A method and system for predicting the lifetime of a semiconductor quartz piece

Abstract

The application discloses a semiconductor quartz piece life prediction method and system, belongs to the field of semiconductor quartz piece life management and industrial process monitoring, and is used for solving the problem that the residual life of a quartz piece is difficult to determine under the conditions of thermal cycling, pickling, ion pollution adsorption and micro-crack accumulation. The application collects multi-source observation of the quartz piece, observation of witness pieces and zero point pieces of the same batch, calibrates bias, drift and missing data, constructs five types of monotonic damage states, outputs residual life and risk by using a multi-state degradation equation and a time sequence residual life prediction head which are constrained by monotonicity, and generates life upper and lower bounds with coverage and replacement or retest decisions by using a preservation interval calibration.

Description

Technical Field

[0001] This invention relates to the field of semiconductor quartz component life management and industrial process monitoring, and particularly to a method and system for predicting the life of semiconductor quartz components. Background Technology

[0002] In the semiconductor manufacturing process, quartz boats, quartz tubes, quartz cavities, and quartz fixtures need to undergo repeated high-temperature thermal cycles, acid pickling and cleaning, and process atmosphere effects. On-site replacement is usually arranged according to the number of uses, maintenance cycle, or manual visual inspection.

[0003] Existing methods cannot simultaneously utilize particle release, ion or metal adsorption, hydroxyl peaks, size drift, and microcrack proxy signals, and cannot uniformly calibrate different process formulations, spatial positions, and sensor drifts; periodic replacement can also easily lead to premature scrapping of still usable parts, or allow degraded parts to continue operating.

[0004] Therefore, there is a need for a method and system for predicting the lifespan of semiconductor quartz components that can overcome the shortcomings of the existing technologies. Summary of the Invention

[0005] One objective of this invention is to propose a method and system for predicting the lifetime of semiconductor quartz components. This system provides a continuous technical solution for particle release, dimensional drift, and abnormal surface conditions caused by thermal cycling, acid washing, ion contamination adsorption, and microcrack accumulation. The solution encompasses multi-source observation calibration, monotonic damage state updates, time-series remaining lifetime prediction, interval calibration, and maintenance write-back. This solution utilizes the residuals of witness components from the same batch and zero-point wafers to correct batch bias and sensor drift, handles missing observations with channel reliability, and maintains the directionality of the degradation state through physical constraints.

[0006] This invention provides a method for predicting the lifetime of semiconductor quartz components, comprising:

[0007] S1. Collect multi-source observations of a single semiconductor quartz component, process context, and observations of witness components and zero-point wafers in the same batch. Use the residuals of the witness components to calibrate the observation bias, drift, and missing measurement status, and output a calibration observation sequence with channel quality weights.

[0008] S2. Extract thermal, chemical, ionic, particle, geometric and crack features from the calibration observation sequence, establish five types of monotonic sub-states, and output the multi-state damage vector as the degradation update input;

[0009] S3. Input the multi-state damage vector and process context into the multi-state degradation equation constrained by monotonicity, update the thermal damage state with temperature integral or Arrhenius constraint, update the structural damage state with crack growth rate upper bound constraint, process the missing channel according to the channel quality weight, and output the damage state trajectory and state uncertainty.

[0010] S4. Input the damage state trajectory, state uncertainty and process context into the time-series remaining lifetime prediction header, and output the point estimate of the remaining cycle number or remaining hours, as well as the risk of particle, size and component failure.

[0011] S5. Utilize the residual prediction of witness parts from the same batch to perform order-preserving interval calibration based on point estimation and risk distribution. Output the upper and lower bounds of remaining lifetime, particle and size risks with target coverage. Generate a replacement decision when the lower bound of remaining lifetime reaches the replacement threshold. Generate a high-resolution retest decision when the interval width reaches the retest threshold. Write the results back to the maintenance system.

[0012] Optionally, S1 includes:

[0013] The batch residual and zero-point residual are obtained by subtracting the observations of the witness pieces from the zero-point reference pieces in the same process batch, the same formula and the same spatial location.

[0014] The median of the residuals within the preset sliding window is used as the bias correction amount, the time difference of the median of the residuals is used as the drift correction amount, and the quality weight of the channels with no missing measurements is set to one, and the quality weight of the channels with missing measurements is set to a preset value between zero and one, and the calibration observation sequence is output.

[0015] Furthermore, it also includes: grouping batch residuals according to formula identification, cavity location and quartz type, using the median and interquartz distance of the group residuals to perform domain offset correction on the sensor output, and normalizing the reciprocal of the interquartz distance to the channel confidence level, the channel confidence level being multiplied by the quality weight and then used for feature calculation in step S2.

[0016] Optionally, S2 includes:

[0017] The cumulative thermal cycling amount is obtained by integrating the temperature time series; the effective cleaning dose is obtained by combining the pickling solution concentration, temperature and action time; the adsorption cumulative amount is obtained by integrating the ion or metal signal over time; the particle count per unit time and the change of the hydroxyl peak relative to the zero point plate are used as particle and hydroxyl features, respectively; the signed difference of the size measurement value relative to the initial size is used as the size drift feature; and the slope of the adjacent window of the microcrack proxy signal is used as the crack growth feature.

