An artificial intelligence-based closed-loop automated chemical synthesis method and system

Abstract

This disclosure belongs to the field of reaction monitoring technology. This disclosure provides a closed-loop automated chemical synthesis method and system based on artificial intelligence. The method includes: acquiring multi-source data of each reaction step in the chemical synthesis process in real time; and optimizing and controlling the synthesis parameters of the reaction step based on the multi-source data.

Description

Technical Field

[0001] This disclosure belongs to the field of reaction monitoring technology, and in particular relates to a closed-loop automated chemical synthesis method and system based on artificial intelligence. Background Technology

[0002] Chemical synthesis is a common method for obtaining target compounds in fields such as drug development, materials preparation, and fine chemicals. The process typically involves multiple reaction steps, and the reaction conditions at each step (such as temperature, feed ratio, reaction time, and catalyst type) significantly affect the purity and yield of the final product. Taking peptide drug synthesis as an example, from amino acid activation and coupling to deprotection, dozens or even hundreds of reaction steps are required, and the efficiency of each step cumulatively affects the quality of the final product.

[0003] Traditional chemical synthesis optimization relies heavily on researchers' experience, employing a "trial and error" approach to adjust parameters one by one. This method has the following inherent drawbacks: First, the optimization process lacks real-time data support. Traditional methods typically only analyze the purity of the final product after the entire synthetic route is completed, using offline detection techniques (such as high-performance liquid chromatography), and cannot detect the reaction status of intermediate steps. Even if a problem occurs in a certain step, it can only be discovered hours or even days later, leading to a waste of reagents and time. Taking peptide synthesis as an example, if a coupling reaction is inefficient, the problem is often only detected by mass spectrometry after the entire peptide chain synthesis is completed, at which point it is impossible to trace which specific step went wrong.

[0004] Second, parameter optimization relies on human experience, which is inefficient and difficult to reproduce. Faced with multiple adjustable parameters such as temperature, feed ratio, reaction time, and catalyst type, manual trial and error is not only extremely time-consuming, but the optimization results also vary from person to person, making it difficult to form generalizable knowledge. Even with accumulated historical experimental data, there is a lack of effective means to uncover the correlation between key parameters and product quality. Taking peptide synthesis as an example, optimizing the synthesis conditions of a pentapeptide may require dozens of experiments, and for longer peptide chains, the number of experiments increases exponentially.

[0005] Third, existing automated synthesis systems only automate the operational process, not the decision-making process. Although automated synthesizers exist, they can only mechanically execute preset programs and lack the ability to autonomously adjust parameters based on real-time reaction conditions. Once process conditions change (such as batch differences in raw materials or fluctuations in ambient temperature), the system cannot adaptively adjust and still requires manual intervention for empirical judgment and adjustment.

[0006] Therefore, there is an urgent need in this field for a technical solution that can acquire multi-dimensional reaction data in real time, autonomously analyze key influencing factors, and intelligently optimize synthesis parameters, so as to break through the efficiency bottleneck of traditional trial and error methods and realize the paradigm shift of chemical synthesis from "experience-driven" to "data-driven". Summary of the Invention

[0007] This disclosure provides a closed-loop automated chemical synthesis method and system based on artificial intelligence, which can effectively solve the above-mentioned problems.

[0008] This disclosure is implemented as follows: In a first aspect, this disclosure provides a closed-loop automated chemical synthesis method based on artificial intelligence, the method comprising: Acquire multi-source data of each reaction step in the chemical synthesis process in real time. The multi-source data includes process data reflecting the reaction progress and result data reflecting the quality of the reaction products. Based on the multi-source data, the synthesis parameters of this reaction step are optimized and controlled, including: Based on the interpretable machine learning algorithm, the historical synthesis parameters are analyzed, the contribution weight of each synthesis parameter to the corresponding result data is calculated, and key parameters are identified based on the contribution weight to construct a dimensionality-reduced parameter search space. The synthesis parameters include static process parameters and the corresponding process data. Based on the Bayesian optimization algorithm, an iterative search is performed within the parameter search space to generate recommended combinations of key parameters; Based on reinforcement learning algorithms, the corresponding static process parameters are dynamically adjusted according to the process data during each reaction step.

[0009] Secondly, this disclosure provides a closed-loop automated chemical synthesis system based on artificial intelligence, the system comprising: The multi-source data acquisition module is used to acquire multi-source data of each reaction step in the chemical synthesis process in real time. The multi-source data includes process data reflecting the reaction progress and result data reflecting the quality of the reaction products. The intelligent decision-making module is used to optimize and control the synthesis parameters of the reaction step based on the multi-source data, including: The parameter parsing module analyzes historical synthetic parameters based on an interpretable machine learning algorithm, calculates the contribution weight of each synthetic parameter to the corresponding result data, and identifies key parameters based on the contribution weight to construct a dimensionality-reduced parameter search space. The synthetic parameters include static process parameters and the corresponding process data. The parameter optimization module, based on the Bayesian optimization algorithm, performs an iterative search within the parameter search space to generate recommended combinations of key parameters; The parameter control module, based on a reinforcement learning algorithm, dynamically controls the corresponding static process parameters according to the process data during each reaction step.

[0010] Thirdly, this disclosure provides an electronic device, including: Memory, the memory storing execution instructions; and A processor that executes execution instructions stored in the memory, causing the processor to perform the method described in the first aspect.

[0011] Fourthly, this disclosure provides a readable storage medium storing executable instructions, which, when executed by a processor, are used to implement the method described in the first aspect.

[0012] Compared with the prior art, the beneficial effects of this disclosure are: This disclosure provides a closed-loop automated chemical synthesis method and system based on artificial intelligence. By acquiring multi-source data of each reaction step in the chemical synthesis process in real time, it achieves dual real-time perception of reaction progress and product quality, transforming the synthesis process from a "black box" to "transparent" and providing a rich data foundation for subsequent intelligent decision-making. By analyzing historical synthesis data based on interpretable machine learning algorithms, it calculates the contribution weight of each synthesis parameter to the result data and identifies key parameters. It can autonomously mine key factors affecting product quality and their contribution from historical data, breaking through the limitations of traditional methods that rely on human experience. By performing iterative search in the dimensionality-reduced key parameter space based on Bayesian optimization algorithms, it generates recommended optimal key parameter combinations, achieving efficient and interpretable parameter optimization. Only 5-6 iterations are needed to achieve intermediate purity exceeding 90%, with optimization efficiency significantly higher than traditional trial-and-error methods. By dynamically adjusting synthesis parameters based on real-time process data during each reaction step using reinforcement learning algorithms, it achieves millisecond-level real-time adaptive adjustment, effectively compensating for minor perturbations within batches and ensuring that the reaction process always proceeds on the optimal trajectory. Through the synergy of the above steps, a complete data closed loop of "perception-analysis-optimization-regulation" is constructed. Attached Figure Description

[0013] Figure 1 This is a flowchart of S100, a closed-loop automated chemical synthesis method based on artificial intelligence provided in an embodiment of this disclosure.

[0014] Figure 2 This is a conductivity curve of a typical experiment of an amino acid activation reaction provided in the embodiments of this disclosure.

[0015] Figure 3 This is a comprehensive color index curve of a typical experiment of an amino acid coupling reaction provided in the embodiments of this disclosure.

[0016] Figure 4 This is a convergence curve of the Bayesian optimization algorithm provided in this embodiment of the present disclosure in different parameter spaces of peptide synthesis.

[0017] Figure 5 This is a schematic diagram of the structure of an artificial intelligence-based closed-loop automated chemical synthesis system 1000 provided in an embodiment of this disclosure.

[0018] Figure 6 This is a schematic diagram of the structure of an artificial intelligence-based closed-loop automated peptide synthesis system provided in an embodiment of this disclosure.

[0019] Figure 7 This is a schematic diagram of the synthesis mechanism of elastin tripeptide.

[0020] Figure 8 This is a mass spectrum of the elastin tripeptide synthesized by the polypeptide synthesis system provided in the embodiments of this disclosure.

[0021] Figure 9 This is a liquid chromatogram of the crude elastin tripeptide synthesized by the polypeptide synthesis system provided in this embodiment of the present disclosure and the post-processed product.

