A film preparation system of a recipe language model driven closed intelligent robot box and a control method thereof

Abstract

This invention provides a thin film preparation system and its control method based on a formula language model-driven enclosed intelligent robot box. The thin film preparation system includes: a robot hardware execution device comprising a consumable storage mechanism, a gantry transfer mechanism, a capping and clamping mechanism, a spin coating mechanism, a vacuum-induced crystallization mechanism, an annealing mechanism, and an in-situ spectral detection mechanism; an intelligent cognitive device deploying a formula domain vertical model, a language intelligent agent module, a literature learning closed-loop module, and an experimental learning closed-loop module, wherein the language intelligent agent module is configured to coordinate the data interaction between the literature learning closed-loop module and the experimental learning closed-loop module; and an intelligent agent isolation device that, after confirming that the recommended formula instructions conform to physical boundary constraints, parses them into low-level control instructions and sends them to the robot hardware execution device. This invention achieves autonomous acquisition and transformation of formula knowledge, secure verification of semantic instructions, and self-evolution of underlying mechanisms and self-optimization of formulas during the thin film preparation process.

Description

Technical Field

[0001] This invention relates to a thin film preparation system and its control method, specifically to a thin film preparation system and its control method driven by a formulation language model and a closed intelligent robot box, belonging to the field of artificial intelligence and automated material synthesis technology. Background Technology

[0002] The efficient fabrication of thin-film materials is one of the core technological challenges in the field of new energy devices. In recent years, researchers have conducted large-scale combinatorial explorations on thin-film optoelectronic devices, represented by perovskite single-junction cells and crystalline silicon-perovskite tandem cells, with over 100,000 formulations used to improve device performance. However, the synthesis formulations of these thin-film materials are highly complex, involving a vast chemical space comprised of components, numerous interface modifiers, additives, and passivators. Furthermore, the thin-film crystallization process is highly sensitive to the environment, significantly increasing the fabrication difficulty, and the related physicochemical mechanisms remain to be elucidated. Therefore, related research is still largely constrained by time-consuming trial-and-error synthesis and labor-intensive device fabrication processes, urgently requiring a shift towards next-generation scientific tools that can integrate robotic hardware for systematic high-throughput synthesis, fabrication, and characterization.

[0003] As a promising alternative, AI-powered thin film preparation systems integrating preparation and characterization capabilities can achieve high-throughput experiments and data acquisition. To date, various robotic systems have been introduced into the field of materials synthesis and characterization; however, the numerical datasets generated by these systems often lack sufficient analysis, making it difficult to provide effective feedback for semantic-level formulation optimization. The reasons for this are mainly as follows.

[0004] First, existing thin film preparation systems suffer from a translation gap between formulation and process. Traditional materials research and development heavily relies on human experience, and formulation data and process parameters recorded in external literature are often presented in non-standardized natural language, making it difficult to directly and automatically translate them into machine instructions executable by the underlying hardware. This translation gap leads to low utilization of literature knowledge, and a large amount of valuable prior experience cannot be efficiently reused in new experimental designs, severely restricting the efficiency of materials research and development.

[0005] Secondly, automated systems pose black-box security risks. In recent years, solutions that directly interface large language models with the hardware execution layer have been proposed. However, large language models inherently suffer from the illusion problem, potentially recommending process parameters that exceed the physical limits of the hardware or generating logical conflicts in action timing, leading to serious safety hazards such as equipment collisions, execution anomalies, and even device damage. Existing solutions generally lack mechanisms for systematically verifying physical boundaries before semantic instructions are issued to the hardware, making the implementation of industrial-grade artificial intelligence extremely risky.

[0006] Third, there is a lack of in-situ mechanism closed-loop capability. Most existing high-throughput thin film preparation systems operate in an open-loop mode, meaning the equipment is only responsible for executing preset instructions and outputting results data, but cannot correlate the dynamic physical information collected by in-situ characterization methods during the preparation process with the reasoning process of artificial intelligence models. This means that the system cannot autonomously generate mechanism explanations based on real experimental phenomena, nor can it transform mechanism knowledge into a basis for formulation optimization, resulting in a significant waste of the profound value of experimental data. Summary of the Invention

[0007] Based on the above background, the purpose of this invention is to provide a thin film preparation system and its control method that uses a formulation language model to drive a closed intelligent robot box. With a formulation domain vertical model and a collaborative language agent as the core, an intelligent preparation platform with a dual closed-loop architecture is constructed. Both numerical formulations and semantic formulations can be continuously learned and optimized from literature corpora and experimental corpora, thereby iteratively fine-tuning the formulation domain vertical model. This enables autonomous acquisition and transformation of formulation knowledge, secure verification and accurate parsing of semantic instructions, and self-evolution of underlying mechanisms and self-optimization of formulations based on in-situ spectral data during thin film preparation.

[0008] To achieve the above-mentioned objectives, the present invention provides the following technical solution:

[0009] A thin film preparation system driven by a formulation language model and a closed intelligent robotic box includes:

[0010] The robot hardware execution device includes a closed execution chamber and a consumable storage mechanism, a gantry transfer mechanism, a capping clamping mechanism, a spin coating mechanism, a vacuum-induced crystallization mechanism, an annealing mechanism, and an in-situ spectral detection mechanism disposed inside the closed execution chamber.

[0011] An intelligent cognitive device is communicatively connected to the robot hardware execution device. The intelligent cognitive device is equipped with a formulation domain vertical model, a language intelligent agent module, a literature learning closed-loop module, and an experimental learning closed-loop module. The language intelligent agent module is configured to coordinate the data interaction between the literature learning closed-loop module and the experimental learning closed-loop module. The literature learning closed-loop module is configured to automatically download vertical domain literature through data mining, extract formulation and process parameters from it, and generate a first semantic formulation corpus. The first semantic formulation corpus is then used to fine-tune the formulation domain vertical model. The experimental learning closed-loop module includes a mechanism generation unit, which is configured to receive dynamic spectral data collected by the in-situ spectral detection mechanism during the heating process of the annealing mechanism, associate the dynamic spectral data with the currently executed formulation and process parameters, generate a mechanistic text description explaining the experimental phenomena, and convert the mechanistic text description into a second semantic formulation corpus. The second semantic formulation corpus is then used to fine-tune the formulation domain vertical model.

[0012] An intelligent agent isolation device is communicatively connected to the intelligent cognitive device and the robot hardware execution device. The intelligent agent isolation device is equipped with a process rule verification module. The intelligent agent isolation device is configured to receive recommended formula instructions output by the vertical model of the formula domain. After the process rule verification module confirms that the recommended formula instructions meet the physical boundary constraints, it parses the recommended formula instructions into low-level control instructions and sends them to the robot hardware execution device.

[0013] By integrating the above-mentioned functional mechanisms into a closed execution chamber, the entire process of thin film preparation can be completed in an inert atmosphere or under controlled environment.

[0014] By collaboratively deploying a formulation domain vertical model, a language agent module, a literature learning closed-loop module, and an experimental learning closed-loop module in an intelligent cognitive device, the system simultaneously possesses the dual capabilities of continuously acquiring prior formulation knowledge from external literature and continuously accumulating empirical mechanistic knowledge from real experimental processes. As a result, the formulation domain vertical model, which serves as a domain-specific formulation language model, continuously evolves with the deepening of formulation exploration, gradually evolving from general language capabilities into a specialized agent with profound domain understanding.

[0015] By setting up a closed-loop module for literature learning, the knowledge of process routes, material systems, and parameter windows contained in external literature can be systematically transformed into structured corpus that the model can learn. This enables vertical models in the formulation field to generate reasonable initial formulation recommendations without the need for extensive local experiments.

[0016] By setting up a mechanism generation unit and associating in-situ spectral data with formulation variables to generate mechanism text descriptions, the system can be equipped with the ability to autonomously analyze physicochemical mechanisms such as thin film nucleation kinetics, transform experimental data into interpretable mechanism knowledge, and continuously feed this knowledge back into the formulation field vertical model in the form of a corpus.

[0017] By setting up an intelligent agent isolation device and deploying a process rule verification module between the intelligent cognitive device and the robot hardware execution device, the physical boundary constraints of the recommended formula instructions generated by the vertical model in the formula domain can be verified and intercepted in advance before being sent to the hardware layer, thus eliminating the equipment safety risks that may be caused by the illusion of a large model.