[0018] The cumulative amount of thermal cycling is classified into the thermal damage sub-state, the effective cleaning dose and hydroxyl characteristics are classified into the chemical damage sub-state, the cumulative amount of adsorption is classified into the ionic damage sub-state, the particle characteristics are classified into the particle damage sub-state, and the size drift characteristics and crack growth characteristics are classified into the structural damage sub-state.

[0019] The five monotonic states are updated based on the non-negative increments of each feature and the channel quality weights.

[0020] Optionally, S3 includes:

[0021] The temperature integral is input into the Arrhenius function, which is determined by the activation energy and the gas constant. The activation energy is taken from the batch calibration record of the quartz part material, and the gas constant is a preset constant. The function output is used as the non-negative increment of the thermal damage state.

[0022] The slope of the crack growth characteristic is limited to a preset upper bound, which is taken from the crack growth rate quantile value of the historical failure record of the same type of quartz component. The increment exceeding the upper bound is truncated to the upper bound.

[0023] The five monotonic substates at each time step are subject to constraints no less than those of the previous time step, and the increments of each state are weighted and summed using channel quality weights to obtain the physically consistent damage state trajectory.

[0024] Furthermore, the heterogeneous graph node attributes of the five types of monotonic sub-state components, process formulas, and spatial locations are established between nodes based on batch usage relationships, formula relationships, and adjacent spatial relationships. The state difference between adjacent nodes is used as the graph constraint residual, and the graph constraint residual and the channel quality weight are jointly input into the degradation state update in step S3.

[0025] Optionally, S4 includes:

[0026] The damage state trajectory, uncertainty and process context of multiple consecutive time windows are encoded as time-series features. The time-series remaining lifetime prediction head is pre-trained with the historical records of quartz parts with failure markers. The remaining lifetime point estimates measured in cycle number and hours are output respectively. The corresponding risk quantities are output according to the conditions of particle count exceeding particle threshold, size drift exceeding size threshold and damage state reaching failure boundary.

[0027] The particle threshold is taken from the particle limit specified in the process formula, the size threshold is taken from the dimensional tolerance of the quartz part, and the failure boundary is taken from the preset quantile value of the historical failure state.

[0028] Optionally, S5 includes:

[0029] First, the actual failure time of the witnesses in the same batch is converted into the same cycle number or hour unit of measurement as the remaining lifetime point estimate. Then, the difference between the converted actual failure time and the point estimate is used to construct the remaining lifetime inconsistency score. The difference between the actual particle risk, size risk and predicted risk of the witnesses is used to construct particle inconsistency score and size inconsistency score respectively. The quantile values ​​of the three types of inconsistency scores are selected according to the preset target coverage. The remaining lifetime point estimate is subtracted from the remaining lifetime quantile value to form the lower bound of the remaining lifetime, and the quantile value is added to form the upper bound of the remaining lifetime. The particle risk point estimate and the size risk point estimate are subtracted from the corresponding quantile value to form the lower bound of particle risk and the lower bound of size risk respectively, and the corresponding quantile value is added to form the upper bound of particle risk and the upper bound of size risk respectively. The target coverage, the three types of quantile values, the unit of measurement and the sample batch identifier are recorded.

[0030] When the witness document is not invalid within the observation window, the end point of the observation window is used as the right censoring time and is only used for the upper inconsistency score. When the witness document's invalidation time is unavailable, the inconsistency score of the most recent valid batch of the same formula is called. When the coverage verification fails, the interval calibration is stopped and a high-resolution retest decision is generated.

[0031] A replacement decision is generated when the lower bound of the remaining lifetime is less than or equal to the replacement threshold, and a high-resolution retest decision is generated when the difference between the upper and lower bounds is greater than the retest threshold.

[0032] Furthermore, the decision results written back by the maintenance system and the subsequent retest results are used as new samples to update the channel drift correction amount on the edge side. The inconsistency scores are summarized in batches and the quantile values ​​are updated in the cloud. The updated correction amount and quantile values ​​are only used for prediction in the next time window after the witness coverage is verified.

[0033] On the other hand, the present invention also provides a semiconductor quartz component lifetime prediction system, comprising:

[0034] The multi-source observation acquisition and witness calibration module is used to acquire observations from semiconductor quartz components, witness components, and zero-point wafers and output calibration observation sequences with channel quality weights; the feature extraction and monotonic state construction module is used to generate thermal, chemical, ionic, particle, and structural features and five types of monotonic sub-states; the degradation state update module is used to execute multi-state degradation equations constrained by monotonicity and output damage state trajectories; the time-series remaining lifetime prediction module is used to output point estimates of the remaining cycle number or remaining hours and failure risks; the order-preserving interval calibration and maintenance decision module is used to output the upper and lower bounds of the remaining lifetime with coverage, particle and size risks, generate replacement or high-resolution retest decisions, and write back to the maintenance system.

[0035] The beneficial effects of this invention are:

[0036] 1. By observing the offset, drift, and missing measurement status through the residual of the witness piece and the zero-point plate reference calibration, the impact of batch migration and sensor drift on the life judgment of quartz pieces can be reduced.

