A perovskite laser process knowledge graph construction and formula self-recommendation method

Abstract

The application discloses a perovskite laser process knowledge graph construction and formula self-recommendation method, and relates to the perovskite laser processing field.The method comprises the following steps: forming a process evidence segment of a binding process context key, a device control quantity, a workpiece surface action vector and a measurement record, and writing into a graph; determining a state according to residual and damage indexes; establishing a parameter responsibility edge; forming a double-sided conversion interval from the responsibility edge of insufficient removal to qualified and qualified to protected layer damage; determining a boundary margin according to a smaller normalized distance of candidate to double-sided failure evidence value, selecting a candidate and outputting device control quantity and supporting evidence of the candidate; when the feedback segment is the same key and the measurement is complete, edge formation directional update and qualified new candidate are carried out; when the feedback segment is different keys, a version branch is established; and other conditions that do not meet the access control are isolated.The application reduces the misjudgment caused by directly copying cross-device values and abnormal feedback pollution of an activity window.

Description

Technical Field

[0001] This invention relates to the field of perovskite laser processing, and more particularly to a method for constructing a knowledge graph of perovskite laser processes and a self-recommendation method for formulations. Background Technology

[0002] Laser scribing and selective delamination of perovskite photovoltaic devices typically rely on equipment settings, test records, empirical rules, and process knowledge bases to determine processing formulas. The processing results are evaluated based on the removal of the target layer and damage to adjacent functional layers. However, because film stacking, optical path conditions, workpiece surface interaction conditions, and quality measurement apertures can all vary, historical records from different sources are difficult to compare on a unified basis. Formulas based on similar records or overall statistical relationships may not accurately reflect the processing boundaries under current operating conditions, easily leading to target layer residue, damage to protected layers, or distorted judgments of formula suitability.

[0003] In the data acquisition and correlation stage, traditional methods often directly describe the laser effect using equipment settings such as average power, scanning speed, and pulse frequency. The influence of spot state, optical path changes, and incident conditions on the energy distribution on the workpiece surface is not consistently incorporated, leading to the same nominal settings potentially corresponding to different actual operating states. Processing logs and quality measurement records may also be mismatched due to insufficient correspondence between processing areas, inspection times, film layer states, or measurement apertures. Simultaneous changes in multiple control quantities make it difficult to distinguish the local connections between these changes and insufficient removal of the target layer, damage to the protected layer, and composite anomalies. Different results are further compressed into vague "pass" or "fail" labels, which can mask differences in physical processes, making it easy to establish subsequent process relationships based on incomparable or ambiguous data.

[0004] In the judgment and output stages, traditional recommendation methods often select candidate formulations based on similarity, empirical weights, or global fitting results. When continuous observations are lacking across limited historical records, unverified parameter locations are easily mistaken for usable ranges. Furthermore, it's difficult to simultaneously explain whether the candidate poses a residual risk near the target layer or a risk of damage to the protective layer. Recommendation results often only provide equipment control variables or comprehensive scores, failing to adequately express applicable film states, optical path conditions, measurement basis, and the completeness of evidence. When applied across different equipment, deviations may arise due to variations in control settings and their interaction with the workpiece surface. Subsequent processing feedback is written back to the same model or knowledge base without distinguishing between applicable conditions and measurement validity. Abnormal records and changes in conditions may affect the original judgment, causing unreasonable drift in the recommended state, thereby increasing the risk of misjudgments, omissions, and untimely feedback control. Summary of the Invention

[0005] The technical problem to be solved by this invention is how to form a transition zone between the insufficient removal side and the damaged protected layer side from locally comparable laser processing evidence under the constraints of film layer, equipment calibration and quality measurement versions, and output a formulation that can be updated in a targeted manner from existing qualified evidence.

[0006] This invention provides a method for constructing a knowledge graph of perovskite laser processes and for self-recommendation of formulations, including: Obtain perovskite laser processing records, equipment logs, optical path calibration records, and quality measurement records to form process evidence fragments that bind process context keys, equipment control quantities, workpiece surface action vectors, and quality measurement records, and write them into the process knowledge graph; Based on the target layer residual index and the protected layer damage index, each process evidence fragment is classified into insufficient removal state, qualified state, protected layer damaged state, combined failure state and invalid measurement state; For process evidence segments with the same process context key, where only the change of one workpiece surface action component reaches the corresponding resolution and the other components satisfy the corresponding tolerance, sort them according to the component and establish parameter responsibility edges based on the state transitions of adjacent valid segments. The parameter responsibility edges from the insufficiency state to the qualified state and from the qualified state to the damaged state of the protected layer are respectively formed into the lower and upper transition intervals. The existing qualified segments sandwiched between the two intervals are taken as the formulation candidates. The normalized boundary margin of each formulation candidate is the smaller of the normalized distances to the failure evidence values ​​on both sides. Based on this, the formulation candidates are selected, and the recommended evidence records are formed and output according to their equipment control quantities and supporting evidence. When obtaining feedback process evidence fragments formed by laser processing according to the recommended evidence records, if the process context key of the feedback process evidence fragment is consistent and the quality measurement is complete, it is written into the temporary window version; for those that can form parameter responsibility edges, only the transition interval corresponding to their state transition is updated, and those in the qualified state are added as new formula candidates; for feedback process evidence fragments with consistent process context keys but not meeting the quality measurement completeness or window submission conditions, they are written into the isolation set; for feedback process evidence fragments with inconsistent process context keys, they are not written into the isolation set, but an independent version branch is created according to their own process context key.

[0007] In summary, the present invention has at least one of the following beneficial technical effects: This invention binds the film stack, target removal layer, protected layer, optical path calibration, and quality measurement scheme item by item through process context keys. It converts equipment control quantities into workpiece surface action vectors through same-version calibration, and then maps the two types of quality measurements corresponding to the processing area into inadequate removal state, qualified state, protected layer damaged state, combined failure state, and invalid measurement state. Relying solely on context version constraints can only eliminate mismatched records and cannot identify local parameter responsibility when multiple control quantities change simultaneously. Relying solely on the result state cannot guarantee the comparability of segments involved in boundary formation. Therefore, this invention uses the same process evidence segment as a common action object. Under the condition of consistent context and only one action component undergoing a distinguishable change, it forms a single-component comparable group. The parameter responsibility edge is transferred to the lower and upper transition intervals according to the state transition of adjacent segments. Then, the two transition intervals sandwich existing qualified segments and complete the candidate selection with normalized boundary margin, forming a recommended evidence record carrying context, bilateral boundaries, supporting segments, and window version. This continuous processing prevents incomparable records, compound failures, and local sequence breakpoints from being directly used to move boundaries, and avoids using unverified parameter locations as acceptable recipes. This allows recommendations to be traced back to evidence of opposite failures on both sides and the corresponding existing acceptable fragments.

