An automated synthesis and testing system and method for rapid development of membrane materials

Abstract

The application discloses a kind of rapid development membrane material's automated synthesis test system and method.The application includes automated synthesis subsystem, automated test subsystem, machine learning optimization module, in addition, it is equipped with data mining module, handling subsystem, automated control module;Method is in data mining module gives initial formula, process parameter, input into automated synthesis subsystem and is synthesized to prepare membrane material sample, the performance of membrane material sample is tested by automated test subsystem, machine learning optimization module is given more optimal formula process parameter using analysis algorithm reasoning according to performance test result, and is input into automated synthesis subsystem and is synthesized to prepare;Constantly cyclic iteration operation is implemented, and the automated synthesis test of membrane material is realized.The application simultaneously integrates synthesis function and test function to the production of membrane material, also links synthesis function and test function together closely combined feedback, realizes the efficient, optimization of membrane material production.

Description

Technical Field

[0001] This invention relates to the field of membrane material preparation and testing, and in particular to an automated synthesis and testing system and method for rapid development of membrane materials. Background Technology

[0002] Currently, materials science research and development mainly relies on experience-based experimental design and manual experiments. However, manual experiments still have some limitations and drawbacks in materials research and development. First, manual experiments require significant time and resources, and often necessitate multiple experiments to obtain reliable results. Second, manual experiments are susceptible to human factors, such as the instability of experimental procedures and the subjective biases of the experimenter. Furthermore, manual experiments often only cover a limited range of material combinations and experimental conditions, making it difficult to comprehensively explore the properties and behaviors of materials. Finally, manual experiments often provide only a limited amount of data, which is insufficient to meet the demands of large-scale data analysis and machine learning technologies. Materials research and development involves numerous variables; relying solely on limited manual experiments and experience-based experimental design makes it difficult to find ideal material formulations in a large space with high variable dimensions. Summary of the Invention

[0003] To address the problems existing in the background technology, the present invention provides an automated synthesis and testing system and method for rapid research and development of membrane materials. It integrates synthesis and testing functions for membrane material production, and links the synthesis and testing functions together for close feedback, thereby achieving high efficiency and optimization in membrane material production.

[0004] The technical solution adopted in this invention is:

[0005] I. An automated synthesis and testing system for rapid research and development of membrane materials:

[0006] This invention proposes an automated synthesis and testing system for rapid research and development of membrane materials, comprising an automated synthesis subsystem for membrane material samples (hereinafter referred to as the automated synthesis subsystem), an automated testing subsystem for membrane material samples (hereinafter referred to as the automated testing subsystem), a data mining module for initial formulation and process parameter design of membrane materials (hereinafter referred to as the data mining module), a machine learning optimization module for formulation and process parameter optimization (hereinafter referred to as the machine learning optimization module), a membrane material sample handling subsystem (hereinafter referred to as the handling subsystem), and an automated control module with a front-end interface for automated driving and operational status monitoring of the entire platform (hereinafter referred to as the automated driving module).

[0007] The automated synthesis subsystem communicates with the data mining module. Both the automated synthesis subsystem and the automated testing subsystem communicate with the machine learning optimization module. The automated driving module is specifically a communication control mechanism that connects to the automated synthesis subsystem, automated testing subsystem, transport subsystem, data mining module, and machine learning optimization module through communication.

[0008] In practice, the automated synthesis subsystem, the automated testing subsystem, and the machine learning optimization module are first connected into a closed loop.

[0009] The entire system includes:

[0010] The automated synthesis subsystem is used for the automated synthesis of membrane material samples according to the formula and process parameters; the automated synthesis subsystem can prepare membrane material samples according to the specified formula and process parameters.

[0011] The automated testing subsystem is used for automated testing of membrane material samples synthesized by the automated synthesis subsystem.

[0012] The machine learning optimization module is used to optimize and update the formulation and process parameters of membrane materials during automated synthesis.

[0013] And also includes:

[0014] The data mining module is used to generate the initial formula and process parameters when the membrane material is synthesized in an automated manner, and to send the initial formula and process parameters to the automated synthesis subsystem.

[0015] The transport subsystem is used for transporting membrane material samples during synthesis and testing; specifically, it is used for transporting samples during the synthesis process within the automated synthesis subsystem and during the testing process from the automated synthesis subsystem to the automated testing subsystem.

[0016] The automation drive module is used for automated drive control and operational status monitoring of the automated synthesis subsystem and the automated testing subsystem, and is displayed through the front-end interface.

[0017] The automated synthesis subsystem comprises three modules: an automated solution preparation and powder distribution module, an automated spin coating module, and an automated annealing module.

[0018] The three modules each implement a corresponding function:

[0019] The automated solution preparation and powder distribution module is used to prepare the original membrane material solution by mixing the raw materials (powder or solution) according to the formula by mass / volume.

[0020] An automated spin coating module is used to coat the prepared membrane material stock solution onto the substrate surface of the membrane material.

[0021] The automated annealing module is used to perform heating / cooling annealing operations on a substrate with a solution coated on its surface to obtain a film material sample.

[0022] The membrane material stock solution obtained by the automated liquid preparation and powder distribution module is transported to the automated spin coating module for coating by the transport subsystem. The substrate obtained by coating by the automated spin coating module is transported to the automated annealing module for annealing by the transport subsystem.

[0023] The automated synthesis subsystem described above can flexibly adjust the formula and process parameters in real time when preparing membrane material samples.

[0024] The machine learning optimization module includes analysis algorithm software, which is mainly based on the Bayesian optimization algorithm framework. By improving the form of the acquisition function and replacing it with a neural network, the module constructs a simulated experimental environment based on different datasets and different machine learning models. The neural network is trained using reinforcement learning methods, thereby obtaining an analysis algorithm that integrates information contained in the prior data and has higher sampling efficiency, achieving rapid optimization within a specified number of iterations.

[0025] Specifically, the machine learning optimization module includes:

[0026] 1) First, define the domain, range, and constraints of the membrane material sample optimization problem to form a configuration file;

[0027] The defined domain is the adjustable space of the formulation and process parameters during the synthesis of membrane material samples;

[0028] The value range refers to the adjustable range of test performance parameters during the synthesis of membrane material samples;

[0029] The constraints mentioned are a priori conditions, such as: "The heating temperature should not exceed 200℃, otherwise the material will decompose" and "The rotation speed of the spin coater should be in the range of 3000 to 5000 RPM" and other basic conditions that limit the formulation and process parameters.

[0030] 2) Using the configuration file as input, a combination of parameters from one or a batch of membrane material samples is sampled in the domain of the configuration file using a random sampling algorithm. These parameters are used as initial configuration parameters and then transmitted to the automated synthesis subsystem.

[0031] 3) The automated synthesis subsystem uses the received initial configuration parameters as input X to synthesize and prepare membrane material samples. Then, the automated testing subsystem tests the prepared membrane material samples to obtain the target performance parameters Y corresponding to the initial configuration parameters, thereby obtaining a dataset {X,Y} of configuration parameters and performance parameters.

[0032] 4) For the current dataset {X,Y} of configuration parameters and performance parameters, use the surrogate probability model as the black box function approximation to obtain the posterior distribution of the black box function. Then, use the posterior distribution of the black box function to predict any point in the domain to obtain the expectation and uncertainty of the black box function. The expectation is the value in the corresponding range of any point, and the uncertainty is the variance.

[0033] 5) Based on the expected value and uncertainty predicted by the posterior distribution of the black-box function, different acquisition functions are used to collect points in the domain. The collected points in the domain are evaluated based on the expected value and uncertainty predicted by the posterior distribution of the black-box function. The evaluation is to the extent that the algorithm can approach the global optimum of the black-box function.