[0022] Figure 10 This is a purity improvement curve of the SHAP-BO collaborative optimization paradigm provided in this embodiment of the present disclosure in iterative experiments. Detailed Implementation

[0023] The present disclosure will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the disclosure. Furthermore, it should be noted that, for ease of description, only the parts relevant to the present disclosure are shown in the accompanying drawings.

[0024] It should be noted that, where there is no conflict, the embodiments and features described in this disclosure can be combined with each other. The technical solutions of this disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0025] Unless otherwise stated, the exemplary implementations / embodiments shown are to be understood as providing exemplary features of various details that provide ways in which the technical concepts of this disclosure can be implemented in practice. Therefore, unless otherwise stated, the features of various implementations / embodiments may be additionally combined, separated, interchanged and / or rearranged without departing from the technical concepts of this disclosure.

[0026] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of this disclosure. The singular forms “a,” “the,” and “the” used in the embodiments of this disclosure and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0027] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0028] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0029] The terms "first" and "second" used herein are merely to distinguish similar objects and do not represent a specific ordering of the objects. Understandably, the specific order or sequence of "first" and "second" can be interchanged where permitted. It should be understood that the objects distinguished by "first" and "second" can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein.

[0030] Example 1 Please refer to Figure 1 This disclosure provides a closed-loop automated chemical synthesis method S100 based on artificial intelligence.

[0031] Specifically, method S100 includes: S102, acquire multi-source data of each reaction step in the chemical synthesis process in real time, the multi-source data including process data reflecting the reaction progress and result data reflecting the quality of the reaction products; S104, Based on the multi-source data, the synthesis parameters of this reaction step are optimized and controlled, including: S1042, Based on the interpretable machine learning algorithm, the historical synthesis parameters are analyzed, the contribution weight of each synthesis parameter to the corresponding result data is calculated, and key parameters are identified according to the contribution weight to construct a dimensionality-reduced parameter search space, wherein the synthesis parameters include static process parameters and the corresponding process data; S1044, Based on the Bayesian optimization algorithm, perform iterative search within the parameter search space to generate recommended key parameter combinations; S1046, Based on reinforcement learning algorithm, the corresponding static process parameters are dynamically adjusted according to the process data during each reaction step.

[0032] The method S100 forms a data closed loop of "perception-analysis-optimization-control" through the real-time acquisition of multi-source data and the coordination of intelligent decision-making, thereby realizing the autonomous optimization of chemical synthesis parameters.

[0033] In some embodiments, the chemical synthesis process includes peptide synthesis. Those skilled in the art will understand that peptide synthesis, as one of the most complex multi-step reactions in chemical synthesis, demonstrates the versatility of the disclosed technical solution in various types of chemical synthesis. For other types of chemical synthesis (such as nucleic acid synthesis, small molecule synthesis, polymer synthesis, etc.), the same closed-loop optimization effect can be achieved simply by adjusting the monitoring parameters and optimization targets according to the specific reaction characteristics.

[0034] Specifically, the polypeptide synthesis process involves the following three key reaction steps: amino acid activation, amino acid coupling, and deprotection. Those skilled in the art will understand that these three reaction steps have the greatest impact on product quality. Other steps, such as washing, are performed sequentially during the synthesis reaction and do not affect the implementation of the embodiments of this disclosure.

[0035] Step S102 is the foundation for building a data-driven closed-loop optimization, which aims to capture the dynamic information and static results of each reaction step in the chemical synthesis process in real time through multimodal online characterization, so as to provide comprehensive and accurate data support for subsequent intelligent decision-making.

[0036] The multi-source data is divided into two categories based on the dimensions it reflects: Process data refers to physical or chemical signals acquired in real time during a reaction, characterizing the dynamic changes in the reaction process, and continuously recorded in the form of a time series. This type of data has high temporal resolution and can reflect stages such as the start, progress, equilibrium, or end of the reaction. For example, real-time monitoring of the ion concentration change curve of the reaction system using conductivity electrodes is particularly suitable for amino acid activation steps, because the activation process is accompanied by the generation or consumption of ionic groups, and the conductivity time series curve can accurately indicate the start and completion of the activation reaction. Real-time monitoring of the color change trajectory of the reaction system using color sensors is particularly suitable for amino acid coupling steps, because coupling reactions are often accompanied by the generation or decolorization of chromogenic groups, and the color time series change can reflect the progress and rate of the coupling reaction.

[0037] Results data refer to data acquired after the reaction step or at a specific time that can quantitatively characterize the quality of the reaction product. This type of data typically has high accuracy and is considered the gold standard for evaluating synthetic efficacy. For example, online high-performance liquid chromatography (HPLC) can be used to monitor the purity and composition of the reaction product in real time. HPLC can separate and quantify target products and byproducts, providing key quality indicators such as intermediate purity, the number and distribution of impurity peaks.

[0038] Multi-source data collection relies on the automated execution of the multimodal online characterization module, eliminating the need for manual sampling or offline detection. Specifically, during each reaction step of chemical synthesis (such as amino acid activation, coupling, and deprotection in peptide synthesis), the multimodal online characterization module continuously or at preset frequencies acquires multidimensional data of the reaction system through embedded sensors (such as conductivity electrodes and optical sensors) and online analytical instruments (such as online HPLC). The real-time acquired data, after signal processing and data alignment, forms a timestamped multidimensional dataset, ensuring that process data and result data are temporally correlated and traceable.

[0039] Compared with traditional methods, this step has the following significant advantages: Eliminating hysteresis: Traditional methods typically perform offline monitoring after the reaction is complete, failing to detect anomalies during the reaction process. This step, with its real-time data acquisition, enables the system to detect deviations from expectations immediately, creating conditions for real-time control.

[0040] Enriching data dimensions: Traditional monitoring is often limited to a single indicator (such as measuring only pH), while this step integrates multi-dimensional data such as electrical, optical, and chromatographic data to characterize the reaction state from different perspectives, making the synthesis process "transparent" instead of "black box".

[0041] Ensuring data integrity: Automated real-time data collection avoids errors and omissions that may be introduced by manual recording, ensuring the integrity and consistency of historical data and providing high-quality training samples for subsequent interpretable machine learning analysis.

[0042] The multi-source data collected in this step serves as input for all subsequent intelligent decisions: process data is used for real-time control of reinforcement learning algorithms and for extracting dynamic features (such as conductivity plateau time and color change rate), which can assist in interpretable machine learning analysis. The result data will serve as the target value for Bayesian optimization and as some or all of the historical synthetic parameters, constituting the training dataset for interpretable machine learning algorithms, used to identify key parameters and their contributions.

[0043] In some embodiments, multi-source data of each reaction step in the chemical synthesis process are acquired in real time. This multi-source data includes process data reflecting the reaction progress and result data reflecting the quality of the reaction products, including: The reaction progress is determined by real-time monitoring of ion concentration changes in the amino acid activation reaction solution using a conductivity monitoring module. The color monitoring module is used to monitor the color change of the amino acid coupling reaction solution in real time to determine the reaction progress. The purity of the products from each reaction step was monitored using an online high-performance liquid chromatography module.

[0044] Specifically, the conductivity monitoring module includes a high-precision conductivity electrode, an embedded signal processor, and an endpoint determination model. The electrode is directly immersed in the reaction solution to sense changes in ion concentration in real time; the signal processor filters, amplifies, and performs analog-to-digital conversion on the raw conductivity signal; the endpoint determination model determines the endpoint or extracts features based on the processed data stream. Using conductivity to monitor changes in ion concentration to track the progress of chemical reactions is a well-known technique in the field (for example, it has wide applications in titration analysis and polymerization reaction monitoring). This disclosure innovatively applies this technique to the online, real-time endpoint determination of the specific step of amino acid activation in peptide synthesis, forming a closed loop with subsequent intelligent decision-making modules (such as SHAP analysis and reinforcement learning), constituting part of the overall solution. This is achieved through the following precise signal conversion link: the electrode generates a continuous millivolt-level analog electrical signal. This analog signal is converted into a digital signal by the signal processor at a sampling frequency of 10 Hz and transmitted to the endpoint determination model.

