AI-driven x-ray spectroscopy experiment active optimization and decision system and method

Abstract

The application discloses an AI-driven X-ray spectroscopy experiment active optimization and decision system and method, and belongs to the technical field of X-ray analysis. The system comprises a spectroscopy measurement hardware, a standardized equipment abstract interface layer, a real-time spectrum feature extraction module, a first-principle knowledge graph database and a meta-decision fusion engine. The meta-decision fusion engine fuses the macro-path planning ability of deep reinforcement learning and the micro-parameter search ability of Bayesian optimization, and embeds a physical knowledge graph as a constraint, can dynamically evaluate the scientific value according to the real-time measured spectrum data, autonomously decide the optimal next measurement action (such as adjusting the energy step, switching the sample or changing the environmental conditions), realize the paradigm shift from "passive script execution" to "active optimization analysis". The application significantly improves the measurement efficiency and scientific discovery probability of X-ray spectroscopy analysis, and is especially suitable for complex measurement scenes such as catalyst screening and in-situ phase change dynamics.

Description

Technical Field

[0001] This invention belongs to the field of intelligent control technology for high-precision scientific instruments, specifically designing an AI-driven active optimization and decision-making system and method for X-ray spectroscopy experiments. This invention achieves active parameter optimization and real-time command decision-making for synchrotron radiation sources and laboratory X-ray spectroscopy measurement equipment through a standardized hardware interface. Background Technology

[0002] X-ray spectroscopy (such as X-ray absorption fine structure spectroscopy (XAFS) and X-ray emission spectroscopy (XES)) is a key characterization method for resolving the local electronic structure and geometric configuration of materials, and has wide applications in catalysis, energy materials, environmental science, and other fields. Currently, whether it is a spectroscopic beamline relying on large scientific facilities such as synchrotron radiation facilities or a benchtop spectrometer in the laboratory, the experimental mode generally follows an open-loop process of "preset parameters - static execution - offline analysis". Measurement parameters (such as energy range, scan step size, and sampling time) are statically set entirely based on the researcher's prior knowledge. For example, in XANES measurements, the energy region is usually roughly divided into the front edge region, the edge-edge region, and the back edge region, and a fixed step size is set for each. This coarse-grained division method is very likely to cause the fine features in front of the edge to be completely missed, or to cause distortion of the characteristic peak shape due to improper step size settings.

[0003] In recent years, several technological explorations have emerged in this field to improve the efficiency of scientific discovery. On the one hand, automation and high-throughput technologies have been initially applied. Driven by the Materials Genome Initiative, researchers have established "experimental factories" that integrate robots, automated characterization, and batch data processing, enabling the rapid preparation and screening of large numbers of samples. For example, the Computer Network Information Center of the Chinese Academy of Sciences developed a high-throughput integrated computing and data management method and system for materials genome engineering (CN106372329A), realizing a high-throughput automated workflow for first-principles calculation tasks. Shanghai University proposed a method to accelerate the construction of a multi-component materials thermodynamic database (CN119069042A), achieving an automated workflow for database construction through an active learning strategy. However, the "intelligence" of such systems is mainly reflected in parallelization and workflow automation; their experimental sequences are usually driven by fixed scripts and lack the ability to perform real-time, global strategy optimization based on intermediate results.

[0004] On the other hand, artificial intelligence is beginning to be introduced into spectroscopic data analysis and experimental design. In spectral analysis, the University of Electronic Science and Technology of China proposed a Bayesian regularization-based method for quantitative analysis of heavy metals in soil (CN113866204A), which utilizes Bayesian regularization to optimize the hyperparameters of a BP neural network, improving the accuracy of quantitative analysis using X-ray fluorescence spectroscopy. In spectral simulation, Guo et al. established a first-principles calculation database containing 2681 sulfur K-edge XAS spectra, covering 66 structural models of crystalline and glassy lithium thiophosphate, making it one of the largest open-source computational spectroscopic databases in this field (Guo H, et al. Simulated sulfur K-edge X-ray absorption spectroscopy database of lithium thiophosphate solid electrolytes. Sci Data 10, 349, 2023). In the field of adaptive sampling and measurement design, Bayesian optimization methods have made significant progress. Ranjan et al. applied Bayesian optimization to fit the crystal field multistate theory parameters of Kβ X-ray emission spectra, reliably optimizing four key parameters, including the Slater integral reduction factor and crystal field splitting, significantly improving the accuracy of XES analysis (Ranjan R, et al. Kβ X-ray Emission Spectra Analysis Using Bayesian Optimization. Photon Science 2026 1 (1), 12-23). ​​Regarding the application of deep reinforcement learning for experimental control, Niezen et al. applied deep reinforcement learning to liquid chromatography method development, studying the impact of reward function design and experimental budget. The results showed that in high-dimensional parameter tasks such as five-segment gradient program optimization, the success rate of the PPO agent (31%) was superior to Bayesian optimization (24%) (Niezen LE, et al. Reinforcement learning for automated method development in liquidchromatography: insights in the reward scheme and experimental budget selection. Journal of Chromatography A, 2025, 1748: 465845).