[0018] By driving the collaborative operation of the dual closed loop through a unified intelligent agent coordination mechanism, prior knowledge from literature and empirical knowledge from experiments can be integrated and converged in the same domain-specific model, thereby achieving continuous iterative improvement in the formulation recommendation capability.

[0019] Preferably, the gantry transfer mechanism includes an X-axis moving assembly, a linearly movable Y-axis moving assembly mounted on the X-axis moving assembly, multiple independently linearly movable Z-axis moving assemblies mounted on the Y-axis moving assembly, a gripper assembly mounted on at least one Z-axis moving assembly, and a pipette assembly mounted on at least one Z-axis moving assembly; the gripper assembly is configured to transfer the reagent bottle to the capping clamping mechanism for fixing and gripping the cap of the removed reagent bottle; the pipette assembly is configured to extend into the reagent bottle to aspirate liquid and reset after aspiration; the gripper assembly is also configured to close the cap after the pipette assembly resets;

[0020] The vacuum-induced crystallization mechanism includes a vacuum crystallization cavity and a vacuum pump; the gantry transfer mechanism is configured to transfer the substrate to be treated into the vacuum crystallization cavity after the spin coating mechanism completes the basic film formation of the substrate to be treated; the vacuum pump is configured to reduce the pressure in the vacuum crystallization cavity to a target vacuum level; the gantry transfer mechanism is also configured to transfer the substrate to be treated into the annealing mechanism after crystallization is completed;

[0021] The in-situ spectral detection mechanism includes a spectral probe assembly and a transfer mechanism for driving the spectral probe assembly to translate and move up and down. The transfer mechanism is configured to retract the spectral probe assembly to a standby position during non-detection periods, and to move the spectral probe assembly to the area to be tested above the substrate after it enters the annealing mechanism. The spectral probe assembly is configured to continuously acquire steady-state photoluminescence or absorption spectra of the thin film on the substrate during the annealing process as the dynamic spectral data.

[0022] Preferably, the literature learning closed-loop module includes a literature learning unit, a parameter generation unit, a literature corpus unit, a literature-side model fine-tuning unit, a literature-side inference unit, a literature-side evaluation unit, and a literature-side optimization unit. The literature learning unit is used to collect process routes and material system data from external databases. The parameter generation unit is used to extract and construct a set of formulation and process parameters from the material system data. The literature corpus unit is used to convert the set of formulation and process parameters into the first semantic formulation corpus. The literature-side model fine-tuning unit is used to perform model fine-tuning based on the first semantic formulation corpus. The literature-side inference unit is used to generate the recommended formulation instruction based on prior knowledge. The literature-side evaluation unit is used to score the quality of the generated formulation. The literature-side optimization unit is used to optimize the vertical model of the formulation domain based on the scoring results.

[0023] The experimental learning closed-loop module further includes a data feedback unit, an experimental learning unit, an experimental corpus unit, an experimental-side model fine-tuning unit, an experimental-side inference unit, an experimental-side evaluation unit, and an experimental-side optimization unit. The data feedback unit is used to transmit physical representation data and equipment operating status back to the system. The experimental learning unit is used to receive the transmitted data and coordinate with the mechanism generation unit to process the data. The experimental corpus unit is used to construct the generated mechanism text description into the second semantic recipe corpus. The experimental-side model fine-tuning unit is used to perform model fine-tuning based on the second semantic recipe corpus. The experimental-side inference unit is used to generate iterative recommended recipe instructions based on experimental feedback. The experimental-side evaluation unit is used to score the experimental mechanism and iterative recipe. The experimental-side optimization unit is used to update the weight parameters of the recipe domain vertical model based on the scoring results.

[0024] The intelligent agent isolation device includes an agent unit and a device adaptation unit. The process rule verification module is deployed in the agent unit to perform pre-boundary verification and interception. The device adaptation unit is used to parse and translate the verified recommended recipe instructions into the underlying control instructions.

[0025] The robot hardware execution device also includes a device execution control unit, which is configured to receive the underlying control commands issued by the device adapter unit and drive the corresponding mechanism actions.

[0026] A control method for a closed intelligent robot box-driven thin film preparation system based on the above-described formulation language model, the method comprising the following steps:

[0027] S1. The language intelligent agent module coordinates the data interaction between the literature learning closed-loop module and the experimental learning closed-loop module, and generates recommended formula instructions containing formula and process parameters based on the fine-tuned formula domain vertical model.

[0028] S2. The intelligent agent isolation device receives the recommended formula instruction. After the process rule verification module confirms that the recommended formula instruction conforms to the physical boundary constraints, it parses the recommended formula instruction into a low-level control instruction and sends it to the robot hardware execution device.

[0029] S3. The robot hardware execution device, according to the underlying control command, sequentially completes opening the lid, reagent absorption, spin coating film formation on the substrate to be treated in the spin coating mechanism, vacuum crystallization on the substrate to be treated in the vacuum induced crystallization mechanism, and heat treatment on the substrate to be treated in the annealing mechanism.

[0030] S4. During the heat treatment process of the annealing mechanism, the in-situ spectral detection mechanism collects dynamic spectral data and transmits the physical characterization data containing the dynamic spectral data back to the experimental learning closed-loop module.

[0031] S5. The mechanism generation unit of the experimental learning closed-loop module transforms the physical representation data into a mechanism text description explaining the experimental phenomenon, and transforms the mechanism text description into a second semantic recipe corpus, and uses the second semantic recipe corpus to fine-tune the recipe domain vertical model.

[0032] Preferably, the literature learning closed-loop module includes a literature learning unit, a parameter generation unit, a literature corpus unit, and a literature-side model fine-tuning unit. Before step S1, the method further includes the following steps:

[0033] The literature learning unit collects process routes and material system information from external databases;

[0034] The parameter generation unit extracts and constructs a set of formulation and process parameters from the material system data. The variables in the set of formulation and process parameters include precursor components, interface modification molecules, additives, passivating agents and corresponding preparation parameters.

[0035] The document corpus unit transforms the set of formulas and process parameters into the first semantic formula corpus;

[0036] The document-side model fine-tuning unit performs model fine-tuning based on the first semantic formula corpus.

[0037] Preferably, the intelligent agent isolation device includes a proxy unit and a device adaptation unit, the process rule verification module is deployed in the proxy unit, and step S2 specifically includes:

[0038] The agent unit, which is equipped with the process rule verification module, extracts the specific process parameters from the recommended formula instruction and calls the preset task whitelist.

[0039] The specific process parameters extracted are compared to see if they exceed the physical limits of the robot hardware execution device, and whether timing conflicts occur between the various action commands.

[0040] If the physical limit boundary is exceeded or the timing conflict is triggered, the agent unit intercepts the recommended recipe instruction, generates an out-of-bounds error message, and triggers a regeneration instruction.

[0041] If the verification passes, the device adaptation unit will translate the semantic action mapping in the recommended recipe instruction into the underlying control instruction that can be executed by the robot hardware execution device.

[0042] Preferably, the gantry transfer mechanism is equipped with a gripper assembly and a pipette assembly. In step S3, the spin coating process of the substrate to be treated by the spin coating mechanism specifically includes:

[0043] The gripper assembly is driven to place the substrate to be processed onto the spin coating mechanism and perform vacuum adsorption fixation.

[0044] The pipette assembly is driven to quantitatively add the thin film precursor solution;

[0045] Thin film nucleation control can be achieved by either adding an anti-solvent during spin coating or by vacuum-induced crystal formation after spin coating.

[0046] Preferably, in step S5, the mechanism generation unit converts the physical characterization data into a mechanistic text description explaining the experimental phenomenon, specifically including:

[0047] Extract the burst nucleation time, fluorescence decay slope, and blue or red shift of characteristic peaks from the dynamic spectral data, and combine them with the type of interface modification molecules or the concentration of additives to infer the degree of acceleration of heterogeneous nucleation or the release state of residual stress.