[0045] The conductivity monitoring module collects real-time process data as a conductivity-time curve, which fully records the dynamic evolution of ion concentration during the reaction. Several key features can be extracted from this curve, such as: conductivity peak: marking the complete addition of the activator, corresponding to the reaction start time; conductivity decrease rate after peak: reflecting the rate of ion pair consumption due to the formation of side reaction products, used for dynamic assessment of side reaction progress; stabilization time of the decrease rate: the moment when the conductivity decrease trend becomes gradual, used to determine the endpoint where the side reaction level is lowest and the next reaction step can proceed. These features are not only used for real-time endpoint determination but also serve as input to reinforcement learning algorithms to construct the state space, helping the model perceive the current reaction state and make real-time control decisions. Taking the activation reaction of Fmoc-Pro-OH (amino acid raw material) and HOSu / DIC (activator) as an example: Peak conductivity (with activator fully added): When HOSu / DIC (activator) is added to Fmoc-Pro-OH, the reaction occurs rapidly, generating a large number of O-acylisourea (positively charged) plasma species. At this point, the concentration of charged species in the system reaches its maximum. This moment signifies that all raw materials have been attacked by DIC and transformed into highly active charged intermediates. However, this does not mean that the final product Fmoc-Pro-OSu has been formed.

[0046] Rate of decline after peak (consumption by side reactions): O-Acylisourea is highly unstable. It has two main pathways: the main reaction pathway, where it is nucleophilically attacked by HOSu to form the final neutral product Fmoc-Pro-OSu (uncharged and contributing no conductivity); and the side reaction pathway, where it undergoes rearrangement to form the stable, uncharged byproduct N-acylurea. As these charged ion-pair intermediates gradually transform into the neutral final product (OSu ester) or byproduct (N-acylurea), the number of charge carriers in the solution decreases, thus the conductivity begins to decline. The rate of decline reflects how quickly the charged intermediates are consumed.

[0047] The rate of decline stabilizes (reaction endpoint): When the decreasing trend of conductivity begins to level off, it indicates that most of the charged intermediates (O-acylisourea) in the solution have been completely converted. The system mainly consists of neutral Fmoc-Pro-OSu, neutral byproduct DCU (diisopropylurea), and possibly excess HOSu. At this point, the degree of side reactions is minimal, making it the optimal time for the intermediates of the amino acid activation process to proceed to the next reaction step for amino acid coupling.

[0048] Figure 2 A real-time conductivity monitoring curve of a typical amino acid activation reaction process provided in this embodiment of the present disclosure is shown. The conductivity curve marked "Pass" by the system in the figure exhibits the following characteristic stages: Rapid rise phase (t0) t1): With the addition of the activator, the ion concentration in the system increases rapidly, and the conductivity rises sharply to its peak value.

[0049] Rapid descent phase (t1) t2): The activation reaction mainly occurs during the period when the charged intermediate (O) is in its prime. When acylisourea is converted into a product, the ion concentration decreases and the conductivity drops rapidly.

[0050] Gradual descent / plateau segment (t2) t3): As the reaction nears completion, the rate of decrease in conductivity slows significantly. Within a continuous time period Δt (1 minute), the rate of change is ≤0.7 µS·cm. - ¹·min - When ¹, the system determines that "the downward trend is leveling off", meaning that the reaction has reached its endpoint and the side reactions are minimized.

[0051] This curve visually demonstrates how to determine the reaction endpoint online and quantitatively by observing the trend changes in real-time conductivity data.

[0052] In some implementations, these features are compiled and stored in a historical database, which, together with static process parameters and result data, constitutes an analysis dataset used to quantify the contribution of each parameter to product purity.

[0053] The color monitoring module includes a high-frame-rate optical acquisition module, an embedded signal processor, and an endpoint determination model. The high-frame-rate optical acquisition module captures images of the reaction solution at a fixed frequency; the signal processor preprocesses the raw reaction solution images and extracts RGB predicted values ​​based on a convolutional neural network; simultaneously, it calculates a composite color index using traditional image processing methods (such as weighted fusion based on RGB mean). After fusion, a comprehensive color index is obtained, serving as a quantitative indicator of the reaction progress; the endpoint determination model then determines the endpoint or extracts features based on the processed data stream. The process data acquired in real time by the color monitoring module is a curve showing the comprehensive color index changing over time. This curve fully records the dynamic evolution of amino acid concentration during the reaction. Several key features can be extracted from this curve, such as: the magnitude of the initial index change, reflecting the initial kinetic characteristics of the reaction; and the plateau phase of the index curve: the moment when the rate of change of the comprehensive color index significantly decreases and tends to flatten, used to determine the endpoint when the coupling reaction is nearing completion and can proceed to the next coupling reaction step. Please refer to [reference needed]. Figure 3 This disclosure provides a typical example to demonstrate the indicative role of the plateau phase of the exponential curve in indicating the reaction endpoint. In the figure, after approximately 25 minutes, the exponential rate of change converges below the threshold and remains so for a period. HPLC sampling verification shows that amino acid residue is <1% and product purity is >95%, confirming a high degree of consistency between the indicated endpoint and the actual endpoint. Similar to conductivity data, these characteristics are used both for real-time monitoring of the coupling reaction process and to provide key variables reflecting the reaction state for intelligent decision-making models.

[0054] The online high-performance liquid chromatography (HPLC) module includes an injection pump, a mobile phase pump, a multi-channel quantitative valve, a chromatographic column, a UV detector, and a waste collection device. The system automatically takes a trace sample from the reaction system, dilutes or derivatizes it appropriately, and injects it into the column. The UV detector monitors the eluent at a specific wavelength (e.g., 210–220 nm, targeting peptide bonds), thus obtaining a chromatogram in real time. The core output of the online HPLC module is the chromatogram of the products from each reaction step. From the chromatogram, we can calculate: the peak area of ​​the target product and its proportion of the total peak area, i.e., the purity of the intermediate; the number, retention time, and relative peak area of ​​impurity peaks, reflecting the by-product situation; and the resolution between the target peak and adjacent impurity peaks, indicating the separation effect. In subsequent steps, the purity value will be used as the target variable for Bayesian optimization, while the impurity spectral characteristics will serve as additional input for interpretable machine learning analysis, helping to understand the causes of side reactions.

[0055] Step S104 is the core step in building a data closed loop, aiming to achieve autonomous optimization and dynamic control of synthesis parameters through an intelligent decision-making module based on real-time collected multi-source data. In traditional chemical synthesis, parameter optimization and control are two separate steps: optimization relies on offline experimental design during process development, while control relies on the real-time experience and judgment of operators. This step unifies the two under a single intelligent decision-making framework. Through machine learning analysis of historical synthesis data, implicit process experience is transformed into a quantifiable parameter contribution model; based on the current reaction process data, execution parameters are dynamically adjusted to ensure the reaction always runs on the optimal trajectory; and complete data from each reaction is fed back to the system, allowing the optimization model to continuously improve with data accumulation.

[0056] This step employs a stepwise optimization strategy, constructing independent optimization and control loops for each key reaction step in chemical synthesis (such as amino acid activation, coupling, and deprotection). The core of this design lies in the fact that the influencing factors of different reaction steps are fundamentally different, and mixing data from each step for analysis would lead to causal confusion. Furthermore, the quality of the final product in a multi-step synthesis is the product of the efficiency of each step; only by ensuring the purity of the intermediates at each step can the quality of the final product be fundamentally guaranteed. Stepwise optimization isolates data at the step level, allowing each optimization model to learn only the causal relationships of its own step, avoiding cross-step interference. By decomposing the global optimization problem into several low-dimensional sub-problems, optimization efficiency is significantly improved. Experiments show that this strategy can achieve intermediate purity exceeding 90% within 5-6 iterations.

[0057] In some implementations, the synthesis parameters of the reaction step are optimized and controlled based on the multi-source data, including: Based on the process data, kinetic characteristic parameters reflecting the real-time state of the reaction are extracted.

[0058] Specifically, the raw process data stream, such as the conductivity time-series signal, is acquired in real time through a multimodal online characterization module. This data is continuously sampled at a high frequency of 10 Hz-100 Hz to record the dynamic changes in ion concentration in the reaction system. The raw signal is susceptible to instrument noise, ambient light fluctuations, and stirring disturbances. Therefore, preprocessing is required before feature extraction: for example, median filtering (window width of 5-11 sampling points) is used to eliminate impulse noise; moving average smoothing is applied to the conductivity and temperature signals; white balance correction is performed on the color signal based on the gray-world assumption to eliminate light source color temperature drift; baseline correction is used to eliminate sensor zero-point drift and ensure consistent signal starting references.

[0059] In some implementations, recurrent neural networks are used to learn dynamic features from raw time-series data.