[0005] However, when these technological advancements are applied to X-ray spectroscopy measurements, they still suffer from a significant "triple disconnect" problem, failing to achieve true closed-loop intelligence. First, there is a disconnect between decision-making and execution. Most existing AI algorithms, including Bayesian optimization and reinforcement learning methods, typically operate as "offline proposers," requiring manual interpretation and setup of the optimization results onto the device, or batch execution via offline scripts. Even those that achieve online adaptive sampling limit their decision-making to the selection of measurement points for a single spectral line, failing to form a closed loop with macroscopic measurement strategies (such as sample switching and changes in environmental conditions). The system cannot establish a millisecond-level real-time control closed loop with the measurement hardware, making it difficult to dynamically respond to emerging phenomena during measurement, such as transient intermediates or non-equilibrium phase transitions in catalytic reactions.

[0006] Second, there is a disconnect between knowledge and data. There is a lack of organic connection between the massive amounts of first-principles calculations and real-time measurement data. Although large-scale computational spectroscopic databases have been constructed, and machine learning models based on computational data have been developed, these knowledge bases are typically static and offline, unable to be invoked in real-time during measurements to constrain and guide the AI's search direction. Purely data-driven optimization algorithms may lead to search paths with ambiguous physical meaning or even violate fundamental laws, and their decision-making processes are difficult for scientists to understand and trust.

[0007] Third, there is a disconnect between macroscopic and microscopic perspectives. Existing optimization algorithms typically address only single-level problems, such as optimizing acquisition parameters for a single spectral line or control strategies for a single measurement process. However, they lack a unified framework capable of handling multi-level decision-making tasks simultaneously. In complex materials exploration tasks, the system needs to determine not only "the next measurement point for the current spectral line" at the microscopic level but also "when to switch samples," "when to change reaction conditions," and "when to switch from coarse scanning mode to fine scanning mode" at the macroscopic level. The timescales and objective functions of these two types of decisions differ, making it difficult for a single algorithm to simultaneously address both.

[0008] Therefore, current X-ray spectroscopy measurement methods, when faced with a multidimensional and complex parameter space composed of composition, structure, environment, and time, still heavily rely on researchers' experience and trial and error. This results in a significant waste of valuable machine time on acquiring redundant data with low information entropy, easily leading to the omission of crucial scientific information. Deeply embedding artificial intelligence into the real-time control loop of the measurement process, and developing autonomous intelligent measurement systems with closed-loop capabilities of "perception-cognition-decision-execution," has become a core competitive area for breaking through existing research models and seizing the commanding heights of the next generation of scientific discovery. Summary of the Invention

[0009] This invention aims to provide an AI-driven active optimization system and method for X-ray spectroscopy experiments. By constructing a real-time hardware measurement and control closed loop, it solves the problems of rigid control logic, slow response of heterogeneous hardware, and low resource utilization in traditional experimental equipment.

[0010] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides an AI-driven active optimization and decision-making system for X-ray spectroscopy experiments, comprising: The spectroscopic measurement hardware subsystem performs specific physical measurement operations. This hardware includes an X-ray source actuator (such as a monochromator motor), a multi-degree-of-freedom precision sample environment unit (such as a variable-temperature, variable-pressure in-situ cell), and signal acquisition and detection components. A standardized device abstraction interface layer connects bidirectionally to the hardware subsystem and has a built-in multi-protocol adapter. Its core function is to encapsulate heterogeneous hardware operations (such as specific brand driver pulse settings) into standardized atomic measurement and control commands, and to map the physical feedback from hardware sensors (such as count rate and voltage signals) into real-time status signals in a unified format.

[0011] The real-time spectral feature extraction module performs online denoising, normalization, and physical feature vectorization on the state signal, generating feature vectors (including edge peak positions, white line peak intensities, etc.) representing the current measurement and control state in real time. A physically constrained intelligent inference engine incorporates a spectral-structure mapping model based on a first-principles knowledge graph. This engine infers the microstructure assumptions and their confidence levels of the current sample in real time based on the feature vectors, providing hard constraints on the physical search boundaries for the decision-making system.

[0012] The meta-decision fusion engine serves as the core control hub of the system. Based on feature vectors, physical constraint parameters, and preset technical measurement and control objectives, it generates the next hardware control strategy in real time using a hierarchical decision-making strategy that integrates deep reinforcement learning (DRL) and Bayesian optimization (BO). The closed-loop feedback execution module connects the engine to the interface layer. It is responsible for translating the abstract control strategy into specific low-level drive pulses or control command sequences in real time, driving the hardware subsystems to execute actions, thus forming a complete millisecond-level measurement and control closed loop.

