, and obvious black precipitate as 2. System homogeneity: Heterogeneous experimental phenomenon recorded as 0, and homogeneous as 1. Deboronization byproducts: No deboronized aromatic spots observed as 0, and visible deboronized aromatic spots as 1.
[0049] The structured observation data is organized into a wide matrix table, with each row corresponding to all the encoded information of a complete experiment.
[0050] In some embodiments, structured observation data is input into a multi-task Gaussian process model, and the training of the multi-task Gaussian process model to obtain a yield predictor parameter package includes: Construct a latent yield function, input the descriptor vector into the latent yield function, and obtain the logit transformation value of the true yield; construct a latent phenomenon function, input the descriptor vector into the latent phenomenon function, and obtain the potential severity of each qualitative phenomenon. Construct a one-sided biased likelihood function by inputting the logit transformation values of the true yield and the observed logit transformation values into the one-sided biased likelihood function to obtain the probability of observing the observed logit transformation value at the true yield; construct a censoring likelihood function by inputting the logit transformation values of the true yield, the censoring direction marker, and the boundary logit transformation values into the censoring likelihood function to obtain the probability of the censored record appearing at the true yield; combine the one-sided biased likelihood function and the censoring likelihood function to obtain the composite likelihood function; Construct an ordinal likelihood function and set a segmentation threshold for each level of a qualitative phenomenon. The segmentation threshold is a boundary value used to divide adjacent qualitative phenomenon level intervals. Input the potential severity of the qualitative phenomenon, the numerical level of the qualitative phenomenon label, and the segmentation threshold into the ordinal likelihood function to obtain the probability of observing the level label under the potential severity of the phenomenon. A task-related matrix is constructed, and the latent functions of yield and phenomenon are correlated through the task-related matrix. The total likelihood is constructed by the composite likelihood function and the ordinal likelihood function. Joint inference is performed by maximizing the total likelihood. The task-related matrix and model parameters are jointly learned to obtain the predictor parameter package.
[0051] It is understandable that ordinary single-task regression Gaussian processes can only handle precise continuous numerical samples. They cannot accommodate the one-way observation bias of yield caused by separation loss in chemical experiments, nor can they utilize censored interval records and ordered qualitative grade labels. Furthermore, there is an objective correlation between yield and byproducts such as tarring and catalyst precipitation; separate modeling would lose this constraint information. This step constructs a multi-task Gaussian process framework, setting yield latent functions and phenomenon latent functions separately. The statistical correlation between yield and qualitative phenomena is mined through the task correlation matrix. One-sided biased likelihood, censored likelihood, and ordinal likelihood are specifically designed to adapt to the three types of heterogeneous observations. Multi-task joint inference is achieved by maximizing the total likelihood, thereby fully mining the correlation information inherent in limited chemical experimental data and improving the accuracy and reliability of yield prediction under small sample conditions.
[0052] In this embodiment, the training process of the multi-task Gaussian process model will be described in detail: S3-1: Organize the training data matrix Organize all the structured observation data output by S2 into a training dataset by rows.
[0053] This represents the descriptor vector corresponding to the i-th experiment, and the feature dimension is determined by the materialized descriptor table. This represents the task number to which the i-th observation belongs. A zero number indicates that the observation belongs to the yield task, while other positive integer numbers correspond to different qualitative phenomenon tasks. This represents the observation value of the i-th observation. When assigned to a yield task, the observation value is the measured logit transform value, or the logit transform value of the censored boundary combined with the censoring direction marker. When assigned to a qualitative phenomenon task, the observation value is the numerical level of the qualitative phenomenon label.
[0054] A single experiment may simultaneously produce productivity observations and multiple phenomenon observations. In this case, the experiment will generate multiple training samples, all of which use the same set of descriptor vectors but are assigned different task numbers. This embodiment focuses on the palladium-catalyzed Suzuki coupling reaction, setting up four qualitative phenomena: degree of tarring, palladium black precipitation, system homogeneity, and deboronization byproducts.
[0055] S3-2: Constructing a multi-task Gaussian process prior The yield latent function outputs the corresponding value of the true yield in the logit domain; the phenomenon latent function outputs the potential severity of each phenomenon. The covariance structure adopts a linear model with co-regionalization.