[0060] Specifically, for each reaction step, a time-series feature extraction model based on a Long Short-Term Memory (LSTM) network is constructed. The model's input is a multivariate time series, and the feature vector at each time step contains multidimensional process data collected by the multimodal online representation module at the current moment, such as conductivity, reaction temperature, and stirring rate. The model structure employs a two-layer LSTM with a hidden layer dimension of 64 to capture long-range dependencies in the time-series data. The output of the LSTM layer is mapped to a 32-dimensional feature vector via a fully connected network, serving as the kinetic feature parameters for that reaction step. During training, the model uses the intermediate purity after the reaction step as a supervisory signal, performing supervised learning by minimizing the mean squared error (MSE), enabling the LSTM to learn to map the original time-series signal to a low-dimensional feature space related to product quality. The Adam optimizer is used during training, with an initial learning rate of 10. -3 The batch size is 32, the training epochs are 100, and an early stopping mechanism (patience value 10) is set to prevent overfitting. After training, the LSTM model can be used for real-time inference: the real-time process data of the current step is input into the model, and a 32-dimensional feature vector is output. This vector is concatenated with the static process parameters to form a complete augmented feature vector. This vector fully describes the process settings and real-time dynamic response of the current reaction step, and can be used as input features for subsequent interpretable machine learning (such as SHAP analysis), or directly as the state space of the reinforcement learning model.

[0061] By transforming raw process data into kinetic feature parameters with explicit physical meaning, a synergistic effect of multiple technologies is achieved. Taking the input of raw process data into an LSTM model to learn a 32-dimensional implicit feature vector as an example, firstly, the high-frequency continuous waveform is compressed into a low-dimensional dense vector, effectively avoiding the curse of dimensionality while providing an information-dense input representation for subsequent machine learning models; secondly, although the vector itself does not have explicit physical meaning, its 32 dimensions encode the kinetic patterns related to the reaction result in the original time-series signal. Through algorithms such as SHAP, the marginal contribution of each input sensor signal (such as conductivity, comprehensive color index, etc.) to the purity of the final product can be further analyzed, thereby indirectly revealing the nonlinear correlation between the reaction dynamic process and the synthesis result; more importantly, these real-time updated 32-dimensional feature vectors directly reflect the deviation of the reaction path, providing accurate state input for the subsequent reinforcement learning module, enabling it to compensate for batch disturbances in real time by fine-tuning the feed rate or temperature setting, ensuring that the reaction trajectory always approaches the optimal curve.

[0062] In step S1042, the historical synthesis dataset is a complete experimental record accumulated during the historical synthesis process. Each record should include: synthesis parameters, which serve as the constituent elements of the high-dimensional feature space, including: static process parameters, such as preset reaction conditions such as reaction temperature, feed ratio, reaction time, and catalyst type; kinetic feature parameters, such as implicit feature vectors extracted from process data; and result data, which serve as the gold standard for measuring the synthesis effect, such as the purity of intermediates or the purity of final products measured after each reaction step.

[0063] All parameters together constitute a high-dimensional feature space (e.g., 10-30 dimensions), providing a complete set of input features for subsequent interpretable machine learning analysis. The system automatically stores the complete record of each synthesis into a historical database. As the number of runs increases, the dataset continues to expand, and the model accuracy continuously improves.

[0064] In the initial stage of system operation, when there is insufficient proprietary data, literature data or public databases can be introduced as initial training sets to enable the model to quickly acquire basic analytical capabilities. Historical synthesis data can be obtained from publicly available academic literature, patent documents, or technical reports. For example, a research team's paper on optimizing a specific peptide synthesis process may record detailed parameters such as temperature, feed ratio, and reaction time, along with the corresponding product purity, from dozens of experiments. After processing, this data can be directly used as input for interpretable machine learning analysis. Although datasets can be obtained from external sources such as literature, the data collected in real-time through the multimodal online characterization module has unique value that literature data cannot replace, which is one of the core innovations of this disclosure. Online characterization data can provide millisecond-level process feedback, providing continuous input information for real-time control, thereby revealing the deep correlation between reaction progress and product quality—information dimensions that are difficult to systematically provide through literature data. More importantly, by collecting and accumulating historical synthesis data in real-time through the multimodal online characterization module, a self-driven data ecosystem of "automatic data accumulation—continuous model evolution—optimization feedback closed loop" is constructed, enabling the system to have the self-evolving ability of continuous learning and self-optimization.

[0065] This step analyzes historical synthesis parameters using interpretable machine learning algorithms, quantifying the marginal contribution of each parameter to product purity and revealing the intrinsic relationship between process parameters and synthesis results. This allows for the accurate selection of a subset of key parameters that dominate the reaction process. This process not only effectively reduces the dimensionality of subsequent optimizations and avoids the curse of dimensionality by eliminating redundant variables, but also makes the parameter selection criteria transparent and interpretable, enhancing the credibility and interpretability of optimization decisions.

[0066] In some implementations, the interpretable machine learning algorithm employs the SHAP algorithm.

[0067] Based on interpretable machine learning algorithms, historical synthesis parameters are analyzed to calculate the contribution weight of each synthesis parameter to the corresponding result data. Key parameters are then identified based on these contribution weights to construct a dimensionality-reduced parameter search space, including: The Shapley value of each historical synthesis parameter is calculated using the SHAP algorithm. Synthesis parameters whose average absolute value of Shapley value is lower than a preset percentage are removed from the search space of the current dimension.

[0068] The core idea of ​​SHAP is to treat each feature as a participant in a "cooperative game" and quantify the contribution of each feature to the final prediction result by calculating the average of the marginal contribution of each feature to all possible combinations of features.

[0069] First, a high-precision prediction model (such as an XGBoost regressor) is trained using a historical dataset. This model takes the synthesis parameters as input and the corresponding product purity as output. After training, SHAP analysis is applied to the model to obtain the SHAP value of each parameter for each sample. This model is used to accurately fit the nonlinear mapping relationship between synthesis parameters and product purity in historical data, providing a reliable interpretive basis for subsequent SHAP analysis.

[0070] For a trained machine learning model, SHAP assigns a contribution value to each feature of each sample in the following manner: Construct an additive explanation model: ,in Represents a simplified feature vector. Representation of features SHAP value, This represents the predicted mean of all samples.

[0071] By solving for Shapley values ​​that satisfy local accuracy, missing values, and consistency, we ensure... It can fairly reflect the characteristics. The contribution to the deviation of the current sample's predicted value from the mean.

[0072] In some implementations, the TreeExplainer from the SHAP library (version 0.45+) is used to interpret and analyze the trained XGBoost model.

[0073] Two key types of information can be obtained through SHAP analysis: Global feature importance: The average of the absolute values ​​of SHAP across all samples yields the average contribution of each parameter across the entire historical dataset. This value reflects the overall influence of the parameter on product purity.

[0074] Importance of local features: For a specific sample, the sign and magnitude of the SHAP values ​​of each parameter can be observed to understand whether each parameter positively promotes or negatively inhibits purity in this experiment.

[0075] Based on the importance of global features, the system sorts all parameters from highest to lowest contribution. The contribution weights are usually normalized to the [0, 1] interval to facilitate subsequent threshold setting.

[0076] To identify the key parameters that truly dominate the reaction results, the system introduces a preset contribution threshold (e.g., 0.05). The threshold can be set based on the following considerations: statistical significance: determining a value that can eliminate noisy variables through cross-validation or rules of thumb; computational resource constraints: adjusting the threshold according to the dimensionality tolerance of subsequent Bayesian optimization to ensure the number of key parameters is controlled within an acceptable range (e.g., 3-5); process experience: fine-tuning the threshold based on domain knowledge to avoid accidentally deleting important parameters.

[0077] The specific identification strategy is as follows: Parameters with a contribution value below the threshold are classified as "non-critical parameters." These parameters have a negligible impact on product purity and can be temporarily ignored from an optimization perspective. Parameters with a contribution value above the threshold are retained as "critical parameters," forming the core variables for subsequent optimization.

[0078] The parameter search space after dimensionality reduction is constructed as follows: The original search range of key parameters is kept unchanged to ensure that optimization is carried out within the feasible region; For parameters that show a positive contribution in SHAP analysis ( >0), biases can be injected into the mean function of a Gaussian process optimized by Bayesian optimization, for example, let ,in Fitting based on the average SHAP value allows the acquisition function to preferentially sample in high contribution regions; For parameters that contribute negatively or have process limitations, set hard constraints (such as an upper limit not exceeding the historical best value) to prevent the algorithm from exploring known unfavorable regions.