[0037] 2. By using a multi-state degradation update with five types of monotonic damage sub-states and thermal constraints and upper bound constraints on crack growth rate, the remaining life prediction is made consistent with the actual degradation direction of the quartz component.

[0038] 3. By calibrating the sequence interval, the lifetime upper and lower bounds with target coverage are output, and replacement and high-resolution retest are triggered by the lifetime lower bound and interval width, respectively, reducing the risk of invalid replacement and use with defects. Attached Figure Description

[0039] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0040] Figure 1 This is a flowchart of a semiconductor quartz component lifetime prediction method according to the present invention.

[0041] Figure 2 This is a flowchart of the S5 sequence-preserving interval calibration and maintenance decision of the present invention. Detailed Implementation

[0042] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0043] refer to Figures 1-2 A method for predicting the lifetime of a semiconductor quartz component, comprising:

[0044] S1. Collect multi-source observations of a single semiconductor quartz component, process context, and observations of witness components and zero-point wafers in the same batch. Use the residuals of the witness components to calibrate the observation bias, drift, and missing measurement status, and output a calibration observation sequence with channel quality weights.

[0045] S2. Extract thermal, chemical, ionic, particle, geometric and crack features from the calibration observation sequence, establish five types of monotonic sub-states, and output the multi-state damage vector as the degradation update input;

[0046] S3. Input the multi-state damage vector and process context into the multi-state degradation equation constrained by monotonicity, update the thermal damage state with temperature integral or Arrhenius constraint, update the structural damage state with crack growth rate upper bound constraint, process the missing channel according to the channel quality weight, and output the damage state trajectory and state uncertainty.

[0047] S4. Input the damage state trajectory, state uncertainty and process context into the time-series remaining lifetime prediction header, and output the point estimate of the remaining cycle number or remaining hours, as well as the risk of particle, size and component failure.

[0048] S5. Utilize the residual prediction of witness parts from the same batch to perform order-preserving interval calibration based on point estimation and risk distribution. Output the upper and lower bounds of remaining lifetime, particle and size risks with target coverage. Generate a replacement decision when the lower bound of remaining lifetime reaches the replacement threshold. Generate a high-resolution retest decision when the interval width reaches the retest threshold. Write the results back to the maintenance system.

[0049] In this specific embodiment, S1 includes:

[0050] The life prediction controller establishes observation tasks based on the component number, cavity location, quartz component type, formula identifier, and batch identifier of a single quartz boat. After each process cycle, it reads temperature, pickling solution concentration, ion or metal signal, particle count, dimensional measurement value, and microcrack proxy signal from the edge collector, and simultaneously reads witness pieces and zero-point records of the same process batch. Each original record is written into an original observation table containing timestamp, channel identifier, unit, sampling status, and source device number. Missing fields are saved with a missing measurement mark instead of being replaced with zero value to ensure that subsequent status updates can distinguish between true zero value and unmeasured value.

[0051] The controller is grouped according to formula identification, cavity location, and quartz component type. Within the same process batch, the same formula, and the same spatial location, the observations of the witness piece are subtracted from the zero-point reference to obtain the batch residual and the zero-point residual; for the first... The first channel and the first At each sampling time, the batch residual is recorded as... The zero-point residual is denoted as Both maintain their original units of measurement; when the timestamp deviation of the witness or zero-point slice record exceeds one sampling period, the record is marked as unpairable and removed from the current window, and the reason for removal is saved in the observation table.

[0052] In length of Within the sliding window, the controller calculates the median of the residuals by group, and uses the difference between the median of the batch residuals and the median of the zero-point residuals as the bias correction. The slope of the median residual over time is used as the drift correction. ;in Take the 20 most recent effective cycles of the same formula. The unit is the same as the channel measurement value. The unit is measurement units per hour, and the window length and minimum effective sample size are specified in the versioned calibration configuration table. The configuration table provides fields including grouping key, window length, update time, and version number.

[0053] For channels with batch residuals and zero-point residuals, the controller will return the original value. according to Calibration is performed, among which The current window start time is used, with time differences in hours; for channels with only zero-point slices and no witnesses, the median of the zero-point residuals is used for estimation. And Set to the valid value of the same group in the previous version; for channels where neither of the two types of benchmarks is available, retain the original value but set their calibration status to pending review, and prohibit the channel from triggering a replacement decision on its own.

[0054] The controller calculates the effective sample rate for each channel after calibration. ,in This represents the number of valid observations within the current calibration window that have passed the unit, range, time pairing, and quality mark checks. The number of observations to be collected within this window, both of which are dimensionless non-negative integers. And calculate the interquartile range of the residuals. ;by Indicate data integrity and define stability terms. ,in Consistent with the normalization scale; the two items are weighted according to the calibration configuration table. and merge and crop to To obtain the channel credibility ,and ;

[0055] For missing channels, the controller does not use interpolated values ​​in bias estimation, but instead sets their confidence level to a preset lower limit in subsequent feature calculations. For two consecutive windows that are both below The system generates a channel retest flag and uses the most recent valid calibration value as a read-only reference to prevent missing measurements from being misinterpreted as damage degradation; preset lower limits, consecutive window numbers, and the validity period of the retest flag are all written into the system. The table can be used with recipe identifier and channel type as the query key;

[0056] The controller writes calibration values, confidence levels, bias corrections, drift corrections, missing measurement status, reference batches, and configuration versions into the calibration observation sequence table. The primary key of the sequence consists of the component number, cycle number, and timestamp. The same sequence retains both the original value and the calibration value to facilitate subsequent verification and drift updates. Each channel in the sequence has a unit and quality status field.