[0034] 6) Based on the different performance parameters output by each acquisition function and corresponding to different values ​​in the domain, combine and transform them to obtain a batch of formula and process parameters in the domain, and send them to the automated synthesis subsystem and automated testing subsystem for synthesis and testing, to obtain a new dataset of configuration parameters and performance parameters {X',Y'}.

[0035] 7) Then merge the new dataset of configuration parameters and performance parameters {X',Y'} with the previously obtained dataset of configuration parameters and performance parameters, and send it back to the algorithm software, that is, return to step 4). Iterate through steps 4)-6) to optimize until the performance parameters of the membrane material sample reach the expected satisfactory value, or the best performance among all synthesized membrane material samples within a certain number of rounds no longer improves, then the iterative optimization ends.

[0036] In step 5), for a collection function, an optimization method is used to find a set of extreme values ​​of the collection function within the domain, and the extreme value most likely to make the black-box function achieve the global optimum and its corresponding performance parameters in the value domain are found.

[0037] The optimization method mentioned is, for example, a gradient-based method.

[0038] In the algorithm software of the machine learning optimization module of this invention, in order to minimize the number of optimization rounds (i.e., minimize the optimization cost) and make full use of the experimental data collected before executing the optimization task, the algorithm software introduces neural acquisition functions in the form of neural networks into the acquisition function set. The training of the neural acquisition functions is completed by the machine learning teaching model and PPO reinforcement learning algorithm obtained based on the existing experimental data.

[0039] The machine learning optimization module also includes optimizing the acquisition function in the following manner:

[0040] T1. Select data with corresponding domain and range from known public databases to form a batch of prior datasets {Dn};

[0041] T2. Divide the prior dataset according to confidence level;

[0042] T3. Using the prior dataset as input, a batch of different machine learning models are established. The teaching model is used as a surrogate probability model, and the hyperparameters are adjusted and selected using random or optimization algorithms. The prior dataset is input into different machine learning models for training. Different teaching models are obtained after training.

[0043] Training different machine learning models requires hyperparameters as input. For a given machine learning model, the algorithm software will use random or optimization algorithms to adjust and select hyperparameters to obtain the machine learning model trained under the optimal hyperparameters. The machine learning model achieves the lowest prediction error on the test set of the prior dataset.

[0044] T4. Then, a neural network is used as the neural acquisition function and combined with each teaching model to form different overall models. The PPO reinforcement learning algorithm is used for training to obtain the trained neural acquisition function, which then replaces at least one acquisition function in step 5).

[0045] Through the aforementioned algorithm processing, it is possible to iteratively guide the design of experimental parameters for robotic experimental systems using prior data, and to find the optimal material composition, process, and other parameters as quickly as possible and at the lowest cost.

[0046] The batch of machine learning models trained and optimized on different prior datasets were used as teaching models, with the domain of the black-box function as input and the range as output. They have a similar functional relationship to the black-box function for optimizing actual materials, and can quickly obtain the function value for any input. Therefore, they can be used as a low-cost alternative to the black-box function.

[0047] Based on this batch of black-box functions, a batch of teaching optimization tasks can be established by imitating the framework process of Bayesian optimization. These teaching optimization tasks use different teaching models as black-box functions and the same neural network as the neural acquisition function. The neural acquisition function takes the surrogate model's prediction of the function value of a certain point, variance, the function value of that point, and the current iteration number as inputs, and outputs the degree to which that point can be used as a recommendation parameter for the next round.

[0048] Based on this batch of teaching optimization tasks, a neural acquisition function is trained using reinforcement learning, taking the classic PPO algorithm as an example. The learned agent policy is the neural acquisition function, the environment / task is the teaching optimization task, the state is the input of the function, the behavior is the behavior of the acquisition function in a whole optimization iteration task, and the feedback function is a comprehensive function that integrates parameters such as the optimal material performance obtained after the optimization iteration task and the number of rounds to achieve this performance.

[0049] In a single PPO optimization step, the agent performs optimization tasks multiple times under different teaching optimization tasks with different hyperparameters and optimization budgets. The penalty or feedback values ​​obtained from these tasks are recorded, so a neural network can be used to fit the probability distribution of the policy function in the state-behavior. Based on this distribution, the expected reward of the current neural acquisition function policy in the environment can be calculated. By differentiating the expected reward with respect to the parameter values ​​of the neural acquisition function, the parameters of the neural acquisition function are updated using the gradient ascent method.

[0050] After PPO training is completed, it is believed that the obtained neural acquisition function has been trained for a long time on a teaching function similar to the actual black box function, and has learned the functional relationships and properties therein, that is, it has obtained the relationship law between material parameters and properties contained in the prior data.

[0051] Finally, the trained neural acquisition function will be put into the actual optimization task, participating in the actual BO optimization task described in the framework process (I), to process and optimize material parameters.

[0052] The automated testing subsystem includes a spectral testing module for testing the optical properties of the film material and / or an electrical testing module for testing the electrical properties of the film material, which can realize the testing and characterization of the optical / electrical properties of the film material sample synthesized by the automated synthesis subsystem.

[0053] The spectral testing module is used to test the optical properties of the film material. The main optical properties tested are absorption / transmission / reflection / fluorescence spectra.

[0054] The electrical testing module is used to test the electrical properties of membrane materials. The main electrical properties tested are resistivity, capacitance, and inductance.

[0055] The core function of the data mining module is to extract relevant parameters from literature using a material information retrieval system based on knowledge graphs and vector databases, providing a starting point for subsequent robot experiments. These parameters are then passed to the automated testing subsystem for execution.

[0056] The data mining module searches for publicly available literature information on the Internet, and analyzes and processes the literature information in conjunction with a large language model to obtain the initial membrane material formulation and process parameters.

[0057] The formulation includes solid / liquid raw materials, etc. Specifically, it refers to the type / weight / volume of the solution and powder used when preparing the original membrane material solution, and the process parameters refer to the membrane thickness and annealing temperature / time when preparing the membrane material sample.

[0058] The formula and process parameters are initially provided by the data mining module and subsequently optimized by the machine learning module.

[0059] The data mining module employs a material information retrieval method combining knowledge graphs and vector databases for analysis and processing, specifically:

[0060] S1. Search publicly available literature from internet databases and use algorithms to filter out literature related to membrane material preparation;

[0061] S2. From the abstracts of the selected documents, extract knowledge triples using a large language model, and perform node scoring and pruning operations to construct a knowledge graph;

[0062] S3. Again, using the large language model, search the knowledge graph for the top k articles most relevant to membrane material preparation;

[0063] S4. Convert the k searched articles into vector form and store them in a vector database;

[0064] S5. Utilize vector databases for query operations and interact with knowledge graphs to achieve deep retrieval.

[0065] The large language model mentioned above can be, for example, the GPT model.

[0066] The transport subsystem mainly includes a robotic arm and an actuator at the end of the robotic arm.

[0067] II. An automated synthesis and testing method for rapid development of membrane materials:

[0068] The method involves inputting initial formulations and process parameters from the data mining module into an automated synthesis subsystem to synthesize membrane material samples. An automated testing subsystem then tests the functionality and performance of these samples. A machine learning optimization module uses analytical algorithms to deduce better formulations and process parameters based on the performance test results, which are then input back into the automated synthesis subsystem to synthesize more membrane material samples. This iterative process is repeated continuously, achieving automated synthesis and testing of membrane materials. Therefore, this invention enables the rapid development of high-performance membrane material formulations and their corresponding preparation process parameters.