[0079] Based on the key parameters identified by SHAP analysis, subsequent optimization will primarily focus on a parameter subspace with significantly reduced dimensionality. To illustrate the necessity of using SHAP analysis for parameter dimensionality reduction in the initial optimization stage, this disclosure compares the performance of the Bayesian optimization (BO) algorithm in parameter spaces of different dimensions through simulation experiments. Please refer to [reference needed]. Figure 4 The results show that the optimization efficiency in the 4D parameter space is significantly higher than that in the 30D parameter space, verifying the key role of dimensionality reduction in improving optimization efficiency.

[0080] Taking the activation reaction step of a certain amino acid as an example, the historical dataset contains 100 complete experimental records. First, an XGBoost model is trained with parameters set to 300 weak learners, a maximum depth of 6, and a learning rate of 0.1. After training, SHAP analysis is applied to the model to obtain the global contribution of each parameter: temperature 0.35, activator ratio 0.28, reaction time 0.15, stirring rate 0.02, and pump speed 0.01. If the threshold is set to 0.05, stirring rate and pump speed are removed, retaining temperature, activator ratio, and reaction time as key parameters, reducing the search space to 3 dimensions. Bias is injected into the Bayesian-optimized Gaussian process mean function for positive contribution parameters, causing the acquisition function to prioritize searching for high-contribution regions.

[0081] The SHAP analysis and dimensionality reduction optimization method quantifies the marginal contribution of each process parameter to product purity, accurately screens key parameters, and then seamlessly connects the dimensionality-reduced parameter space with Bayesian optimization. By using strategies such as bias injection, it achieves targeted search and significantly improves optimization efficiency. This method has good universality and scalability, is applicable to different reaction systems, and can work in conjunction with the kinetic feature extraction modules of each step to jointly construct a complete closed loop of "perception-analysis-optimization".

[0082] In step S1044, by applying the Bayesian optimization algorithm to the dimensionality-reduced key parameter search space, sequential decision-driven intelligent parameter optimization is achieved.

[0083] In some implementations, an iterative search is performed within the parameter search space based on a Bayesian optimization algorithm to generate recommended combinations of key parameters, including: A proxy model between the parameter search space and the result data is constructed using a Gaussian process, and a collection function is called to find the parameter point that maximizes the expected gain in the proxy model to generate the recommended key parameter combination.

[0084] The system first uses a Gaussian process to construct a surrogate model between the parameter search space and the result data. A Gaussian process is a nonparametric probabilistic model that can simultaneously provide predicted values ​​and uncertainty estimates, making it particularly suitable for sequential optimization problems under small sample conditions.

[0085] Specifically, assuming it has been completed Group experimental data are ,in Indicates the first The key parameter combinations for this experiment (such as temperature, activator ratio, reaction time, etc.) This represents the corresponding result data (such as intermediate purity). The Gaussian process model assumes an objective function... It follows a Gaussian process prior: ,in This represents the mean function, and is usually set to a constant or zero; This represents the covariance function (kernel function), used to measure the similarity between points with different parameters. In some implementations, the squared exponential kernel commonly used in Gaussian processes is employed. This kernel function controls the smoothness of each input dimension through a length scale parameter, and characterizes the overall amplitude of the function output and the intensity of observed noise through signal variance and noise variance, respectively. These hyperparameters can be learned from historical data by maximizing marginal likelihood.

[0086] Based on the Gaussian process model, for any unsampled point, its posterior predicted distribution can be calculated: ; ,in This represents the predicted mean, indicating the expected purity achieved under this parameter combination. This represents the prediction variance, indicating the uncertainty of the prediction. Represents the covariance matrix among historical samples; This represents the covariance vector between historical samples and new samples.

[0087] After obtaining the posterior prediction distribution of the Gaussian process surrogate model, the system calls the acquisition function to search for the parameter point that maximizes the expected gain in the parameter search space. The acquisition function strikes a balance between "utilizing" (selecting regions with high prediction mean) and "exploring" (selecting regions with high prediction uncertainty) to efficiently approximate the global optimum.

[0088] In some implementations, an expected improvement (EI) acquisition function is used, the expression of which is: ,in , and Let represent the cumulative distribution function and probability density function of the standard normal distribution, respectively. This indicates an exploration-utilization balance parameter (e.g., a value of 0.01). The EI acquisition function takes a larger value in regions with high prediction mean and high prediction uncertainty, which can effectively balance exploration and utilization.

[0089] Those skilled in the art will understand that, in addition to EI, other standard acquisition functions such as probability boosting (PI) or upper confidence limit (UCB) can be selected according to actual needs. Their implementation methods are well known in the art and will not be described in detail here.

[0090] The system determines the next set of key parameter combinations for the experiment by maximizing the value of the selected acquisition function. ,in This serves as the search space for key parameters after dimensionality reduction (e.g., temperature 20-60℃, activator ratio 0.8-1.5, coupling time 5-60 min). This recommended parameter combination will be used in the next round of experiments.

[0091] After completing a new round of experiments, the system will combine the newly obtained parameters. and its corresponding result data Add to historical dataset The Gaussian process surrogate model is updated, and the above iterative process is repeated. As the number of iterations increases, the accuracy of the surrogate model continuously improves, and the points recommended by the acquisition function gradually converge towards the global optimum.

[0092] Taking a specific amino acid coupling reaction step as an example, the historical dataset contains key parameter combinations and purity results from completed experiments. Initial values ​​for the hyperparameters of the Gaussian process kernel function are set as follows: signal variance 1.0, length scale 0.5, and noise variance 0.1, adaptively adjusted by maximizing marginal likelihood. The acquisition function chosen is Expected Boosting (EI), and the exploration-equilibrium parameters are... The value was set to 0.01. An iterative search was performed within the 3D parameter space (temperature, activator ratio, coupling time). The first round of recommended parameters were (temperature 35℃, activator ratio 1.2, coupling time 25 min), with an actual experimental purity of 92.5%. After adding this data to the historical set, the surrogate model was updated. The second round of recommendations were (temperature 38℃, activator ratio 1.3, coupling time 20 min), with a purity of 93.8%. This iterative process continued until, after 5 rounds, the purity stabilized above 94.5%, meeting the process requirements.

[0093] In the aforementioned Bayesian optimization iterative search mechanism, the Gaussian process surrogate model can obtain reliable predictions and uncertainty estimates under small sample conditions, typically converging to the optimal parameter combination in only 5-10 iterations, significantly outperforming traditional grid search or random search. The design of the acquisition function strikes a balance between "utilizing" high-value regions and "exploring" unsampled regions, effectively avoiding getting trapped in local optima. This mechanism seamlessly integrates with the SHAP dimensionality reduction step, fully leveraging the efficient search capabilities of Bayesian optimization in the low-dimensional key parameter space. Simultaneously, the prediction variance output by the Gaussian process intuitively reflects the uncertainty of the recommended parameters, providing operators with a quantifiable reference for decision confidence.

[0094] In step S1046, the reinforcement learning algorithm is used to dynamically adjust the reaction process based on real-time process data, thereby achieving millisecond-level instant compensation for minor disturbances within the batch (such as temperature fluctuations and raw material differences), ensuring that the actual reaction trajectory always approaches the ideal curve. This mechanism compensates for the deviation between the static process parameter settings and the dynamic reaction process, enabling each reaction step to operate under optimal conditions, thus ensuring the consistency and stability of intermediate quality from the source.

[0095] In some implementations, based on reinforcement learning algorithms, the corresponding static process parameters are dynamically adjusted according to the process data during each reaction step, including: Using the weighted sum of the real-time feature values ​​of the process data and the reaction time as the reward function, the correction amount of the synthesis parameters is calculated in real time using a reinforcement learning algorithm. The convergence of the current reaction step is optimized by adjusting the real-time value of the synthesis parameters.

[0096] This step models the reaction process as a Markov decision process, where: state These are composed of real-time feature values ​​from the process data at the current moment, i.e., the aforementioned dynamic feature parameters. For example, LSTM feature vectors can be directly used as state inputs for reinforcement learning.