[0008] This invention uses recommended evidence records as a common intermediate result for output verification and feedback updates. When the target device and the evidence device are the same, the action vector and release constraints of the device control quantity are verified using the same version calibration relationship. When the target devices are different, the solution is reversed only within the applicable domain of the measured calibration, and is constrained by control resolution, safety limits, and unique decision conditions. Outputting historical device control quantities alone cannot eliminate calibration differences between devices, and performing a single mapping verification cannot prevent incorrect versions or invalid measurement feedback from changing the active window. This invention uses the device logs and quality measurements generated after processing based on the recommended evidence records as feedback process evidence fragments. It adopts contextual consistency, measurement completeness, and single-component adjacency gates, so that feedback from the side with insufficient removal only updates the lower conversion interval, and feedback from the side with damage to the protected layer only updates the upper conversion interval. The closed and conflict-free new window version is used for subsequent candidate formation through temporary storage, recalculation, and submission. When calibration goes out of domain, mapping has no solution, multiple solutions are not adjudicated, feedback conditions are not met, or recalculation conflicts occur, a rejection state is output and an independent version branch is isolated or established. Therefore, the recommendation that evidence recording, conditional formulation release, and window version feedback be continuously closed around the same bilateral boundary can reduce misjudgments caused by direct copying of cross-device values ​​and abnormal feedback contamination activity windows, and ensure that subsequent recommendations are consistent with the current film layer, optical path calibration, and measurement conditions. Attached Figure Description

[0009] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a flowchart of a method for constructing a knowledge graph and self-recommending formulations for perovskite laser processes, provided by an embodiment of the present invention.

[0011] Figure 2 This is a semantic topology graph of the process knowledge graph provided in the embodiments of the present invention.

[0012] Figure 3 This is a schematic diagram of single-component local evidence and bilateral feasible window qualitative analysis provided in an embodiment of the present invention.

[0013] Figure 4 This is a feedback window version status transition diagram provided in an embodiment of the present invention. Detailed Implementation

[0014] The following description, in conjunction with the implementation of the present invention, is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the concept of the invention, and all such modifications and additions should fall within the protection scope of the present invention.

[0015] Example 1: This invention is applicable to laser scribing and selective delamination of perovskite photovoltaic devices or modules. The technical object is a film stack with a target removal layer and a protected layer. The processing data comes from the process route, laser equipment log, optical path calibration record and quality measurement record corresponding to the processing area.

[0016] In this operating condition, different film stack versions, equipment optical path states, and measurement schemes can render nominal control quantities and quality results incomparable. Simultaneous changes in multiple laser parameters can also prevent ordinary historical formula retrieval from determining which component supports the local transformation of the insufficient removal side or the damaged protected layer side. Therefore, this implementation method forms a traceable formula through versioned evidence fragments, single-component local adjacent responsibility edges, and two-sided transformation intervals, and explicitly rejects branches to prevent incomparable data and incorrect version feedback from entering the current window.

[0017] The method to be protected in this application will be further described below with reference to the accompanying drawings. The following embodiments are used to illustrate the acquisition, processing, judgment, output, and abnormal paths of various data, and do not indicate that the applicant has completed production verification using a certain equipment model, material size, laser power, or quality threshold. The actual film structure, detection equipment, quality threshold, and control resolution are determined by the specific perovskite device process and are recorded in this method in the corresponding version.

[0018] This application discloses a method for constructing a knowledge graph of perovskite laser processes and for self-recommendation of formulations.

[0019] Reference Figures 1-4 A method for constructing a knowledge graph and self-recommending formulations for perovskite laser processes, specifically including: Obtain perovskite laser processing records, equipment logs, optical path calibration records, and quality measurement records to form process evidence fragments that bind process context keys, equipment control quantities, workpiece surface action vectors, and quality measurement records, and write them into the process knowledge graph; Specifically, this method aligns records of the same laser processing session using workpiece identifiers, batch identifiers, process identifiers, processing area identifiers, and timestamps. Then, it binds segment identifiers, process context keys, equipment control quantities, workpiece surface action vectors obtained through optical path calibration of the same version, and quality measurement records corresponding to the processing area into process evidence segments. Relationships between these process evidence segments and their bound objects are established in the process knowledge graph. If necessary records are missing, the corresponding regions are not unique, or the versions are inconsistent, the corresponding records are not included in the active evidence set. It should be further noted that the segment identifier is generated based on the workpiece identifier, batch identifier, process identifier, processing area identifier, and timestamp, and is used to distinguish and trace process evidence segments corresponding to the same laser processing session.

[0020] This method acquires and processes records related to the same laser processing from the process expert database, experimental record database, laser equipment log database, equipment optical path calibration record database, and quality inspection record database.

[0021] The process expert database provides information on material name, layer order, process definition, target removal layer, protected layer, and quality criteria version. Experimental records provide information on workpiece, batch, processing area, formula, and test results. Equipment logs provide equipment identification, optical path or splitter branch identification, average power, pulse repetition frequency, scanning speed, focal position, trajectory spacing, number of scan passes, scanning direction, processing time, and operating status. Optical path calibration records provide information on the power response acting on the workpiece surface, spot size, focal plane position, calibration application area, control resolution, and calibration version. Quality inspection records provide information on the detection items, detection values, range, area, time, measurement scheme version, and quality criteria version for target layer residue and protected layer damage.

[0022] This method uses workpiece identifier, batch identifier, process identifier, processing area identifier, and timestamp to align the above records to the same laser processing. The processing area can be mapped to the calibrated transformation relationship between the equipment trajectory coordinates and the detection area coordinates. If the transformation relationship is not obtained, the area correspondence is not unique, or the detection time point is later than a subsequent process that would change the state of the area, the corresponding detection result is not bound to this processing, but the record is written to the isolated set to be completed.

[0023] This method is used for laser processing of perovskite photovoltaic film stacking with a target removal layer and a protected layer. Each actual process registers the process identifier, target removal layer, and protected layer in the process route. The specific layer names and hierarchical relationships are determined based on the current film stack version; layer names from different film systems are not directly applied.

[0024] Furthermore, the process context key includes processing step identifier, film stack version, target removal layer identifier, protected layer identifier, substrate type, laser wavelength type, pulse time type, incident direction type, equipment optical path calibration version, and quality measurement scheme version. Only process evidence fragments with identical items are included in the active evidence set corresponding to the same process context key, while process evidence fragments with missing required items are written into an isolated set.

[0025] In one specific embodiment, the method first forms a process context key. The process context key includes at least the processing step identifier, film stack version, target removal layer identifier, protected layer identifier, substrate type, laser wavelength type, pulse time type, incident direction type, equipment optical path calibration version, and quality measurement scheme version. If the layer exposure state, edge morphology state, or electrical isolation state formed in the previous process would change the current target removal layer and protected layer, the measured version of the previous process state can also be written into the process context key.