[0069] This invention provides a closed-loop approach for the rapid development of membrane materials based on synthesis, testing, and machine learning. Its key feature is that the entire system integrates three main functions: membrane material sample synthesis, performance testing, and machine learning-based formulation and process parameter optimization, forming a closed loop. This approach enables rapid iterative optimization of membrane materials. The synthesis subsystem is a robot-based automated synthesis system capable of performing common experimental operations such as raw material preparation, spin coating, and annealing. The performance testing subsystem is an automated testing system that detects the spectral response and electrical parameters of the membrane material. The formulation optimization module is a machine learning optimization module based on a modified Bayesian optimization algorithm. This module infers the formulation and synthesis process parameters of membrane materials that may produce better performance based on existing experimental data in position space and inputs them into the automated synthesis subsystem. These three subsystems form a closed loop, enabling rapid development of membrane materials.

[0070] The beneficial effects of this invention are:

[0071] This invention, based on high-throughput robotic experiments combined with machine learning for iterative optimization, offers the following advantages compared to traditional manual material formulation development:

[0072] High efficiency: Robots and high-throughput experiments can rapidly prepare, test, and characterize large numbers of material samples, thereby accelerating the materials research and development process.

[0073] Accuracy: Robotics and high-throughput experiments can provide high-precision, highly repeatable data, thereby improving the accuracy and reliability of materials research and development.

[0074] Automation: Robots and high-throughput experiments can automate experimental procedures, reducing human intervention and thus minimizing human error.

[0075] Combining machine learning: Machine learning can analyze large amounts of material data to discover relationships and patterns among materials, guiding material development pathways. Combined with the machine learning optimization software included in this invention, algorithms such as Bayesian optimization can be used to efficiently search for and infer optimal values ​​for material formulations and processes based on the large amounts of high-quality data generated from high-throughput experiments. This can help researchers discover new materials and optimize existing ones more quickly. Attached Figure Description

[0076] Figure 1 This is a flowchart of the system of the present invention;

[0077] Figure 2 This allows the robotic arm to switch between gripping and suction modes.

[0078] Figure 3 This is a schematic diagram of the spectral detection module;

[0079] Figure 4 This is a schematic diagram of different spectral detection states of the spectral detection module;

[0080] Figure 5 This is a schematic diagram of the automated driving software.

[0081] Figure 6 A diagram illustrating software optimization for machine learning;

[0082] Figure 7 This is a schematic diagram of the data mining module;

[0083] Figure 8 A schematic diagram of the machine learning optimization module.

[0084] Figure 2 In the middle: 1. Six-axis robotic arm; 2. Gripper; 3. Suction cup module; 3-1. Suction cup arm; 3-2. Suction cup base; 3-3. Fixed motor; 3-4. Air pump;

[0085] Figure 3 In the middle: 4. Light source; 5. Optical fiber; 6. Transmission and reflection measurement bracket; 7. Dual-channel optical path switcher; 8. Housing; 9. Spectrometer. Detailed Implementation

[0086] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0087] like Figure 1 As shown, the specific embodiments and implementation process of the present invention are as follows:

[0088] This invention comprises six modules / subsystems: a data mining module, an automated synthesis subsystem, an automated testing subsystem, a machine learning optimization module, and a transport subsystem. Its operation includes multiple stages: initial experimental design, solution preparation, spin coating, annealing, detection, and machine learning optimization.

[0089] The system operation process is as follows: Figure 1 As shown:

[0090] The various subsystems constituting this invention will be described separately below, and the entire implementation process of the invention will be explained step by step.

[0091] (A) Automated Synthesis Subsystem:

[0092] The automated synthesis subsystem comprises three modules: an automated solution preparation and powder distribution module, an automated spin coating module, and an automated annealing module.

[0093] The three modules each implement a corresponding function:

[0094] The automated solution preparation and powder distribution module is used to prepare the original membrane material solution by mixing the raw materials (powder or solution) according to the formula by mass / volume.

[0095] An automated spin coating module is used to coat the prepared membrane material stock solution onto the substrate surface of the membrane material.

[0096] The automated annealing module is used to perform heating / cooling annealing operations on a substrate with a solution coated on its surface to obtain a film material sample.

[0097] The membrane material stock solution obtained by the automated liquid preparation and powder distribution module is transported to the automated spin coating module for coating by the transport subsystem. The substrate obtained by coating by the automated spin coating module is transported to the automated annealing module for annealing by the transport subsystem.

[0098] This system uses the synthesis of perovskite thin film materials as an example, but it can also be extended to other thin film materials synthesized based on liquid / solid raw materials. The following raw materials are used in the synthesis of perovskite thin film materials:

[0099] (1) Raw materials:

[0100] Disposable plastic pipette tips are used in conjunction with pipette modules and are installed at the tip of the pipette module for drawing solutions.

[0101] The pipette module draws liquid feedstocks, including but not limited to: solutions used in the synthesis of membrane materials. This example includes solvents: DMF and DMSO in a 9:1 volume ratio.

[0102] Mixed liquids; solutions of MAI, FAI, PbI2, CsI, SnI2, MABr, FABr, CsBr, PbBr2, and SnBr2 dissolved in solvents. The specific types of liquid raw materials vary depending on the application and experimental requirements.

[0103] Solid raw materials include, but are not limited to, powdered raw materials used in the synthesis of membrane materials. This example includes organic / inorganic additives: urea, DIAI, caffeine, and PbCl2, which can be varied depending on the application and experimental requirements.

[0104] Substrate: A solid sheet used to support the thin film material. In this example, it is a clean glass slide 75mm long and 25mm wide. The substrate can be varied depending on the application and experimental requirements.

[0105] (2) Includes the following parts:

[0106] Tip holder: Contains disposable plastic pipette tips, which can be used and discarded by the pipette module. Used to aspirate liquid during pipetting.

[0107] Waste bin: Used to discard used disposable plastic gun heads, as well as other waste materials and liquids generated during experiments.

[0108] Multi-well plate: Used to hold / mix liquid feedstocks. This example uses a standard 96-well multi-well plate.

[0109] Headspace vials: Used for holding / mixing liquid and solid feedstocks. In this example, a 2mL headspace vial is used, with a filter paper membrane capped with a cross-shaped opening. During pipetting operations, the disposable plastic pipette tip can penetrate the filter paper membrane opening to draw solution below the liquid surface inside the headspace vial.

[0110] In the specific implementation, two types of headspace bottles were used, one for holding liquid raw materials and the other for liquid preparation operations. The difference between the two headspace bottles is whether or not they have a filter paper membrane with a cross-shaped opening.

[0111] Test tube rack: Used to place / secure headspace bottles.

[0112] Pipettor Module: This pipette module, along with the gripper, is mounted on the end effector of the robotic arm and features a controllable lifting mechanism. When the pipette module descends to a height lower than the gripper at the end effector, the robotic arm gains pipetting functionality and can access objects on the platform that require interaction with the pipette module, such as pipette tip boxes, waste containers, headspace vials, and multi-well plates.

[0113] An automated pipette with controllable aspiration rate (µL / s) and aspiration volume (µL) that connects to a host computer and can be controlled by the host computer. It includes a movable metal sleeve that can be pressed down to discard the disposable plastic tip attached to the pipette module.

[0114] Powder distribution module: This module is used to add a specified weight of powdered solid raw material into the headspace bottle.

[0115] Spin coater: Used to coat the prepared film material stock solution onto the substrate surface.