[0013] Furthermore, the meta-decision fusion engine includes: a deep reinforcement learning macro-planning unit, used to output macro-hardware actions based on long-period observation states. Actions include, but are not limited to: automatic displacement switching between samples, adjustment of the reaction atmosphere ratio, or switching from "coarse scan mode" to "feature region fine scan mode." A Bayesian optimization micro-search unit, responding to macro-actions, uses Gaussian process regression to calculate the optimal micro-observation control parameters in real time within a specific parameter space. Parameters include: the non-uniform step size distribution of the energy scan and the detector's adaptive integration time window.

[0014] Furthermore, the deep reinforcement learning macro-planning unit integrates a multi-objective reward function calculation module; the multi-objective reward function calculation module is used to calculate the reward signal R according to the following formula based on the physical information entropy increment of the newly acquired data, the hardware execution cycle duration, and the signal noise level: R=W1×S+W2×(1 / T)-W3×A, where S is the physical information entropy increment, T is the hardware execution cycle duration, A is the signal noise level, and W1, W2, and W3 are preset technical weight coefficients.

[0015] Furthermore, the standardized device abstract interface layer is provided with a multi-protocol adapter, which is logically connected to the communication protocol stack of the interface layer and supports at least one of EPICS, Tango, Modbus and TCP / IP protocols to realize the communication adaptation between decision signals and underlying physical hardware.

[0016] Furthermore, the physically constrained guided intelligent inference engine is configured with an experimental feedback interface, which is set at the database input end of the engine and connected to an external data verification end. This interface is used to feed back the verified measured spectrum data to the intelligent inference engine to update the spectrum-structure mapping model parameters.

[0017] Furthermore, the real-time spectral feature extraction module integrates a hardware state anomaly monitoring unit. The hardware state anomaly monitoring unit is located in the signal monitoring layer of the extraction module and is used to calculate the signal saturation in real time and send a hardware interruption protection signal to the meta-decision fusion engine when a signal overload feature is detected.

[0018] Secondly, this invention provides an active optimization and decision-making method for X-ray spectroscopy experiments based on the above system, comprising the following steps: S1. Hardware initialization and state pre-scan: Control the spectroscopic measurement hardware subsystem to perform a fast global scan of the target under test to obtain initial physical state distribution data; S2. Real-time state perception and physical mapping: The real-time spectrum feature extraction module processes the feedback signal to generate a feature vector, and the intelligent inference engine is used to map the current physical perception range. S3. Macroscopic Measurement and Control Decision Generation: Utilizing the deep reinforcement learning macroscopic planning unit in conjunction with the current physical cognition interval, output a macroscopic hardware action control sequence; S4. Online optimization of micro-control parameters: The Bayesian optimization micro-search unit is used to calculate and issue the optimal single-point motion step size or time integral parameter online within the macro-action range. S5. Hardware closed-loop execution and policy feedback: Convert macroscopic actions and microscopic parameters into atomic instructions and send them to the spectroscopic measurement hardware subsystem for execution, and calculate reward values ​​based on the newly collected data to update the reinforcement learning policy network.

[0019] Furthermore, in step S4, when the measurement target is to capture transient physical processes, the Bayesian optimized microscopic search unit uses time resolution as an optimization constraint and automatically reduces the motion step size of the spectroscopic measurement hardware subsystem when the rate of change of the feature signal exceeds a preset threshold.

[0020] Furthermore, the method is equipped with a measurement and control termination judgment logic: when the inference confidence of the measured physical structure reaches a preset threshold, or the volatility of the real-time feature vector is lower than a preset convergence threshold, the execution of loop steps S2 to S5 is forcibly stopped.

[0021] Furthermore, in step S2, if the hardware status anomaly monitoring unit detects that the signal response delay or intensity is lower than a preset safety threshold, it will forcibly skip steps S3 and S4, and the closed-loop feedback execution module will directly issue an emergency shutter closing command to the spectral measurement hardware subsystem.

[0022] Compared to existing technologies, this invention offers the following advantages: 1. It achieves autonomous control in the field of scientific instruments: By directly embedding AI decision-making into the hardware measurement and control loop, the system possesses the ability to dynamically intervene in the underlying hardware based on the evolution of physical signals, significantly improving the accuracy of capturing non-equilibrium and transient phase transition processes. 2. It significantly reduces beam resource waste: Through a hierarchical decision-making mechanism, the system can autonomously identify and quickly pass through "valueless energy regions," while automatically increasing sampling density in "high-value feature regions," resulting in experimental efficiency more than 10 times higher than traditional methods. 3. It solves the problem of collaborative control of heterogeneous equipment: The standardized interface layer design achieves complete decoupling between the algorithm and hardware, enabling the system to be quickly ported to different energy ranges and types of spectroscopic beamlines. Attached Figure Description

[0023] Figure 1 The diagram shows the overall hardware architecture and signal closed-loop connection of the intelligent measurement and control system provided by this invention.