[0026] The above fields are compared item by item by identifier or version, without fuzzy merging based on similarity thresholds. Only records that are identical item by item are included in the same active evidence set. If the membrane stack version changes, a new process context key is created even if the equipment model and nominal power are the same. If only the version identifier in the process record is entered incorrectly, the original record, the corrected record, and the substitution relationship between them are retained. The new process evidence fragment after correction is re-examined, without directly overwriting the original fragment.

[0027] Each completed laser processing step forms a process evidence segment. A process evidence segment is at least bound to a segment identifier, process context key, equipment control variables, workpiece surface action vector, processing area, processing time, target layer residual quality measurement, protected layer damage quality measurement, discrete result status, source record identifier, and segment status. The segment status can at least include active, temporary, isolated, and replaced. This method establishes relationships between process evidence segments and process context, equipment control variables, workpiece surface action vector, quality measurement records, and discrete result status within a process knowledge graph.

[0028] Knowledge graphs can be implemented using graph databases capable of storing nodes, relationships, attributes, and versions. The specific storage format of the graph database does not limit this method. Parametric responsibility edges in the graph are not directly generated by graph embedding models or general relationship predictions, but are generated by single-component comparability and local adjacency rules described later; graph queries are only responsible for retrieving evidence objects that satisfy the context, version, and graph relationship conditions.

[0029] Furthermore, the workpiece surface interaction vector is formed, specifically including: Obtain the average power, pulse repetition frequency, and scanning speed from the equipment log; obtain the effective spot area on the workpiece surface and the effective spot size along the scanning direction from the same equipment optical path calibration version. The single-pulse surface dose is obtained by dividing the average power by the product of the pulse repetition frequency and the effective spot area on the workpiece surface. The pulse overlap rate along the scanning direction is obtained by subtracting the product of the scanning speed, the pulse repetition frequency, and the effective spot size along the scanning direction. The effective single-pulse surface dose and the pulse overlap rate along the scanning direction are then written into the workpiece surface action vector.

[0030] In a specific embodiment, the device control quantity represents the control settings received by the device, and the workpiece surface action vector represents the quantity actually used to describe the laser action on the workpiece surface after calibration with the same version of the optical path. The two should not be directly confused. Components that can be written into the workpiece surface action vector include single-pulse surface dose, pulse overlap rate along the scanning direction, lateral trajectory overlap rate, normalized defocus amount, number of scan passes, and scanning direction category. For each component actually used in a certain process, the name, unit, direction attribute, resolution, effective domain, acquisition rules, and calibration version are all registered; unregistered components do not participate in the formation of responsibility edges.

[0031] When the average power, pulse repetition frequency, and effective spot area on the workpiece surface are all effective, this method uses these three factors to form a single-pulse surface dose characterizing the effect of a single pulse on a unit area of ​​the workpiece surface. The single-pulse surface dose is determined according to... Confirmed. Among them, This is a single-pulse surface dose, measured in joules per square centimeter. The average power given in the device log or power calibration record, in watts; The pulse repetition frequency is expressed in Hertz (Hz). The effective spot area measured on the workpiece surface under the same optical path calibration version is expressed in square centimeters. For finite nonnegative numbers, and It is a finite positive number. or If the aforementioned conditions are not met, the input exceeds the calibration range, any input is missing, or the value is non-finite, this method will... Marked as invalid, the corresponding fragment is not included in the single-component comparable group where the single-pulse surface dose is the component to be analyzed.

[0032] When the scanning speed, pulse repetition frequency, and effective spot size along the scanning direction are all effective, and the center-to-center distance between adjacent pulses does not exceed the effective spot size, this method generates a pulse overlap rate along the scanning direction that characterizes the coverage of adjacent pulses along the trajectory. The pulse overlap rate along the scanning direction is calculated according to... Confirmed. Among them, The pulse overlap rate along the scanning direction is a dimensionless quantity. The scan speed given for the device logs; The pulse repetition frequency; The effective spot size along the scanning direction for the same optical path calibration version. and Use the same length per time dimension. For finite nonnegative numbers, and For finite positive numbers. When equal hour, It is zero; when Greater than When the pulse interval is recorded, the calculated negative number is not used as the overlap rate for sorting; if the denominator is invalid or the input exceeds the calibration range, the component is marked as invalid.

[0033] The lateral trajectory overlap rate can be formed according to the similar relationship between trajectory spacing and lateral effective spot size; the normalized defocus amount can be obtained from the difference between the current focal position and the calibrated focal plane position relative to the calibrated depth of focus. Since the defocus amount may have different effects on both sides of the focal plane, the defocus direction can be used as a category field during implementation, and local evidence can be formed in each category, without assuming its monotonicity across the entire domain. The number of scan passes is a discrete count component, and the corresponding responsibility edge can only be established when other conditions between adjacent passes meet the tolerance.

[0034] Based on the target layer residual index and the protected layer damage index, each process evidence fragment is classified into insufficient removal state, qualified state, protected layer damaged state, combined failure state and invalid measurement state; Specifically, this method first checks the measurement integrity, range, finiteness, measurement scheme version, and processing area correspondence of the target layer residual index and the protected layer damage index. When the measurement is valid, the two indexes are compared with their respective quality criteria. The state of insufficient removal, qualified state, protected layer damage state, or composite failure state is determined according to the qualified combination of the two indexes. When the measurement is invalid, it is determined as an invalid measurement state.

[0035] Furthermore, each process evidence fragment is categorized into insufficient state, qualified state, protected layer damaged state, combined failure state, and invalid measurement state, including: An invalid measurement state is determined when any of the following conditions are met: missing required measurement, exceeding the range, non-finite value, invalid measurement scheme version, or non-corresponding processing area. When the measurement is valid and the residual index of the target layer is unqualified while the damage index of the protected layer is qualified, it is judged as an insufficient removal state. When both indexes are qualified, it is judged as a qualified state. When the residual index of the target layer is qualified but the damage index of the protected layer is unqualified, it is judged as a damaged state of the protected layer. When both indexes are unqualified, it is judged as a combined failure state.

[0036] In a specific embodiment, the target layer residual quality index is used to determine whether the target removal layer meets the removal requirements of the current process, and the protected layer damage quality index is used to determine whether the protected layer maintains the structural, electrical, or morphological state required by the current process. The former can be derived from residual area ratio, residual continuity, composition detection, or electrical isolation detection; the latter can be derived from protected layer material defects, cracks, penetration, conductivity, or insulation performance detection. The choice of which detection item, threshold, and retesting rules are used is determined by the quality measurement scheme and the version is recorded.

[0037] This method first checks the measurement completeness and validity of two indicators. If any required measurement is missing, exceeds the measurement range, is a non-finite value, the measurement scheme version is invalid, the detection area and processing area do not correspond, or the specified retest still cannot obtain a unique state, the segment is determined to be an invalid measurement state. Invalid measurement states do not generate parameter responsibility edges.

