Chemical synthesis closed-loop autonomous optimization system and method based on in-situ spectroscopic feedback
The closed-loop autonomous optimization system for chemical synthesis, based on in-situ spectral feedback, identifies the reaction process in real time and automatically generates multi-parameter collaborative intervention commands. This solves the problems of insufficient real-time feedback and low optimization efficiency in automated chemical synthesis systems, achieving efficient multi-parameter collaborative intervention and cross-batch self-learning, thereby improving the yield and quality of target products.
Patent Information
- Application Number
- CN202610429840.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-04-02
- Publication Date
- 2026-07-10
AI Technical Summary
Existing automated chemical synthesis systems lack real-time feedback and dynamic intervention capabilities, resulting in fragmented monitoring and control, low optimization efficiency, and limited parameter adjustment, making it difficult to achieve multi-parameter collaborative intervention and cross-batch self-learning.
A closed-loop autonomous optimization system for chemical synthesis based on in-situ spectral feedback is adopted, which includes a reaction execution module, an in-situ spectral detection module, an AI real-time analysis module, a dynamic intervention control module, and an automatic convergence optimization module. The system uses a deep learning model to identify the reaction process in real time, predict the trend of by-product formation, and automatically generate multi-parameter collaborative intervention instructions. It also combines a Bayesian optimization algorithm to perform cross-batch self-learning optimization.
It enables real-time sensing of chemical reactions and coordinated intervention of multiple parameters, improving the yield of target products and the stability of product quality, and increasing the efficiency of process development.
Smart Images

Figure SMS_1
Abstract
Description
Technical Field
[0001] This invention relates to the field of automated chemical synthesis technology, and more specifically, to a closed-loop autonomous optimization system and method for chemical synthesis based on in-situ spectral feedback. Background Technology
[0002] With the rapid development of chemical engineering, drug development, and new material synthesis, automation and intelligentization of chemical synthesis processes have become key technological pathways for improving R&D efficiency, ensuring product quality, and reducing production costs. Traditional automated synthesis systems mainly rely on preset program control, where operators pre-set fixed parameters such as time-temperature curves and constant material dropping rates. The system can only execute these mechanically and cannot dynamically adjust according to actual changes during the reaction process. This open-loop control mode is highly susceptible to problems such as changes in raw material batches, environmental fluctuations, or reaction kinetics when facing complex and variable chemical reactions. These changes can lead to deviations from expectations, resulting in increased byproducts and decreased yields of the target product.
[0003] In recent years, advancements in in-situ spectroscopy techniques (such as infrared spectroscopy and Raman spectroscopy) have provided powerful tools for real-time monitoring of reaction processes. Researchers have attempted to immerse spectral probes directly into the reaction system to continuously acquire spectral information of reactants, products, and intermediates, thus visualizing the reaction process. For example, CN121034441A discloses a fermentation optimization method based on online Raman spectroscopy, which monitors the fermentation process through spectral data. However, most of these technologies remain at the monitoring level, with their core function being data acquisition and post-analysis. Spectral data is not used for real-time, automated closed-loop control of reaction parameters; that is, when an anomaly is detected, the system cannot intervene autonomously and still requires operator judgment and manual adjustment. This results in a significant time delay between the occurrence of an anomaly and effective intervention, making it difficult to suppress rapidly forming side reactions.
[0004] On the other hand, optimizing reaction conditions is the core of chemical synthesis research. Traditional methods heavily rely on researchers' experience, searching for optimal process parameters through iterative design-experimentation-analysis processes. This process is time-consuming, inefficient, and has poor reproducibility. Although some intelligent control systems have attempted to address this issue, they have significant shortcomings. For example, CN120893330A discloses an artificial intelligence-based textile design optimization method, but it has deficiencies in the effective fusion and feature extraction of multimodal data, making it difficult to achieve comprehensive perception and optimization of complex reaction processes. Similarly, CN121314494A discloses a photocatalytic synthesis device that employs a reinforcement learning algorithm, but its algorithm parameter optimization and control accuracy still need improvement, failing to establish a stable and efficient autonomous optimization mechanism.
[0005] In summary, existing technologies in the field of automated chemical synthesis have the following main drawbacks: 1. Open-loop control mode, lacking real-time feedback and dynamic intervention capabilities: Traditional automated systems rely on preset programs and cannot make real-time adjustments based on actual changes in the reaction process (such as abnormal generation of by-products), which makes the reaction prone to deviation and makes it difficult to guarantee product quality and yield.