[0056]
[0057] In the formula, The covariance operator represents the degree of linear correlation between the values of two latent functions. The function represents the latent function corresponding to the k-th task; x represents a set of reaction condition input vectors. In this application, task k can be a qualitative / quantitative observation task such as yield prediction, tarring grade, or precipitation phenomenon. Indicates the first The latent function corresponding to each task; For another set of reaction condition input vectors, k and... It can be the same task or different tasks, thus describing the relationship within or between tasks; Represents the task relevance matrix The kth row of the middle The matrix elements of the column, i.e., task k and task The cross-task correlation coefficients, a matrix learned by the model, are used to quantify the strength of the correlation between different tasks. When a statistical correlation exists between the yield task and the phenomenon labeling task... By taking non-zero values, the ability to infer yield from phenomenon labels can be achieved; The fundamental kernel function representing the input space (single-task kernel) describes two sets of reaction conditions x and The similarity between them is determined only by the distance between process parameters and is independent of the task type. Common forms include radial basis core (RBF).
[0058] The covariance of two latent functions equals the correlation strength between the tasks multiplied by the similarity between the two sets of response conditions. If the two sets of response conditions are similar and the two tasks are highly correlated, the corresponding latent function values are more correlated; if the tasks are independent, then... The cross-task covariance is 0.
[0059] S3-3: Constructing a triple similarity function One-sided bias likelihood employs an exponentially modified Gaussian distribution to accommodate unidirectional product loss caused by chemical experimental separation; censoring likelihood uses a cumulative probability form to handle left and right censored records; the two are combined to obtain a composite likelihood function. Ordinal likelihood employs a probit cumulative linkage model to learn the segmentation thresholds for each phenomenon level, handling ordered but unequally spaced qualitative level labels.
[0060] S3-4: Jointly solving all model parameters based on variational inference Since there is a non-Gaussian likelihood, there is no analytical solution in the posterior. This embodiment uses sparse variational inference; it can be replaced by the Laplace approximation or expectation propagation.
[0061] The first step is to introduce sparse variational induced points to reduce the overall computational complexity. M induced points are selected in the descriptor vector space. The corresponding induced latent function value As a variational free variable, the number of induced points M is much smaller than the number of training samples N; in this embodiment, M is taken as 128. Induced point locations. It can also be optimized as a learnable parameter.
[0062] The second step is to construct the Gaussian variational posterior distribution. The mean vector and covariance matrix of the variational posterior distribution are used as variational parameters, denoted as the variational mean. Variational covariance .
[0063] The third step is to calculate the lower bound of evidence, which consists of two terms: a KL divergence regularization term between the variational posterior and the latent function prior, and an expected term of the log-likelihood of each training sample. For non-Gaussian likelihoods, the expected term is approximated using Monte Carlo sampling. This involves sampling latent function values from the variational posterior, substituting them into the corresponding likelihood function, calculating the log-likelihood, and then averaging the results. The value of the lower bound of evidence is then used as the model optimization objective.
[0064] The fourth step is to jointly optimize all the parameters to be determined. These parameters include the length scale and amplitude of each dimension of the input kernel function, all independent elements of the task correlation matrix, the mean parameter of the one-way loss of the one-sided bias likelihood and the variance of the measurement noise, the ordinal segmentation thresholds for each qualitative phenomenon, and the location of the induction point. and variational parameters , The optimization process uses the Adam optimizer and the LBFGS optimizer to iterate alternately; when the relative change of the lower bound of evidence is less than the preset convergence threshold in multiple consecutive iterations, the model training is considered to have converged.
[0065] In this embodiment, the initial induction point position is selected by performing KMeans clustering on the training descriptor vector; the task relevance matrix is initialized to an identity matrix; the length scale of each kernel is initialized to the value range of the corresponding feature dimension; and the ordinal segmentation threshold is initialized to an equidistant distribution.
[0066] S3-5: Packing Yield Forecaster Parameter Package After training convergence, the kernel function hyperparameters, task relevance matrix, likelihood correspondence parameters, ordinal splitting threshold, variational posterior parameters, and inducement points are all packaged into a yield predictor parameter package. The prediction interface receives descriptor vectors. As input, the model calculates the latent yield function based on the variational posterior. After processing the posterior mean and posterior variance through the inverse logit transformation, the yield prediction distribution and the corresponding uncertainty can be obtained.
[0067] When only yield prediction is required, the prediction results of the phenomenon latent function are not output. If subsequent steps require using phenomenon information to filter candidate conditions, the phenomenon latent function prediction interface corresponding to the same parameter package can be called to output the probability distribution of each qualitative phenomenon falling at each level.