[0038] When the measurement is valid, this method compares each of the two indicators with its respective quality criteria. If the target layer residual indicator is unqualified but the protected layer damage indicator is qualified, the segment is classified as insufficiently removed; if both indicators are qualified, the segment is classified as qualified; if the target layer residual indicator is qualified but the protected layer damage indicator is unqualified, the segment is classified as protected layer damaged; if both indicators are unqualified, the segment is classified as a combined failure state. The combined failure state retains two types of non-compliance labels, but these are not forcibly used as endpoints of either transition interval.

[0039] When the measurement result falls within the uncertainty range specified in the quality measurement scheme, this method performs a retest according to the same scheme. The detection location, number of repetitions, and aggregation rules for the retest are given by the scheme and cannot be changed temporarily by the recommended process. If a unique state is obtained after the retest, a new measurement record is generated and the original measurement record is retained; if a unique state still cannot be obtained, the invalid measurement state is maintained.

[0040] Dead zone, linewidth, insulation resistance, trajectory continuity, and yield, among other indicators, can be used as constraints for formulation release. Only indicators that correspond to the same processing area, testing time point, and measurement version can be used for release checks. Other indicators do not change the definitions of the above five states to avoid a single threshold simultaneously bearing the two different physical meanings of target layer removal and protected layer damage.

[0041] For process evidence segments with the same process context key, where only the change of one workpiece surface action component reaches the corresponding resolution and the other components satisfy the corresponding tolerance, sort them according to the component and establish parameter responsibility edges based on the state transitions of adjacent valid segments. Specifically, this method selects a workpiece surface action component to be analyzed within the same process context key, and uses the condition that the component reaches the registered resolution and the other numerical components do not exceed the corresponding tolerance and the other category components are the same as the comparability conditions; then sorts according to the component values ​​to be analyzed, isolates indistinguishable conflicts of the same value, checks adjacent valid segments in turn, and establishes parameter responsibility edges only for local state transitions from insufficient state to qualified state and from qualified state to damaged state of protected layer.

[0042] Furthermore, for process evidence fragments with the same process context key, where only the change of one workpiece surface action component reaches the corresponding resolution and the remaining components satisfy the corresponding tolerance, the fragments are sorted according to this component, and parameter responsibility edges are established based on the state transitions of adjacent valid fragments, including: For the same process context key and the surface action component of the workpiece to be analyzed, a single component comparable group is formed based on the other numerical components satisfying the corresponding tolerance and the other categories of components being the same. The valid process evidence fragments within the group are sorted according to the values ​​of the surface action components of the workpiece to be analyzed, and fragments with the same component value corresponding to different states that are indistinguishable are written into an isolated set. After sorting, two adjacent valid segments are obtained in sequence. When the low-value segment is in a state of insufficient removal and the high-value segment is in a state of qualified status, a parameter responsibility edge supporting the lower transition interval is established. When the low-value segment is in a state of qualified status and the high-value segment is in a state of damage to the protected layer, a parameter responsibility edge supporting the upper transition interval is established. The remaining states are converted and registered as local sequence breakpoints.

[0043] In one specific embodiment, this method establishes a single-component comparable group for a process context key and a workpiece surface action component to be analyzed. Process evidence fragments belonging to the same group should meet the following requirements: the process context keys are identical item by item; all components to be analyzed are valid; the distinguishable difference of the component to be analyzed reaches the registered resolution of that component; the difference of the remaining numerical components does not exceed the corresponding tolerance; the remaining category components are identical; and the quality measurements come from the same scheme version and can be mapped to comparable processing areas.

[0044] The tolerances for each fixed component are derived from equipment control resolution, optical path calibration repeatability, log quantization accuracy, and region alignment error. During implementation, a tolerance that covers measurement repetition fluctuations and does not cross process setting levels can be selected based on the impact of these errors on the same fixed component, and the determination criteria and tolerance versions should be saved together. When traceable evidence is lacking, a fundamentally identical tolerance with uncertain meaning is not used as a substitute; instead, the corresponding segment is registered as a multi-component difference segment.

[0045] This method sorts the valid fragments within a group according to the values ​​of the components to be analyzed. Fragments with the same value or a difference less than the resolution are formed into a group of identical evidence. When the states within a group of identical evidence are consistent, all sources are retained and that state is used for adjacency checks; when different states appear within a group of identical evidence, a conflict group is formed and written into the isolation set, without the majority state covering the minority state.

[0046] After removing conflicting groups of the same value, this method sequentially checks two adjacent valid evidence values ​​after sorting, establishing long-distance edges without crossing intermediate valid evidence values. When the smaller value segment is in a state of insufficient removal and the larger value segment is in a qualified state, a parameter responsibility edge supporting the lower transition interval is established; when the smaller value segment is in a qualified state and the larger value segment is in a state of damage to the protected layer, a parameter responsibility edge supporting the upper transition interval is established. When adjacent segments are in the same state, only local support for that state is added; when adjacent states include composite failure states, invalid measurement states, or state changes inconsistent with the above two types of directions, a local sequence breakpoint is established without generating a boundary responsibility edge.

[0047] The smaller and larger value segments mentioned above correspond to the smaller and larger values ​​of the two actual evidence values ​​within the current group, respectively, after being sorted numerically. Each parameter responsibility edge only covers the local observation direction between its two endpoints and does not indicate that the action component is monotonic across the entire parameter domain. If multiple qualified segments appear in different numerical ranges within the same group, this method processes each local segment separately and does not fit multiple local segments into a single global curve.

[0048] The parameter responsibility edges from the insufficiency state to the qualified state and from the qualified state to the damaged state of the protected layer are respectively formed into the lower and upper transition intervals. The existing qualified segments sandwiched between the two intervals are taken as the formulation candidates. Specifically, this method identifies the largest qualified segment composed of consecutive qualified segments from the sorted valid segments. It defines the lower and upper transition intervals by removing insufficient state evidence directly on the lower side and the damaged state evidence of the protected layer directly on the higher side, respectively. Only segments that are sandwiched by the two transition intervals and have qualified measurement records are selected as formulation candidates. The corresponding candidate is stopped when there is no failure evidence on either side, the largest qualified segment is empty, or there are unresolved local sequence breakpoints.