[0006] 2. Monitoring and control are disconnected and fail to form a closed loop: Existing in-situ spectroscopy technology is only used for monitoring, and spectral data is mainly used for offline analysis. It fails to form a closed loop of monitoring-analysis-decision-control with the reaction execution module, and cannot provide real-time and automated intervention for abnormal situations.
[0007] 3. Low optimization efficiency and lack of cross-batch self-learning capability: Traditional process optimization relies on manual iteration, which is inefficient and has poor reproducibility. Existing intelligent systems have shortcomings in multimodal data fusion, feature extraction, and adaptive algorithm optimization, making it difficult to achieve the intelligent leap from passive execution to proactive optimization.
[0008] 4. Limited parameter adjustment and insufficient collaborative control capabilities: Existing systems often use single parameter adjustment, such as adjusting only the temperature, during intervention. They fail to incorporate multiple key parameters, such as temperature, dropping speed, and stirring rate, into a unified collaborative intervention logic, resulting in limited intervention effects.
[0009] Therefore, how to achieve real-time perception of reaction processes in chemical synthesis, autonomous decision-making and execution of multi-parameter collaborative intervention, and possess cross-batch self-learning capabilities has become a technical problem that urgently needs to be solved. Summary of the Invention
[0010] To address the aforementioned technical problems in related technologies, this invention proposes a closed-loop autonomous optimization system and method for chemical synthesis based on in-situ spectral feedback, which can overcome the above-mentioned shortcomings of existing technologies.
[0011] To achieve the above-mentioned technical objectives, the technical solution of the present invention is implemented as follows: A closed-loop autonomous optimization system for chemical synthesis based on in-situ spectral feedback, comprising: The reaction execution module is used to execute chemical reactions; The in-situ spectral detection module is used to acquire the spectral data of the reaction system in the reaction execution module in real time; The AI real-time analysis module is connected to the in-situ spectral detection module and has a built-in trained deep learning model to receive and analyze the spectral data in order to identify the reaction process and predict the generation trend of by-products in real time. A dynamic intervention control module, connected to the AI real-time analysis module and the reaction execution module, is used to automatically generate and execute adjustment instructions according to preset collaborative intervention logic when the AI real-time analysis module predicts an abnormal trend in by-product generation, thereby adjusting the reaction parameters of the reaction execution module; and An automatic convergence optimization module connects multiple sets of parallel-running reaction execution modules. It is used to collect yield data of each batch of reaction and automatically optimize the reaction conditions of subsequent batches based on the optimization algorithm, thereby achieving cross-batch self-learning optimization.
[0012] Furthermore, the reaction execution module includes a reaction vessel, a temperature control system, a material dropping system, and a stirring system; the reaction parameters include at least one of reaction temperature, material dropping rate, and stirring rate.
[0013] Furthermore, the in-situ spectral detection module includes at least one of an in-situ infrared spectrometer and a Raman spectrometer, with its detection probe immersed in the reaction system; the spectral data includes infrared spectral data and Raman spectral data.
[0014] Furthermore, the deep learning model built into the AI real-time analysis module is a combination of convolutional neural networks and recurrent neural networks, used to extract spectral features and predict temporal changes; the output of the AI real-time analysis module includes target product concentration, conversion rate, by-product generation rate, and by-product generation trend prediction.
[0015] Furthermore, the collaborative intervention logic of the dynamic intervention control module includes: when the by-product generation rate exceeds a preset threshold, first reduce the reaction temperature; if the by-product generation rate does not decrease after temperature adjustment, reduce the material dropping speed; when catalyst sedimentation or agglomeration is detected, increase the stirring rate.
[0016] Furthermore, the automatic convergence optimization module employs a Bayesian optimization algorithm or a reinforcement learning algorithm; the Bayesian optimization algorithm recommends the response conditions for the next round of experiments by constructing a Gaussian process regression model and maximizing the acquisition function.
[0017] Furthermore, the AI real-time analysis module is also used to predict the by-product generation trend in the future. When it is predicted that the by-product generation rate will exceed a preset threshold, the dynamic intervention control module executes the adjustment instruction in advance. The deep learning model in the AI real-time analysis module is pre-trained through transfer learning and fine-tuned using experimental data of a specific reaction system. The system also has a model iteration update function, which is used to incrementally train the model based on newly generated experimental data.