[0068] Technical benefits: By leveraging task relevance, yield can be constrained solely by phenomenon labels; one-sided bias likelihood corrects the system underestimation caused by separation loss; censoring and ordinal likelihood fully utilize various heterogeneous experimental records to improve model reliability under small sample sizes.
[0069] In some embodiments, the step of selecting a recommended experimental list from the candidate response condition set with the goal of maximizing the acquisition function specifically includes: Each candidate reaction condition in the set of candidate reaction conditions is input into the joint model to obtain the mean yield, uncertainty, and deterministic cost. Monte Carlo sampling was performed on the mean yield and uncertainty, and combined with deterministic costs to form a set of simulation data; Based on the current Pareto front, the expected hypervolume improvement of the simulation dataset to the current Pareto front is calculated and used as the acquisition function value; The candidate reaction conditions with the largest acquisition function values will be output as a list of recommended experiments.
[0070] It is understandable that chemical reaction optimization is a multi-objective problem, requiring the simultaneous pursuit of high yield and low cost. Ordinary single-objective greedy algorithms only pursue the highest predicted yield, easily ignoring cost constraints and prone to getting trapped in local optima. Furthermore, artificial laboratories typically perform multiple sets of experiments in parallel per round, necessitating the batch output of candidate conditions. This step uses a joint model to simultaneously acquire yield probability information and determine costs, employing Monte Carlo sampling to numerically process prediction uncertainty. Using expected hypervolume improvement as the acquisition function, it comprehensively considers the expansion benefits to the Pareto front, balancing the development of known high-quality conditions with the exploration of highly uncertain unknown regions. It batch selects a set of candidates with the optimal overall benefit, outputting a recommended experimental list suitable for the actual working conditions of batch experiments in artificial laboratories.
[0071] In this embodiment, the joint model consists of two coupled parts. One part is a trained multi-task Gaussian process yield predictor that outputs the complete probabilistic posterior distribution of the yield. The other part is a deterministic cost analysis module that reads the preset reagent unit price based on parameters such as catalyst equivalent, base equivalent, substrate concentration, and solvent type in the candidate reaction conditions, and directly calculates the material cost through analytical formulas. The cost output is a deterministic value and there is no prediction noise.
[0072] The candidate reaction condition set is obtained by sampling across the entire reaction condition value space. The sampling method can combine Latin hypercube sampling and random sampling to generate a sufficiently large candidate set for subsequent optimization calculations of the acquisition function. For each candidate reaction condition within the candidate set, it is fed into a joint model for processing. The model outputs the corresponding mean yield prediction, yield prediction uncertainty, and the deterministic material cost corresponding to that set of conditions.
[0073] Monte Carlo sampling is performed on the posterior distribution of the yield for each candidate condition. The number of samplings can be set according to computing resources; in this embodiment, the number of samplings for a single candidate point is two thousand. Each sampling yields a set of yield sample values, which are virtual experimental results at the probabilistic level. The corresponding material cost directly uses the determined cost output by the joint model and does not participate in the sampling variation. After all sampling is completed, all sampling results are aggregated to form a simulated dataset. Each sample point within the simulated dataset consists of a set of virtual yield values and a fixed material cost.
[0074] The current Pareto front, constructed from all historical experimental data, is retrieved. This front consists of a series of non-dominated solutions. Using the current Pareto front as the calculation benchmark, the expected hypervolume improvement corresponding to each candidate condition is statistically analyzed for the simulated dataset. This value is the acquisition function value, calculated using the following formula:
[0075] In the formula, This represents the theoretical expected hypervolume improvement (the theoretical truth value of the acquisition function). It also represents the set of candidate conditions for selection. When conducting the experiment, the mathematical expectation of the hypervolume increment at the Pareto front; This represents the set of candidate reaction conditions to be evaluated, which is a set of candidate process parameters to be evaluated in this round; Represents the mathematical expectation operator, subscript This means that the expectation operation is performed on the latent function f according to the posterior probability distribution output by the Gaussian process model, which is used to take into account the uncertainty in the yield prediction itself; The hypervolume operator takes a set of target solutions as input and outputs the hypervolume enclosed by the solution set relative to the reference point in the target space, which is used to quantitatively evaluate the comprehensiveness of the solution set. The current Pareto front represents the non-dominated process solution set obtained from historical experimental data; This represents the new solution set obtained after incorporating the target points corresponding to the candidate conditions into the existing Pareto front P; This represents the increase in volume resulting from adding new candidate conditions in a single sampling scenario; the expectation is to average this increase over the entire posterior distribution.