[0049] Furthermore, the parameter responsibility edges from the deficient state to the qualified state and from the qualified state to the damaged state of the protected layer are respectively used to form lower and upper transition intervals. Existing qualified segments sandwiched between these two intervals are taken as formulation candidates, including: Identify the largest qualified segment consisting of consecutive qualified state segments in a single-component comparable group; When the adjacent segment directly below the maximum qualified segment is in a state of insufficient removal and the adjacent segment directly above the maximum qualified segment is in a state of protected layer damage, the lower conversion interval is defined by the minimum value from the adjacent segment directly below the maximum qualified segment, the upper conversion interval is defined by the maximum value from the maximum qualified segment to the adjacent segment directly above the maximum qualified segment, and the qualified segments already in the maximum qualified segment are used to form formulation candidates. When multiple discontinuous maximum qualified segments exist, recipe candidates are formed respectively. When any one of the following conditions is met: missing adjacent failure segments on either side, the maximum qualified segment is empty, or there is an unresolved local sequence breakpoint, the corresponding recipe candidate is stopped from being output.

[0050] In one specific embodiment, this method identifies the largest qualified segment consisting of consecutive qualified state evidence values ​​within a single-component comparable group. The largest qualified segment is a set of consecutive segments that, after being sorted by the component to be analyzed, are not separated by local sequence breakpoints, and whose adjacent values ​​are all qualified.

[0051] When the direct lower-side adjacent evidence of the maximum qualified segment is in a state of insufficient removal, the direct lower-side evidence value and the minimum evidence value of the maximum qualified segment together define the lower-side transition interval; when the direct upper-side adjacent evidence of the maximum qualified segment is in a state of protected layer damage, the maximum evidence value of the maximum qualified segment and the direct upper-side evidence value together define the upper-side transition interval. Only when both sides exist and the endpoint order is valid can a bilaterally closed locally feasible window be formed.

[0052] Within the locally feasible window, only process evidence segments with valid measurement records within the largest acceptable segment are considered as production formulation candidates. Untested points in the transition interval are not automatically considered acceptable. When a new action value needs to be verified within the window, this method only registers it as a verification candidate. Only after the candidate completes actual processing, generates a complete measurement, and passes the same access control is it converted into a production formulation candidate.

[0053] When multiple discontinuous maximum qualified segments exist, different local feasible window versions are created, and their respective responsibility edges are saved. Different windows are not merged, and a global window is selected without using the average value. When conflicts of the same value or repeated state transitions occur in the same context, this method first checks whether the context, unit, calibration version, and measurement version have been mistakenly merged; if it can be further split according to fixed components, processing areas, or function categories, more refined single-component comparable groups are created; if it cannot be split, the affected local segments are marked as evidence conflicts and their production formula output is stopped.

[0054] When there is a lack of evidence of failure on the lower adjacent side, a lack of evidence of failure on the upper adjacent side, or no existing qualified segment within the window, the output will show an insufficient boundary evidence status. When there is no qualified segment within the group, the output will show a no qualified candidate status. When the order of the endpoints on both sides is invalid, there is no intersection between the qualified sides, the window width does not exceed the resolution of the active component, or the unresolved breakpoint is located in the largest qualified segment, the output will show an empty window or evidence conflict status. In all of the above states, the production formula will not be output, and the output will indicate the missing boundary category, the component to be analyzed, the comparison group, and the type of evidence that needs to be supplemented.

[0055] The normalized boundary margin of each formulation candidate is the smaller of the normalized distances to the failure evidence values ​​on both sides. Based on this, the formulation candidates are selected, and the recommended evidence records are formed and output according to their equipment control quantities and supporting evidence. Specifically, this method calculates two normalized distances between the value of the component to be analyzed and the failure evidence values ​​on both sides for each formulation candidate, and uses the smaller value as the normalization boundary margin. First, the candidate with the largest margin is selected. If they are still tied, they are decided in turn according to the completeness of evidence, the total width of the transition interval, the distance of the action vector, and the stability identifier of the process evidence fragment. Then, the complete equipment control quantity and supporting evidence of the selected candidate are packaged into a recommended evidence record and output.

[0056] Furthermore, the normalized boundary margin is determined by the smaller normalized distance from each candidate formulation to the failure evidence values ​​on both sides, and candidate formulations are selected accordingly, including: The distances between the candidate formulation component value and the evidence value of insufficient removal on the direct low side, and the distances between the evidence value of damage to the protected layer on the direct high side and the component value to be analyzed are obtained respectively. The distances between the failure evidence values ​​on both sides are normalized respectively, and the smaller value of the two normalization results is taken. First, select the formulation candidate with the largest normalized boundary margin. When there are ties, determine the selected formulation candidate in ascending order according to the following criteria: higher completeness of evidence on both sides, smaller total width of the transition interval on both sides, smaller distance of the action vector from the most recently verified qualified formulation, and process evidence fragment identifier.

[0057] In a specific embodiment, for each existing qualified candidate in a bilaterally closed local window, this method obtains the distance from the candidate component value to the direct low-side insufficient evidence value and the distance from the direct high-side protected layer damage evidence value to the candidate value. This distance is then normalized using the distance between the two failure evidence values ​​to form a normalized boundary margin characterizing the degree to which the candidate approaches the closer of the two opposing failure evidence values. The normalized boundary margin is calculated according to the following formula:

[0058] in, Candidate formulation The normalized boundary margin is a dimensionless quantity. as a candidate The values ​​of the components to be analyzed; To directly remove insufficient evidence values ​​from the low side; This is the direct evidence value for damage state of the protected layer on the higher side; This indicates taking the smaller of the two values. , and Having the same unit and satisfying .when Not greater than If any input is a non-finite value or the candidate is not an existing qualified segment, no calculation is performed. And mark the corresponding window as invalid.

[0059] This method first selects the candidate formulation with the largest normalized boundary margin. When the margins are the same, for each side of directly adjacent evidence, four conditions are set as one: all necessary measurements are complete, all measurements are within their range and are finite values, the measurement scheme version is valid, and the processing area corresponds. The sum of these conditions on both sides is taken as the evidence integrity value, and the candidate with the larger value is selected first. If the evidence integrity values ​​are still the same, only under the same component to be analyzed and the same unit, the candidate with the smaller total width is selected by the sum of the widths of the conversion intervals on both sides. If the components to be analyzed are different or the total widths are still the same, the absolute value of the difference between each component of the candidate action vector and the corresponding component of the most recently verified qualified formulation of the current equipment is divided by the comparison tolerance of that component, and the maximum value of the resulting ratio is taken as the dimensionless action vector distance, and the candidate with the smaller distance is selected. If any required comparison field, comparison tolerance, or reference formulation is missing, the corresponding comparison item is skipped; if it is still impossible to distinguish, a unique candidate is determined by ascending order of the stability identifier of the process evidence fragment. The counting, normalization, and stability identifier sorting are only executable determination rules and do not indicate that actual experimental verification has been obtained.

[0060] When a complete formulation is constrained by local windows of multiple active components, it can be tailored to each component. Calculate the normalized boundary margin for each component, and then take the minimum value among the component margins as the margin of the complete formulation. Only when all component values ​​come from the same existing qualified process evidence segment is the complete formulation considered as a production candidate; individual components from different segments cannot be arbitrarily spliced ​​together to form an unverified new formulation.