[0018] Furthermore, the system also includes a human-computer interaction interface for displaying real-time analyzed response curves, historical intervention records and optimization suggestions, and allows operators to manually adjust intervention thresholds and optimization parameters.
[0019] Furthermore, the system's parameter configuration is determined through time-domain response analysis, specifically including: analyzing the dominant frequency of byproduct generation rate fluctuations through Fourier transform, setting the spectral acquisition interval based on the Nyquist sampling theorem; and statistically analyzing historical intervention success rates to set early warning thresholds and high-risk thresholds.
[0020] A closed-loop autonomous optimization method for chemical synthesis based on in-situ spectral feedback, applied to the aforementioned system, includes the following steps: S1: Initialize and set the basic reaction parameters and optimization objectives; S2: Real-time acquisition of multimodal spectral data of the reaction system; S3: Utilize a trained deep learning model to analyze the spectral data, identify the reaction process in real time, and predict the generation trend of by-products; S4: When an abnormal trend in byproduct formation is predicted, an adjustment instruction is automatically generated and executed based on the synergistic intervention logic to adjust the reaction parameters; S5: After the batch is completed, calculate the yield of this batch and archive the data; S6: After several batches are completed, the reaction conditions of subsequent batches are automatically optimized based on the yield data of each batch using the optimization algorithm, so as to achieve self-learning optimization across batches.
[0021] The beneficial effects of this invention are as follows: This invention uses an AI real-time analysis module to perform real-time inference and trend prediction on in-situ spectral data, enabling the dynamic intervention control module to automatically execute multi-parameter collaborative intervention before or in the early stages of abnormal by-product generation. This transforms traditional open-loop execution into closed-loop autonomous regulation of monitoring-analysis-decision-control, effectively suppressing side reactions and significantly improving the yield of the target product. By incorporating temperature, dropping rate, and stirring rate into a unified intervention logic, the intervention effect of multi-parameter collaborative regulation is superior to that of single-parameter regulation, achieving refined control of the reaction process. Through an automatic convergence optimization module, Bayesian optimization or reinforcement learning algorithms are used to iteratively optimize multiple batches of parallel experimental data, enabling the system to have cross-batch self-learning capabilities, continuously accumulating experience and autonomously optimizing the feeding ratio. This achieves an intelligent upgrade from passive execution to active optimization, significantly improving process development efficiency and product quality stability. Detailed Implementation
[0022] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the present invention.
[0023] According to an embodiment of the present invention, a closed-loop autonomous optimization system for chemical synthesis based on in-situ spectral feedback includes a reaction execution module, an in-situ spectral detection module, an AI real-time analysis module, a dynamic intervention control module, and an automatic convergence optimization module.
[0024] The reaction execution module includes at least one reaction vessel, a temperature control system, a material dropping system, and a stirring system. The reaction vessel can be a batch reactor, a tubular reactor, or a tank reactor; the temperature control system employs a dual-layer temperature control design, including an outer circulating water bath and an inner temperature sensor, enabling precise control of the reaction temperature; the material dropping system uses an injection pump or a metering pump to achieve precise material dropping; and the stirring system uses a mechanical stirrer or an ultrasonic stirrer to ensure uniform mixing of the reactants.
[0025] The in-situ spectral detection module includes an in-situ infrared spectrometer and a Raman spectrometer. Its detection probes are either Attenuated Total Reflectance (ATR) probes or transmission probes, immersed in the reaction system. The infrared spectrometer has an optical path length of 2-5 mm and a resolution of 4-8 cm⁻¹; the Raman spectrometer uses a laser wavelength of 532 nm or 633 nm and a beam radius of 2-5 mm. The module also includes an automatic brightness enhancement system and a spectral calibration system to ensure the accuracy and stability of the spectral data.
[0026] The AI real-time analysis module is connected to the in-situ spectral detection module and incorporates a trained deep learning model. This model uses a convolutional neural network to extract spectral features and combines it with a recurrent neural network to predict temporal changes. The model's input is real-time spectral data, and its output includes the conversion curve of the target reactant, the formation rate of key byproducts, and other reaction parameters. The model's training data includes historical experimental data and reference spectra of standard substances, and it is trained and optimized using an iterative backpropagation algorithm.