[0097] action This refers to the correction amount for static process parameters, such as temperature fine-tuning values ​​and stirring speed correction amounts. A continuous motion space can be used to support fine-tuning.

[0098] award Used to evaluate the immediate effect of the current control action and guide the learning algorithm to optimize the control strategy.

[0099] This disclosure presents a reward function that integrates process state and reaction time. ,in express The comprehensive health score of time-process data can be obtained by adding a linear layer or a small fully connected network on top of the LSTM feature vectors when using LSTM feature vectors. This maps the 32-dimensional features to a health score between 0 and 1. This network can be jointly trained with a reinforcement learning policy network or fitted offline based on historical data. This indicates the time elapsed since the current reaction began. The time term is introduced to encourage the algorithm to complete the reaction as quickly as possible while maintaining quality, avoiding unnecessary delays. Weighting coefficients. , To balance quality and efficiency, in processes that pursue high purity, a setting can be made... Much larger For example, values ​​of 1.0 and 0.05 are taken respectively. This reward function design enables reinforcement learning algorithms to learn the optimal control strategy online: when the response state is good (…). High) and fast response ( High rewards are given for short responses; conversely, low rewards or even penalties are given for responses that deviate from the ideal trajectory or take too long.

[0100] Reinforcement learning algorithms suitable for continuous action spaces, such as Deep Deterministic Policy Gradient (DDPG) or Proximal Policy Optimization (PPO), can be employed. The algorithm uses historical response data (including LSTM feature vectors, actual control actions, and final intermediate purity) as training samples and learns a mapping policy network from state to action through offline training.

[0101] The training process is as follows: Data collection: Utilize historical reaction data to build an experience replay pool, where each experience includes state, action, reward, and next moment's state.

[0102] Policy training: By maximizing the expected cumulative reward, the policy network parameters are iteratively updated, enabling the algorithm to learn to take the optimal control action in various states.

[0103] Model Deployment: After training, the strategy network is deployed to the real-time control system. During each reaction step, the system collects the current state at a fixed frequency (e.g., 10 Hz), inputs it into the strategy network, outputs the action (i.e., the correction amount of each process parameter), and executes it after superimposing it onto the static setpoint.

[0104] Through the above design, this step can perceive state changes in real time during the reaction process and make optimal control decisions based on the reward function, effectively compensating for batch-level disturbances, ensuring that process data closely approximates the preset kinetic target, and that the reaction trajectory always closely approximates the ideal curve, thereby improving the stability and consistency of intermediate quality. This mechanism, in conjunction with the aforementioned kinetic feature extraction step, forms an information closed loop of "real-time perception - feature analysis - dynamic control." It should be noted that the control commands not only correct static process parameters but also directly drive the process data to evolve according to the expected trajectory, achieving real-time observability and verifiability of the control effect.

[0105] The method disclosed herein constructs a complete data closed loop of "perception-analysis-optimization-regulation" through deep synergy of real-time acquisition of multi-source data, interpretable machine learning analysis, Bayesian optimization search, and reinforcement learning dynamic regulation: The perception layer acquires process and result data at millisecond-level frequency, making the entire reaction process transparent and observable; the analysis layer quantifies the marginal contribution of each parameter to product purity through interpretable machine learning, accurately selecting key parameters from high-dimensional parameters; the optimization layer uses Bayesian optimization iterative search within the dimensionality-reduced key space to quickly approximate the globally optimal parameter combination with a small number of experiments; the regulation layer dynamically fine-tunes static parameters based on real-time process data during each reaction step, instantly compensating for batch-to-batch perturbations and ensuring that the reaction trajectory always closely approximates the ideal curve. These four layers work together to form a data-driven, self-evolving intelligent system, realizing a paradigm shift in chemical synthesis from experience-driven to data-driven. Taking peptide synthesis as an example, experiments show that the method can improve purity to over 99%, while reducing raw material consumption by 80% and shortening the development cycle by 93%. Furthermore, the method does not depend on a specific reaction system or synthesis device and can be widely applied to various chemical synthesis processes such as peptide synthesis, nucleic acid synthesis, small molecule drug synthesis, and polymer material preparation. It also supports flexible expansion from laboratory research and development to industrial production, providing a general technical solution for the intelligent upgrading of chemical synthesis.

[0106] Example 2 This disclosure provides an artificial intelligence-based closed-loop automated chemical synthesis system 1000.

[0107] The synthesis system 1000 may include corresponding modules that execute one or more steps in the flowchart of the AI-based closed-loop automated chemical synthesis method described above. Therefore, each or more steps in the flowchart can be executed by a corresponding module, and the synthesis system 1000 may include one or more of these modules. A module may be one or more hardware modules specifically configured to execute a corresponding step, or implemented by a processor configured to execute a corresponding step, or stored in a readable storage medium for processor implementation, or implemented through some combination thereof.

[0108] Specifically, such as Figure 5 As shown, the synthesis system 1000 includes: The multi-source data acquisition module 1002 is used to acquire multi-source data of each reaction step in the chemical synthesis process in real time. The multi-source data includes process data reflecting the reaction progress and result data reflecting the quality of the reaction products. The intelligent decision-making module 1004 is used to optimize and control the synthesis parameters of the reaction step based on the multi-source data, including: The parameter parsing module 10042 analyzes historical synthetic parameters based on an interpretable machine learning algorithm, calculates the contribution weight of each synthetic parameter to the corresponding result data, and identifies key parameters based on the contribution weight to construct a dimensionality-reduced parameter search space, wherein the synthetic parameters include static process parameters and the corresponding process data. The parameter optimization module 10044, based on the Bayesian optimization algorithm, performs an iterative search within the parameter search space to generate recommended combinations of key parameters; The parameter control module 10046, based on a reinforcement learning algorithm, dynamically controls the corresponding static process parameters according to the process data during each reaction step.

[0109] In some embodiments, the synthesis system further includes: An automated synthesis platform is used to automate the entire process of chemical synthesis reactions, from raw material input to product output. The multimodal online characterization module is used to collect multi-source data in real time during the reaction process executed by the automated synthesis platform.

[0110] The synergy between the automated synthesis platform and the multimodal online characterization module forms the physical foundation of the closed-loop automated chemical synthesis system: while the platform executes the reaction, the characterization module quantifies and records the reaction state at each moment in real time, achieving synchronization between perception and execution; furthermore, the collected process data is acquired by the multi-source data acquisition module, and after being analyzed by the intelligent decision-making module, optimization parameters are output. The platform's central control system then adjusts the execution parameters in real time, forming a real-time closed loop of "perception-analysis-optimization-execution"; at the same time, all the full-process data (including static process parameters, real-time process data, and result data) from each platform run are completely stored in the historical database, continuously expanding the high-quality training set for interpretable machine learning analysis, making the parameter contribution quantification of the analysis module continuously more refined, the search strategy of the optimization module increasingly efficient, and the action output of the control module more precise.

[0111] This disclosure also provides an electronic device, including: a memory storing execution instructions; and a processor or other hardware module executing the execution instructions stored in the memory, causing the processor or other hardware module to execute the above-described visual method for real-time monitoring of homogeneous liquid phase reactions.

[0112] This disclosure also provides a readable storage medium storing execution instructions, which, when executed by a processor, are used to implement the above-described visual method for real-time monitoring of homogeneous liquid phase reactions.

[0113] The hardware architecture of the synthesis system 1000, implemented using a processor-based hardware approach, can be implemented using a bus architecture. The bus architecture can include any number of interconnect buses and bridges, depending on the specific application and overall design constraints of the hardware. Bus 1100 connects various circuits including one or more processors 1200, memory 1300, and / or hardware modules. Bus 1100 can also connect various other circuits 1400 such as peripheral devices, voltage regulators, power management circuits, external antennas, etc.

[0114] Bus 1100 can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Component Architecture (EISA) bus, etc. Bus 1100 can be divided into address bus, data bus, control bus, etc. For ease of representation, only one connection line is used in this diagram, but this does not indicate that there is only one bus or one type of bus.

[0115] Any process or method description in the flowcharts or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of preferred embodiments of this disclosure includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of this disclosure pertain. The processor performs the various methods and processes described above. For example, the method embodiments of this disclosure may be implemented as software programs tangibly contained in a machine-readable medium, such as memory. In some embodiments, part or all of the software program may be loaded and / or installed via memory and / or a communication interface. When the software program is loaded into memory and executed by the processor, one or more steps of the methods described above may be performed. Alternatively, in other embodiments, the processor may be configured to perform one of the methods described above by any other suitable means (e.g., by means of firmware).