[0076] Since the analytical expectation cannot be directly obtained into a closed-form solution, a Monte Carlo numerical approximation is used. If a total of S samples are taken, the sample set is obtained. Then the approximate calculation formula is:
[0077] In the formula, Let s be the sample drawn from the posterior yield for the sth time. The Monte Carlo estimate of the desired hypervolume improvement is given by the acquisition function value, which represents the expected expansion of the Pareto front in the target space if the experiment is conducted under the selected candidate conditions.
[0078] The optimization calculation distinguishes between continuous and discrete reaction variables. For continuous variables such as temperature, catalyst equivalent, and substrate concentration, a multi-starting-point gradient ascent optimization strategy is adopted, setting multiple sets of random initial starting points to carry out gradient iteration, reducing the risk of getting trapped in local extrema. For discrete variables such as solvent type, an enumeration comparison method is used to traverse all available solvent types.
[0079] After optimization, the data acquisition function values are sorted from highest to lowest. A specified number of candidate conditions at the top of the sorting are selected to form the recommended experimental list for this round. Within the list, candidate conditions are automatically considered in terms of spatial distribution to avoid outputting multiple sets of highly similar or redundant experimental conditions. Simultaneously, the expected peak return α corresponding to this batch of recommended experimental conditions is output; this value represents the maximum expected hypervolume improvement that this set of candidate conditions can bring.
[0080] In some embodiments, the user configuration information further includes a single operation cost, the termination condition includes a termination confirmation round and a termination threshold, and the determination of whether the termination condition is met includes: Input the cost of a single operation and the deterministic cost, calculate the total batch budget and unit cost return rate required to execute the recommended experiment list, and determine that the termination condition is met when the unit cost return rate of the consecutive termination confirmation rounds is lower than the termination threshold.
[0081] It is understandable that the timing of iterative termination in traditional process optimization relies entirely on the subjective experience of researchers, which easily leads to two drawbacks: premature termination, missing out on better process conditions; or continuing ineffective experiments with extremely low marginal returns, wasting material, labor, and time costs. This step incorporates material costs and fixed operating costs per experiment into the total batch budget, establishes a unit cost return rate index, and quantifies the expected optimization benefits obtained per unit of investment. Then, through multiple rounds of sliding window judgment, it filters out misjudgments caused by random fluctuations in a single round, forming a data-driven objective termination criterion, achieving a balance between optimization benefits and experimental costs.
[0082] Before proceeding with the termination decision, three preset parameters need to be read from the system initialization configuration package: single operation cost, number of termination confirmation rounds, and termination threshold. The single operation cost is preset by the user based on the actual operating costs of the laboratory. In this embodiment, the single operation cost is set to 200 yuan, including the combined costs of laboratory personnel's operating hours, product separation and processing, and chromatographic instrument analysis and testing. The number of termination confirmation rounds represents the number of iteration rounds that need to be continuously observed. In this embodiment, it is set to two rounds by default. This parameter is used to construct a sliding window judgment logic to avoid erroneous termination caused by a random low return on unit cost due to sampling fluctuations in a particular round. The termination threshold represents the minimum acceptable return on unit cost, which is set by the user in conjunction with the project budget and research and development goals.
[0083] After obtaining the recommended experiment list and the corresponding expected peak return α in each round, first calculate the deterministic material cost of each candidate condition in the list, sum the material costs, and obtain the total material cost of all experiments in this round. The formula for calculating the total material cost is as follows:
[0084] In the formula, This is the set of candidate conditions for this round of recommendations. This represents the material cost corresponding to a single candidate condition x.
[0085] Read the total number of recommended experimental conditions in this round, and then... Multiply by the cost per operation The total fixed operation cost for all experiments in this round is obtained as follows:
[0086] Add the total material cost to the total fixed operating cost to obtain the total batch budget:
[0087] Using the expected peak return of this round Divide by the total batch budget to calculate the unit cost return rate for this iteration. The calculation formula is:
[0088] The return on unit cost characterizes the expected Pareto front hypervolume improvement that can be achieved for every unit of experimental funding invested.