[0027] The dynamic intervention control module is connected to the AI real-time analysis module and the reaction execution module. When the by-product generation rate analyzed by the AI real-time analysis module exceeds a preset threshold, the module automatically generates an adjustment command. First, the reaction temperature is reduced by 2-5°C. If the by-product generation rate still does not decrease, the material dropping speed is simultaneously reduced to 60-80% of the original speed. The module also includes an intelligent sleep function, which automatically enters a low-power mode when the reaction is stable.
[0028] The automatic convergence optimization module connects multiple parallel-running reaction execution modules and employs Bayesian optimization or reinforcement learning algorithms for optimization. Based on the yield data of the first batch of reactions, this module automatically optimizes the feed ratio for the second batch. The objective function is the target product yield or selectivity, and the algorithm converges to the optimal solution through multiple iterations. This module also has data storage capabilities, which can be used for long-term experimental data analysis and model updates.
[0029] The system also includes a human-computer interaction interface (HCI) for displaying real-time analyzed response curves, historical intervention records, and optimization suggestions. The interface uses a touchscreen design for intuitive operation. Through the interface, operators can manually adjust intervention thresholds, optimization parameters, and system settings, enabling both remote control and local operation of the system.
[0030] The system can also be equipped with optional analytical instruments such as gas chromatographs and mass spectrometers to achieve online monitoring and analysis of the reaction gas phase and products. The system features a modular design, facilitating functional expansion and system upgrades to meet different needs.
[0031] The core of this invention lies in its closed-loop logic processing engine. After the system starts, it automatically runs according to the closed-loop process of "collection-analysis-decision-execution-learning," and the specific logical steps are broken down as follows: Step 1: Initialization and Target Setting Operators set basic reaction parameters (initial temperature, dropping rate) and optimization targets (such as target product yield > 85%, byproduct formation rate threshold < 0.05 mmol / L·s) through the human-machine interface. The system performs a self-check to ensure the spectral probe is clean and the temperature control system is functioning properly.
[0032] Step 2: Real-time acquisition of multimodal data After the reaction is initiated, the in-situ spectral detection module acquires infrared and Raman spectral data alternately or simultaneously at a preset frequency (e.g., 30 seconds / time). After automatic baseline calibration and noise reduction, the spectral data forms a standardized data stream, which is then input to the AI real-time analysis module.
[0033] Step 3: Real-time AI analysis (state awareness) The AI real-time analysis module incorporates a deep learning model that performs real-time inference on the spectral data stream. This model can not only identify the current concentration of the target product, but more importantly, it can predict the trend of byproduct formation within the next 5-10 minutes. For example, if the model identifies that the slope of a byproduct characteristic peak (such as the C=C peak of a dimer) is rising sharply, even if the current absolute value has not yet exceeded the limit, the system will mark it as a "potential risk state".
[0034] Step 4: Dynamic intervention decision-making (closed-loop control) The dynamic intervention and control module makes multi-level decisions based on the AI analysis results: Normal state (green zone): Byproduct rate < 80% of the threshold, the system continuously monitors, the intelligent sleep function is activated, and the sampling frequency is reduced to save resources.
[0035] Warning status (yellow zone): When the by-product rate reaches 80%-100% of the threshold, the system issues an interface prompt and prepares to intervene.
[0036] Intervention status (red zone): Byproduct rate ≥ threshold (e.g., 0.05 mmol / L·s). The system immediately executes the collaborative intervention logic: First priority (temperature control): Decrease the temperature at a rate of 0.5℃ / step, with a maximum temperature drop of no more than 5℃, in order to reduce the reaction kinetic rate and suppress side reactions; Second priority (dropping rate adjustment): If the byproduct rate does not show a decreasing trend within 30 seconds after cooling, the system initiates dropping rate adjustment, reducing the dropping rate in steps of 0.1 mL / min, down to a minimum of 60% of the original rate; Third priority (stirring adjustment - optional): If Raman spectroscopy detects sedimentation or agglomeration of solid catalyst, the system instantaneously increases the stirring rate (e.g., by 200-300 rpm) for 15 seconds to break up the agglomerates before resuming.