[0061] Components that do not form bilateral evidence retain their original values ​​as candidate segments; if they cannot retain their original values, output insufficient data.

[0062] Furthermore, the recommended evidence record includes a process context key, a workpiece surface action vector, a lower transition interval, an upper transition interval, a normalized boundary margin, process evidence fragment identifiers supporting both transition intervals, a quality measurement scheme version, and a window version. Confidence characterization is formed by whether the two transition intervals are closed, the integrity level of the evidence on both sides, the normalized boundary margin, and the effective status of cross-device mapping.

[0063] In one specific embodiment, this method encapsulates the complete equipment control variables, workpiece surface action vectors, process context keys, lower transition intervals, upper transition intervals, normalized boundary margins, supporting responsibility edges, supporting evidence fragment identifiers, quality measurement scheme versions, and window versions bound to the selected candidate into a recommended evidence record. The recommended evidence record also stores whether both sides are closed, the evidence integrity level, and the effective status of cross-equipment mapping, to constitute a traceable confidence representation. This confidence representation is not based on arbitrary probabilities given by similarity or black-box models.

[0064] When the target device and its optical path calibration version match the selected candidate, this method obtains the complete device control quantity bound to the candidate fragment and uses the currently valid calibration relationship of the same version to forward calculate the workpiece surface action vector. If the forward result is still within the local window of the candidate and other published quality constraints have valid evidence of compliance, the device control quantity and recommended evidence record are output. If the forward result no longer falls within the local window, the calibration version has expired, or other published quality constraints lack valid evidence, the recipe is output as unpublishable.

[0065] The output formulation always includes the applicable process, film stack, target removal layer, protected layer, equipment optical path calibration version, and quality measurement scheme version. This method does not provide single power, speed, or frequency values ​​outside of these context fields. If any required context field is missing, the corresponding candidate will be marked as an unpublishable formulation.

[0066] Furthermore, when the target device is different from the device corresponding to the selected formula candidate, check the calibration relationship and calibration application domain of the control quantity of the target device to the action vector on the workpiece surface; When the workpiece surface action vector and target operating conditions of the candidate formulation are within the calibration applicable domain, and the target control quantity that uniquely satisfies the target equipment control resolution and safety limit can be obtained by reverse calculation, the target equipment formulation is output. If any of the following conditions are met: exceeding the calibration application range, no solution in reverse solving, or multiple solutions in reverse solving that fail to be decided according to the predetermined minimum control quantity change rule, output the corresponding rejection status and do not output the target equipment recipe.

[0067] In one specific embodiment, when the target device differs from the device corresponding to the candidate evidence, this method first examines the measured calibration relationship between the target device and the workpiece surface action vector, as well as its applicable domain. The calibration relationship should record the input control quantity range, operating condition range, workpiece surface action output range, control resolution, calibration error, version, and validity period.

[0068] Only when all necessary components of the target action vector and the target operating conditions are within the calibration applicable domain, this method forms a finite number of control quantity candidates according to the target equipment control resolution, excludes candidates that exceed the control limits and safety limits, calculates the remaining candidates into action vectors through a positive calibration relationship, and retains control quantity candidates whose differences from each component of the target action vector do not exceed the calibration allowable error.

[0069] If only one control variable candidate meets the conditions, that candidate is output as the target equipment formula. If multiple candidates meet the conditions, for each candidate, the absolute value of the difference between each component of its forward calibration action vector and the corresponding component of the target action vector is divided by the calibration allowable error of that component. The maximum value of the resulting ratio is taken as the dimensionless action vector deviation, and the candidate with the smallest deviation is retained first. If there are still multiple candidates, the absolute value of the difference between each control variable and the control variable corresponding to the current stable formula of the target equipment is divided by the corresponding control resolution. The maximum value of the resulting ratio is taken as the dimensionless control variable change, and the candidate with the smallest change is retained. If they are still tied, a unique candidate is determined by ascending order of the control variable stability identifier. If any required calibration allowable error, control resolution, or reference stable formula is missing or not positive, or if the stability identifier still cannot be uniquely sorted, a mapping multi-solution rejection state is output. If the target action vector exceeds the applicable domain, a cross-device domain rejection state is output. If no control variable candidate meets the conditions, a target equipment unreachable state is output. No target equipment production formula is output in any rejection state. The above normalized decision is only used for deterministic ranking of a finite set of candidates and does not indicate that cross-device mapping has been actually verified.

[0070] Cross-device conversion is merely a conditional step based on an existing locally feasible window and selected action vector; it is not a mandatory step recommended for the same device. After the target device completes controlled verification and generates a qualified measurement, the result serves as a new process evidence fragment with the target device's calibration version. It enters the corresponding context according to the feedback gating of this method, without directly rewriting the original device's window. If the target device's calibration version fails, the verification measurement is invalid, or the context is inconsistent, the resulting result does not enter the original window. The rejection result only saves the failure reason and calibration version; it does not generate a production formula for the target device.

[0071] When obtaining feedback process evidence fragments formed by laser processing according to the recommended evidence records, if the process context key of the feedback process evidence fragment is consistent and the quality measurement is complete, it is written into the temporary window version; only the transition interval corresponding to its state transition is updated for those that can form parameter responsibility edges, and those in the qualified state are added as new formula candidates; for feedback process evidence fragments with consistent process context keys but not meeting the quality measurement completeness or window submission conditions, they are written into the isolation set; for feedback process evidence fragments with inconsistent process context keys, they are not written into the isolation set, but an independent version branch is created according to their own process context key; Specifically, this method obtains feedback equipment logs and quality measurement records based on recommended evidence records, and generates feedback process evidence fragments according to the initial evidence formation rules. Only when its process context key is consistent with the window to be updated and the quality measurement is complete, is it written into the temporary window version of the window to be updated, and parameter responsibility edge-oriented updates or new formulation candidate processing are performed according to the discrete result status. After recalculation, it is submitted or isolated according to the window submission conditions. When its process context key is inconsistent, it is not written into the temporary window version or isolation set of the window to be updated, but an independent version branch is established according to its own process context key. When its process context key is consistent but the quality measurement is incomplete, the original active window is maintained and the feedback fragment is written into the isolation set.

[0072] Furthermore, feedback process evidence fragments are updated, isolated, or used to create independent version branches, including: Write the feedback process evidence fragment that is consistent with the process context key and has complete quality measurement to the temporary window version; When it forms adjacent parameter responsibility edges to remove insufficient states from qualified states, only the lower transition interval is updated; When the adjacent parameter responsibility edge from the qualified state to the damaged state of the protected layer is formed, only the upper transition interval is updated; If it is in a qualified state, it is considered as a new formulation candidate, and the closure conditions on both sides are re-examined; When the transition intervals on both sides are closed after recalculation and there is no state conflict, submit a new window version. If one of the following conditions is met after recalculation: empty window, state conflict, or insufficient boundary evidence, maintain the original window version and write the new fragment into the isolation set. When the process context keys of the feedback process evidence fragments are inconsistent, an independent version branch is established according to its process context key. After the independent version branch forms the conversion intervals on both sides, it is used as a new window version for the next formulation candidate formation.