[0048] Receive diffraction data, extract the peak intensity and full width at half maximum (FWHM) of characteristic precursor residues or intermediate phases, and map them to generate a textual description of the mechanism of additive regulation of crystallization or surface passivation agent release of lattice stress.

[0049] Receive thin film images, extract film coverage and grayscale values, and map them to generate a mechanistic text description of solvent removal and supersaturation kinetics.

[0050] Preferably, the literature learning closed-loop module includes a literature-side evaluation unit and a literature-side optimization unit. After the recommended recipe instruction is generated in step S1, the literature-side evaluation unit scores the quality of the generated recipe, and the literature-side optimization unit optimizes the vertical model of the recipe domain based on the scoring results.

[0051] The experimental learning closed-loop module also includes an experimental evaluation unit and an experimental optimization unit. After the mechanism text description is generated in step S5, the experimental evaluation unit scores the experimental mechanism and iterative formulation. The experimental optimization unit updates the weight parameters of the formulation domain vertical model based on the scoring results.

[0052] Preferably, this method is deployed and executed in a distributed manner based on an edge cloud collaborative computing architecture, specifically including:

[0053] The process of fine-tuning the vertical model of the formulation domain using the first semantic formulation corpus and the second semantic formulation corpus is deployed on a cloud server to perform high-performance training and large-scale data processing.

[0054] The reasoning process of the language intelligence module generating the recommended recipe instruction and the verification process of the process rule verification module confirming physical boundary constraints are deployed on edge computing devices to perform real-time joint debugging of local model reasoning verification and pre-boundary verification interception.

[0055] The process of parsing and issuing the underlying control commands is deployed on a local industrial control host to achieve deterministic real-time control of the robot hardware execution device.

[0056] Compared with the prior art, the present invention has the following advantages:

[0057] This invention discloses a thin film preparation system and its control method driven by a formulation language model and a closed-loop intelligent robot box. By constructing a dual closed-loop architecture of literature and experiment, the entire system can simultaneously absorb prior knowledge from external literature and local experimental mechanism knowledge. This allows the recommendation capability of the formulation domain vertical model set within the system to continuously enhance with the increase of experimental rounds, effectively solving the translation gap problem from formulation to process. This invention introduces an intelligent agent isolation device and a process rule verification module to perform dual verification of physical boundaries and temporal conflicts before the formulation recommended by the formulation domain vertical model enters the hardware execution layer, eliminating the equipment safety risks caused by the illusion of a large model. This invention integrates in-situ spectral detection data with a mechanism generation unit, enabling the system to autonomously analyze and express physicochemical mechanisms such as thin film nucleation kinetics and lattice stress state, realizing the underlying mechanism evolution and formulation optimization based on in-situ spectral data. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0059] Figure 1 This is a schematic diagram of the overall architecture of the thin film preparation system driven by the formulation language model of the present invention.

[0060] Figure 2 This is a schematic diagram of the overall structure of the robot hardware execution device provided in an embodiment of the present invention;

[0061] Figure 3This is a schematic diagram of the internal functional structure layout of the robot hardware execution device provided in an embodiment of the present invention;

[0062] Figure 4 This is a schematic diagram of the architecture of the intelligent agent isolation device provided in an embodiment of the present invention;

[0063] Figure 5 This is a schematic diagram showing the changes in burst nucleation time and fluorescence decay slope with formula variables in the dynamic spectral data collected by the in-situ spectral detection mechanism provided in the embodiments of the present invention;

[0064] Figure 6 This is a schematic diagram showing the variation of the characteristic peak intensity of the two-dimensional perovskite phase and the residual stress of the thin film with the type of passivating agent in the dynamic spectral data provided by the embodiments of the present invention;

[0065] Figure 7 This is a schematic diagram illustrating the cumulative growth of the training corpus size produced by the literature learning closed-loop module and the experiment learning closed-loop module provided in the embodiments of the present invention as the experiment progresses;

[0066] Figure 8 This is a schematic diagram illustrating the evolution of the evaluation scores of the formulation domain vertical model provided in the embodiments of the present invention in the two dimensions of formulation recommendation and mechanism reasoning as the dual closed loop continues to operate;

[0067] Figure 9 This is a schematic diagram illustrating the change in power conversion efficiency of a thin film preparation system driven by a formulation language model in a closed intelligent robot box during the controllable preparation process, as the formulation exploration stage progresses.

[0068] Figure 10 This is a schematic diagram illustrating the relationship between the importance of formulation variables and device performance in a thin film preparation system driven by a formulation language model in a closed intelligent robot box, as provided in an embodiment of the present invention.

[0069] Figure 11 This is a schematic diagram illustrating the evolution of device open-circuit voltage, short-circuit current density, and fill factor in the four formulation optimization stages of the thin film preparation system driven by the formulation language model of the embodiment of the present invention.

[0070] In the diagram: 1. Gantry transfer mechanism; 2. Capping and clamping mechanism; 3. Spin coating mechanism; 4. Vacuum-induced crystallization mechanism; 5. Annealing mechanism; 6. In-situ spectral detection mechanism; 7. Equipment execution control unit; 8. Consumable storage mechanism; 9. Material transfer mechanism; 101. X-axis moving assembly; 102. Y-axis moving assembly; 103. Z-axis moving assembly; 104. Gripper assembly; 105. Pipette assembly; 401. Vacuum crystallization chamber; 402. Vacuum pump; 601. Spectrometer probe assembly; 602. Transfer mechanism. Detailed Implementation

[0071] The technical solution of the present invention will be further described in detail below through specific embodiments and in conjunction with the accompanying drawings. It should be understood that the implementation of the present invention is not limited to the following embodiments, and any modifications and / or alterations made to the present invention will fall within the protection scope of the present invention.

[0072] In this invention, unless otherwise specified, all parts and percentages are by weight, and the equipment and raw materials used are commercially available or commonly used in the art. Unless otherwise specified, the methods in the following embodiments are conventional methods in the art. Unless otherwise specified, the components or equipment in the following embodiments are general standard parts or components known to those skilled in the art, and their structures and principles can be learned by those skilled in the art through technical manuals or conventional experimental methods.

[0073] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings. In this detailed description, numerous specific details are set forth to facilitate explanation and provide a thorough understanding of the embodiments of the present invention. However, one or more embodiments may be practiced by those skilled in the art without these specific details.

[0074] like Figure 1-4 As shown, an embodiment of the present invention discloses a thin film preparation system driven by a formula language model and a closed intelligent robot box, including a robot hardware execution device, an intelligent cognitive device, and an intelligent agent isolation device.

[0075] The robot hardware execution device includes a closed execution chamber and, within the closed execution chamber, a gantry transfer mechanism 1, a capping clamping mechanism 2, a spin coating mechanism 3, a vacuum-induced crystallization mechanism 4, an annealing mechanism 5, an in-situ spectral detection mechanism 6, a consumable storage mechanism 8, and a material transfer mechanism 9. The closed execution chamber physically forms a sealed preparation space isolated from the external environment, maintaining an inert atmosphere or controlled humidity conditions inside. Each mechanism is modularly arranged within the closed execution chamber. The robot hardware execution device also includes a device execution control unit 7, deployed in the bottom control compartment of the closed execution chamber. This unit integrates servo drivers, programmable logic controllers, industrial computers, and motion control cards, and is responsible for receiving low-level control commands from the intelligent agent isolation device and driving each execution mechanism to complete the corresponding physical actions.

[0076] The gantry transfer mechanism 1 includes an X-axis moving assembly 101, a Y-axis moving assembly 102 linearly movable on the X-axis moving assembly 101, multiple Z-axis moving assemblies 103 independently linearly movable on the Y-axis moving assembly 102, a gripper assembly 104 on one of the Z-axis moving assemblies 103, and a pipette assembly 105 on the other Z-axis moving assembly 103. During solution dispensing, the gripper assembly 104 transfers the reagent bottle from the medicine bottle buffer to the capping clamping mechanism 2. After the capping clamping mechanism 2 positions and clamps the reagent bottle, the gripper assembly 104 grasps and screws on the cap to open it. The pipette assembly 105 then extends into the reagent bottle to quantitatively aspirate the film precursor solution, and resets after aspiration. After the pipette assembly 105 resets, the gripper assembly 104 screws the cap back onto the reagent bottle, completing the capping operation. The spin coating mechanism 3, centered on the spin coater main unit, is equipped with an integrated vacuum adsorption fixation system and a waste liquid suction and recovery device. After the gripper assembly 104 places the substrate to be treated on the spin coating stage of the spin coating mechanism 3, the vacuum adsorption system applies a vacuum negative pressure to fix the substrate. The pipette assembly 105 quantitatively adds the film precursor solution to the surface of the substrate, and the spin coater starts its rotation program to complete the basic spin coating film formation. Within the set time window of the spin coating mechanism 3's rotation, the pipette assembly 105 further quantitatively adds an anti-solvent, promoting uniform nucleation through an anti-solvent-assisted process.