[0057] Once the calibration observation sequence is written, the controller transmits the component number, batch identifier, recipe identifier, spatial location, and sequence version to the feature extraction module of S2. S2 only reads the calibration values ​​and their confidence levels, and does not reread the original sensor data, thus ensuring that bias, drift, and missing measurement processing remain consistent throughout the lifetime prediction chain.

[0058] The calibration grouping is confirmed by the batch management table each time the batch is switched. The batch management table must save at least the formula identifier, cavity number, spatial coordinates, quartz type, witness number, and zero point number. The controller closes the old window and opens the new window with the effective time of the batch management table as the boundary to avoid mixing the residuals of different formulas into the same median calculation.

[0059] The channel confidence calculation results are simultaneously written to the edge cache and the cloud calibration database. The edge cache stores the original residual summaries of the two most recent windows. The effective sample rate and version number are stored in the cloud database, which saves the historical distribution of the groups. When the version on the edge side is inconsistent with the version on the cloud side, the edge side continues to use the version that has recently passed the coverage verification and issues a synchronization event.

[0060] For channels with units such as temperature, concentration, and size, the calibration module performs a unit consistency check before writing, comparing the Celsius temperature, mass concentration, millimeter size, and hourly drift with the channel metadata respectively; records with inconsistent units or out-of-range values ​​are only entered into the anomaly queue and do not participate in bias, drift, or confidence calculations. The anomaly queue records the source device, timestamp, and reviewer.

[0061] Before outputting the calibration observation sequence, the controller sorts the components by number and checks the continuity of the cyclic numbering. If a cyclic missing number occurs, a placeholder record with a missing measurement mark is inserted into the sequence without falsifying the measurement value. At the same time, the missing number interval is passed to S2 so that the feature extraction module can calculate the integral and slope according to the real time interval.

[0062] The calibration configuration table is generated from the pairing records of witness pieces and zero-point pieces from the six most recent valid batches. The generation task calculates the window midpoint, interquartile range, and time slope according to the grouping key, and verifies whether the residuals converge after correction on the reserved batches. The configuration record uses the formula identifier, cavity location, and quartz piece type as the query key. The value fields include window length, minimum sample size, bias range, drift range, and effective time. If the query does not find a match, only the most recently verified record of the same type is allowed to be used, and the rollback version and pending review status must be written.

[0063] In this specific embodiment, S2 includes:

[0064] The feature extraction module reads the calibration observation sequence output by S1 and rearranges the time axis with the process cycle as the sampling unit; it constructs a temperature time series for the temperature channel using the cycle start and end times and sampling intervals; it constructs a chemical record for pickling solution concentration, temperature, and treatment time; it constructs a time integral record for ion or metal signals; it counts the particle channel by unit time; it saves the current size and initial size for the geometric channel; and it saves the slope and peak position of adjacent windows for microcrack proxy signals; all features are associated with the component number, formula version, cavity position, and channel confidence level.

[0065] For the Temperature sequence of a process cycle The module calculates the thermal cycle integral. ,in The batch calibration temperature is given in degrees Celsius, and the integration results are stored in degrees Celsius per hour; pickling solution concentration. ,temperature and duration of action Then according to To form an effective cleaning dose, among which The difference between the start and end times of the pickling process, in hours, is obtained directly from the process event timestamp; samples with abnormal concentrations exceeding the process formula range are marked as chemical channel anomalies.

[0066] Ion or metal signal Integrating over time Indicates the cumulative amount of adsorption, particle characteristics Hydroxyl spectrum characteristics obtained by subtracting from the zero-point plate reference based on the particle count per unit time. The change in the hydroxyl peak in this cycle relative to the corresponding hydroxyl peak on the zero-point plate is obtained, with units consistent with the measured peak value. When this non-negative peak change exists, it is used as the observation input for the chemical state increment and written into the database. Other missing or negative changes are handled by missing mask; when the number of valid sampling points in the integration interval is less than the minimum proportion specified in the configuration table, the relevant integration results are not padded with zeros, but a missing flag is output and the corresponding quality weight is passed into the subsequent state construction.