[0024] Figure 2 This is the control logic diagram for the interaction between the meta-decision fusion engine (DRL+BO) and the physical constraint database.

[0025] Figure 3 This is a flowchart illustrating the real-time control process for hardware closed-loop active optimization in this invention. Detailed Implementation

[0026] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the protection scope of the present invention.

[0027] Example 1 This embodiment provides an AI-driven active optimization system for X-ray spectroscopy experiments. (Refer to...) Figure 1 The AI-driven X-ray spectroscopy experimental active optimization and decision-making system of this invention consists of six main parts: a spectroscopic measurement hardware subsystem, a standardized equipment abstract interface layer, a real-time spectral feature extraction module, a physically constrained guided intelligent inference engine, a meta-decision fusion engine, and a closed-loop feedback execution module. The standardized equipment abstract interface layer, serving as the communication hub of the entire system, is located on the communication link between the intelligent computing terminal and the physical hardware. Its internally nested multi-protocol adapter is located in the underlying kernel of this interface layer and achieves bidirectional communication with the spectroscopic measurement hardware subsystem via an Ethernet or fiber optic physical link. The spectroscopic measurement hardware subsystem is located at the bottom controlled end of the standardized equipment abstract interface layer and is arranged in the physical space of the experimental beamline according to the optical path sequence. Specifically, it includes components located at the physical beginning of the beamline and used for adjustment... The system comprises an X-ray source assembly for controlling the X-ray wavelength, an adjustable sample environment assembly located downstream of the source assembly and used to carry the sample under test and adjust the temperature and pressure conditions, a spectrometer and detector assembly located on the light path of the sample environment assembly with the detection window aligned with the sample to convert photoelectric signals, and a low-level device controller located at the electrical drive end of each of the above components and used to execute atomic measurement and control commands. The real-time spectral feature extraction module is connected to the data output end of the standardized device abstraction interface layer to perform feature vectorization processing on the state signal stream fed back by the hardware. Furthermore, a hardware state anomaly monitoring unit is integrated in the signal bypass position within this module. This unit is used to calculate the detector count rate saturation in real time and send an emergency interrupt command to the downstream meta-decision fusion engine when an overload feature is detected, thereby protecting the precision hardware from damage.

[0028] Reference Figure 2The meta-decision fusion engine of this invention achieves deep interaction with a physically constrained intelligent inference engine through a closed-loop feedback loop. The physically constrained intelligent inference engine is located at the logic output of the real-time spectral feature extraction module. Internally, it includes a first-principles knowledge graph database storing massive density functional theory calculated spectra, a spectrum-structure mapping engine electrically connected to the database output, and an experimental feedback interface located at the database read / write interface and externally connected to the verification terminal. The meta-decision fusion engine is connected to the parameter output of the intelligent inference engine and the feature output of the feature extraction module via signal lines. Its internal local structure is further refined into a top-level decision position within the engine used for output sample switching. The system includes a deep reinforcement learning (DRL) macro-planning unit for switching scanning modes, a multi-objective reward function calculation module attached to the DRL unit's policy evolution loop for calculating the scientific gain reward signal, and a Bayesian optimization (BO) micro-search unit connected in series at the DRL unit's action output port for online correction of the energy step size and integral window based on Gaussian process regression. The closed-loop feedback execution module is set on the feedback instruction path between the meta-decision fusion engine and the standardized device abstraction interface layer. It is used to reorganize the control strategy generated by the above units into a pulse sequence that can be directly recognized by the underlying device controller, and send it to the spectroscopic measurement hardware subsystem in real time through the physical interface of the interface layer, realizing millisecond-level mapping from decision scheme to physical action.

[0029] Specifically, (a) The spectroscopic measurement hardware subsystem is the foundation for the system to perform physical measurements, including: X-ray source components: These can be synchrotron radiation beamlines or laboratory X-ray tubes, used to generate X-ray beams with adjustable energy or continuous wavelength.

[0030] Adjustable sample environment component: Located in the optical path, this component carries the sample. It integrates a high-precision temperature controller, a pressure controller, and a mass flow meter, providing precise and controllable reaction conditions for the sample, ranging from low temperatures (e.g., 2.5K) to high temperatures, and from vacuum to atmospheric environments (e.g., H2, O2, Ar).