[0089] The system maintains a round count variable, denoted as . This is used to record the cumulative number of rounds in which the return rate continuously falls below the termination threshold. The judgment logic satisfies:
[0090]
[0091] If the unit cost return rate calculated in this round is less than the termination threshold, the cumulative round count is incremented by one; if the unit cost return rate in this round is greater than or equal to the termination threshold, the cumulative round count is reset to zero, and the next round of iteration optimization is executed.
[0092] When the cumulative number of rounds reaches the preset termination confirmation round, it means that the input-output ratio of continuing the experiment for multiple consecutive rounds does not meet the project's preset minimum requirements. At this point, the termination condition is met, the iteration loop stops, and the latest Pareto front is output. If the termination confirmation round has not yet been reached, the recommended experiment list is delivered to the experimenters to carry out the actual experiments, obtain the latest experimental data, and return to the observation-unified transformation step to start a new round of iteration.
[0093] In some embodiments, the step of screening representative Pareto optimal conditions from the latest Pareto frontier and extracting target response conditions from the representative Pareto optimal conditions includes: Input the latest Pareto frontier, extract the reaction conditions, the mean predicted yield corresponding to the reaction conditions, the predicted uncertainty interval, and the deterministic material cost from the latest Pareto frontier, and sort them from high to low yield; and select representative Pareto optimal conditions based on at least one of the following: uncertainty threshold, experimental support strength, condition discrimination degree, and qualitative phenomenon label, and extract the target reaction conditions from the representative Pareto optimal conditions.
[0094] It's understandable that the Pareto front generates a large number of non-dominated solutions, all of which represent trade-offs between yield and cost. However, some solutions suffer from high prediction uncertainty, lack of experimental data support, severe side reactions, and high similarity among solutions. Providing all of these solutions to researchers would create a selection burden and lack practical process reference value. This step first sorts all Pareto front solutions by yield, and then filters them based on prediction confidence, experimental support, condition discrimination, and side reaction phenomena. Solutions with poor engineering practicality are filtered out, retaining a small number of representative Pareto optimal conditions. This allows researchers to quickly select target reaction conditions based on their project budget and process requirements.
[0095] In this embodiment, after the iteration loop terminates, the system obtains the complete latest Pareto front point set, where each point corresponds to a set of reaction conditions. First, information extraction is performed on all points within the Pareto front. For each set of reaction conditions, the complete process parameters corresponding to that condition are read. Simultaneously, the average predicted yield output from the multi-task Gaussian process model and the prediction uncertainty interval corresponding to the 95% confidence level are retrieved. Then, the deterministic cost calculation module is called to recalculate the deterministic material costs corresponding to that set of reaction conditions.
[0096] After information extraction, the conditions within all Pareto frontiers are sorted in descending order based on the mean predicted yield. Let the latest Pareto frontier contain M candidate conditions, and let the mean predicted yield corresponding to the i-th condition be denoted as... The value of i iterates through all condition numbers from 1 to M. After sorting in descending order, a new sequence of condition indices is obtained. The index sequence satisfies:
[0097]
[0098] All Pareto conditions are rearranged based on this index sequence, and the k-th condition after rearrangement is denoted as... ,satisfy After rearrangement, the conditional sequences satisfy the relationship between the mean yields:
[0099]
[0100] This results in an ordered list of Pareto conditions. Entries at the beginning of the list correspond to smaller values of k, representing process schemes with higher predicted yields; entries at the end of the list correspond to larger values of k, representing process schemes with relatively lower predicted yields but lower material costs.
[0101] Next, perform representative criteria filtering. Filtering can use a single filtering dimension or a combination of multiple dimensions.
[0102] First, screening is based on uncertainty thresholds. An uncertainty threshold is preset to limit the acceptable confidence level of the prediction results. The prediction uncertainty range of each set of conditions is compared. If the range exceeds the preset uncertainty threshold, it means that the model prediction for that condition fluctuates greatly and is not reliable enough. This condition is directly screened out, and conditions with sufficiently reliable prediction results are retained.
[0103] Second, screening is based on experimental support strength. Experimental support strength measures the density of historical experimental samples actually completed near the reaction condition. The characteristic distance between each Pareto condition and all historical completed experimental samples is calculated. If there are not sufficiently close measured experimental points around a Pareto optimal condition, it means that the scheme is far from existing experimental data, the model prediction extrapolation is high, and the experimental support is weak. Such conditions can be screened out.