[0037] Recovery logic: When the AI analysis shows that the byproduct rate is below the threshold for 1 minute, the system gradually restores the parameters to the initial set value in the order of "first restore the drip rate, then restore the temperature".
[0038] Step 5: Batch End and Data Archiving After the reaction is completed, the automatic convergence optimization module automatically calculates key performance indicators such as the target product yield, total by-products, number of interventions, and energy consumption for this batch, and packages them with the spectral data and operating parameters of the entire process into the historical database.
[0039] Step 6: Cross-batch self-learning and model update After all parallel experiments in multiple batches have been completed, the system initiates cross-batch optimization: For the Bayesian optimizer: update the Gaussian process prior using the yield data from the new batch, recalculate the acquisition function, and recommend experimental parameters for the next round (such as the feed ratio).
[0040] For the AI analysis model: every month or after every 50 batches of reactions, the system calls historical data to incrementally train the deep learning model, optimize the model weights, and enable it to identify new and rare side reaction paths.
[0041] To facilitate understanding of the above technical solutions of the present invention, the following detailed description of the above technical solutions of the present invention will be provided through specific usage methods.
[0042] Example 1: This embodiment provides a closed-loop autonomous optimization system for chemical synthesis based on in-situ spectral feedback, used to optimize the synthesis reaction of a certain drug intermediate. This reaction is temperature-sensitive and prone to generating dimer byproducts due to excessively high local concentrations. The reaction execution module uses a 100mL jacketed reactor equipped with a circulating water bath temperature control system (accuracy ±0.5℃) and a syringe pump dispensing system. The in-situ spectral detection module uses an immersion ATR-FTIR probe (silicon crystal, 3mm optical path), with a spectral acquisition range of 4000-650cm⁻¹, a resolution of 4cm⁻¹, and acquires spectral data every 30 seconds.
[0043] The AI real-time analysis module's built-in deep learning model, trained on 50 batches of reaction data, accurately identifies the carbonyl characteristic peak of the target product at 1680 cm⁻¹ and the C=C characteristic peak of the byproduct at 980 cm⁻¹. The model outputs the target product concentration and byproduct formation rate in real time. When the byproduct formation rate exceeds the threshold of 0.05 mmol / L·s, the dynamic intervention control module triggers intervention: first, the reaction temperature is reduced from the set value of 60℃ to 58℃; if the byproduct rate still does not drop below the threshold after 30 seconds, the dropping rate is reduced from 0.5 mL / min to 0.35 mL / min. Intervention continues until the byproduct rate returns to normal, after which the original parameters are gradually restored.
[0044] To ensure the accuracy and generalization ability of the AI real-time parsing module, the model construction follows a refined process: 1. Pre-training phase (transfer learning) Data sources: Publicly available chemical spectral databases (such as the NIST Chemistry WebBook) and measured spectral data of standard substances (pure target product, pure byproduct).
[0045] Training objective: To enable convolutional neural network layers to learn to recognize common chemical bond characteristic absorption peaks (such as C=O, C=C, OH, etc.) and establish preliminary feature extraction capabilities.
[0046] 2. Fine-tuning stage (for specific responses) Data augmentation: For the target reaction system, conduct 5-10 "exploratory" experiments, deliberately changing conditions such as temperature and concentration to generate rich byproduct spectra. Perform data augmentation operations on these spectral data, such as random shifting, adding Gaussian noise, and baseline tilting, to simulate various disturbances in a real-world environment.
[0047] Labeling: Based on the results of offline liquid chromatography-mass spectrometry analysis, the "true" concentration of byproducts at each time point in the spectral data is accurately labeled.
[0048] Training and Optimization: The Adam optimizer was used with an initial learning rate of 0.001. The entire network (convolutional neural network + recurrent neural network) was fine-tuned using labeled data. The loss function was a weighted sum of mean squared error and structural similarity index, ensuring the model could accurately predict concentration values while precisely capturing subtle changes in the spectral curve.
[0049] 3. Model iteration (continuous learning) Triggering mechanism: The system automatically monitors the deviation between the AI analysis results and the offline analysis results. If the deviation exceeds 5% for three consecutive batches, a model update alert will be triggered.
[0050] Incremental learning: Without losing existing knowledge, the model is incrementally trained using newly generated high-confidence data, enabling the model to adapt to the effects of long-term drift such as changes in raw material batches and catalyst aging.