[0116] The spin coater in this case study consists of a main spin coater unit with an absolute encoder mounted on its rotating shaft. This encoder transmits the current angle of the spin coater's turntable, allowing the calculation of the current angle of the substrate slide. This angle is fed back to a six-axis robotic arm, enabling the arm to pick up the spin-coated substrate slide. The spin coater in this case study features an automatically opening and closing flip cover. The flip cover is driven by a stepper motor. A 3cm diameter circular hole is located in the center of the flip cover.

[0117] During the spin coating process, after the spin coater turntable starts rotating, the pipette module can drop the original film material solution onto the substrate slide adsorbed on the spin coater turntable through the hole. The spin coater turntable has a hole connected to the vacuum pump of the spin coater. When the vacuum pump is started, it can tightly adsorb the substrate slide placed on the surface of the spin coater turntable, preventing it from falling off during spin coating.

[0118] Heating stage: A heating stage whose temperature can be adjusted via a knob. Used in annealing operations to heat a glass slide whose surface is coated with a thin film of the original solution.

[0119] Cold stage: A 1cm thick metal plate used to cool heated substrate slides during annealing.

[0120] Sample holder: Used for batch storage of clean substrate glass slides and synthesized-tested membrane material samples.

[0121] (B) Automated testing system:

[0122] The automated testing subsystem includes a spectral testing module for testing the optical properties of film materials and / or an electrical testing module for testing the electrical properties of film materials, enabling the testing and characterization of the optical / electrical properties of film material samples synthesized by the automated synthesis subsystem.

[0123] The spectral testing module is used to test the optical properties of film materials. The main optical properties tested are absorption, transmission, reflection, and fluorescence spectra.

[0124] The electrical testing module is used to test the electrical properties of membrane materials. The main electrical properties tested are resistivity, capacitance, and inductance.

[0125] The membrane material samples obtained by the automated annealing module of the automated synthesis subsystem are transported to the automated testing subsystem via the transport subsystem for testing of optical and electrical properties.

[0126] like Figure 3 As shown, this system detects the spectral / electrical response of membrane material samples to obtain the performance parameters of the membrane material samples synthesized in the automated synthesis subsystem. The spectral response detection is achieved through a spectral detection module, which includes:

[0127] Light source 4: The light source used to provide spectral detection. In this application case, a deuterium-tungsten lamp is used as the light source, with a wavelength range of 190-400nm (deep ultraviolet deuterium lamp) and 360-2000nm (halogen tungsten lamp).

[0128] Fiber 5: Used to build optical paths;

[0129] Transmission and reflection measurement bracket 6: used to build the optical path;

[0130] Dual-channel optical path switcher 7: used for switching transmission / reflection / excitation optical paths for different wavelength light sources;

[0131] Outer shell 8: Blocks out external stray light, creating a dark chamber. Mounted on a transmittance and reflectance measurement bracket, with a horizontal slit on the surface into which the membrane material sample to be tested can be inserted.

[0132] Spectrometer 9: Used to detect transmission / reflection / fluorescence spectra.

[0133] Specifically, a membrane material sample is placed inside the outer casing 8, and the detector end of the spectrometer 9 is connected to the outer casing 8 via optical fiber 6.

[0134] A light source 4 is introduced to the upper side of the membrane material sample via an optical fiber 6 to emit light from the upper side onto the membrane material sample to excite fluorescence. The fluorescence is then received by the spectrometer 9 via the optical fiber 6 of the spectrometer 9 to obtain the fluorescence spectrum.

[0135] Another light source 4 is connected to the input of a dual-channel optical path switcher 7, which has two outputs:

[0136] One output of the dual-channel optical path switcher 7 is introduced above the membrane material sample via optical fiber 6, and is used to emit light from above onto the membrane material sample and then reflect it. The reflected light is then incident on the spectrometer 9 via optical fiber 6 and received to obtain the reflection spectrum.

[0137] The other output of the dual-channel optical path switcher 7 is introduced to the bottom of the membrane material sample via optical fiber 6, and is used to emit light from below onto the membrane material sample and then transmit it. The light transmitted through the membrane material sample is then incident on the spectrometer 9 via optical fiber 6 and received to obtain the transmission spectrum.

[0138] like Figure 4 As shown, in the specific implementation, two light sources are set. One light source is guided by an optical fiber to the upper side of the membrane material sample for excitation light path. The other light source is connected to the input end of the dual-channel optical path switcher by an optical fiber. One output end of the dual-channel optical path switcher is guided by an optical fiber to the upper side of the membrane material sample for emission light path. The other output end of the dual-channel optical path switcher is guided by an optical fiber to the lower side of the membrane material sample for transmission light path.

[0139] The electrical response detection is achieved through an electrical detection module, which includes:

[0140] Four-probe instrument: used to test the resistivity, dielectric constant and other properties of film material samples;

[0141] Probe: The probe is designed for use with a four-probe instrument and has four raised, parallel-arranged electrodes on its surface.

[0142] Support: Used to hold the probe.

[0143] (C) Transport Subsystem:

[0144] The transport subsystem's main function is to transport membrane material samples to the required locations during the operation of the automated synthesis subsystem and the automated testing subsystem.

[0145] The handling subsystem mainly consists of a robotic arm and an actuator at the end of the robotic arm. The actuator grips / picks up substrate or membrane material samples and transfers them through the movement of the robot.

[0146] The automated synthesis subsystem is used in conjunction with the transportation subsystem. For example... Figure 2 As shown, the handling subsystem of this invention includes a six-axis robotic arm 1, a gripper 2, and a suction cup module 3.

[0147] Pipette module: Installed at the end of a six-axis robotic arm.

[0148] Grippers: are mounted on the end of a six-axis robotic arm.

[0149] The suction cup module includes a suction cup arm 3-1, a suction cup base 3-2, a fixed motor 3-3, a microcontroller, an air pump 3-4, and an air hose. The suction cup module operates as follows: the suction cup is mounted on the suction cup arm 3-1, and the air hose connects the air pump 3-4 to the suction cup. The microcontroller controls the start and stop of the air pump 3-4. When the air pump 3-4 is running, the suction cup draws air to pick up the substrate slide. When the air pump 3-4 stops, the suction cup stops drawing air and releases the substrate slide. When not in use, the suction cup arm 3-1 is placed on the suction cup base 3-2. The fixed arm of the fixed motor 3-3 rotates and presses down on the surface of the suction cup arm 3-1, locking the suction cup arm 3-1 onto the suction cup base 3-2. In use, the robotic arm moves to the washing module position, the gripper at the end of the robotic arm clamps the suction cup arm 3-1, the fixed motor 3-3 rotates to unlock, the gripper holds the suction cup arm 3-1, and the robotic arm gains the ability to pick up objects. In this case, the robotic arm uses the suction cup arm 3-1 to pick up the substrate glass slide and move it between various functional modules.

[0150] In the handling subsystem, the robotic arm can switch between gripping and suction modes. The six-axis robotic arm is equipped with a gripper at its end, and is normally in gripping mode.

[0151] During the switching process, firstly, the gripper at the end of the six-axis robotic arm moves to the suction cup module position; then, the gripper grasps the suction cup arm 3-1 of the suction cup module, with the grasping position shown as the "grabbing point" in the diagram; next, the fixed motor 3-3 rotates to unlock the suction cup arm 3-1; then, the gripper at the end of the six-axis robotic arm lifts the suction cup arm 3-1, at which point the suction cup arm 3-1, the six-axis robotic arm, and the gripper are considered as a single unit, and the six-axis robotic arm switches to suction mode. The entire switching process is as follows: Figure 2 As shown.