[0031] Spectroscopic and Detection Components: Used for energy analysis and intensity detection of X-rays after they have interacted with the sample. For example, for XAFS, this may include a monochromator and an ionization chamber / solid-state detector; for XES, this may include a crystal analyzer and a detector.

[0032] The underlying device controller includes a motor controller, a temperature controller, a counter calibrator, etc., which are electrically connected to the execution and sensing parts of the above components. It is used to receive underlying instructions and drive hardware actions, while also providing feedback on the hardware status.

[0033] (ii) The standardized equipment abstract interface layer is a bridge connecting the virtual world and the physical world.

[0034] The standardized equipment abstraction interface layer incorporates a multi-protocol adapter, supporting protocols such as EPICS (a common protocol for large scientific facilities), Tango, Modbus, or PLC. Through this adapter, heterogeneous communication interfaces from underlying device controllers (such as motors from different brands) are uniformly encapsulated into a standardized set of RESTful APIs or message queue interfaces. Atomic commands such as "move_sample(x, y)" issued by the decision layer are translated by this interface layer into serial port commands or network data packets that the specific controller can understand. Simultaneously, raw data such as voltage, temperature, and position fed back by hardware sensors are mapped into a unified JSON-formatted status data stream and broadcast to upper-layer modules. This layer's design decouples the core AI logic from specific hardware brands and models, which is crucial for the system's adaptability to various spectrometers (such as XAFS, XES, and XRD).

[0035] (III) Real-time Spectral Feature Extraction Module The real-time spectral feature extraction module connects to the standardized device abstraction interface layer, continuously receiving raw spectroscopic data streams. Internally, it runs a lightweight real-time data processing pipeline: 1. Preprocessing: Denoise the raw data (e.g., using Savitzky-Golay filters), remove background, and normalize the data.

[0036] 2. Feature Extraction: A specialized algorithm library is invoked to identify and quantify key physical features online from the preprocessed spectrum. Taking XAFS as an example, extracted features may include: the energy position and integral intensity of edge-front peaks, the peak energy and full width at half maximum (FWHM) of white line peaks, and the oscillation frequency of the extended X-ray absorption fine structure spectrum. These features are combined into a low-dimensional "feature vector" as a digital representation of the current measurement state.

[0037] 3. Anomaly Detection: This module also includes an anomaly detection unit based on statistical or lightweight models. It monitors the signal-to-noise ratio of the spectrum, the rationality of characteristic peaks, and the hardware status data stream in real time. Once a sharp decline in data quality, the appearance of unexplained spectral peaks, or a hardware fault alarm is detected, it immediately sends an interrupt signal to the meta-decision fusion engine.

[0038] (iv) The first-principles knowledge graph database is the "physical brain" of the system, endowing AI with scientific common sense. Its construction and operation mechanism is as follows: 1. Theoretical Calculation Data Layer: Utilizing first-principles methods such as density functional theory, a massive amount of theoretical X-ray absorption or emission spectra are pre-calculated for experimentally relevant elements and systems (such as Fe-NC catalysts). Each theoretical spectrum is precisely associated with its corresponding atomic structure model (such as the coordination number, bond length, and spin state of Fe), electronic structure information, and calculation parameters.

[0039] 2. Spectrum-Structure Mapping Engine: This is a deep learning model trained on a graph neural network or convolutional neural network. It learns a nonlinear mapping relationship from spectral feature vectors to microstructure parameters. When it receives a "feature vector" from the real-time spectral feature extraction module, the mapping engine quickly infers and outputs the most probable microstructure hypothesis for the current sample and its confidence level (e.g., "The target Fe site has an 85% probability of being tetracoordinated with a valence state of +2.5"). This hypothesis provides a crucial physical constraint for subsequent decisions.

[0040] 3. Dynamic Evolution Interface: The database is not static. The system includes a feedback interface to receive high-quality, experimentally validated data verified through closed-loop experiments. This data (especially data from complex environments that are difficult to accurately simulate through theoretical calculations) is cleaned and labeled manually or automatically before being incorporated into the database for incremental learning or fine-tuning of the spectrum-structure mapping engine. This allows the system's ability to understand real-world complex systems to continuously evolve with the number of measurements.

[0041] (v) The meta-decision fusion engine is the core decision-making unit of the system, such as Figure 2 As shown, it employs a unique two-layer architecture internally: 1. Deep Reinforcement Learning (DRL) Macro-planning Unit: This unit resides above the decision-making level, responsible for the "strategic" aspect. It models the entire measurement process as a Markov decision process. Its state space consists of real-time feature vectors, structural hypothesis confidence, and the coverage of explored regions. The action space is a set of macro-level measurement operations, such as switching to the next candidate sample, entering the variable-temperature fine-scan mode for the current sample, and changing the reaction atmosphere. Its policy network (such as DuelingDQN or PPO) learns to output the optimal macro-level action based on the current state through pre-training and online learning. The learning is driven by the reward signal provided by the multi-objective reward function calculation module, which integrates scientific information gain (such as the emergence of new features), goal achievement (such as the degree of matching with the target theoretical spectrum), and time cost. This macro-planning capability enables the system to explore extensively in the early stages of measurement and focus on utilization in the later stages, effectively solving the problem that pure local search methods such as Bayesian optimization are difficult to use for long-term sequence planning.