[0051] The automatic convergence optimization module ran four reactions in parallel, each with a different initial feed ratio (1.0:1.2, 1.0:1.5, 1.0:1.8, and 1.0:2.0). After the first batch of reactions, using yield as the optimization objective, a Bayesian optimization algorithm was used to recommend the optimal feed ratio for the second batch. After three iterations, the yield increased from the initial 62% to 84%.
[0052] To achieve efficient cross-batch autonomous optimization, the Bayesian optimization algorithm in the automatic convergence optimization module follows the following mathematical logic: 1. Define the optimization space and objective Parameter space: Assume the optimization variable is the molar ratio x of reactants A and B, with a value range of [0.8, 2.2].
[0053] Objective function: target product yield f(x) Our goal is to find x such that f(x) maximize.
[0054] 2. Initialization (First Batch) The system selects four initial points (e.g., x = 1.0, 1.3, 1.6, 2.0) in the parameter space using Latin hypercube sampling, and runs four sets of reactions in parallel to obtain initial yield data. y1,y2,y3,y4 .
[0055] 3. Iterative optimization (core loop): Building a surrogate model: utilizing existing observation data (x i, yi) The system constructs a Gaussian process regression model. This model can predict any new point. x new average yield μ(x) It can also provide the uncertainty (variance) of the prediction. σ(x) Prediction uncertainty is small near the observed data point and large in areas far from the observation point.
[0056] Maximizing the acquisition function: The system does not simply select... μ(x) Instead of focusing on the largest possible value, the goal is to construct and maximize a collection function (such as Expected Improvement). This function balances "exploitation" (selecting regions with high mean) and "exploration" (selecting regions with high uncertainty).
[0057] For example, the formula is: EI(x)=E[max(f(x)-f best This means that conducting the experiment at x will yield a yield exceeding the current best yield f. best What is the expected value? The system will recommend the next option. EI(x) Largest experimental point x next .
[0058] Execution and Update: The system executes the next batch of reactions according to the recommended parameters to obtain the new yield. y next .Will (x next , y next ) The observed dataset was added, and the Gaussian process regression model was updated to provide a more accurate understanding of the parameter space.
[0059] Convergence Judgment: Repeat the above process. When the recommended experimental points do not change significantly over multiple rounds, and the maximum value of the acquisition function approaches 0, the algorithm is considered to have converged to the optimal solution. In this patent embodiment, after 3 iterations (using a total of 4 + 3 * 4 = 16 batches of data), the yield increased from 62% to 84%, proving the effectiveness of the algorithm.
[0060] Example 2: The difference between this embodiment and Embodiment 1 is that the in-situ spectral detection module also integrates a Raman spectrometer to monitor the phase changes of the solid catalyst in the reaction system. The AI real-time analysis module integrates infrared and Raman spectral data to improve the accuracy of byproduct identification.
[0061] The dynamic intervention control module adds the stirring rate as an adjustable parameter. When catalyst agglomeration is detected, which causes a decrease in the reaction rate, the stirring rate is automatically increased by 200 rpm and maintained for 30 seconds before returning to the original value.
[0062] System configuration example Experimental results show that, within the optimal parameter range, the system's intervention response time (from detecting the abnormality to executing the adjustment) is less than 5 seconds, the byproduct inhibition efficiency reaches more than 75%, and cross-batch optimization increases the average yield by 20-35%.
[0063] Methods for determining and adjusting the optimal range of system parameters: The optimal parameters listed in Table 1 (such as a sampling interval of 30 seconds and a threshold of 0.05 mmol / L·s) are not arbitrarily set, but are obtained through a systematic "time-domain response analysis method" combined with specific reaction kinetic characteristics. The determination method is as follows: Determining the spectral acquisition interval: For the target reaction, monitoring is first performed at the highest frequency (e.g., 5 seconds / time), and the dominant frequency of byproduct formation rate fluctuations is analyzed using Fourier transform. According to the Nyquist sampling theorem, the acquisition interval is set to half of this dominant frequency period (e.g., if the dominant period of byproduct fluctuations is 60 seconds, the acquisition interval is set to 30 seconds), ensuring that data redundancy and computational load are reduced without losing key dynamic information.