[0067] Geometric features From the current size Compared with the initial size The signed difference normalization yields ,in Provided by the component's first qualified measurement record, in millimeters; microcrack characteristics Take the slope of the proxy signal in adjacent time windows , For crack proxy signals, the time difference is measured in hours; only when Missing When the value is zero or the current dimension record has a missing measurement flag, the geometric feature will be marked as missing and the channel will be disabled for independent failure assessment; when the measurement is valid and Time Retention Record the measurement as a zero increment according to the gauge resolution without rewriting it as a missing measurement;

[0068] Module , , , , and Normalization to At the same time, retain the untrimmed signed geometric deviations. and its measurement unit; the geometric amplitude field used for model input is denoted as Mapped to the bidirectional range of the dimensional tolerance table S3's bidirectional direction determination only reads... The normalization benchmark is derived from the material batch characteristic range table. The table stores the lower limit, upper limit, unit, source batch, and update time according to component type and recipe version; values ​​exceeding the upper limit are truncated to 1, and values ​​below the lower limit are truncated to 0, and the truncation flag is written to the feature record to prevent extreme values ​​from changing the direction of the monotonic state;

[0069] Five types of monotonic substates are established in the state construction table, namely thermal damage substates. Chemical damage substate Ion damage substate Particle damage sub-state and structural damage substate For each iteration, write the non-negative increment of the corresponding feature into the state input vector. If the feature is a missing test, then a missing test mask is also written. and channel credibility ;

[0070] For multiple features derived from the same physical quantity, the module only applies the deduplication rules in the feature relevance table to duplicate derived fields of the same physical quantity and retains the main feature, recording the merged fields in the table; particle count. With hydroxyl peak Corresponding to the particle damage sub-state and the chemical damage sub-state respectively, even if the conflict direction is given in the same cycle, their respective non-negative increments are retained, and the input of one type of state is not changed to the verification field of another type; the conflict is only used as a correlation and quality verification record, and if necessary, the quality weights of the two channels are reduced and backed up according to the conservative values ​​of their respective states; the non-negative change of the hydroxyl peak is entered according to the chemical increment synthesis rule. The non-negative change in particle count enters The relevant records and rollback reasons are written into the state construction log for use in S3 uncertainty propagation;

[0071] The module writes five types of sub-states, incremental vectors, missing test masks, channel confidence, feature range versions, and cycle timestamps into a multi-state damage vector table, using the component number and cycle number as the primary keys. The output of this table is the input of the degradation equation of S3. S3 reads the same version of the state vector and quality field to avoid scale drift across steps caused by renormalization.

[0072] The feature extraction module establishes a traceable field mapping table for each type of input. The mapping table stores the original channel, derived field, unit, normalization range, zero-point slice source, and feature version. The same original channel must not be mapped to two features with different meanings at the same time. When a conflict occurs, the main mapping is retained and the conflicting field is written to the verification log.

[0073] Temperature integral, cleaning dosage and adsorption accumulation are all calculated using real sampling timestamps instead of a fixed number of points. The timestamps are sorted and then integrated using a trapezoidal shape. When the interval between adjacent samplings is greater than the maximum interval allowed by the formulation, the integral is calculated in segments and a time break mark is set in the feature record. Subsequent state updates will adjust the increment of this segment according to the confidence level.

[0074] Particle and hydroxyl peak characteristics Before entering normalization, geometric dimensions and crack proxy signals are verified against the allowable range of the process formulation and the gauge resolution, respectively. The measurement range, resolution and anomaly flag are transmitted to normalization and chemical increment calculation along the same feature field. Effective changes below the gauge resolution are retained as zero increments and recorded in the resolution field. Changes exceeding the range are clipped to the upper limit of the range and enter the retest candidate to avoid gauge saturation being misjudged as rapid degradation.

[0075] The initial values ​​of the five sub-states are read from the first qualified inspection record of the quartz part. The initial record includes the material batch, size reference, spectral peak reference and particle reference. If the first qualified record is missing, the batch reference of the same type of quartz part is called and the initial uncertainty is increased. The state vector still retains the reference source for the process noise setting of S3.

[0076] The module performs integration on the temperature time series to obtain the thermal cycle accumulation and writes it into the thermal feature field; it uses the signed difference between the size measurement value and the initial size as the size drift feature, and the adjacent window slope of the microcrack proxy signal as the crack growth feature. All three types of features retain the original units, normalized results and calculation window numbers.

[0077] When classifying features, size drift features and crack growth features are classified into structural damage sub-states, and thermal cycling accumulation is classified into thermal damage sub-states. The classification table records the feature name, target sub-state, increment direction, and mass weight fields. The classification table version is passed to S3 along with the multi-state damage vector.

[0078] The normalization module applies each feature according to Perform the mapping, where For the current eigenvalue, and Provided by a table of component types, formulation versions, and material batch ranges; when The batch default value of this feature is used and written to the degradation flag, and it is not used in the monotonic state calculation with infinite values;

[0079] The material batch characteristic range table is constructed from the initial qualified record, stable operation record, and record before actual failure. The module calculates the effective lower limit and upper limit according to the characteristic unit and removes samples with gauge failure flags. The range table is recalculated for each new batch that passes the review. The old version is still bound to the historical state vector. When the runtime query does not find the missing feature, it keeps the missing measurement and increases the uncertainty. The range is forced to normalize without using other units.

[0080] In this specific embodiment, S3 includes:

[0081] The degradation update module reads the multi-state damage vector and process context of S2 according to the component number, and establishes a five-dimensional state vector for each cycle. The process context includes formulation temperature, pickling concentration, reaction time, ion power, chamber location, batch identifier, and cycle interval. All quantities are recorded in the context record with units, source equipment, and configuration version.