[0042] 2. Bayesian Optimization (BO) Micro-Search Unit: This unit, controlled by the macro-planning unit, is responsible for the "tactical" level. The BO unit is activated when the DRL unit decides to perform the macro-action of "performing a fine XANES scan at the current temperature." The BO unit's task is to find the optimal micro-measurement parameters (such as energy step size and integration time) that maximize information acquisition efficiency within the parameter space defined by this action (such as the energy range E_min to E_max). It utilizes Gaussian process regression to construct a posterior probability model of the objective function (such as a combined index of spectral smoothness and signal-to-noise ratio) based on the small number of data points already acquired in the current scan. Then, it determines the next measurement point with the most information by maximizing the acquisition function (such as the desired increase in EI). This allows the system to automatically reduce the step size and increase the integration time in regions with severe spectral oscillations to obtain high-precision data during the fine scanning phase, and vice versa in smooth regions, achieving an adaptive balance between data quality and acquisition speed.

[0043] (vi) The closed-loop feedback execution module is the final output port of the decision. It receives composite instructions from the meta-decision fusion engine (e.g., macroscopic action = "fine scan sample A", microscopic parameter = "dynamic optimization of energy step size") and converts them into an ordered list of atomic instruction tasks. Then, by calling the API of the standardized device abstraction interface layer, it sequentially issues instructions such as "move monochromator to energy E1", "count detector for t seconds", "read data", and "move to the next energy point" to drive the hardware to complete the measurement. The newly acquired data will trigger the next round of perception, cognition and decision-making, forming a continuous "perception-cognition-decision-execution" intelligent closed loop.

[0044] Example 2 This embodiment provides an active optimization method for X-ray spectroscopy experiments based on the system described in Embodiment 1. (Refer to...) Figure 3 The method includes the following steps: The standardized equipment abstraction interface layer sends an initialization signal to the underlying equipment controller, driving the X-ray source component to perform a rapid coarse scan of the sample to establish initial recognition. Subsequently, the real-time spectral feature extraction module processes the detection feedback in real time and extracts feature vectors containing the position of the edge-front peak and the intensity of the white line peak. Simultaneously, the hardware state anomaly monitoring unit uses real-time sensing of the detector saturation to determine whether to skip subsequent steps and directly perform an emergency reset. Then, refer to... Figure 2The meta-decision fusion engine calls the physical constraint parameters in the intelligent inference engine. The DRL macro-planning unit, combined with the gain evaluation of the multi-objective reward function, outputs macro-action instructions focused on specific high-value samples. In response to these instructions, the BO micro-search unit searches for the optimal measurement point in the local parameter space defined by the macro-instruction through the acquisition function. The monochromator motion step size is automatically fine-tuned from the coarse scan state to a precision measurement level of 0.02 eV and below. Finally, the closed-loop feedback execution module sends the optimized atomic instructions to the spectroscopic measurement hardware subsystem for execution through the standardized equipment abstraction interface layer. If there is a deviation between the measured data and the preset confidence level, the above closed-loop process of perception, decision-making and optimization is repeated until the threshold preset by the measurement and control termination judgment logic is met to ensure the beam stability and measurement efficiency of the monochromator during long-term operation.

[0045] The specific steps are as follows.

[0046] S1. Initialization and Global Coarse Scan: After system startup, the closed-loop feedback execution module issues commands through the standardized interface layer to control the hardware to perform a rapid, low-resolution scan of all test samples or a wide parameter space (such as a wide energy range and coarse step size). This stage aims to quickly obtain an overview of the system and provide a data foundation for establishing initial understanding.

[0047] S2. Real-time Spectral Perception and Knowledge Mapping: The real-time spectral feature extraction module processes the coarse scan data online to extract feature vectors. The first-principles knowledge graph database receives these vectors and, through the spectral-structure mapping engine, infers the microstructure hypothesis and confidence level of the current sample, forming the system's "current cognitive state".

[0048] S3. Macroscopic Measurement Path Planning: The DRL unit in the meta-decision fusion engine takes the current cognitive state and the preset scientific goal (such as "finding the most active catalyst") as input, uses its policy network to reason, and outputs the next macroscopic measurement action. For example, when the spectral characteristics of a sample are highly similar to the theoretical structure of high activity, the DRL unit may decide to "focus on this sample for detailed study".