[0077] The vacuum-induced crystallization mechanism 4 includes a vacuum crystallization chamber 401 and a vacuum pump 402. After the substrate to be treated completes the basic film formation by spin coating mechanism 3, gantry transfer mechanism 1 transfers the substrate to be treated into the vacuum crystallization chamber 401. Vacuum pump 402 reduces the pressure inside the vacuum crystallization chamber 401 to the target vacuum level, inducing rapid crystallization of the precursor solute through a low-pressure environment. After the crystallization process is completed, gantry transfer mechanism 1 transfers the substrate to be treated to the annealing mechanism 5.

[0078] Annealing mechanism 5 is equipped with multiple independently temperature-controlled annealing units, each supporting independent temperature control. The annealing temperature, holding time, and heating rate can be set according to the underlying control commands. The parallel configuration of multiple independent annealing units provides the hardware foundation for simultaneous annealing of multiple samples and high-throughput processing of continuous batches.

[0079] The in-situ spectral detection mechanism 6 includes a spectral probe assembly 601 and a transfer mechanism 602 for driving the translation and retraction of the spectral probe assembly 601. During non-detection periods, the transfer mechanism 602 retracts the spectral probe assembly 601 to a standby position. After the substrate to be processed enters the annealing mechanism 5, the transfer mechanism 602 moves the spectral probe assembly 601 above the test area of ​​the substrate. The spectral probe assembly 601 continuously acquires steady-state photoluminescence or absorption spectra of the thin film on the substrate during the annealing process, as dynamic spectral data. After detection, the transfer mechanism 602 drives the spectral probe assembly 601 back to the standby position.

[0080] The consumable storage mechanism 8 is used to store and position substrates (such as glass slides, silicon wafers, etc.), medicine bottles, and TIP heads. The material transfer mechanism 9 serves as a transition chamber between the inside and outside of the closed execution chamber, used to transfer materials between the inside and outside of the closed execution chamber.

[0081] The intelligent cognitive device is connected to the robot hardware execution device. The intelligent cognitive device is equipped with a vertical model of the formulation domain, a language intelligent agent module, a literature learning closed-loop module, and an experimental learning closed-loop module.

[0082] The language agent module is configured to coordinate the data interaction between the literature learning closed-loop module and the experimental learning closed-loop module. On the literature closed-loop side, the language agent module is responsible for encoding process routes and material system data extracted from external databases into machine-readable structured formulation sequences, and scheduling the execution flow of formulation recommendation, quality assessment, and preference alignment optimization. On the experimental closed-loop side, the language agent module is responsible for coordinating the collection and feedback of experimental data, the generation and organization of mechanistic texts, the construction of experimental corpora, and the iterative updating of model weights. Both closed loops are uniformly coordinated and scheduled by the language agent module, enabling the effective integration of prior literature knowledge and experimental empirical knowledge in the formulation domain vertical model.

[0083] While the literature learning closed-loop module and the experimental learning closed-loop module differ in data sources and feedback formats, they complement each other in shaping the vertical model of the formulation domain through the unified coordination of the language intelligent agent module. The literature learning closed-loop module injects broad-spectrum prior knowledge extracted from a large number of published studies into the vertical model of the formulation domain. This knowledge covers statistical correlations between different material systems, different process routes, and different performance results, enabling the model to have a reasonable starting point estimation ability in the early stages of experimental exploration, avoiding repeated exploration within formulation ranges that have been proven ineffective by the literature. The experimental learning closed-loop module injects mechanistic text descriptions into the vertical model of the formulation domain based on in-situ characterization data collected during the actual hardware execution of the system. This type of knowledge has strong equipment and process specificity, containing dynamic process information that is usually not available in the literature, enabling the model to establish inference capabilities adapted to the specific process environment of the system.

[0084] The knowledge injection of the literature learning closed-loop module is driven by the accumulation of external literature, and its update frequency is related to the publication rhythm of literature, making it suitable for providing a stable prior foundation for the model in the early stages of experimental exploration. The knowledge injection of the experimental learning closed-loop module is driven by each batch of experimental data actually executed by the system. It triggers a model fine-tuning after each round of annealing, spectral detection, and mechanism generation, with a higher update frequency, making it suitable for rapid localization correction of the model in the mid-to-late stages of experimental exploration. The language agent module can dynamically schedule the triggering priority of the two closed loops according to the current stage of the experiment. In the early stages of experimental exploration, the literature learning closed-loop module runs first to establish the basic cognitive ability of the vertical model in the formulation domain; while as local experimental data accumulates, the triggering frequency of the experimental learning closed-loop module gradually increases, continuously correcting the model's inference bias with mechanistic text descriptions, so that the accuracy of the recommended formulations continuously converges with the increase of experimental rounds. Figure 5 As shown, with the advancement of experimental rounds, the proportion of second semantic recipe corpus, with mechanism text description as its core, in the model training corpus continues to increase. The semantic corpus containing mechanism information gradually evolved from a state with an extremely low proportion in the first stage to a major component in the fourth stage, with a cumulative total of 578 million tokens. This intuitively reflects the continuous shaping effect of the experimental learning closed-loop module on the vertical model of the recipe domain as the experiment progresses. Figure 6 As shown, the evaluation scores of the formulation domain vertical model in both formulation recommendation and mechanism reasoning dimensions systematically improved with the continuous operation of the dual closed loop, gradually increasing from a low level in the initial stage to a level close to that of experienced researchers. This quantitatively verifies the effect of the dual closed loop synergistically driving the continuous evolution of the formulation domain vertical model.

[0085] The literature learning closed-loop module is configured to extract formulation and process parameters from external literature and generate a first semantic formulation corpus, which is then used to fine-tune the formulation domain vertical model. The literature learning closed-loop module includes a literature learning unit, a parameter generation unit, a literature corpus unit, a literature-side model fine-tuning unit, a literature-side inference unit, a literature-side evaluation unit, and a literature-side optimization unit. The literature learning unit collects process routes and material system data from external databases. The parameter generation unit extracts and constructs a set of formulation and process parameters from the material system data. The variables in the formulation and process parameter sets include precursor components, interface modification molecules, additives, passivating agents, and corresponding preparation parameters. The literature corpus unit transforms the set of formulation and process parameters into a first semantic formulation corpus. The literature-side model fine-tuning unit performs model fine-tuning based on the first semantic formulation corpus, incrementally injecting domain knowledge into the basic large model through efficient parameter fine-tuning methods. The literature-side inference unit generates recommended formulation instructions based on prior knowledge. The literature-side evaluation unit scores the quality of the generated formulations. The literature-side optimization unit optimizes the formulation domain vertical model based on the scoring results to achieve preference alignment.

[0086] Taking the preparation of perovskite thin films as an example, the set of formulation and process parameters constructed by the parameter generation unit explicitly defines each formulation variable using discrete structured units, covering five main functional modules: perovskite precursor components, interface modification molecules mixed into the perovskite precursor, additives mixed into the perovskite precursor, passivating agents applied to the perovskite surface, and interface modification molecular layers applied to the perovskite buried interface. Under each functional module, the set of formulation variables includes both material composition variables and corresponding preparation process parameter variables. Taking the perovskite precursor components as an example, the composition variables include the molar fractions of cesium, methylamine, and formamidinium in the A-site cation and the stoichiometric coefficients of iodine and bromine in the X-site halogen and the total precursor concentration. The preparation process parameter variables include the two-step spin coating speed and duration, the timing and amount of antisolvent addition, and the annealing temperature and annealing time. Taking the interface-modified molecular layer as an example, the compositional variables are the concentrations of various self-assembled molecules, while the preparation process parameters include spin-coating speed, spin-coating time, annealing temperature, and annealing time. Each variable is assigned a clear numerical range constraint within the parameter set, serving as a common reference benchmark for semantic transformation of subsequent literature corpus units and physical boundary verification of the intelligent agent isolation device. As the experimental phase progresses, the dimensions of variables covered by the parameter set expand: in the early stages of system operation, the formulation and process parameter set only covered a few variables related to the perovskite matrix; after introducing interface-modified molecules and additive engineering, the number of variables increased significantly; and after further introducing passivating agent engineering and independent interface-modified molecular layers, the dimensions of variables further expanded, systematically expanding the exploration space of the vertical model in the formulation field and continuously improving the formulation recommendation capability.