[0152] (D) Automation Drive Module:

[0153] The composition and structure of the automation drive module in the system are as follows: Figure 5 As shown.

[0154] The automation drive module includes control modules that drive the robots and automated equipment in the above system to acquire status information, perform relevant experimental operations, and respond to instructions from the host computer.

[0155] The automation-driven module is mainly divided into the driver layer, service layer, application layer, and front-end interactive interface. This module system ensures the coordinated operation of each subsystem and the stable operation of the entire system.

[0156] Driver layer: Receives instructions and parameters from the upper layer, controls hardware devices, executes relevant operations supported by the hardware; or queries relevant status information of the device from within the hardware.

[0157] Service layer: Establishes communication with the device's driver layer through the communication protocol defined by the driver layer (such as MODBUS serial port protocol, MQTT protocol, etc.), and exposes the functions of the relevant communication interfaces as a unified HTTP service for upper-layer applications to call;

[0158] Application Layer: Manages the logical control of machine learning optimization, data analysis, and the entire synthesis and testing process; breaks down the operational steps required for the experiment into corresponding instructions for each device and issues and confirms these instructions; schedules, allocates, and records subtasks in the experimental process; enables the storage, retrieval, modification, and deletion of experimental parameters and test data in the database; and implements basic user management.

[0159] Front-end interactive interface: Provides a user-friendly graphical user interface that enables operators to monitor and control system operation, while also receiving suggestions from machine learning optimization.

[0160] (E) Data Mining Module:

[0161] The data mining module searches for publicly available literature information on the Internet, and analyzes and processes the literature information in conjunction with a large language model to obtain the initial membrane material formulation and process parameters.

[0162] The formulation includes solid / liquid raw materials, etc. Specifically, it refers to the type, weight, and volume of the solution and powder used when preparing the original membrane material solution. The process parameters refer to the membrane thickness and annealing temperature / time when preparing the membrane material sample.

[0163] The formula and process parameters are initially provided by the data mining module and subsequently optimized by the machine learning module.

[0164] The data mining module employs a material information retrieval method combining knowledge graphs and vector databases for analysis and processing. Specifically:

[0165] S1. Search publicly available literature from internet databases and use algorithms to filter out literature related to membrane material preparation;

[0166] S2. From the abstracts of the selected documents, extract knowledge triples using a large language model, and perform node scoring and pruning operations to construct a knowledge graph;

[0167] S3. Again, using the large language model, search the knowledge graph for the top k articles most relevant to membrane material preparation;

[0168] S4. Convert the k searched articles into vector form and store them in a vector database;

[0169] S5. Utilize vector databases for query operations and interact with knowledge graphs to achieve deep retrieval. Employ large language models to extract structured data from unstructured text, such as experimental substances, dosages, and operational procedures, and output the data to a file, for example, a .json file.

[0170] Large language models can be, for example, the GPT model.

[0171] The specific structure of the data mining module and the process for providing initial formula / process parameters are as follows: Figure 7 As shown:

[0172] The data mining module automatically acquires relevant literature information by interacting with the API interface of the literature database. This information includes, but is not limited to, key elements such as article titles, authors, and abstracts. To ensure the high quality of the information provided, the system further filters out highly cited literature. These highly cited documents typically represent core and authoritative content in the field, providing a solid foundation for subsequent knowledge graph construction and information extraction.

[0173] The data mining module utilizes large language models, such as the GPT series models, to automatically construct a knowledge graph. This knowledge graph not only contains basic information about the literature but also key knowledge points such as the relationships, sources, and properties of chemical substances. To ensure the accuracy and relevance of the knowledge graph, the system also performs scoring and pruning operations to ensure that only the content most relevant to the topic is retained.

[0174] The data mining module also employs techniques such as Contextual Compression and MultiQueryRetriever to deeply compress and extract document content. This allows for the extraction of the core parts of articles from multiple perspectives, storing the core content in a vector database. The vector database design makes query operations more efficient, enabling users to quickly find the information they need.

[0175] The data mining module utilizes a vector database for querying and interacts with a knowledge graph. This interactive approach not only extracts structured data but also enables deep retrieval, helping users find initial formulas and process parameters.

[0176] (F) Machine Learning Optimization Module:

[0177] In the machine learning optimization module, specifically:

[0178] 1) First, based on experience or by analyzing past literature and text data using a large language model, set the domain, range, and constraints of the black-box function for the membrane material sample optimization problem in JSON / YAML / XML formats to form the configuration file for the optimization task;

[0179] The domain is defined as the adjustable space of the formulation and process parameters during the synthesis of membrane material samples;

[0180] The value range represents the adjustable range of the test performance parameters during the synthesis of membrane material samples;

[0181] The constraints are some prior conditions based on the experience of the experimenters, such as: "The heating temperature should not exceed 200℃, otherwise the material will decompose" and "The rotation speed of the spin coater should be in the range of 3000 to 5000 RPM" and other basic conditions that limit the formulation and process parameters.

[0182] 2) The configuration file is processed by the algorithm software as input. The parameter combination of one or a batch of membrane material samples is sampled in the domain of the configuration file using a certain random sampling algorithm as the initial configuration parameters and transmitted to the automated synthesis subsystem.

[0183] 3) The automated synthesis subsystem uses the received initial configuration parameters as input X to synthesize and prepare membrane material samples. Then, the automated testing subsystem tests the prepared membrane material samples to obtain the target performance parameters Y corresponding to the initial configuration parameters, thereby obtaining a dataset {X,Y} of configuration parameters and performance parameters.

[0184] 4) For the current dataset {X,Y} of configuration parameters and performance parameters, the algorithm software uses a surrogate probability model (e.g., Gaussian process) as a black box function approximation to obtain the posterior distribution of the black box function. Then, it uses the posterior distribution of the black box function to predict any point in the domain to obtain the sum and uncertainty of the black box function. The uncertainty is the variance, where any point corresponds to a value in the range.

[0185] 5) Based on the expected value and uncertainty predicted by the posterior distribution of the black-box function, different acquisition functions are used to collect points in the domain. The collected points in the domain are evaluated based on the expected value and uncertainty predicted by the posterior distribution of the black-box function. The evaluation is to the extent that the algorithm can approach the global optimum of the black-box function.

[0186] In step 5), for a given acquisition function, an optimization method is used to find a set of extrema of the acquisition function within its domain, and from these, the extrema most likely to make the black-box function achieve a global optimum and its corresponding performance parameters in the range are identified. Optimization methods, such as gradient-based methods, are used.

[0187] 6) Based on the different performance parameters output by each acquisition function and corresponding to different values ​​in the domain, combine and transform them to obtain a batch of formula and process parameters in the domain, and send them to the automated synthesis subsystem and automated testing subsystem for synthesis and testing, to obtain a new dataset of configuration parameters and performance parameters {X',Y'}.

[0188] 7) Then merge the new dataset of configuration parameters and performance parameters {X',Y'} with the previously obtained dataset of configuration parameters and performance parameters, and send it back to the algorithm software, that is, return to step 4). Iterate through steps 4)-6) to optimize until the performance parameters of the membrane material sample reach the expected satisfactory value, or the best performance among all synthesized membrane material samples within a certain number of rounds no longer improves, then the iterative optimization ends.

[0189] The machine learning optimization module also includes optimizing the acquisition function in the following ways:

[0190] T1. Based on the black-box function optimization task, select data with corresponding domain and range from known public databases, such as those collected from literature, obtained from simulation calculations, and obtained from experiments, and use them as a batch of prior datasets {Dn}.