[0064] Determining the intervention threshold: This is done by analyzing the critical points between "reversible side reactions" and "irreversible byproduct accumulation" in historical data. In multiple batches of reactions, it is statistically analyzed whether conventional interventions (such as cooling) can reverse the rate of byproducts once it exceeds a certain value. The maximum initial byproduct rate corresponding to 80% of successful intervention cases is set as the warning threshold; the rate point with an intervention success rate below 50% is set as the high-risk threshold.
[0065] When the reaction system changes, those skilled in the art can quickly adapt to the new system by following the standardized process described in this invention, based on the teachings of this invention: 1. Kinetic preliminary experiment: Conduct 2-3 sets of "open-loop" ramp-up experiments (such as linear heating or linear accelerated dropping) to induce side reactions; 2. Coarse parameter adjustment: Using the preliminary experimental data, the sampling frequency and threshold range of the new system are initially estimated through the above-mentioned "time domain response analysis method"; 3. Model cold start: Use pre-experimental data to quickly fine-tune the pre-trained deep learning model (train only the last few layers of the network), so that it can initially identify the feature peaks of the new system within 1-2 days; 4. Adaptive Adjustment: The system enters semi-supervised operation mode. Initially, AI intervention suggestions require operator confirmation. The system automatically adjusts the thresholds based on the manual confirmation. Typically, after 3-5 batches, the system can stably operate within the optimal parameter range of the new system.
[0066] Through the detailed description of the steps and disclosure of the parameter adjustment methods above, those skilled in the art can apply the present invention to different chemical reaction systems without creative effort, which fully demonstrates the practicality and feasibility of the present invention.
[0067] After the system starts, it will run automatically according to the following process: 1. Initialization: The operator inputs the target reactant, expected yield target, and safety threshold range; 2. Reaction Start-up: The reaction execution module starts the reaction according to the initial parameters, and the in-situ spectral detection module begins real-time acquisition; 3. AI Real-time Analysis: The conversion rate curve and by-product rate are updated every 30 seconds, and the data is displayed in real time on the human-machine interface; 4. Dynamic decision-making: If the byproduct rate exceeds the limit, the intervention logic is triggered; if there is no abnormality, the system continues to monitor. 5. Batch End: Calculate the yield of this batch and transfer the data to the automatic convergence optimization module; 6. Cross-batch optimization: After all parallel experiments are completed, update the optimization model and recommend the feed ratio for the next round; 7. Iterative cycle: Repeat steps 2-6 until the preset optimization goal is achieved or the operator terminates the process.
[0068] In summary, by utilizing the technical solutions described above, the AI real-time analysis module performs real-time inference and trend prediction on in-situ spectral data, enabling the dynamic intervention control module to automatically execute multi-parameter collaborative intervention before or at the initial stage of abnormal by-product generation. This transforms traditional open-loop execution into closed-loop autonomous regulation involving monitoring, analysis, decision-making, and control, effectively suppressing side reactions and significantly improving the yield of the target product. By incorporating temperature, dropping rate, and stirring rate into a unified intervention logic, the intervention effect of multi-parameter collaborative regulation is superior to that of single-parameter regulation, achieving refined control of the reaction process. The automatic convergence optimization module uses Bayesian optimization or reinforcement learning algorithms to iteratively optimize multiple batches of parallel experimental data, enabling the system to possess cross-batch self-learning capabilities, continuously accumulating experience and autonomously optimizing the feeding ratio. This achieves an intelligent upgrade from passive execution to active optimization, significantly improving process development efficiency and product quality stability.
[0069] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A closed-loop autonomous optimization system for chemical synthesis based on in-situ spectral feedback, characterized in that, include: The reaction execution module is used to execute chemical reactions; The in-situ spectral detection module is used to acquire the spectral data of the reaction system in the reaction execution module in real time; The AI real-time analysis module is connected to the in-situ spectral detection module and has a built-in trained deep learning model to receive and analyze the spectral data in order to identify the reaction process and predict the generation trend of by-products in real time. A dynamic intervention control module, connected to the AI real-time analysis module and the reaction execution module, is used to automatically generate and execute adjustment instructions according to preset collaborative intervention logic when the AI real-time analysis module predicts an abnormal trend in by-product generation, thereby adjusting the reaction parameters of the reaction execution module; and The automatic convergence optimization module connects multiple sets of parallel-running reaction execution modules to collect yield data for each batch of reaction and automatically optimize the reaction conditions for subsequent batches based on the optimization algorithm, thereby achieving cross-batch self-learning optimization.