[0087] The experimental learning closed-loop module includes a mechanism generation unit, a data feedback unit, an experimental learning unit, an experimental corpus unit, an experimental model fine-tuning unit, an experimental inference unit, an experimental evaluation unit, and an experimental optimization unit. The data feedback unit aggregates multimodal data such as in-situ spectral data, thin film image parameters, device performance parameters, and equipment operating status, and returns this structured data to the upper-level system. The experimental learning unit receives the data returned by the data feedback unit and collaborates with the mechanism generation unit to process the data and extract knowledge. The mechanism generation unit receives dynamic spectral data collected by the in-situ spectral detection mechanism 6 during the heating process of the annealing mechanism 5, correlates the dynamic spectral data with the currently executed formula and process parameters, and generates a mechanistic text description explaining the experimental phenomena.

[0088] Specifically, the mechanism generation unit extracts burst nucleation time, fluorescence decay slope, and blue or red shift of characteristic peaks from dynamic spectral data. Combined with the type of interface-modifying molecules or additive concentration in the current formulation, it infers the acceleration of heterogeneous nucleation or the release state of residual stress, and records the inferences in the form of structured mechanism text. For example, when dynamic spectral data shows a shortened burst nucleation time and a decreased fluorescence decay slope, the mechanism generation unit, combined with the type and concentration of interface-modifying molecules in the current formulation, infers that the preferential precipitation of interface-modifying molecules on the substrate surface promotes the formation of heterogeneous nucleation sites, accelerates and synchronizes the nucleation process, and transforms this mechanism explanation into structured corpus. Similarly, when characteristic peaks show a blue or red shift, the mechanism generation unit, combined with the current passivating agent type, infers the formation state of the surface-confined low-dimensional perovskite phase and the degree of residual stress transformation from tensile stress to compressive stress, generating corresponding mechanism description text. The mechanism generation unit transforms the mechanism text description into a second semantic recipe corpus. It then uses this second semantic recipe corpus to fine-tune the vertical model in the recipe domain, enabling the system to transform experimental data into interpretable mechanism knowledge. This drives the vertical model in the recipe domain to evolve from factual memory to deep inference based on mechanism support.

[0089] The three types of descriptors extracted by the mechanism generation unit each carry different levels of physicochemical information. For example... Figure 7 As shown, the burst nucleation time reflects the time point at which a large number of crystal nuclei suddenly form from a supersaturated state during spin coating or annealing heating of the precursor solution. In the in-situ photoluminescence spectrum, the burst nucleation time is represented by the moment when the intensity of the luminescence signal suddenly increases. Figure 7The study also presents systematic data on the changes in burst nucleation time and fluorescence decay slope, along with corresponding in-situ photoluminescence spectra, under experimental conditions ranging from no interface-modified molecules to single interface-modified molecules and then to co-interface-modified molecules. When interface-modified molecules are introduced into the formulation, they preferentially precipitate on the substrate surface during spin-coating, forming dense heterogeneous nucleation sites. This lowers the supersaturation threshold required for nucleation, thereby shortening the burst nucleation time. The mechanistic generation unit infers the degree of acceleration of heterogeneous nucleation by comparing the changes in burst nucleation time before and after the introduction of interface-modified molecules: the greater the reduction in burst nucleation time, the denser the nucleation sites provided by the interface-modified molecule, and the more synchronized the nucleation process.

[0090] The fluorescence decay slope reflects the rate at which the photoluminescence signal decays over time after an explosive nucleation event, and is directly related to the formation rate of defect states and grain boundary density during grain growth. A smaller fluorescence decay slope indicates a more stable grain growth process, a lower accumulation rate of non-radiative recombination centers at grain boundaries, and a lower overall defect density in the film. The mechanistic generation unit correlates the fluorescence decay slope with the additive concentration in the current formulation. When the additive concentration increases and the fluorescence decay slope decreases, it is inferred that the additive inhibits defect accumulation during grain growth by coordinating with residual lead ions or participating in grain boundary passivation.

[0091] like Figure 8 As shown, the blue shift or red shift of the characteristic peaks reflects the state of lattice stress in the thin film. In perovskite thin films, the lattice stress generated during annealing due to the difference in thermal expansion coefficients at the upper and lower interfaces leads to a slight change in the band structure, which manifests as a shift in the position of the characteristic peaks in the photoluminescence spectrum relative to the stress-free reference state. Figure 8 The evolution of diffraction peak intensity and residual stress in the two-dimensional perovskite phase with varying passivator types is illustrated, visually demonstrating the transformation of residual stress from tensile stress to compressive stress. The mechanism generation unit correlates the shift direction and amount of characteristic peaks with the type of passivator in the current formulation.

[0092] The mechanism generation unit integrates the inferences from the three sources into a coherent mechanistic text description. The generated mechanistic text description is presented in the form of a structured formulation report. Each report includes a comparison of formulation parameters between the current experiment and comparative experiments, in-situ spectral and crystallographic characterization results, mechanistic inferences based on the characterization data, and a summary of the causal relationship between changes in formulation variables and changes in device performance indicators. Taking a typical experiment in the passivator engineering optimization process as an example, the mechanism generation unit compared a single passivator with a mixed dual passivator scheme, recording that the intensity of the two-dimensional perovskite phase diffraction peak at a diffraction angle of approximately 4 degrees increased from 234.75 counts to 972.90 counts, while the residual tensile stress of the thin film decreased from 11 MPa to 2.4 MPa. Based on the mechanism description, the mechanism generation unit inferred that aromatic ammonium passivators induced the formation of a more continuous surface-confined low-dimensional perovskite phase on the film surface, terminating lattice growth and releasing lattice stress through phase separation. Meanwhile, bisammonium passivators provided additional stress buffering through the dual-site coordination of their ammonium groups with adjacent defect sites. The above synergistic effect increased the open-circuit voltage from 1.06 V to 1.13 V, the fill factor from 75.44% to 80.57%, and the power conversion efficiency from 20.00% to 23.57%.

[0093] The experimental corpus unit integrates the aforementioned mechanistic text descriptions along with corresponding formulation parameters, characterization data, and device performance indicators into a second semantic formulation corpus, which is used for fine-tuning the formulation domain vertical model. The experimental model fine-tuning unit performs model fine-tuning based on the second semantic formulation corpus. The experimental inference unit generates iterative recommended formulation instructions based on experimental feedback. The experimental evaluation unit scores the experimental mechanisms and iterative formulations for quality. The experimental optimization unit updates the weight parameters of the formulation domain vertical model based on the scoring results. With the accumulation of experimental rounds, the formulation domain vertical model gradually establishes a complete logical chain of formulation variable combinations, specific descriptor change patterns, mechanism explanations, and performance predictions, enabling its formulation recommendation capability to evolve from extrapolation based on historical data to proactive inference based on physical mechanisms.

[0094] The intelligent agent isolation device communicates with the intelligent cognitive device and the robot hardware execution device, and is equipped with a process rule verification module. The intelligent agent isolation device is configured to receive recommended recipe instructions output from the vertical model of the recipe domain. After the process rule verification module confirms that the recommended recipe instructions conform to physical boundary constraints, it parses the recommended recipe instructions into low-level control instructions and sends them to the robot hardware execution device.

[0095] Specifically, the intelligent agent isolation device includes an agent unit and an equipment adaptation unit, with the process rule verification module deployed in the agent unit.