[0191] T2. The prior dataset has different confidence levels. The prior dataset is divided according to the confidence level and then input into the algorithm software.

[0192] During training, weights are set, with higher weights assigned to higher confidence levels.

[0193] T3. The algorithm software takes a certain prior dataset as input, establishes a batch of different machine learning models, takes the domain and range as input and output, takes the teaching model as the surrogate probability model, and uses random or optimization algorithms to adjust and select hyperparameters. The prior dataset is input into different machine learning models for training, and different teaching models are obtained after training.

[0194] Training different machine learning models requires hyperparameters as input. For a given machine learning model, the algorithm software will use random or optimization algorithms to adjust and select hyperparameters to obtain the machine learning model trained under the optimal hyperparameters. The machine learning model achieves the lowest prediction error on the test set of the prior dataset.

[0195] T4. Then, a neural network is used as the neural acquisition function and combined with each teaching model to form different overall models. The PPO reinforcement learning algorithm is used for training to obtain the trained neural acquisition function, which then replaces at least one acquisition function in step 5).

[0196] Through the aforementioned algorithm processing, it is possible to iteratively guide the design of experimental parameters for robotic experimental systems using prior data, and to find the optimal material composition, process, and other parameters as quickly as possible and at the lowest cost.

[0197] The batch of machine learning models trained and optimized on different prior datasets were used as teaching models, with the domain of the black-box function as input and the range as output. They have a similar functional relationship to the black-box function for optimizing actual materials, and can quickly obtain the function value for any input. Therefore, they can be used as a low-cost alternative to the black-box function.

[0198] The specific implementation of the machine learning optimization module can be structured as follows: Figure 6 As shown, the system includes a prior dataset composed of experimental data obtained through iterative optimization by the system itself, a prior knowledge base composed of external knowledge such as the experience of experimenters and published literature, and a sampling algorithm. The sampling algorithm predicts the recipe and parameters for the next round of experiments based on data inference from the prior dataset and prior knowledge base. After the experiment is executed by the robot platform, the new experimental data is stored in the prior dataset for inference in the next round of sampling algorithm.

[0199] The machine learning optimization module includes core algorithms that support the implementation of automated systems, including pipeline modules for material performance prediction, model selection, and model tuning based on machine learning models, a simulation experiment module based on machine learning models, and a module for recommending experimental parameters for iterative material optimization based on an active learning and Bayesian optimization framework.

[0200] Pipeline module for material property prediction, model selection, and model tuning based on machine learning models: This module enables automatic training, tuning, and model selection of machine learning models based on material parameters and performance databases, allowing for rapid training of material machine learning models; and supporting fast and accurate prediction of material properties.

[0201] Simulation Experiment Module Based on Machine Learning Model: A rapid simulation experiment method based on a materials machine learning model is used to test the robot's autonomous iterative optimization algorithm before formal material performance optimization;

[0202] The material iterative optimization experimental parameter recommendation module based on active learning and Bayesian optimization framework: Based on active learning and Bayesian optimization framework, it analyzes the currently observed data, explores the balance of uncertainty and mines the existing highest points, and realizes black-box optimization in a multi-objective, high-dimensional search space; it replaces the acquisition function in the Bayesian optimization framework with a neural network, and uses reinforcement learning to train the acquisition function of the neural network in the above-mentioned simulated experimental environment based on different datasets and different models, thereby obtaining an improved Bayesian optimization algorithm that integrates prior knowledge about material properties, and realizes fixed-point, rapid material development under a specified iteration budget;

[0203] During the iterative process, the parameters of the acquisition function in the multi-objective Bayesian optimization algorithm are adjusted so that: in the early iterations (1-N), the multi-objective Bayesian optimization algorithm tends to explore when generating the next recipe; in the late iterations (N+1), the multi-objective Bayesian optimization algorithm tends to mine when generating the next recipe. N is a preset threshold.

[0204] This module features include: comprehensive scoring of multiple performance parameters for various samples based on activation functions, and the ability to perform biased Bayesian optimization screening. For example, in the case of perovskite sample synthesis, this machine learning optimization module will apply different activation functions such as sigmoid to parameters such as film uniformity, conductivity, and bandgap, and then weight them to obtain a total score to characterize the "score" or performance of the film material sample. Simultaneously, by manually setting the weighting values ​​of various parameters, human bias can be added to the Bayesian optimization process. For example, increasing the weighting value of the conductivity parameter increases its proportion in the final total score, and the Bayesian optimization will tend to screen for formulations with better conductivity performance.

[0205] This invention treats uniformity as a logical parameter and electrical / optical performance as a numerical parameter. For the logical parameter, a Sigmoid function is used to binarize its score into 0 and 1, which is then multiplied by the numerical parameter value. This significantly reduces the score of samples with substandard logical parameters, while samples with satisfactory logical parameters have a smaller impact on the final score. In this way, the Sigmoid function addresses the difficulty in achieving good results in multi-objective optimization processes, specifically in the multi-objective optimization of membrane material performance and perovskite membrane material performance.

[0206] This invention designs and uses an activation function (Sigmoid) to transform multiple technical parameters of membrane materials. This transformation relationship is a numerical expression of the R&D personnel's experience, which can help the optimization algorithm understand the prior knowledge of the technical personnel in the membrane material R&D scenario. The transformed target value can be used for multi-objective evaluation and simultaneous optimization of membrane materials, and quickly develop membrane materials that meet the objectives.

[0207] The present invention avoids giving high scores to materials with substandard logical parameters but good other properties through the above-mentioned processing, and also avoids investing too many resources to optimize uniformity when the uniformity has already met the standard, thereby making the optimization process more rational and directional.

[0208] Examples of implementations of this invention are as follows:

[0209] (I) Preparation and Testing of Perovskite Samples

[0210] Take the preparation of a perovskite thin film sample with the following formula parameters: PbI2 1.1mol*50uL + MAI 2.2mol*100uL + urea 20mg, and the following process parameters: rotation speed 500rpm for 15 seconds, rotation speed 4000rpm for 45 seconds, and annealing temperature 140℃ as an example.

[0211] S1 Solution Preparation Stage:

[0212] S101. The pipette module is lowered by the lifting mechanism until it is below the gripper at the end of the robotic arm, at which point the robotic arm acquires the pipetting function.

[0213] S102. The robotic arm moves the pipette module to the tip box position, causing the pipette module to descend. The disposable plastic tip is pressed and fixed to the top of the pipette module by the pressure of the robotic arm.

[0214] S103. The robotic arm moves the pipette module to the headspace vial position. The height of the robotic arm end drops, the height of the pipette module drops, and the disposable plastic tip at the top pierces the cross opening of the headspace vial pre-filled with PbI2 stock solution and extends below the liquid surface inside the vial to extract the liquid.

[0215] After the pipette module aspirates 50 μL of solution, the pipette module rises, the robotic arm rises, and it detaches from the headspace bottle.

[0216] S104. The robotic arm moves to the empty headspace vial, and similar to the process in S103, the pipette module descends to the headspace vial opening with a disposable plastic pipette tip. The pipette module dispenses 50 μL of the original PbI2 solution.

[0217] S105. The robotic arm moves above the waste bin, and the metal sleeve on the top of the automatic pipette of the pipette module slides, squeezing out the disposable plastic pipette tip nested on the top of the automatic pipette. After detaching from the automatic pipette, it falls into the waste bin.

[0218] S106. Following the procedure in S102, the robotic arm goes to the gun head box to retrieve a clean disposable plastic gun head.