[0082] For thermal damage conditions, the module reads the activation energy from the material batch calibration table. and with Arrhenius's head Calculate the thermal damage rate, where The pre-calibration factor is in hourly increments. In Joules per mole The preset gas constant is joules / (molar Kelvin). The temperature is in Celsius; the thermal state increment within one cycle is... The calibration table uses material batch and formulation version as keys and saves the fitted samples, version number and effective time;

[0083] For both chemical and ionic states, the module calculates the normalized dimensionless effective cleaning dose based on the output of S2. Hydroxyl spectrum characteristics With normalized dimensionless adsorption cumulative amount Calculate nonnegative increments , ,in , , The regression coefficients are non-negative and dimensionless, and are provided by the monotonic regression coefficient table of the same batch of historical failure records. , and All use the normalized field of S2, and the original unit value is only saved as an audit field to make the obtained increment dimensionless; when the formula version is not found in the coefficient table, the most recent valid version of the same formula is called and a downgrade flag is written. If there is still no valid version, the state increment is frozen and the state uncertainty is increased.

[0084] For particle and structural states, the module converts the normalized dimensionless particle features and sign-preserving geometric features output by S2 into non-negative increments, and then uses them respectively. and ,in This represents the non-negative dimensional degradation increment obtained by mapping bidirectional dimensional tolerances; for unidirectional dimensional tolerances, the dimensional tolerance table provides direction coefficients and assigns the corresponding direction... Mapped to For two-way tolerances, the absolute value increment of the signed geometric deviation is used, and zero drift is maintained at zero increment; where , and All are non-negative and dimensionless normalization calibration coefficients, provided by the calibration table of component type and formula version, which makes the normalization feature generate dimensionless increments; when the dimensional tolerance direction configuration is not hit, the dimensional increment is frozen and marked for retesting; This is the preset upper quantile value for the crack growth rate in historical failure records of similar quartz components, expressed in hours. For the cyclic interval, the total increment of size drift and crack exceeding the upper limit is truncated to the upper limit and the truncation mark is saved;

[0085] The module first combines the graph constraint residuals with edge quality weights to form a correction, and then applies this correction to states with valid adjacent edges during the state update phase. Figure 1 Consistency constraints; for missing or low-quality edges, reduce the correction amount and retain only the observation increment of this component; then apply monotonicity constraints at each time step. ,in When the input increment becomes negative due to noise, the increment is set to zero without reverting to the existing state, ensuring that the five states of thermal, chemical, ionic, particle, and structural do not decrease over time; the upper limit of the state is given by the state configuration table, and states exceeding the upper limit are kept at the upper limit and a saturation flag is recorded.

[0086] For channels with missing tests, the module uses quality weights. For the weighted increment By performing weighted summation, we obtain the weighted increment for entering the monotonic state update. ; The weight is determined by the channel reliability of S1 and the feature missing mask of S2. When the missing mask is 1, the weight takes the conservative value specified in the configuration table. When the continuous missing test exceeds two cycles, the process noise variance of this state is increased at the same time.

[0087] The module also establishes a heterogeneous graph based on components, recipes, and spatial locations. Graph node fields include five states, component type, batch, and cavity coordinates, while edge fields include usage relationships within the same batch, relationships within the same recipe, and adjacent spatial relationships. For adjacent nodes... , Calculate state difference And form graph-constrained residuals according to the constraint coefficients of the edge type. ,in The non-negative dimensionless edge weights are obtained from the same batch, same formula, or adjacent spatial relationship configuration table. The source and effective version are saved along with the edge record. When an adjacent node is missing, the edge is deleted but the node status is retained. Zero difference replacement is not used.

[0088] The module writes various state increments, state trajectories, graph constraint residuals, saturation flags, missing measurement masks, and process noise variances into the damage state trajectory table, and uses the cycle number, state version, and graph configuration version as joint version identifiers; the state uncertainty is synthesized by the weighted incremental variance and the graph constraint residual variance, and then passed to the time-series remaining lifetime prediction head of S4 along with the trajectory and process context; the graph constraint residuals and their channel quality weights participate in the state correction as correction and constraint terms during state updates, and monotonicity and state upper limit checks are performed after correction;

[0089] Each coefficient of the degradation equation carries the material batch, formulation version, and effective time. The module locks the coefficient version used in the current cycle at the beginning of the cycle and allows configuration updates only after the cycle ends. This way, even if the cloud publishes new calibration results during operation, the current trajectory is still generated by the same set of parameters, which facilitates reproduction and auditing.

[0090] For loops where both temperature integral and Arrhenius rate exist, the module prioritizes the rate integral with time resolution and uses the temperature integral as a cross-validation measure; when the difference between the two exceeds the proportion specified in the validation table, the rate integral result is retained but the thermal state uncertainty is increased, and a thermal model inconsistency flag is written into the trajectory.

[0091] The coefficients of the graph constraint residuals are set with different dimension normalization factors according to the edge type. The same batch usage relationship is used to constrain the consistency of the degradation direction, the same formula relationship is used to constrain the similarity of material response, and the adjacent spatial relationship is used to identify local anomalies in the cavity. Each edge saves the generation conditions, weight source and deletion reason, and the graph version number is added when the graph structure changes.