[0096] The process rule verification module in the agent unit receives recommended recipe instructions output by the vertical model of the recipe domain, extracts the specific process parameters, and calls a preset task whitelist to compare whether the extracted specific process parameters exceed the physical limit boundaries of the robot hardware execution device, and whether there are timing conflicts between the various action instructions. If the physical limit boundaries are exceeded or timing conflicts are caused, the agent unit intercepts the recommended recipe instructions, generates an out-of-bounds error message, and triggers the regeneration of instructions. If the verification passes, the device adaptation unit translates the semantic actions in the recommended recipe instructions into low-level control instructions executable by the robot hardware execution device.

[0097] The process rule verification module performs two types of verifications: physical limit boundary verification and timing conflict verification, each targeting different types of out-of-bounds risks. Physical limit boundary verification addresses situations where a single process parameter exceeds the rated capacity of the robot hardware execution device. The task whitelist pre-sets upper and lower limits for each type of controllable parameter of the robot hardware execution device, specifically including: the upper limit of the motion speed and travel boundary of each axis of the gantry transfer mechanism 1, the clamping force range of the gripper assembly 104, the upper and lower limits of the single liquid aspiration volume of the pipette assembly 105, the upper limit of the rotation speed and acceleration of the spin coating mechanism 3, the lower limit of the target vacuum level achievable by the vacuum pump 402 and the upper limit of the pumping rate, the upper limit of the temperature and the upper limit of the heating rate of each annealing unit, and the transfer travel boundary of the in-situ spectral detection mechanism 6. When a parameter recommended by the vertical model in the formulation domain exceeds any of the above constraints, the corresponding instruction item is marked as out of bounds, and the agent unit immediately terminates the further transmission of the recommended formulation instruction.

[0098] Timing conflict verification addresses situations where multiple action commands exhibit logical contradictions in the time dimension. It primarily includes the following typical conflict scenarios: First, spatial occupancy conflict, where the movement trajectory of the gantry transfer mechanism 1 and the transfer mechanism 602 of the in-situ spectral detection mechanism 6 are simultaneously instructed to move to positions that interfere with each other spatially; second, state prerequisite conflict, where one action command requires the completion state of another action as a prerequisite, and the two commands are arranged to be executed in parallel, for example, the pipette assembly 105 is instructed to perform aspiration before the capping clamping mechanism 2 has completed its opening action; third, resource contention conflict, where two actions requiring exclusive use of the same physical resource are arranged to be executed within overlapping time windows, for example, two spin coating commands are instructed to simultaneously occupy the same spin coating mechanism 3. The process rule verification module parses the timestamps and resource occupancy declarations of all action commands in the recommended formula instructions, constructs an action dependency graph, and detects circular dependencies or resource conflicts in the graph to identify the aforementioned conflict scenarios.

[0099] When a violation is detected in either the physical limit boundary check or the timing conflict check, the proxy unit generates an out-of-bounds error message containing the specific violation parameter name, violation value, and corresponding upper and lower constraint limits. This error message, along with the original recommended recipe instruction, is returned to the recipe domain vertical model of the intelligent cognitive device, triggering a regeneration process. Upon receiving the out-of-bounds error message, the recipe domain vertical model incorporates it as a constraint into the context of the next inference, generates a corrected recommended recipe instruction, and resubmits it to the process rule verification module for verification until it passes. This iterative mechanism, without restricting the inference freedom of the recipe domain vertical model, confines all instructions that ultimately enter the hardware execution layer to the physically feasible domain.

[0100] The device adapter unit is deployed in an industrial computer environment. Internally, it integrates a motion control card calling interface, device execution flow orchestration logic, a log recording module, a status management module, and communication interfaces with programmable logic controllers, sensors, and spectral modules. The device adapter unit shields the register details and communication protocol differences of the underlying hardware, allowing the upper-level intelligent system to complete task calls and status interactions solely through a unified semantic interface. Low-level control commands are issued to each actuator via the device execution control unit 7, driving the corresponding mechanism to complete deterministic physical actions.

[0101] In terms of logical architecture, the intelligent agent isolation device is located between the intelligent cognitive device and the robot hardware execution device. This device establishes a secure isolation and instruction translation intermediary layer between the intelligent cognitive layer and the hardware execution layer. The pre-interception mechanism of the process rule verification module terminates potential out-of-bounds recommendations from the vertical model of the formulation domain before they reach the hardware. This is a key difference between this invention and large-model-driven hardware solutions.

[0102] Furthermore, the thin film preparation system driven by the formulation language model of the closed intelligent robot box is deployed and executed in a distributed manner based on the edge cloud collaborative computing architecture, forming a three-layer computing power collaborative system of cloud, edge and local terminals.

[0103] The cloud server undertakes the high-computing-power, large-data-volume model training task, specifically including the literature-side fine-tuning training process of the vertical model in the formulation domain using the first semantic formulation corpus, and the experimental-side fine-tuning training process using the second semantic formulation corpus.

[0104] Edge computing devices undertake inference and verification tasks with high real-time and security requirements. Specifically, this includes the inference process of the language agent module generating recommended recipe instructions, and the verification process of the process rule verification module confirming physical boundary constraints. Edge computing devices communicate with the robot's hardware execution device via a low-latency local network, completing real-time joint debugging of model inference verification and pre-emptive boundary verification and interception locally.

[0105] The local industrial control host is responsible for parsing and issuing low-level control commands. Through industrial-grade control components such as motion control cards and programmable logic controllers, it drives various execution units, including the gantry transfer mechanism 1, the spin coating mechanism 3, and the annealing mechanism 5, to complete specific physical actions. The local industrial control host interacts with the edge computing device through a standardized communication interface to achieve real-time issuance of control commands and timely reporting of execution status.

[0106] Embodiments of the present invention also provide a control method for a thin film preparation system driven by a closed intelligent robot box based on the above-described formulation language model, the method comprising the following steps:

[0107] S1, the language intelligent agent module coordinates the data interaction between the literature learning closed-loop module and the experimental learning closed-loop module, and generates recommended formula instructions containing formula and process parameters based on the fine-tuned formula domain vertical model.

[0108] S2. The intelligent agent isolation device receives the recommended formula instruction. After the process rule verification module confirms that the recommended formula instruction meets the physical boundary constraints, it parses the recommended formula instruction into a low-level control instruction and sends it to the robot hardware execution device.

[0109] S3. The robot hardware execution device sequentially completes the following steps according to the underlying control instructions: opening the lid, absorbing the reagent, spin coating the substrate to be treated in the spin coating mechanism 3, vacuum crystallizing the substrate to be treated in the vacuum induced crystallization mechanism 4, and heat treating the substrate to be treated in the annealing mechanism 5.

[0110] S4. During the heat treatment process in the annealing mechanism 5, the in-situ spectral detection mechanism 6 collects dynamic spectral data and transmits the physical characterization data containing the dynamic spectral data back to the experimental learning closed-loop module.

[0111] S5, the mechanism generation unit of the experimental learning closed-loop module transforms physical representation data into mechanistic text descriptions that explain experimental phenomena, and transforms the mechanistic text descriptions into second semantic recipe corpus, using the second semantic recipe corpus to fine-tune the vertical model of the recipe domain.

[0112] When the system initiates a new thin film preparation task, the literature learning closed-loop module runs first. The literature learning unit collects literature data from external databases, the parameter generation unit extracts and standardizes the process routes and material system parameters, the literature corpus unit transforms the standardized parameter set into first semantic formulation corpus, and the literature-side model fine-tuning unit uses this to efficiently fine-tune the parameters of the formulation domain vertical model. The literature-side inference unit then generates recommended formulation instructions, including precursor components, interface modification molecules, additives, passivators, and corresponding preparation parameters, based on the fine-tuned formulation domain vertical model and given target performance conditions. The literature-side evaluation unit performs multi-dimensional quality scoring on the recommended formulation instructions, and the literature-side optimization unit feeds back the scoring results to the model for preference alignment optimization.

[0113] After the recommended recipe instruction enters the agent unit of the intelligent agent isolation device, the process rule verification module performs a comprehensive physical boundary and timing logic verification on it. The verified recommended recipe instruction is transmitted to the equipment adaptation unit, which translates the semantic recipe into underlying deterministic control instructions, and then distributes them to each actuator via the equipment execution control unit 7.