[0116] The logic and / or steps represented in the flowchart or otherwise described herein may be specifically implemented in any readable storage medium for use by, or in conjunction with, an instruction execution system, apparatus or device (such as a computer-based system, a processor-based system or other system that can fetch and execute instructions from, an instruction execution system, apparatus or device).

[0117] For the purposes of this specification, a "readable storage medium" can be any means capable of containing, storing, communicating, propagating, or transmitting a program for use in or in conjunction with an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of readable storage media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable read-only memory (CDROM). Furthermore, a readable storage medium can even be paper or other suitable media on which a program can be printed, since a program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in memory.

[0118] It should be understood that various parts of this disclosure can be implemented in hardware, software, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0119] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0120] Furthermore, the functional units in the various embodiments of this disclosure can be integrated into a single processing module, or each unit can exist physically separately, or two or more units can be integrated into a single module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a readable storage medium. The storage medium can be a read-only memory, a disk, or an optical disk, etc.

[0121] Example 3 Please refer to Figure 6 This disclosure provides an artificial intelligence-based closed-loop automated peptide synthesis system, wherein the automated synthesis platform includes: a reagent storage module for storing reactants such as amino acids, coupling reagents, activators, deprotection reagents, and solvents; a fluid delivery module composed of multiple controllable pumps and valves for precisely delivering reagents to corresponding reactors according to a preset program; a temperature control and stirring module for adjusting the reaction temperature and providing homogeneous mixing; a reaction unit including at least an activation reactor, a coupling reactor, and a deprotection reactor, each for executing corresponding reaction steps; an ultraviolet detection flow path system for timely introducing the reaction solution into an online high-performance liquid chromatography module for analysis; and a central control system for coordinating the operation of the above modules to achieve fully automated operation from raw material injection and reaction execution to product output.

[0122] The multimodal online characterization module, as described above, includes: a conductivity monitoring module for real-time monitoring of ion concentration changes in the amino acid activation reaction solution to determine the reaction progress; a color monitoring module for real-time monitoring of color changes in the amino acid coupling reaction solution to determine the reaction progress; and an online high-performance liquid chromatography module for monitoring the purity of the products in each of the aforementioned reaction steps.

[0123] The peptide synthesis system constructs a complete intelligent synthesis closed loop through the efficient collaboration of four core modules. The automated synthesis platform executes peptide synthesis, the multimodal online characterization module collects multi-source data from each reaction step in real time, the multi-source data acquisition module acquires this characterization data, and the intelligent decision-making module analyzes the characterization data in real time and autonomously optimizes the synthesis parameters, driving the peptide synthesis system to achieve fully unmanned operation from automatic injection of amino acid raw materials to output of high-yield, high-purity peptide products.

[0124] This embodiment of the disclosure uses the above-described system to synthesize an elastin tripeptide with the following tripeptide sequence: H-Pro-Gly-Gly-OH. The synthesis mechanism is as follows: Figure 7 As shown.

[0125] The specific steps include, where steps 1-7 are automated synthesis steps, and step 8 is a synthesis optimization step. The meanings of reagent abbreviations in the following text are shown in Table 1.

[0126] Table 1. Reagent Abbreviations

[0127] 1. Startup Steps An automated synthesis platform is constructed based on the liquid-phase synthesis method. The fundamental difference between liquid-phase synthesis and solid-phase synthesis lies in the elimination of the resin swelling step. Once the automated synthesis platform is started, the fluid delivery module will quantitatively transfer a specific ratio of raw material liquid and solvent from the reagent storage module according to preset synthesis parameters and deliver it to the reaction module to initiate the synthesis reaction.

[0128] 2. Initial amino acid activation steps The "DMF-Fmoc-Pro-OH" pump in the fluid delivery module quantitatively delivers DMF solvent and initial amino acids to the amino acid activation reaction module. When the system determines that the amount of amino acid raw material meets the preset parameter requirements for this synthesis task, this pump automatically shuts off. Subsequently, the "EA-DIC," "EA-HOSU," and "EA" delivery pumps transfer HOSU, DIC, and EA to the reaction flask according to preset molar ratios, respectively. After delivery, the system activates the material mixing and temperature control module, allowing the mixed solution to undergo an activation reaction at a preset temperature and stirring rate. During this process, the physical property data of the reaction solution are collected in real time using a high-precision conductivity electrode and simultaneously uploaded to the cloud-based endpoint determination model for dynamic analysis. When the model determines that the side reaction level is below a set threshold, the system automatically delivers the product of this step to the amino acid coupling module.

[0129] During amino acid activation, the ion pairs introduced by the addition of the activator cause a sharp increase in solution conductivity and the appearance of a characteristic peak. It should be noted that this conductivity peak only indicates the complete addition of the activator and does not signify the end of the main reaction (formation of the active ester). Subsequently, the formation of the byproduct diisopropylurea (DIPU) consumes the ion pairs, causing the ionic strength, i.e., conductivity, to gradually decrease. Therefore, changes in the online conductivity curve can intuitively reflect the extent of the byproduct. The model dynamically assesses the progress of the byproduct by analyzing the rate of decrease in conductivity after the peak in real time. When the rate of decrease in conductivity is below a threshold, the system determines that the byproduct is at its lowest level, and the product of the amino acid activation process can proceed to the next reaction step.

[0130] 3. The coupling step of amino acid at position 2 After initial amino acid activation, the system automatically switches to the amino acid coupling stage. First, the activated amino acids are quantitatively transferred to the coupling reaction flask via a "DMF-Fmoc-Pro-OSU" pump, while a "DCM-H-Pro-OH" pump precisely delivers the coupling raw materials according to a preset stoichiometric ratio. Subsequently, "Na2CO3-H2O" and "NaCl-H2O" solutions are quantitatively transferred using a high-precision pump valve assembly according to preset parameters. During this process, an immersion pH sensor monitors the pH changes of the reaction solution in real time. When the pH value stabilizes within the preset range of 8.0±0.2, the system automatically terminates the buffer solution transfer and simultaneously activates the temperature-controlled mixing module to enter the isothermal reaction stage. Throughout this process, the color monitoring module dynamically analyzes the comprehensive color index.

[0131] When the rate of change of the exponential change converges to below the threshold (0.03 units / minute) and remains below it for a period of time (3 minutes), the system determines that all free amino acids have participated in the coupling and delivers the amino acid coupling product to the subsequent steps.

[0132] 4. Washing steps for coupling reaction flasks After the amino acid at position 2 is coupled, the system will activate the "CH3CH2OH-H2O" extraction pump at a flow rate of 240 mL / min to inject the ethanol / water mixture into the coupling reaction flask at high speed. This pump will be turned off after 6 seconds. Simultaneously, a high-speed stirrer will be activated to thoroughly wash the flask for 30 seconds. Then, the waste liquid pump will be activated at a flow rate of 240 mL / min and turned off after 10 seconds, transferring all the washing waste liquid from the coupling reaction flask to a waste bottle. This process will be repeated three times. Finally, the system will activate the temperature control module to reach the preset temperature and dry the coupling reaction flask.

[0133] 5. The coupling step of amino acid at position 3 Steps 3 and 4 are repeated using modules such as fluid transport to complete the coupling of the amino acid at position 3.

[0134] 6. Fmoc Protecting Group Removal Steps After all amino acid coupling is complete, the system automatically switches to the peptide deprotection stage. First, the tripeptide product with the Fmoc protecting group is transferred to the deprotection reaction flask via the "DMF-Fmoc-Pro-Gly-Gly-OH" pump, while simultaneously, the deprotection reagent is quantitatively transferred to the deprotection reaction flask via the "DMF-(C2H5)2NH" extraction pump. During this process, the system activates the HPLC injection pump and HPLC mobile phase pump in forward flow mode, with the injection pump flow rate set to 0.33 mL / min and the mobile phase pump flow rate set to 1.00 mL / min. Both pumps drive the mobile phase to first deliver the quantitative reaction solution to the chromatographic column for separation, and then to the UV detector for analysis. The analytical data is synchronously uploaded to the intelligent decision module for dynamic analysis. Figure 8 As shown, it can be confirmed that the target product obtained by isolation is elastin tripeptide.