[0092] Before outputting, the module performs cycle-by-cycle monotonicity, upper bound, and missing test propagation checks on the five types of state trajectories. If any check fails, the calculated trajectory is not deleted. Instead, the cycle is marked as pending review and the conservative state of the previous cycle is output. At the same time, the reason for the exception and the corresponding input field are sent to S4 as a window mask.

[0093] The process noise variance is jointly determined by the residual variance of each type of feature within the calibration window and the channel confidence. When the quality weight is reduced, the uncertainty is increased without reversely changing the damage increment. The module writes the variance, missing measurement duration and graph constraint residual into the state uncertainty field, so that S4 can distinguish between measurement noise and actual degradation within the time window.

[0094] When the Arrhenius parameter, crack growth upper bound, or graph constraint coefficient undergoes a version switch, the module establishes a new version segment in the trajectory table and saves the last state of the old version; the new version continues to apply non-negative increments from this state, and the version switch point is used as the time series feature of S4 to prevent the parameter update from being mistaken by the model as a sudden damage to the component.

[0095] At the end of each loop, the module writes the five states, each state increment, quality weight, process noise and graph residual into an immutable trajectory record. The verification task asynchronously checks whether the state is monotonic, whether it exceeds the upper bound, and whether there are any unexplained mutations. Only trajectories that pass the verification are entered into the S4 window cache, while records that fail retain their original values ​​and carry a flag indicating that they are pending verification.

[0096] The state trajectory table also stores the physical quantity sources and unit mappings corresponding to each state, where the thermal state corresponds to the Celsius-hour integral, the chemical state corresponds to the effective cleaning dose, the ionic state corresponds to the adsorption accumulation, the particle state corresponds to the unit time count, and the structural state corresponds to the size and crack characteristics; this mapping is used for input normalization of S4 and result interpretation by maintenance personnel.

[0097] The trajectory output simultaneously records the start time, end time, and actual duration in hours of this cycle, allowing S4 to maintain the same time reference between cycle measurement and hourly measurement.

[0098] In addition to storing the variances of each state, the state uncertainty also stores a summary of the covariance between the five types of states and the set of valid channels that generated the summary. When graph constraints associate two adjacent components, only the uncertainty summary is updated and the state values ​​of adjacent components are not directly copied, thereby avoiding the abnormal state of one component from contaminating other components in the same batch. The covariance summary and the trajectory version are locked together and transmitted to S4.

[0099] In this specific embodiment, S4 includes:

[0100] The time-series remaining lifetime prediction head reads the damage state trajectory, state uncertainty, and process context of S3, and generates continuous time windows by component number; each window contains the most recent The five states of each cycle, the increment of each state, the missing test mask, the graph constraint residual, the process code, and the time interval. Take 32 cycles. If the window is insufficient, fill it with the nearest valid window of the same recipe and mark the filling position.

[0101] The prediction head employs a gated recurrent network with time interval encoding and a linear output layer. The input tensor fields include... , , , With process context encoding, output the remaining lifetime point estimate for the cycle unit. Estimated remaining lifespan in hourly units And risks of particle, size and component failure. , , The model structure, input field order, and normalized version are written into the inference checklist. ;

[0102] The model training data consists of historical quartz component records, witness component failure markers, and actual replacement times in the maintenance system, and is divided into training and validation sets according to component batches. The training objective consists of the logarithmic error of the remaining cycle number or remaining hours, the binary cross-entropy of the risk, and the monotonicity penalty. The monotonicity penalty constrains that the predicted remaining lifetime must not increase when the damage state increases. The model version is only released when the time error, risk calibration error, and monotonicity violation rate of the validation set are all lower than the threshold in the model configuration table.

[0103] The prediction head encodes the component's formulation, cavity location, and quartz component type into a context vector, and uses the state uncertainty of S3 as an additional input. It uses masked attention on missing fields within the window to prevent them from participating in effective feature aggregation, while simultaneously passing the uncertainty to the output layer, ensuring the risk output includes a conservative term due to insufficient observation. If the effective state dimension is lower than the minimum dimension and there is a previous effective window for the same formulation, the output of that window is used, and the prediction state is set to "to be retested" with a "historical window rollback" flag. When S5 receives this flag, it does not use the rollback point estimate for interval calibration or automatic replacement comparison, but only generates a high-resolution retest decision. New window prediction is only resumed after the retest passes quality and coverage verification. If there is no previous effective window for the same formulation, no lifetime point estimate or risk interval is generated; only the "to be retested / insufficient data" status, the reason for the missing data, and the high-resolution retest task flag are output. When S5 receives this status, it stops interval calibration and automatic replacement comparison, retaining only the high-resolution retest decision and audit records.

[0104] Particle risk is determined by comparing the predicted particle state with the particle threshold specified in the process formulation; dimensional risk is determined by comparing the geometric state with the dimensional tolerances of the quartz component; and component failure risk is determined by the combined distance between the five particle states and historical failure boundaries. All risk values ​​are tailored to... The larger the value, the higher the probability of failure. The risk threshold is queried from the risk threshold table maintained by formula version. If it is not hit, the most recent effective version of the same formula is used and the version rollback is recorded.