[0114] After the underlying control command is issued, the system enters the automated physical execution stage. Based on the underlying control command, the gantry transfer mechanism 1 drives the gripper assembly 104 and the pipette assembly 105 to complete the transfer, opening, quantitative liquid aspiration, and closing of the reagent bottle. The gripper assembly 104 transfers the substrate to be treated to the spin coating mechanism 3. After vacuum adsorption and fixation, the pipette assembly 105 sequentially adds the film precursor solution and antisolvent to complete the spin coating. After spin coating, the gantry transfer mechanism 1 transfers the substrate to be treated to the vacuum-induced crystallization mechanism 4. The vacuum pump 402 reduces the chamber pressure to the target vacuum level to induce crystallization. After crystallization, the substrate is transferred to the annealing mechanism 5 for heat treatment.

[0115] While the annealing heat treatment is underway, the transfer mechanism 602 of the in-situ spectral detection mechanism 6 positions the spectral probe assembly 601 above the test area of ​​the substrate to be treated. The spectral probe assembly 601 continuously acquires dynamic spectral data of the thin film during the annealing process. This in-situ detection captures the evolution of the thin film crystallization kinetics over time, rather than the static information of the final state, and contains first-hand physical evidence of the thin film nucleation and growth mechanism. The dynamic spectral data is collected and organized by the data feedback unit, and together with the thin film image parameters, device performance data, and equipment operating status, it is structured and uploaded to the experimental learning closed-loop module.

[0116] After receiving the returned data, the experimental learning unit initiates the mechanism analysis process. The mechanism generation unit extracts burst nucleation time, fluorescence decay slope, and blue or red shifts of characteristic peaks from the dynamic spectral data. It then combines this with the currently executed interface modification molecule type or additive concentration to infer causal relationships and generate a mechanism text description. The experimental corpus unit integrates the mechanism text description with corresponding formulation parameters and experimental data into a second semantic formulation corpus. The experimental model fine-tuning unit uses this corpus to further fine-tune the formulation domain vertical model, incrementally injecting mechanism knowledge based on real experiments into the model. The experimental inference unit then generates iterative recommended formulation instructions based on the updated model and the mechanism feedback from this experiment, for reference in the next round of preparation.

[0117] The synergistic advantages of the thin film preparation system and its control method driven by the formulation language model and the closed intelligent robot box are reflected in the following aspects. The dual drive of literature-based and experimental closed-loop systems allows the vertical model in the formulation field to benefit simultaneously from the joint shaping of prior knowledge from literature and real experimental mechanistic knowledge. The former provides breadth for formulation exploration, while the latter provides depth and accuracy in mechanistic understanding. The pre-verification mechanism of the intelligent agent isolation device complements the dual-loop formulation optimization mechanism. The verification mechanism ensures system safety under any model capability state, while the dual-loop mechanism continuously improves model capabilities to reduce the probability of out-of-bounds recommendations. The deep integration of the in-situ spectral detection mechanism 6 with the mechanism generation unit ensures that each experiment is not only an accumulation of data points but also a deepening of the understanding of physicochemical mechanisms; with the increase of experimental rounds, the depth of the system's mechanistic understanding and the accuracy of formulation recommendations show a synergistic improvement trend. Figure 9 As shown, during the process of completing over 50,000 robot experiments, the power conversion efficiency of the system showed a significant increasing trend in each stage of the formulation exploration phase. The performance distribution gradually narrowed from a broad and discrete distribution in the first stage to a high-performance range, ultimately reaching a power conversion efficiency of 27.0%. This directly reflects the continuous improvement in formulation recommendation accuracy under the dual-closed-loop collaborative drive. Figure 10 As shown, the distribution of the importance of each formulation variable to device performance evolved from broad and divergent to concentrated in a few key variables as the experiment progressed, indicating that the system's formulation exploration strategy, under the coordination of the language agent module, shifted from stochastic exploration to deterministic optimization. Figure 11As shown, the open-circuit voltage, short-circuit current density, and fill factor all exhibit a progressively increasing trend in the four optimization stages. This corresponds to the sequential improvement in interface recombination, film quality, and stress state brought about by the introduction of different formulation variables. From the perspective of device performance, this verifies the practical effect of the mechanism generation unit in feeding back the vertical model in the formulation domain and driving continuous formulation optimization through mechanism text description. Furthermore, the three-layer distributed deployment architecture places high-computing training, real-time inference verification, and deterministic control tasks at the most suitable computational levels, enabling the entire system to maintain good scalability while balancing the flexibility of artificial intelligence with the reliability of industrial control.

[0118] This article uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. It should be noted that those skilled in the art can make several improvements and modifications to the present invention without departing from the principles of the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

Claims

1. A thin film preparation system driven by a formulation language model and a closed intelligent robot box, characterized in that: The formulation language model drives a closed-loop intelligent robotic box-based thin film preparation system, which includes: The robot hardware execution device includes a closed execution chamber and a consumable storage mechanism, a gantry transfer mechanism, a capping clamping mechanism, a spin coating mechanism, a vacuum-induced crystallization mechanism, an annealing mechanism, and an in-situ spectral detection mechanism disposed inside the closed execution chamber. An intelligent cognitive device is communicatively connected to the robot hardware execution device. The intelligent cognitive device is equipped with a formulation domain vertical model, a language intelligent agent module, a literature learning closed-loop module, and an experimental learning closed-loop module. The language intelligent agent module is configured to coordinate the data interaction between the literature learning closed-loop module and the experimental learning closed-loop module. The literature learning closed-loop module is configured to automatically download vertical domain literature through data mining, extract formulation and process parameters from it, and generate a first semantic formulation corpus. The first semantic formulation corpus is then used to fine-tune the formulation domain vertical model. The experimental learning closed-loop module includes a mechanism generation unit, which is configured to receive dynamic spectral data collected by the in-situ spectral detection mechanism during the heating process of the annealing mechanism, associate the dynamic spectral data with the currently executed formulation and process parameters, generate a mechanistic text description explaining the experimental phenomena, and convert the mechanistic text description into a second semantic formulation corpus. The second semantic formulation corpus is then used to fine-tune the formulation domain vertical model. An intelligent agent isolation device is communicatively connected to the intelligent cognitive device and the robot hardware execution device. The intelligent agent isolation device is equipped with a process rule verification module. The intelligent agent isolation device is configured to receive recommended formula instructions output by the vertical model of the formula domain. After the process rule verification module confirms that the recommended formula instructions meet the physical boundary constraints, it parses the recommended formula instructions into low-level control instructions and sends them to the robot hardware execution device.

2. The thin film preparation system of a closed intelligent robot box driven by a formulation language model according to claim 1, characterized in that: The gantry transfer mechanism includes an X-axis moving assembly, a linearly movable Y-axis moving assembly mounted on the X-axis moving assembly, multiple independently linearly movable Z-axis moving assemblies mounted on the Y-axis moving assembly, a gripper assembly mounted on at least one Z-axis moving assembly, and a pipette assembly mounted on at least one Z-axis moving assembly. The gripper assembly is configured to transfer the reagent bottle to the capping clamping mechanism for fixation and to grasp the cap of the removed reagent bottle. The pipette assembly is configured to extend into the reagent bottle to aspirate liquid and reset after aspiration. The gripper assembly is also configured to close the cap after the pipette assembly resets. The vacuum-induced crystallization mechanism includes a vacuum crystallization cavity and a vacuum pump; the gantry transfer mechanism is configured to transfer the substrate to be treated into the vacuum crystallization cavity after the spin coating mechanism completes the basic film formation of the substrate to be treated; the vacuum pump is configured to reduce the pressure in the vacuum crystallization cavity to a target vacuum level; the gantry transfer mechanism is also configured to transfer the substrate to be treated into the annealing mechanism after crystallization is completed; The in-situ spectral detection mechanism includes a spectral probe assembly and a transfer mechanism for driving the spectral probe assembly to translate and move up and down. The transfer mechanism is configured to retract the spectral probe assembly to a standby position during non-detection periods, and to move the spectral probe assembly to the area to be tested above the substrate after it enters the annealing mechanism. The spectral probe assembly is configured to continuously acquire steady-state photoluminescence or absorption spectra of the thin film on the substrate during the annealing process as the dynamic spectral data.