[0049] S4. Microscopic Measurement Parameter Optimization: In response to macroscopic actions, the BO unit in the meta-decision fusion engine is activated. Within the measurement framework defined by the macroscopic actions (such as "fine scanning"), the BO unit constructs a Gaussian process surrogate model based on existing data points and solves the acquisition function to calculate a set of microscopic measurement parameters that maximize information gain in real time (such as the step size of temperature variation, the integration time of spectral acquisition, etc.).

[0050] S5. Closed-Loop Execution and Feedback: The closed-loop feedback execution module integrates macroscopic actions and microscopic parameters, converting them into specific atomic instruction sequences to drive hardware to perform measurements and collect new high-value data. After processing, the new data generates reward values ​​by the multi-objective reward function calculation module and feeds them back to the DRL unit for online updates to its policy network (reinforcement learning). The system then loops through S2 to S5 until the measurement termination condition is met (e.g., the confidence level of the objective hypothesis exceeds 95%, or the measurement time is exhausted).

[0051] In a preferred embodiment of this method, for measurement scenarios capturing transient phase transitions, the Bayesian optimization microscopic search unit in step S4 uses "temporal resolution" as one of its core optimization objectives. The Gaussian process model predicts the rate of change of spectral features over time. In regions with high rates of change (such as when a phase transition occurs), the acquisition function automatically suggests a higher sampling frequency (increasing the number of sampling points); in steady-state regions with low rates of change, it suggests reducing the sampling frequency, thereby achieving adaptive focusing of measurement resources in the time dimension.

[0052] In another preferred embodiment of this method, if the real-time spectral feature extraction module in step S2 detects an anomaly (such as sudden beam loss or a sharp drop in the spectral signal-to-noise ratio), it will trigger an interrupt. The system will jump to an interrupt handling sub-process, which will be re-planned by the meta-decision fusion engine. It may issue instructions such as "pause the current scan and perform diagnostic actions" or "repeat the last measurement to confirm," thereby enhancing the robustness of the system.

[0053] Example 3 To verify the technical effectiveness of this invention, the system and method of this invention were used in an exemplary study on the structure-activity relationship of oxygen reduction active centers in Fe-NC single-atom catalysts. The measurement parameter space includes more than 200 theoretically possible Fe-N coordination configurations.

[0054] Using the traditional "trial and error + full spectrum scanning" mode, it is usually necessary to perform a high-precision full spectrum scan (wide energy range, fine step size) for each configuration, which is expected to take several weeks.

[0055] Using the system of this invention, the system first performs a rapid coarse scan (S1) to initially screen all samples. In cycles S3 and S4, the system autonomously decides to perform a fine scan only on approximately 50 configurations exhibiting high potential activity based on their spectral characteristics. During the scan, the BO unit autonomously optimizes the scan step size and automatically densifies sampling points near key absorption edges. Ultimately, the system successfully identified the three most promising coordination configurations in less than 48 hours, improving measurement efficiency by more than 10 times.

[0056] This process fully demonstrates the beneficial effects of the present invention: by combining strategic sample screening (DRL) with tactical parameter optimization (BO) and using physical knowledge (knowledge graph) as guidance, the system can quickly converge to the scientific objective with the fewest total measurements and total machine time, significantly improving the efficiency and probability of scientific discovery.

[0057] Although embodiments of the present invention have been shown and described above, it is understood that these embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions, and alterations to the above embodiments within the scope of the present invention without departing from its principles and spirit. The scope of protection of the present invention is defined by the claims and their equivalents.

Claims

1. An AI-driven active optimization and decision-making system for X-ray spectroscopy experiments, characterized in that, include: The system includes a spectroscopic measurement hardware subsystem, a standardized equipment abstraction interface layer, a real-time spectral feature extraction module, a physically constrained intelligent inference engine, a meta-decision fusion engine, and a closed-loop feedback execution module. The standardized equipment abstract interface layer is bidirectionally connected to the spectroscopic measurement hardware subsystem, and is used to encapsulate heterogeneous hardware communication protocols into standardized atomic measurement and control commands. The spectroscopic measurement hardware subsystem is located at the control end of the standardized equipment abstract interface layer and is used to perform X-ray measurement actions and output raw physical signal streams. The real-time spectral feature extraction module is connected to the signal output end of the standardized device abstract interface layer, and is used to perform online feature vectorization processing on the received state signal stream and generate feature vectors. The physical constraint-guided intelligent inference engine is set at the logic output of the real-time spectral feature extraction module and is used to infer the physical constraint parameters of the sample under test based on the feature vector. The meta-decision fusion engine includes a deep reinforcement learning macro-planning unit and a Bayesian optimization micro-search unit. The deep reinforcement learning macro-planning unit is located in the decision control layer of the meta-decision fusion engine and is used to output macro-hardware actions based on the current state. Its specific process includes sample station switching signals, temperature control field step signals, or scanning mode switching commands. The Bayesian optimization micro-search unit is connected to the execution feedback end of the deep reinforcement learning macro-planning unit and is used to calculate the optimal monochromator motion step size and detector integration window duration using Gaussian process regression within the parameter space defined by the macro-action. The meta-decision fusion engine is electrically connected to the feature output end of the real-time spectrum feature extraction module and the parameter output end of the intelligent inference engine, respectively, and is used to generate hardware control strategies in real time based on the feature vector, physical constraint parameters, and preset technical measurement and control indicators. The closed-loop feedback execution module is connected between the meta-decision fusion engine and the standardized device abstraction interface layer, and is used to convert the hardware control strategy into the underlying driving pulse sequence and drive the spectroscopic measurement hardware subsystem to execute.