[0219] S107. The robotic arm moves to the headspace vial containing the MAI stock solution and aspirates 100 μL of the MAI stock solution, similar to the process in S103.

[0220] S108. The robotic arm moves to the headspace vial that has already been filled with 50 μL of PbI2 solution. Similar to the process in S103, the pipette module descends to the headspace vial opening with a disposable plastic pipette tip. The pipette module then discharges 100 μL of the original MAI solution.

[0221] S109. Following the procedure in S105, the robotic arm discharges the used disposable plastic gun head into the waste bin.

[0222] S110, the lifting mechanism drives the pipette module to rise above the end gripper of the robotic arm, at which point the robotic arm gains the gripping function.

[0223] S111: The robotic arm grasps the headspace vial containing PbI2 and MAI stock solution, and moves it to the powder distribution module via a test tube rack. The powder distribution module continuously adds urea powder to the headspace vial, and monitors the weight change in the headspace vial using an electronic scale at the bottom of the module, stopping when the weight reaches 20mg.

[0224] S112. The robotic arm picks up the headspace vial containing PbI2 stock solution, MAI stock solution, and urea powder raw material, and moves it to the test tube rack.

[0225] S113. Following the procedure in S102, the robotic arm goes to the gun head box to retrieve a clean disposable plastic gun head.

[0226] S114. Following the procedure in S101, the robotic arm acquires the pipetting function.

[0227] S115. The robotic arm moves the pipette module to the headspace vial containing PbI2 stock solution, MAI stock solution, and urea powder raw material. The robotic arm tip descends, lowering the pipette module, with the disposable plastic tip extending below the inner surface of the headspace vial. The pipette module repeats this process 10 times, drawing 50µL of liquid and discharging 50µL of liquid for recycling. A mixing operation is then performed; the mixed liquid is subsequently referred to as the membrane material stock solution.

[0228] S116. The pipette module aspirates 120 μL of the original membrane material solution, then the pipette module rises, and the robotic arm rises as well. At this point, the pipette module is higher than the end gripper of the robotic arm, and the pipette module contains 120 μL of the original membrane material solution.

[0229] S2: Solution preparation is now complete; proceed to the spin coating stage.

[0230] S21. The robotic arm moves to the suction cup module, descends, and the gripper grasps the suction cup arm of the suction cup module. At this point, the robotic arm, gripper, and suction cup module are fixedly connected, and the robotic arm gains suction capability. The fixing motor of the suction cup module rotates, unlocking the fixation of the suction cup arm. The robotic arm rises, and the suction cup arm detaches from the fixing base of the suction cup module.

[0231] S18. The robotic arm moves to the sample holder and picks up a clean substrate slide, starting from the top one.

[0232] S19. The host computer controls the opening of the spin coater's flip cover and reads the current angle value of the spin coater's turntable using an absolute encoder installed on the spin coater's rotating shaft.

[0233] S20. The robotic arm places the clean substrate glass slide on the spin coater turntable at the same angle as the current angle value of the spin coater turntable.

[0234] S21. The vacuum pump of the spin coater is running, evacuating air, and adsorbing and fixing the substrate glass slide onto the spin coater turntable.

[0235] S22, The host computer controls the opening and closing of the coating machine.

[0236] S23. The lifting mechanism lowers the pipette module below the gripper at the end of the robotic arm, at which point the robotic arm acquires the pipetting function. The disposable plastic pipette tip at the top of the automatic pipette in the pipetting module still contains the 120 μL original membrane material solution drawn in step (16). The disposable plastic pipette tip descends through the circular hole at the top of the spin coater's flip-top and hovers vertically 1 cm above the spin coater's turntable.

[0237] S24. According to the process parameters, the spin coater starts to rotate at a speed of 500 rpm, and the whole process lasts for 15 seconds.

[0238] S25. The pipetting module discharges the original solution of the membrane material at a rate of 30uL / s and drops it onto the surface of the substrate glass slide adsorbed on the turntable of the spin coater.

[0239] S26. After 15 seconds, the spin coater speed is increased to 4000 rpm and maintained for 45 seconds. During this process, the original solution of the membrane material is spin-coated onto the surface of the substrate glass slide.

[0240] S2 marks the end of the spin coating operation and the start of the annealing stage:

[0241] S21. The lifting mechanism drives the pipette module to rise above the suction cup arm held by the gripper at the end of the robotic arm, at which point the robotic arm acquires the suction function.

[0242] S22. The spin coater stops, and the flip cover opens automatically. The current angle value of the spin coater turntable is read through an absolute encoder. The robotic arm moves to the corresponding suction angle, descends, and the substrate glass slide (hereinafter referred to as the membrane material sample) with the original membrane material solution spin-coated on its surface is sucked up by the suction cup at the end of the suction cup arm.

[0243] After the suction cup picks up the sample, the vacuum pump of the spin coater stops running and the connection with the membrane material sample is released.

[0244] S23. The robotic arm that picks up the membrane material sample moves to the hot stage, places the membrane material sample on the surface of the hot stage, and anneals and heats the membrane material sample.

[0245] S24. After annealing and heating the membrane material sample for 30 minutes, the robotic arm picks up the heated membrane material sample and moves it to the surface of the cooling stage for cooling. After cooling for 5 minutes, the annealing operation is completed, and the membrane material sample synthesis is complete.

[0246] S3 annealing operation is now complete, proceeding to the testing phase:

[0247] S31. The robotic arm picks up the annealed film material sample from the surface of the cold stage and moves it to the optical inspection module.

[0248] S32. The robotic arm picks up the membrane material sample, inserts it into the gap on the surface of the light shield, and aligns the center point of the membrane material sample with the constructed detection optical path. Transmission / reflection / excitation signals are detected separately, and the spectral response results are input to the host computer.

[0249] S33. The robotic arm extracts the membrane material sample and moves it to the electrical detection module. The center of the sample is placed below the probe, and the robotic arm rises until the 4-pin electrode on the 4-probe of the 4-probe instrument contacts the surface of the sample. The 4-probe instrument reads the resistivity / dielectric constant of the membrane material sample and inputs it to the host computer.

[0250] S4 concludes the detection process. The detection results are then input into the machine learning optimization module to calculate potentially better-performing formulas.

[0251] S41. The robotic arm moves the membrane material sample to the sample holder and inserts it into the storage slot. The air pump stops running, and the suction cup releases the sample.

[0252] S42. The robotic arm moves the suction cup arm to the fixed base of the suction cup module. The gripper at the end of the robotic arm releases the suction cup arm, the fixed motor rotates to fix the suction cup arm, and the suction cup arm returns to its initial position.

[0253] S43. This concludes the process of preparing the solution, spin coating, testing, and storing a single membrane material sample.

[0254] (II) Process optimization flow for perovskite thin films:

[0255] Taking the fabrication of a MAPbI3 perovskite thin film as an example, we optimized the process parameters, including the amount of additives, the amount of back-extractant, the spin coating speed and time, and the annealing temperature and time, to obtain a perovskite thin film sample with the best crystallinity.

[0256] STEP 1: First, a (or batch of) formulation / process parameters are generated by a machine learning algorithm (randomly generated), or a data mining module summarizes relevant information from published literature online to obtain a (or batch of) formulation / process parameters. These parameters are then passed to the automation-driven module in the form of a configuration file. The automated system will synthesize samples and test their performance according to the perovskite sample preparation and testing case flow and the process parameters specified in the configuration file. Specifically, the performance of the membrane material in three aspects is evaluated separately, and then the score value of each performance is corrected by superimposing a nonlinear function to obtain a weighted score.