[0104] Third, selection is based on conditional discriminativeness. Pareto fronts often present multiple sets of process parameters that are highly similar, with minimal differences in yield and cost. These conditions can lead to selection redundancy. The characteristic distance between the parameters of each candidate condition is calculated. If the parameter distance between two sets of conditions is less than a preset distance threshold, the two schemes are considered highly similar. Only one set is retained, and the remaining similar and repetitive conditions are eliminated, ensuring that the process schemes under each representative condition have significant discriminativeness.
[0105] Fourth, screening based on qualitative phenomenon labels. The predicted probability of each qualitative phenomenon corresponding to each Pareto condition is retrieved. For phenomena that are technically unacceptable, such as severe tarring or excessive precipitation, if the predicted probability of a high-level phenomenon for that condition exceeds a preset upper limit, that condition is screened out, eliminating solutions with severe side reactions that are unsuitable for practical process implementation.
[0106] After one or more of the above screenings, the remaining set of conditions constitutes the representative Pareto optimal set. From this set, the complete reaction process parameters corresponding to each condition can be directly extracted as target reaction conditions for researchers to choose from.
[0107] Further visualization output processing can be performed to generate a two-dimensional scatter plot visualization menu of cost and yield. Each representative Pareto optimal condition is marked in the chart, and the accompanying table displays the complete reaction parameters, predicted average yield, uncertainty interval, material cost, and prediction level of each qualitative phenomenon for each set of conditions. R&D personnel can combine project budget, yield indicators, and side reaction control requirements to directly select the target process suitable for implementation.
[0108] In some embodiments, such as Figure 2 As shown, this application also provides a chemical reaction condition optimization system based on deterministic cost constraints, comprising: The data management and human-computer interaction module is used to structure and store the chemical reaction configuration information input by the user. The user configuration information includes the value range of each reaction condition, preset rules, deterministic cost model and initial reaction condition set; experiments are conducted on the initial reaction condition set to obtain experimental data; The observation coding module is used to uniformly transform experimental data through preset rules to obtain structured observation data. The multi-task Gaussian process model training module is used to input structured observation data into the multi-task Gaussian process model, train the multi-task Gaussian process model, and obtain the yield predictor parameter package. The simulation experiment module is used to construct a joint model based on the yield predictor parameter package and the deterministic cost model, sample within the range of reaction conditions to obtain a set of candidate reaction conditions, and conduct experimental simulations on the set of candidate reaction conditions through the joint model to obtain simulation experimental data. The experiment list recommendation module is used to acquire historical experimental data, construct the current Pareto front based on the historical experimental data, construct a data acquisition function based on the current Pareto front, collect simulation experimental data, and select a recommended experiment list from the candidate reaction condition set with the goal of maximizing the data acquisition function. The termination decision module is used to conduct experiments on the recommended experiment list to obtain the latest experimental data. The latest experimental data replaces the initial experimental data, and the results are returned to the observation coding module for unified transformation of the latest experimental data observations. This process continues until the termination condition is met, at which point the loop stops, and the latest Pareto front is obtained. The output filtering module is used to filter out representative Pareto optimal conditions from the latest Pareto frontier and extract target reaction conditions from the representative Pareto optimal conditions.
[0109] It should be noted that the chemical reaction condition optimization system based on deterministic cost constraints provided in this application embodiment and the chemical reaction condition optimization method based on deterministic cost constraints provided in this application embodiment are based on the same inventive concept. Therefore, the specific implementation of this embodiment can refer to the implementation of the aforementioned chemical reaction condition optimization method based on deterministic cost constraints, and the repeated parts will not be described again.
[0110] This application also provides a non-transitory computer storage medium storing a computer program thereon, which, when executed by a processor, implements the aforementioned method for optimizing chemical reaction conditions based on deterministic cost constraints. This computer-readable medium may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into that device / apparatus / system. The aforementioned computer-readable medium carries one or more programs, which, when executed, implement the method as described in the embodiments of this application.
[0111] According to embodiments of this application, a computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media can also be any computer-readable medium other than computer-readable storage media, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wired, optical fiber, radio frequency signals, etc., or any suitable combination thereof.
[0112] In some embodiments, such as Figure 3 An electronic device provided in this application embodiment includes a processor and a memory; the memory stores a computer program, wherein the computer program, when executed by the processor, implements the above-described method for optimizing chemical reaction conditions based on deterministic cost constraints.
[0113] Specifically, the processor may include, for example, a general-purpose microprocessor, an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor may also include onboard memory for caching purposes. The processor may be a single processing unit or multiple processing units for performing different actions of the method flow according to embodiments of this application.