2. The system according to claim 1, characterized in that, The deep reinforcement learning macro-planning unit integrates a multi-objective reward function calculation module. The multi-objective reward function calculation module is used to calculate the reward signal R according to the physical information entropy increment of the newly acquired data, the hardware execution cycle duration, and the signal noise level, as follows: R=W1×S+W2×(1 / T)-W3×A, where S is the physical information entropy increment, T is the hardware execution cycle duration, A is the signal noise level, and W1, W2, and W3 are preset technical weight coefficients.

3. The system according to claim 1, characterized in that, The standardized device abstract interface layer is equipped with a multi-protocol adapter, which is logically connected to the communication protocol stack of the interface layer. It supports at least one of EPICS, Tango, Modbus and TCP / IP protocols to realize the communication adaptation between decision signals and underlying physical hardware.

4. The system according to claim 1, characterized in that, The physical constraint-guided intelligent inference engine is equipped with an experimental feedback interface. The experimental feedback interface is set at the database input end of the engine and connected to the external data verification end. It is used to feed back the verified measured spectrum data to the intelligent inference engine to update the spectrum-structure mapping model parameters.

5. The system according to claim 1, characterized in that, The real-time spectral feature extraction module integrates a hardware state anomaly monitoring unit. The hardware state anomaly monitoring unit is set in the signal monitoring layer of the extraction module and is used to calculate the signal saturation in real time and send a hardware interruption protection signal to the meta-decision fusion engine when a signal overload feature is detected.

6. An AI-driven active optimization and decision-making method for X-ray spectroscopy experiments based on the system described in any one of claims 1-5, characterized in that, Includes the following steps: S1. Hardware initialization and state pre-scan: Control the spectroscopic measurement hardware subsystem to perform a fast global scan of the target under test to obtain initial physical state distribution data; S2. Real-time state perception and physical mapping: The real-time spectrum feature extraction module processes the feedback signal to generate a feature vector, and the intelligent inference engine is used to map the current physical perception range. S3. Macroscopic Measurement and Control Decision Generation: Utilizing the deep reinforcement learning macroscopic planning unit in conjunction with the current physical cognition interval, output a macroscopic hardware action control sequence; S4. Online optimization of micro-control parameters: The Bayesian optimization micro-search unit is used to calculate and issue the optimal single-point motion step size or time integral parameter online within the macro-action range. S5. Hardware closed-loop execution and policy feedback: Convert macroscopic actions and microscopic parameters into atomic instructions and send them to the spectroscopic measurement hardware subsystem for execution, and calculate reward values ​​based on the newly collected data to update the reinforcement learning policy network.

7. The method according to claim 6, characterized in that, In step S4, when the measurement target is to capture transient physical processes, the Bayesian optimized microscopic search unit uses time resolution as an optimization constraint and automatically reduces the motion step size of the spectroscopic measurement hardware subsystem when the rate of change of the characteristic signal exceeds a preset threshold.

8. The method according to claim 6, characterized in that, The method is equipped with a measurement and control termination judgment logic: when the inference confidence of the measured physical structure reaches a preset threshold, or the volatility of the real-time feature vector is lower than a preset convergence threshold, the execution of loop steps S2 to S5 is forcibly stopped.

9. The method according to claim 6, characterized in that, The real-time spectral feature extraction module integrates a hardware state anomaly monitoring unit. This unit is located in the signal monitoring layer of the extraction module and is used to calculate the signal saturation in real time and send a hardware interruption protection signal to the meta-decision fusion engine when a signal overload feature is detected. In step S2, if the hardware state anomaly monitoring unit detects that the signal response delay or intensity is lower than a preset safety threshold, steps S3 and S4 are forcibly skipped, and the closed-loop feedback execution module directly issues an emergency shutter closing command to the spectral measurement hardware subsystem.

Images