[0135] 7. Washing steps After the Fmoc protecting group is completely removed, the H-Pro-Gly-Gly-OH solution is transferred to a sample storage bottle for retention. The system automatically starts the "CH3CH2OH-H2O" extraction pump at a flow rate of 360 mL / min, injecting the ethanol / water mixture at high speed into each reaction flask and tubing. After 60 seconds, the pump is turned off. Simultaneously, the high-speed stirring device is turned on to thoroughly wash each reaction flask for 120 seconds. Then, the waste liquid pump is turned on at a flow rate of 240 mL / min, and after 120 seconds, the pump is turned off, transporting all the washing waste liquid to a waste bottle. The above process is repeated 3 times. The temperature control module is then activated to the preset temperature to dry each reaction flask.

[0136] 8. Product post-processing The system injects a quantitative amount of MTBE into the collected H-Pro-Gly-Gly-OH solution, followed by precipitation, rotary evaporation, and drying to obtain the H-Pro-Gly-Gly-OH tripeptide powder product. For example... Figure 9 As shown, the elastin tripeptide synthesized by the peptide synthesis system has extremely high purity (97.42%) in its crude product, and post-processing has very limited effect on improving the purity (97.44%). This indicates that the peptide synthesis system can efficiently and selectively obtain the target product in one step.

[0137] During steps 1-8 above, various synthesis parameters, including multi-source data collected in real time by the multimodal online characterization module and the final "qualified" / "failed" judgment results for each step, collectively constitute a high-dimensional dataset. The peptide synthesis system employs an intelligent strategy that integrates interpretability analysis and Bayesian optimization to achieve efficient closed-loop optimization.

[0138] First, SHAP analysis was applied to analyze the high-dimensional dataset, quantifying the contribution of different synthetic parameters to product purity and identifying the most critical influencing factors. For example, in analyzing the amino acid activation step, it was found that "reaction temperature" and "substrate ratio" were the dominant parameters, while "activator type" had a relatively weak impact.

[0139] Based on the key parameters identified by SHAP analysis, subsequent optimization will primarily focus on a parameter subspace with significantly reduced dimensionality. For example, when optimizing the amino acid activation process, the system will initiate Bayesian optimization within a subspace composed of key parameters such as "reaction temperature" and "substrate ratio." First, a surrogate model (using a Gaussian process) is constructed based on existing experimental data to correlate product purity with key parameters, and the next set of most promising synthesis parameters is calculated using a data acquisition function. Subsequently, the system executes the synthesis experiments in steps 1-8 according to the recommended parameter combinations, and uses the new data to update the surrogate model, allowing the model to make recommendations again. Through multiple iterative cycles of "modeling-recommendation-experimentation-update," the system gradually approaches the optimal parameter combination. The product purity improvement effect of each reaction step is shown in the figure. Figure 10 As shown.

[0140] Based on this, the system introduces a reinforcement learning dynamic control mechanism: during the execution of a single reaction step in each round of experiments, the process data collected in real time (extracted as LSTM feature vectors) is input into the pre-trained reinforcement learning policy network. The network outputs millisecond-level corrections to static process parameters and process data (such as temperature fine-tuning), instantly compensating for unforeseen disturbances such as raw material batch differences and environmental fluctuations, ensuring that the actual reaction trajectory always approaches the ideal curve set by Bayesian optimization.

[0141] Thus, the system constructs a well-defined collaborative optimization paradigm: SHAP analysis is responsible for identifying the key parameter space, laying the foundation for efficient search; Bayesian optimization performs precise iterative search within this space, exploring the globally optimal parameter combination; and reinforcement learning dynamically adjusts the process in real time during each single-step reaction, ensuring the precise execution of the optimization objective. This deep collaboration among the three significantly improves the efficiency and stability of the peptide synthesis optimization process, achieving a paradigm shift from "experience-driven" to "data-driven."

[0142] Those skilled in the art should understand that the above embodiments are merely for illustrating the present disclosure and are not intended to limit the scope of the disclosure. Those skilled in the art can make other changes or modifications based on the above disclosure, and these changes or modifications still fall within the scope of the present disclosure.

Claims

1. A closed-loop automated chemical synthesis method based on artificial intelligence, characterized in that, The method includes: Acquire multi-source data of each reaction step in the chemical synthesis process in real time. The multi-source data includes process data reflecting the reaction progress and result data reflecting the quality of the reaction products. Based on the multi-source data, the synthesis parameters of this reaction step are optimized and controlled, including: Based on the interpretable machine learning algorithm, the historical synthesis parameters are analyzed, the contribution weight of each synthesis parameter to the corresponding result data is calculated, and key parameters are identified based on the contribution weight to construct a dimensionality-reduced parameter search space. The synthesis parameters include static process parameters and the corresponding process data. Based on the Bayesian optimization algorithm, an iterative search is performed within the parameter search space to generate recommended combinations of key parameters; Based on reinforcement learning algorithms, the corresponding static process parameters are dynamically adjusted according to the process data during each reaction step.

2. The method as described in claim 1, characterized in that, The chemical synthesis process includes peptide synthesis.

3. The method as described in claim 2, characterized in that, Acquire multi-source data for each reaction step in the chemical synthesis process in real time. This multi-source data includes process data reflecting the reaction progress and result data reflecting the quality of the reaction products, including: The reaction progress is determined by real-time monitoring of ion concentration changes in the amino acid activation reaction solution using a conductivity monitoring module. The color monitoring module is used to monitor the color change of the amino acid coupling reaction solution in real time to determine the reaction progress. The purity of the products from each reaction step was monitored using an online high-performance liquid chromatography module.

4. The method as described in claim 1, characterized in that, Based on interpretable machine learning algorithms, historical synthesis parameters are analyzed to calculate the contribution weight of each synthesis parameter to the corresponding result data. Key parameters are then identified based on these contribution weights to construct a dimensionality-reduced parameter search space, including: The Shapley value of each historical synthesis parameter is calculated using the SHAP algorithm. Synthesis parameters whose average absolute value of Shapley value is lower than a preset percentage are removed from the search space of the current dimension.

5. The method as described in claim 1, characterized in that, Based on the Bayesian optimization algorithm, an iterative search is performed within the parameter search space to generate recommended combinations of key parameters, including: A proxy model between the parameter search space and the result data is constructed using a Gaussian process, and a collection function is called to find the parameter point that maximizes the expected gain in the proxy model to generate the recommended key parameter combination.

6. The method as described in claim 1, characterized in that, Based on reinforcement learning algorithms, the corresponding static process parameters are dynamically adjusted according to the process data during each reaction step, including: Using the weighted sum of the real-time feature values ​​of the process data and the reaction time as the reward function, the correction amount of the synthesis parameters is calculated in real time using a reinforcement learning algorithm. The convergence of the current reaction step is optimized by adjusting the real-time value of the synthesis parameters.

7. A closed-loop automated chemical synthesis system based on artificial intelligence, characterized in that, The system includes: The multi-source data acquisition module is used to acquire multi-source data of each reaction step in the chemical synthesis process in real time. The multi-source data includes process data reflecting the reaction progress and result data reflecting the quality of the reaction products. The intelligent decision-making module is used to optimize and control the synthesis parameters of the reaction step based on the multi-source data, including: The parameter parsing module analyzes historical synthetic parameters based on an interpretable machine learning algorithm, calculates the contribution weight of each synthetic parameter to the corresponding result data, and identifies key parameters based on the contribution weight to construct a dimensionality-reduced parameter search space. The synthetic parameters include static process parameters and the corresponding process data. The parameter optimization module, based on the Bayesian optimization algorithm, performs an iterative search within the parameter search space to generate recommended combinations of key parameters; The parameter control module, based on a reinforcement learning algorithm, dynamically controls the corresponding static process parameters according to the process data during each reaction step.

8. The system as described in claim 7, characterized in that, The system also includes: An automated synthesis platform is used to automate the entire process of chemical synthesis reactions, from raw material input to product output. The multimodal online characterization module is used to collect multi-source data in real time during the reaction process executed by the automated synthesis platform.

9. An electronic device, characterized in that, include: The memory stores execution instructions; as well as A processor that executes execution instructions stored in the memory, causing the processor to perform the method according to any one of claims 1-6.

10. A readable storage medium, characterized in that, The readable storage medium stores execution instructions, which, when executed by a processor, are used to implement the method described in any one of claims 1-6.

Images