[0073] In one specific embodiment, this method receives equipment logs and quality measurement records generated from laser processing based on recommended evidence records, and forms feedback process evidence fragments according to the same rules as the initial evidence. Only when the process context key matches the window to be updated and the quality measurement is complete is the fragment written to the temporary window version of the window to be updated. If the process context key is inconsistent, it is not written to the temporary window version or isolated set of the window to be updated; instead, an independent version branch is created based on its own process context key. If the process context key matches but the quality measurement is incomplete, it is written to the isolated set. The temporary window version is then created. The temporary fragment does not change the active window before submission.

[0074] This method checks each step of the feedback segment to ensure that the process, film stack, target removal layer, protected layer, equipment optical path calibration, and quality measurement scheme version are consistent with the window to be updated, checks whether the measurement is complete, and checks whether the segment can form a single-component comparable adjacent pair with the direct predecessor or direct successor in the active evidence sequence.

[0075] When the feedback is to remove the insufficient state and form an adjacent parameter responsibility edge with the qualified state with a higher action value, only write the edge into the support edge set of the lower transition interval; when the feedback is to be a protected layer damaged state and form an adjacent parameter responsibility edge with the qualified state with a lower action value, only write the edge into the support edge set of the upper transition interval; when the feedback is a qualified state, treat it as a new formulation candidate and re-examine the closure conditions on both sides; when the feedback is a composite failure state or an invalid measurement state, do not update either side's transition interval.

[0076] After inserting a new fragment, this method only reconstructs the local responsibility edges between the new fragment and its direct predecessor and direct successor. The old edges between the original direct predecessor and direct successor are marked as having been split and their source is preserved; other contexts, other components, and other local windows are not recalculated.

[0077] This method reconstructs the maximum qualified segment, the transition intervals on both sides, the candidate set, and the normalized boundary margin in the temporary window version. If the recalculation results in closure on both sides, at least one existing qualified candidate, and no unresolved conflicts, a new window version is submitted and used for the next recipe recommendation. If an empty window, state conflict, or insufficient boundary evidence occurs, the original active window version is maintained, the newly added segment and its candidate responsible edge are written into the isolation set, and the rejection reason is recorded.

[0078] When the feedback process context key is inconsistent with the active window, it is not mapped to the old version fragment, but an independent version branch is created according to its own process context key. An independent branch can only become a new active window after forming two conversion intervals and existing qualified candidates. An isolated fragment is only created as a new re-admission fragment and re-checked from the context and measurement gate after missing records are filled, the same area is retested according to the current scheme, traceable record errors are corrected, or single-component adjacent evidence is added under the corresponding version; the original isolated fragment is always retained.

[0079] When input records are duplicated, this method identifies duplicate copies based on the source record identifier and processing time, retains only one active reference, and saves the duplicate source. When input records are out of order, they are first rearranged by processing time and region identifier; if rearrangement to a unique process event is not possible, the records are isolated. When the processing reaches the predetermined resource or time limit, the current input snapshot and completed context keys are saved, and recipes with incomplete decisions are not output. Resumption begins from the same input version.

[0080] When a single comparable group only has deficient states and damaged protected layers but lacks qualified states, this method registers the qualified evidence gap and does not generate a formulation using the midpoints at both ends. When only qualified states are available but lack failure evidence on either side, this method retains the qualified fragment as a verification reference and does not mark it as a production candidate with evidence from both sides.

[0081] If the actual quality criterion version is updated, the old fragment will still belong to the old measurement scheme version. This method can regenerate a derived measurement record and a new fragment with the new version using the new rules when the original detection data is still complete and the applicable conditions of both the old and new rules are met; the new fragment enters an independent context branch, and the original fragment is not overwritten.

[0082] Without altering the continuity of process evidence segments, single-component local responsibility edges, two-sided conversion intervals, existing qualified candidate selection, and feedback-oriented gating, different quality indicators, active components, and graph data storage methods can be selected based on the specific membrane layer and detection conditions. Any replacements should retain the input source, version, domain, abnormal branches, and output purpose.

[0083] The above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described or use similar methods to replace them, as long as they do not deviate from the concept of the invention, they should all fall within the protection scope of the present invention.

[0084] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0085] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A method for constructing a knowledge graph of perovskite laser processes and a self-recommendation method for formulations, characterized in that, include: Obtain perovskite laser processing records, equipment logs, optical path calibration records, and quality measurement records to form process evidence fragments that bind process context keys, equipment control quantities, workpiece surface action vectors, and quality measurement records, and write them into the process knowledge graph; Based on the target layer residual index and the protected layer damage index, each process evidence fragment is classified into insufficient removal state, qualified state, protected layer damaged state, combined failure state and invalid measurement state; For process evidence segments with the same process context key, where only the change of one workpiece surface action component reaches the corresponding resolution and the other components satisfy the corresponding tolerance, sort them according to the component and establish parameter responsibility edges based on the state transitions of adjacent valid segments. The parameter responsibility edges from the insufficiency state to the qualified state and from the qualified state to the damaged state of the protected layer are respectively formed into the lower and upper transition intervals. The existing qualified segments sandwiched between the two intervals are taken as the formulation candidates. The normalized boundary margin of each formulation candidate is the smaller of the normalized distances to the failure evidence values ​​on both sides. Based on this, the formulation candidates are selected, and the recommended evidence records are formed and output according to their equipment control quantities and supporting evidence. Obtain feedback process evidence fragments formed by laser processing according to the recommended evidence records. When the process context key of the feedback process evidence fragments is consistent and the quality measurement is complete, write them into the temporary window version. Only update the transition interval corresponding to its state transition if it can form a parameter responsibility edge. Those in the qualified state are added as new formula candidates. For feedback process evidence fragments that are consistent with the process context key but do not meet the quality measurement integrity or window submission conditions, write them into an isolated set. For feedback process evidence fragments with inconsistent process context keys, they are not written into the isolated set, but rather an independent version branch is created based on their own process context key.

2. The method for constructing a knowledge graph and self-recommending formulations for perovskite laser processes according to claim 1, characterized in that, The process context key includes processing step identifier, film stack version, target removal layer identifier, protected layer identifier, substrate type, laser wavelength type, pulse time type, incident direction type, equipment optical path calibration version, and quality measurement scheme version. Only process evidence fragments with identical items are included in the active evidence set corresponding to the same process context key, while process evidence fragments with missing required items are written into the isolation set.