2. The closed-loop autonomous optimization system for chemical synthesis based on in-situ spectral feedback according to claim 1, characterized in that, The reaction execution module includes a reaction vessel, a temperature control system, a material dropping system, and a stirring system; the reaction parameters include at least one of reaction temperature, material dropping rate, and stirring rate.
3. The closed-loop autonomous optimization system for chemical synthesis based on in-situ spectral feedback according to claim 1, characterized in that, The in-situ spectral detection module includes at least one of an in-situ infrared spectrometer and a Raman spectrometer, with its detection probe immersed in the reaction system; the spectral data includes infrared spectral data and Raman spectral data.
4. The closed-loop autonomous optimization system for chemical synthesis based on in-situ spectral feedback according to claim 1, characterized in that, The deep learning model built into the AI real-time analysis module is a combination of convolutional neural networks and recurrent neural networks, used to extract spectral features and predict temporal changes; the output of the AI real-time analysis module includes target product concentration, conversion rate, by-product generation rate, and by-product generation trend prediction.
5. The closed-loop autonomous optimization system for chemical synthesis based on in-situ spectral feedback according to claim 1, characterized in that, The collaborative intervention logic of the dynamic intervention control module includes: when the by-product generation rate exceeds a preset threshold, the reaction temperature is first reduced; if the by-product generation rate does not decrease after temperature adjustment, the material dropping rate is reduced; when catalyst sedimentation or agglomeration is detected, the stirring rate is increased.
6. The closed-loop autonomous optimization system for chemical synthesis based on in-situ spectral feedback according to claim 1, characterized in that, The automatic convergence optimization module employs either a Bayesian optimization algorithm or a reinforcement learning algorithm. The Bayesian optimization algorithm constructs a Gaussian process regression model and maximizes the acquisition function to recommend response conditions for the next round of experiments.
7. The closed-loop autonomous optimization system for chemical synthesis based on in-situ spectral feedback according to claim 1, characterized in that, The AI real-time analysis module is also used to predict the by-product generation trend in the future. When it is predicted that the by-product generation rate will exceed a preset threshold, the dynamic intervention control module executes the adjustment command in advance. The deep learning model in the AI real-time analysis module is pre-trained through transfer learning and fine-tuned using experimental data of a specific reaction system. The system also has a model iteration update function, which is used to incrementally train the model based on newly generated experimental data.
8. The closed-loop autonomous optimization system for chemical synthesis based on in-situ spectral feedback according to claim 1, characterized in that, The system also includes a human-computer interaction interface for displaying real-time analyzed response curves, historical intervention records and optimization suggestions, and allows operators to manually adjust intervention thresholds and optimization parameters.
9. The closed-loop autonomous optimization system for chemical synthesis based on in-situ spectral feedback according to claim 1, characterized in that, The system's parameter configuration is determined through time-domain response analysis, specifically including: analyzing the dominant frequency of byproduct generation rate fluctuations using Fourier transform, setting the spectral acquisition interval based on the Nyquist sampling theorem; and statistically analyzing historical intervention success rates to set early warning thresholds and high-risk thresholds.
10. A closed-loop autonomous optimization method for chemical synthesis based on in-situ spectral feedback, applied to the system described in any one of claims 1-9, characterized in that, Includes the following steps: S1: Initialize and set the basic reaction parameters and optimization objectives; S2: Real-time acquisition of multimodal spectral data of the reaction system; S3: Utilize a trained deep learning model to analyze the spectral data, identify the reaction process in real time, and predict the generation trend of by-products; S4: When an abnormal trend in byproduct formation is predicted, an adjustment instruction is automatically generated and executed based on the synergistic intervention logic to adjust the reaction parameters; S5: After the batch is completed, calculate the yield of this batch and archive the data; S6: After several batches are completed, the reaction conditions of subsequent batches are automatically optimized based on the yield data of each batch using the optimization algorithm, so as to achieve self-learning optimization across batches.
Citation Information
Patent Citations
Textile design optimization method based on artificial intelligence
CN120893330A
Bi-BDO fermentation pH dissolved oxygen dynamic optimization method based on online Raman spectrum
CN121034441A
Process and device for continuously synthesizing vitamin D3 through photocatalysis
CN121314494A