[0114] Memory can be any medium capable of containing, storing, transmitting, propagating, or transmitting instructions. For example, memory can include, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, instruments, or propagation media. Specific examples of memory include: magnetic storage devices such as magnetic tape or hard disk drives (HDDs); optical storage devices such as optical discs (CD-ROMs); and also random access memory (RAM) or flash memory; and / or wired / wireless communication links.
[0115] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.
[0116] Those skilled in the art will understand that the features described in the various embodiments of this application can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, the features described in the various embodiments of this application can be combined and / or combined in various ways without departing from the spirit and teachings of this application. All such combinations and / or combinations fall within the scope of this application. Therefore, the scope of this application should not be limited to the above embodiments. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.
Claims
1. A method for optimizing chemical reaction conditions based on deterministic cost constraints, characterized in that, include: S1 stores the chemical reaction configuration information input by the user in a structured manner. The user configuration information includes the value range of each reaction condition, preset rules, deterministic cost model, termination conditions, and initial reaction condition set; experiments are conducted on the initial reaction condition set to obtain initial experimental data; S2 uses pre-defined rules to uniformly transform the initial experimental data to obtain structured observation data; S3 inputs structured observation data into a multi-task Gaussian process model, trains the multi-task Gaussian process model, and obtains the yield predictor parameter package. S4 constructs a joint model based on the yield predictor parameter package and the deterministic cost model. It samples the candidate reaction condition set within the range of reaction condition values and uses the joint model to conduct experimental simulations on the candidate reaction condition set to obtain simulation experimental data. S5 acquires historical experimental data, constructs the current Pareto front based on the historical experimental data, constructs a data acquisition function based on the current Pareto front, collects simulation experimental data, and selects a recommended experimental list from the candidate response condition set with the goal of maximizing the data acquisition function. S6 performs experiments on the recommended experimental list to obtain the latest experimental data, replaces the initial experimental data with the latest experimental data, returns to step S2, and continues until the termination condition is met, stops the loop, and obtains the latest Pareto front. S7 selects representative Pareto optimal conditions from the latest Pareto frontier and extracts target response conditions from these representative Pareto optimal conditions.
2. The method for optimizing chemical reaction conditions based on deterministic cost constraints as described in claim 1, characterized in that, The pre-defined rules include a phenomenon vocabulary and a materialized descriptor table. These pre-defined rules are used to uniformly transform experimental data to obtain structured observation data, including: The initial experimental data includes experimental conditions, precise yield, censored records with only interval information, and qualitative descriptions. The experimental conditions are transformed into descriptor vectors based on the materialized descriptor table; The exact yield is transformed using a logit numerical transformation to obtain the measured logit transformed value. For censored records with only interval information, the censored records with only interval information are encoded as left censored or right censored according to the interval constraint direction given by the censored records with only interval information. The boundary value of the interval information is used as the censoring boundary, and the censoring boundary is transformed by logit to obtain the censoring direction mark and the boundary logit transformation value. The terminology of the queried phenomena maps qualitative descriptions to numerical levels, thus obtaining the numerical levels of qualitative phenomenon labels. The descriptor vector, measured logit transform value, censored direction marker and boundary logit transform value, and qualitative phenomenon label numerical level of the same experiment are output as structured observation data.
3. The method for optimizing chemical reaction conditions based on deterministic cost constraints as described in claim 2, characterized in that, Structured observation data is input into a multi-task Gaussian process model. The process of training the multi-task Gaussian process model to obtain a yield predictor parameter package includes: Construct a latent yield function, input the descriptor vector into the latent yield function to obtain the logit transformation value of the true yield, construct a latent phenomenon function, input the descriptor vector into the latent phenomenon function to obtain the potential severity of each qualitative phenomenon; Construct a one-sided biased likelihood function by inputting the logit transformation values of the true yield and the observed logit transformation values into the one-sided biased likelihood function to obtain the probability of observing the observed logit transformation value at the true yield; construct a censoring likelihood function by inputting the logit transformation values of the true yield, the censoring direction marker, and the boundary logit transformation values into the censoring likelihood function to obtain the probability of the censored record appearing at the true yield; combine the one-sided biased likelihood function and the censoring likelihood function to obtain the composite likelihood function; Construct an ordinal likelihood function and set a segmentation threshold for each level of a qualitative phenomenon. The segmentation threshold is a boundary value used to divide adjacent qualitative phenomenon level intervals. Input the potential severity of the qualitative phenomenon, the numerical level of the qualitative phenomenon label, and the segmentation threshold into the ordinal likelihood function to obtain the probability of observing the level label under the potential severity of the phenomenon. A task-related matrix is constructed, and the latent functions of yield and phenomenon are correlated through the task-related matrix. The total likelihood is constructed by the composite likelihood function and the ordinal likelihood function. Joint inference is performed by maximizing the total likelihood. The task-related matrix and model parameters are jointly learned to obtain the predictor parameter package.