3. The thin film preparation system of a closed intelligent robot box driven by a formulation language model according to claim 1, characterized in that: The literature learning closed-loop module includes a literature learning unit, a parameter generation unit, a literature corpus unit, a literature-side model fine-tuning unit, a literature-side inference unit, a literature-side evaluation unit, and a literature-side optimization unit. The literature learning unit is used to collect process routes and material system data from external databases. The parameter generation unit is used to extract and construct a set of formulation and process parameters from the material system data. The literature corpus unit is used to convert the set of formulation and process parameters into the first semantic formulation corpus. The literature-side model fine-tuning unit is used to perform model fine-tuning based on the first semantic formulation corpus. The literature-side inference unit is used to generate the recommended formulation instruction based on prior knowledge. The literature-side evaluation unit is used to score the quality of the generated formulation. The literature-side optimization unit is used to optimize the vertical model of the formulation domain based on the scoring results. The experimental learning closed-loop module further includes a data feedback unit, an experimental learning unit, an experimental corpus unit, an experimental-side model fine-tuning unit, an experimental-side inference unit, an experimental-side evaluation unit, and an experimental-side optimization unit. The data feedback unit is used to transmit physical representation data and equipment operating status back to the system. The experimental learning unit is used to receive the transmitted data and coordinate with the mechanism generation unit to process the data. The experimental corpus unit is used to construct the generated mechanism text description into the second semantic recipe corpus. The experimental-side model fine-tuning unit is used to perform model fine-tuning based on the second semantic recipe corpus. The experimental-side inference unit is used to generate iterative recommended recipe instructions based on experimental feedback. The experimental-side evaluation unit is used to score the experimental mechanism and iterative recipe. The experimental-side optimization unit is used to update the weight parameters of the recipe domain vertical model based on the scoring results. The intelligent agent isolation device includes an agent unit and a device adaptation unit. The process rule verification module is deployed in the agent unit to perform pre-boundary verification and interception. The device adaptation unit is used to parse and translate the verified recommended recipe instructions into the underlying control instructions. The robot hardware execution device also includes a device execution control unit, which is configured to receive the underlying control commands issued by the device adapter unit and drive the corresponding mechanism actions.

4. A control method for a thin film preparation system driven by a closed intelligent robot box based on the formulation language model according to any one of claims 1-3, characterized in that: The method includes the following steps: S1. The language intelligent agent module coordinates the data interaction between the literature learning closed-loop module and the experimental learning closed-loop module, and generates recommended formula instructions containing formula and process parameters based on the fine-tuned formula domain vertical model. S2. The intelligent agent isolation device receives the recommended formula instruction. After the process rule verification module confirms that the recommended formula instruction conforms to the physical boundary constraints, it parses the recommended formula instruction into a low-level control instruction and sends it to the robot hardware execution device. S3. The robot hardware execution device, according to the underlying control command, sequentially completes opening the lid, reagent absorption, spin coating film formation on the substrate to be treated in the spin coating mechanism, vacuum crystallization on the substrate to be treated in the vacuum induced crystallization mechanism, and heat treatment on the substrate to be treated in the annealing mechanism. S4. During the heat treatment process of the annealing mechanism, the in-situ spectral detection mechanism collects dynamic spectral data and transmits the physical characterization data containing the dynamic spectral data back to the experimental learning closed-loop module. S5. The mechanism generation unit of the experimental learning closed-loop module transforms the physical representation data into a mechanism text description explaining the experimental phenomenon, and transforms the mechanism text description into a second semantic recipe corpus, and uses the second semantic recipe corpus to fine-tune the recipe domain vertical model.

5. The control method for a thin film preparation system driven by a formulation language model and a closed intelligent robot box according to claim 4, characterized in that: The literature learning closed-loop module includes a literature learning unit, a parameter generation unit, a literature corpus unit, and a literature-side model fine-tuning unit. Before step S1, the method further includes the following steps: The literature learning unit collects process routes and material system information from external databases; The parameter generation unit extracts and constructs a set of formulation and process parameters from the material system data. The variables in the set of formulation and process parameters include precursor components, interface modification molecules, additives, passivating agents and corresponding preparation parameters. The document corpus unit transforms the set of formulas and process parameters into the first semantic formula corpus; The document-side model fine-tuning unit performs model fine-tuning based on the first semantic formula corpus.

6. The control method for a thin film preparation system driven by a formulation language model and a closed intelligent robot box according to claim 4, characterized in that: The intelligent agent isolation device includes a proxy unit and a device adaptation unit. The process rule verification module is deployed in the proxy unit. Step S2 specifically includes: The agent unit, which is equipped with the process rule verification module, extracts the specific process parameters from the recommended formula instruction and calls the preset task whitelist. The specific process parameters extracted are compared to see if they exceed the physical limits of the robot hardware execution device, and whether timing conflicts occur between the various action commands. If the physical limit boundary is exceeded or the timing conflict is triggered, the agent unit intercepts the recommended recipe instruction, generates an out-of-bounds error message, and triggers a regeneration instruction. If the verification passes, the device adaptation unit will translate the semantic action mapping in the recommended recipe instruction into the underlying control instruction that can be executed by the robot hardware execution device.

7. The control method for a thin film preparation system driven by a formulation language model and a closed intelligent robot box according to claim 4, characterized in that: The gantry transfer mechanism is equipped with a gripper assembly and a pipette assembly. In step S3, the spin coating process of the substrate to be treated by the spin coating mechanism specifically includes: The gripper assembly is driven to place the substrate to be processed onto the spin coating mechanism and perform vacuum adsorption fixation. The pipette assembly is driven to quantitatively add the thin film precursor solution; Thin film nucleation control can be achieved by either adding an anti-solvent during spin coating or by vacuum-induced crystal formation after spin coating.

8. The control method for a thin film preparation system driven by a formulation language model and a closed intelligent robot box according to claim 4, characterized in that: In step S5, the mechanism generation unit converts the physical characterization data into a mechanistic text description explaining the experimental phenomenon, specifically including: Extract the burst nucleation time, fluorescence decay slope, and blue or red shift of characteristic peaks from the dynamic spectral data, and combine them with the type of interface modification molecules or the concentration of additives to infer the degree of acceleration of heterogeneous nucleation or the release state of residual stress. Receive diffraction data, extract the peak intensity and full width at half maximum (FWHM) of characteristic precursor residues or intermediate phases, and map them to generate a textual description of the mechanism of additive regulation of crystallization or surface passivation agent release of lattice stress. Receive thin film images, extract film coverage and grayscale values, and map them to generate a mechanistic text description of solvent removal and supersaturation kinetics.

9. The control method for a thin film preparation system driven by a formulation language model and a closed intelligent robot box according to claim 4, characterized in that: The literature learning closed-loop module includes a literature-side evaluation unit and a literature-side optimization unit. After the recommended recipe instruction is generated in step S1, the literature-side evaluation unit scores the quality of the generated recipe, and the literature-side optimization unit optimizes the vertical model of the recipe domain based on the scoring results. The experimental learning closed-loop module also includes an experimental evaluation unit and an experimental optimization unit. After the mechanism text description is generated in step S5, the experimental evaluation unit scores the experimental mechanism and iterative formulation. The experimental optimization unit updates the weight parameters of the formulation domain vertical model based on the scoring results.

10. The control method for a thin film preparation system driven by a formulation language model and a closed intelligent robot box according to claim 4, characterized in that: This method is based on an edge-cloud collaborative computing architecture for distributed deployment and execution, specifically including: The process of fine-tuning the vertical model of the formulation domain using the first semantic formulation corpus and the second semantic formulation corpus is deployed on a cloud server to perform high-performance training and large-scale data processing. The reasoning process of the language intelligence module generating the recommended recipe instruction and the verification process of the process rule verification module confirming physical boundary constraints are deployed on edge computing devices to perform real-time joint debugging of local model reasoning verification and pre-boundary verification interception. The process of parsing and issuing the underlying control commands is deployed on a local industrial control host to achieve deterministic real-time control of the robot hardware execution device.

Images