3. The method for constructing a knowledge graph and self-recommending formulations for perovskite laser processes according to claim 1, characterized in that, The formation of the workpiece surface interaction vector specifically includes: Obtain the average power, pulse repetition frequency, and scanning speed from the equipment log; obtain the effective spot area on the workpiece surface and the effective spot size along the scanning direction from the same equipment optical path calibration version. The single-pulse surface dose is obtained by dividing the average power by the product of the pulse repetition frequency and the effective spot area on the workpiece surface. The pulse overlap rate along the scanning direction is obtained by subtracting the product of the scanning speed, the pulse repetition frequency, and the effective spot size along the scanning direction. The effective single-pulse surface dose and the pulse overlap rate along the scanning direction are then written into the workpiece surface action vector.

4. The method for constructing a knowledge graph and self-recommending formulations for perovskite laser processes according to claim 1, characterized in that, Each process evidence fragment is categorized into insufficient state, qualified state, protected layer damaged state, combined failure state, and invalid measurement state, including: An invalid measurement state is determined when any of the following conditions are met: missing required measurement, exceeding the range, non-finite value, invalid measurement scheme version, or non-corresponding processing area. When the measurement is valid and the target layer residual index is unqualified while the protected layer damage index is qualified, it is judged as an insufficient removal state. When both indexes are qualified, it is judged as a qualified state. When the target layer residual index is qualified but the protected layer damage index is unqualified, it is judged as a protected layer damage state. When both indexes are unqualified, it is judged as a combined failure state.

5. The method for constructing a knowledge graph and self-recommending formulations for perovskite laser processes according to claim 1, characterized in that, For process evidence fragments with the same process context key, where only one workpiece surface effect component changes to the corresponding resolution and the remaining components satisfy the corresponding tolerance, sort them according to that component, and establish parameter responsibility edges based on the state transitions of adjacent valid fragments, including: For the same process context key and the surface action component of the workpiece to be analyzed, a single component comparable group is formed based on the other numerical components satisfying the corresponding tolerance and the other categories of components being the same. The valid process evidence fragments within the group are sorted according to the values ​​of the surface action components of the workpiece to be analyzed, and fragments with the same component value corresponding to different states that are indistinguishable are written into an isolated set. After sorting, two adjacent valid segments are obtained in sequence. When the low-value segment is in a state of insufficient removal and the high-value segment is in a state of qualified status, a parameter responsibility edge supporting the lower transition interval is established. When the low-value segment is in a state of qualified status and the high-value segment is in a state of damaged protected layer, a parameter responsibility edge supporting the upper transition interval is established. The remaining states are converted and registered as local sequence breakpoints.

6. The method for constructing a knowledge graph and self-recommending formulations for perovskite laser processes according to claim 5, characterized in that, The parameter responsibility edges from the deficient state to the qualified state and from the qualified state to the damaged state of the protected layer are respectively used to form the lower and upper transition intervals. Existing qualified segments sandwiched between these two intervals are taken as formulation candidates, including: Identify the largest qualified segment consisting of consecutive qualified state segments in a single-component comparable group; When the adjacent segment directly below the maximum qualified segment is in a state of insufficient removal and the adjacent segment directly above the maximum qualified segment is in a state of protected layer damage, the lower conversion interval is defined by the minimum value from the adjacent segment directly below the maximum qualified segment, the upper conversion interval is defined by the maximum value from the maximum qualified segment to the adjacent segment directly above the maximum qualified segment, and the existing qualified segments in the maximum qualified segment are used to form formulation candidates. When multiple discontinuous maximum qualified segments exist, recipe candidates are formed respectively. When any one of the following conditions is met: missing adjacent failure segments on either side, the maximum qualified segment is empty, or there is an unresolved local sequence breakpoint, the corresponding recipe candidate is stopped from being output.

7. The method for constructing a knowledge graph and self-recommending formulations for perovskite laser processes according to claim 1, characterized in that, The normalized boundary margin is determined by the smaller normalized distance from each candidate formulation to the failure evidence values ​​on both sides, and candidate formulations are selected accordingly, including: The distances between the candidate formulation component value and the evidence value of insufficient removal on the direct low side, and the distances between the evidence value of damage to the protected layer on the direct high side and the component value to be analyzed are obtained respectively. The distances between the failure evidence values ​​on both sides are normalized respectively, and the smaller value of the two normalization results is taken. First, select the formulation candidate with the largest normalized boundary margin. When there are ties, determine the selected formulation candidate in ascending order according to the following criteria: higher completeness of evidence on both sides, smaller total width of the transition interval on both sides, smaller distance of the action vector from the most recently verified qualified formulation, and process evidence fragment identifier.

8. The method for constructing a knowledge graph and self-recommending formulations for perovskite laser processes according to claim 1, characterized in that, The recommended evidence record includes process context key, workpiece surface action vector, lower transition interval, upper transition interval, normalized boundary margin, process evidence fragment identifier supporting both transition intervals, quality measurement scheme version and window version, and constitutes confidence characterization based on whether the two transition intervals are closed, the integrity level of the two evidences, the normalized boundary margin, and the effective status of cross-device mapping.

9. The method for constructing a knowledge graph and self-recommending formulations for perovskite laser processes according to claim 1, characterized in that, When the target device is different from the device corresponding to the selected formula candidate, check the calibration relationship between the control quantity of the target device and the action vector on the workpiece surface and the calibration application domain. When the workpiece surface action vector and target operating conditions of the candidate formulation are within the calibration applicable domain, and the target control quantity that uniquely satisfies the target equipment control resolution and safety limit can be obtained by reverse calculation, the target equipment formulation is output. If any of the following conditions are met: exceeding the calibration application range, no solution in reverse solving, or multiple solutions in reverse solving that fail to be decided according to the predetermined minimum control quantity change rule, output the corresponding rejection status and do not output the target equipment recipe.

10. The method for constructing a knowledge graph and self-recommending formulations for perovskite laser processes according to claim 1, characterized in that, Feedback on process evidence fragments should be updated, isolated, or used to create independent version branches, including: Write the feedback process evidence fragment that is consistent with the process context key and has complete quality measurement to the temporary window version; When it forms adjacent parameter responsibility edges to remove insufficient states from qualified states, only the lower transition interval is updated; When the adjacent parameter responsibility edge from the qualified state to the damaged state of the protected layer is formed, only the upper transition interval is updated; If it is in a qualified state, it is considered as a new formulation candidate, and the closure conditions on both sides are re-examined; When the transition intervals on both sides are closed after recalculation and there is no state conflict, submit a new window version. If one of the following conditions is met after recalculation: empty window, state conflict, or insufficient boundary evidence, maintain the original window version and write the new fragment into the isolation set. When the process context keys of the feedback process evidence fragments are inconsistent, an independent version branch is established according to its process context key. After the independent version branch forms the conversion intervals on both sides, it is used as a new window version for the next formulation candidate formation.

Images