4. The method for optimizing chemical reaction conditions based on deterministic cost constraints as described in claim 1, characterized in that, The process of selecting a recommended experimental list from the candidate response condition set with the goal of maximizing the acquisition function specifically includes: Each candidate reaction condition in the set of candidate reaction conditions is input into the joint model to obtain the mean yield, uncertainty, and deterministic cost. Monte Carlo sampling was performed on the mean yield and uncertainty, and combined with deterministic costs to form a set of simulation data; Based on the current Pareto front, the expected hypervolume improvement of the simulation dataset to the current Pareto front is calculated and used as the acquisition function value; The candidate reaction conditions with the largest acquisition function values will be output as a list of recommended experiments.
5. The method for optimizing chemical reaction conditions based on deterministic cost constraints as described in claim 1, characterized in that, The user configuration information also includes a single operation cost, and the termination conditions include the number of termination confirmation rounds and a termination threshold. The determination of whether the termination conditions are met includes: Input the cost of a single operation and the deterministic cost, calculate the total batch budget and unit cost return rate required to execute the recommended experiment list, and determine that the termination condition is met when the unit cost return rate of the consecutive termination confirmation rounds is lower than the termination threshold.
6. The method for optimizing chemical reaction conditions based on deterministic cost constraints as described in claim 1, characterized in that, The process involves screening representative Pareto optimal conditions from the latest Pareto frontier, and extracting target response conditions from these representative Pareto optimal conditions, including... Input the latest Pareto frontier, extract the reaction conditions, the mean predicted yield corresponding to the reaction conditions, the predicted uncertainty interval, and the deterministic material cost from the latest Pareto frontier, and sort them from high to low yield; and select representative Pareto optimal conditions based on at least one of the following: uncertainty threshold, experimental support strength, condition discrimination degree, and qualitative phenomenon label, and extract the target reaction conditions from the representative Pareto optimal conditions.
7. A chemical reaction condition optimization system based on deterministic cost constraints, characterized in that, include: The data management and human-computer interaction module is used to structure and store the chemical reaction configuration information input by the user. The user configuration information includes the value range of each reaction condition, preset rules, deterministic cost model and initial reaction condition set; experiments are conducted on the initial reaction condition set to obtain experimental data; The observation coding module is used to uniformly transform experimental data through preset rules to obtain structured observation data. The multi-task Gaussian process model training module is used to input structured observation data into the multi-task Gaussian process model, train the multi-task Gaussian process model, and obtain the yield predictor parameter package. The simulation experiment module is used to construct a joint model based on the yield predictor parameter package and the deterministic cost model, sample within the range of reaction conditions to obtain a set of candidate reaction conditions, and conduct experimental simulations on the set of candidate reaction conditions through the joint model to obtain simulation experimental data. The experiment list recommendation module is used to acquire historical experimental data, construct the current Pareto front based on the historical experimental data, construct the acquisition function based on the current Pareto front, collect simulation experimental data, and select a recommended experiment list from the candidate reaction condition set with the goal of maximizing the acquisition function. The termination decision module is used to conduct experiments on the recommended experiment list to obtain the latest experimental data. The latest experimental data replaces the initial experimental data, and the results are returned to the observation coding module for unified transformation of the latest experimental data observations. This process continues until the termination condition is met, at which point the loop stops, and the latest Pareto front is obtained. The output filtering module is used to filter out representative Pareto optimal conditions from the latest Pareto frontier and extract target reaction conditions from the representative Pareto optimal conditions.
8. An electronic device comprising a processor and a memory, the memory storing computer programs, characterized in that: When the computer program is executed by the processor, it implements the chemical reaction condition optimization method based on deterministic cost constraints as described in any one of claims 1-6.
9. A non-transitory computer storage medium, characterized in that, It stores a computer program, wherein the computer program, when executed by a processor, implements the chemical reaction condition optimization method based on deterministic cost constraints as described in any one of claims 1-6.