[0257] STEP2. After the performance test is completed, the module system records the test results in the database and requests the machine learning algorithm module for the next batch of process parameter recommendations.

[0258] STEP3: The machine learning algorithm accesses the database to obtain the recorded process parameters and performance data.

[0259] STEP4: The machine learning algorithm module uses the specified analysis algorithm to analyze the above data and recommend process parameters that may produce better performance in the next batch.

[0260] The analysis algorithm first uses a Gaussian process as an alternative statistical model to describe the prior distribution and assigns expected values ​​and confidence levels to all unevaluated data points. Then, a gain function is used to evaluate the prior distribution and propose recommendations. In each round, the bias towards mining (preferring higher score values) and exploration (preferring higher uncertainty) is adjusted by setting different gain function parameters, thus generating the next round of formulations. After several iterations, a perovskite thin film formulation with near-globally optimal performance is obtained.

[0261] STEP5: The machine learning algorithm module inputs this batch of process parameters into the automation drive module in the form of a configuration file, and optimizes iteratively according to steps (1-5).

[0262] During the iterative process, the parameters of the acquisition function in the machine learning optimization module are adjusted so that: in the earlier loops, the machine learning optimization module tends to explore when generating the next round of recipes; in the later loops, the machine learning optimization module tends to mine when generating the next round of recipes.

[0263] The timing of the iteration (early or late stage) is determined by a combination of factors including the search space, the number of search rounds, and the number of data points included in each round.

[0264] STEP 6: The machine learning optimization module determines whether the specified number of iterations has been reached or the specified material property target has been achieved, and then interrupts the optimization.

[0265] The stopping condition is defined as follows: (the variance of the scores of the N recipes is less than a preset variance threshold, and the average score of the N recipes is greater than a preset average threshold.)

[0266] STEP7: The machine learning optimization module outputs the process parameters with the best performance, and the optimization is complete.

Claims

1. An automated synthesis and testing system for rapid research and development of membrane materials, characterized in that: The system includes: An automated synthesis subsystem is used to prepare membrane material samples according to specified formulas and process parameters; An automated testing subsystem is used for automated testing of membrane material samples synthesized by the automated synthesis subsystem. The machine learning optimization module is used to optimize and update the formulation and process parameters of membrane materials during automated synthesis. Specifically, the machine learning optimization module includes: 1) First, define the domain, range, and constraints of the membrane material sample optimization problem to form a configuration file; The defined domain is the adjustable space of the formulation and process parameters during the synthesis of membrane material samples; The value range refers to the adjustable range of test performance parameters during the synthesis of membrane material samples; 2) Using the configuration file as input, a combination of parameters from one or a batch of membrane material samples is sampled in the domain of the configuration file using a random sampling algorithm. These parameters are then used as initial configuration parameters and transmitted to the automated synthesis subsystem. 3) The automated synthesis subsystem uses the received initial configuration parameters as input X to synthesize and prepare membrane material samples. Then, the automated testing subsystem tests the prepared membrane material samples to obtain the target performance parameters Y corresponding to the initial configuration parameters, thereby obtaining a dataset {X,Y} of configuration parameters and performance parameters. 4) For the current dataset {X,Y} of configuration parameters and performance parameters, use the surrogate probability model as the black box function approximation to obtain the posterior distribution of the black box function. Then, use the posterior distribution of the black box function to predict any point in the domain to obtain the expectation and uncertainty of the black box function. The expectation is the value in the corresponding range of any point, and the uncertainty is the variance. 5) Based on the expected value and uncertainty predicted by the posterior distribution of the black-box function, different acquisition functions are used to collect points in the domain. The collected points in the domain are evaluated based on the expected value and uncertainty predicted by the posterior distribution of the black-box function. The evaluation is to the extent that the algorithm can approach the global optimum of the black-box function. 6) Based on the different performance parameters output by each acquisition function and corresponding to different values ​​in the domain, combine and transform them to obtain a batch of formula and process parameters in the domain, and send them to the automated synthesis subsystem and automated testing subsystem for synthesis and testing, to obtain a new dataset of configuration parameters and performance parameters {X',Y'}. 7) Then merge the new configuration parameter and performance parameter dataset {X',Y'} with the previously obtained configuration parameter and performance parameter dataset, and send it back to the algorithm software, that is, return to step 4). Iterate through steps 4)-6) to optimize until the performance parameters of the membrane material sample reach the expected satisfactory value, or the performance parameters of all synthesized membrane material samples no longer improve within a certain number of rounds. Then the iterative optimization ends. The machine learning optimization module also includes optimizing the acquisition function in the following manner: T1. Select data with corresponding domain and range from known public databases to form a batch of prior datasets {Dn}; T2. Divide the prior dataset according to the confidence level; T3. Using the prior dataset as input, a batch of different machine learning models are established. The teaching model is used as a surrogate probability model, and the hyperparameters are adjusted and selected using random or optimization algorithms. The prior dataset is input into different machine learning models for training. Different teaching models are obtained after training. T4. Then, a neural network is used as the neural acquisition function and combined with each teaching model to form different overall models. The PPO reinforcement learning algorithm is used for training to obtain the trained neural acquisition function, which then replaces at least one acquisition function in step 5).

2. The automated synthesis and testing system for rapid research and development of membrane materials according to claim 1, characterized in that: It also includes a data mining module, which is used to generate the initial formulation and process parameters when the membrane material is synthesized in an automated manner, and to send the initial formulation and process parameters to the automated synthesis subsystem; The transport subsystem is used for transporting membrane material samples during synthesis and testing. The automation drive module is used for automated drive control and operational status monitoring of the automated synthesis subsystem and the automated test subsystem, and is displayed through the front-end interface.

3. The automated synthesis and testing system for rapid research and development of membrane materials according to claim 1, characterized in that: The automated synthesis subsystem comprises three modules: an automated liquid preparation and powder distribution module, an automated spin coating module, and an automated annealing module. The automated solution preparation and powder distribution module is used to prepare the membrane material raw materials by mixing and adding powder according to the formula in terms of mass and volume, so as to obtain the membrane material stock solution. An automated spin coating module is used to coat the prepared membrane material stock solution onto the substrate surface of the membrane material. The automated annealing module is used to perform annealing operations by heating or cooling a substrate with a solution coated on its surface to obtain a film material sample.

4. The automated synthesis and testing system for rapid research and development of membrane materials according to claim 1, characterized in that: The automated testing subsystem includes a spectral testing module for testing the optical properties of the film material and / or an electrical testing module for testing the electrical properties of the film material.

5. The automated synthesis and testing system for rapid research and development of membrane materials according to claim 2, characterized in that: The data mining module searches for literature information on publicly available internet data, and then analyzes and processes the literature information using a large language model to obtain the initial membrane material formulation and process parameters.

6. The automated synthesis and testing system for rapid research and development of membrane materials according to claim 2, characterized in that: The transport subsystem includes a robotic arm and an actuator at the end of the robotic arm.

7. An automated synthesis and testing method for rapid research and development of membrane materials, applied to the automated synthesis and testing system described in any one of claims 2-6, characterized in that: The method involves inputting the initial formula and process parameters provided by the data mining module into the automated synthesis subsystem to synthesize and prepare membrane material samples. The automated testing subsystem then tests the performance of the membrane material samples. Based on the performance test results, the machine learning optimization module uses analytical algorithms to infer better formulas and process parameters, which are then input into the automated synthesis subsystem to synthesize and prepare membrane material samples. This iterative process is repeated continuously to achieve automated synthesis and testing of membrane materials.

Images