[0105] The prediction head writes the cycle point estimate and hourly point estimate into the lifetime point estimate table, and writes the three types of risks, model version, window start and end time, input state version and uncertainty summary into the risk results table. When the model output has a negative lifetime or exceeds the historical effective range, it is first trimmed according to the boundary and an abnormal output flag and an automatic maintenance hard block flag are generated. This point estimate and risk results must not be entered into the S5 interval calibration center or automatically replaced for comparison.

[0106] For multiple consecutive windows, the prediction header retains the window sequence, model input summary, and output difference to determine whether a sudden change has occurred in the prediction. If the adjacent differences of the point estimate exceed the change limit set in the model configuration table, the prediction confidence of the window is reduced and a high-resolution retest candidate is triggered without changing the historical results already written.

[0107] After inference is completed, the module passes the lifetime point estimate, particle risk, size risk, component failure risk, prediction uncertainty, risk threshold version, and retest candidate flag to S5; S5 uses these point estimates as the center value of the order-preserving interval calibration and retains the correspondence between the cycle unit and the hour unit.

[0108] Historical training records are sampled by batch and cavity stratification to ensure that witnesses from the same batch do not appear in the training window and verification window at the same time; in addition to the state trajectory, training samples also save formula change points, sensor retest points and actual failure or right deletion marks, and the model list records data segmentation rules and sample versions.

[0109] The time interval encoding of the gated recurrent network uses the logarithm of the recurrent interval, and the missing test mask is embedded independently. The model outputs a hidden state summary and the pre-calibration probabilities of the three types of risks in each window. The hidden state summary is saved together with the start and end times of the window, so that S5 can distinguish whether the change in point estimation comes from the change in damage or from the change in observation uncertainty.

[0110] The risk threshold table simultaneously stores the particle count unit, dimensional tolerance unit, and quantile value of the component failure boundary. The threshold record uses the formula identifier, component type, and process version as the query key. When no query result is found, the most recent effective version of the same formula is used. However, if the effective time of the most recent version is earlier than the current formula switching point, only the "to be retested" flag is output and no automatic replacement suggestion is generated.

[0111] When the prediction window contains completion location or model version rollback flags, the module encodes these flags into the results table and reduces the window confidence to the conservative coefficient in the configuration table. When the prediction window contains abnormal output flags, regardless of whether the window confidence or S5 coverage verification passes, automatic maintenance hard blocking is maintained. Point estimates and risks cannot be used as interval centers or replacement comparison bases. Only high-resolution retest decisions and abnormal audit records are generated. Only when subsequent new windows have no abnormal flags and pass quality, measurement, and coverage verification can point estimates be restored as the center value for maintenance decisions.

[0112] Before each model release, the prediction service performs replay verification, generating cycle lifetime, hour lifetime, and three types of risks one by one in the order of historical windows, and aligning them with actual failure or right censoring records. The replay records save the prediction time, model version, input state version, point estimation error, risk threshold hit results, and inference time. Only when the errors of both cycle and hour lifetime measurements meet the configuration requirements and the three types of risks are in the same direction will the model version enter the production inference list.

[0113] During production inference, the model input window and output result use the same component number and window end time as the association key. After the service completes the inference, it checks whether the two lifetime point estimates are consistent with the cycle conversion relationship. If the conversion deviation exceeds the allowable ratio of the formula, the original output is retained but the window is marked as measurement inconsistency. S5 only generates retest decisions and does not publish new lifetime ranges.

[0114] Before writing to the window, the model input service checks whether the versions of the damage trajectory, state uncertainty, and process context are consistent. If any object is missing a joint version identifier, the splicing window is rejected. The reason for rejection is written to the inference task table and returned to S3 to complete it, so as to avoid inputting the state and context of different parameter versions into the same lifetime prediction.

[0115] After the inference task is completed, the input summary verification value, model file verification value and execution node number are saved. When S5 receives the result, the verification value is recalculated. Any result with inconsistent verification will not be included in the interval calibration and the inference task will be resubmitted. At the same time, the number of task retries and the completion time are recorded.

[0116] In this specific embodiment, S5 includes:

[0117] The interval calibration module reads the lifetime point estimate and risk results of S4, and reads the actual failure time, observation window endpoint, particle measurement risk and size measurement risk from the witness part table of the same batch; the actual failure time of the witness part is first converted into the number of cycles or hours according to the cycle period of the component formulation, so that it uses the same unit of measurement as the point estimate, and the conversion factor, sample batch and failure mark are retained in the record;

[0118] For each witness unit with an actual failure time, the module calculates the lifetime residual. Particle residual Dimensional residuals For witnesses that have not failed within the observation window, the window end point is used as the right censoring time, which is only used to form the upper inconsistency score. It is forbidden to use the censoring time as the actual failure time in the lower bound calculation.

[0119] The module summarizes the three types of inconsistency scores separately within the same batch and sorts them according to the target coverage. Query quantile value , and The quantile table is stored by formula, component type, batch, and sample quantity. Fields include quantile value, unit of measurement, coverage, sample batch identifier, calculation time, and version number. The coverage is set to 0.90. When the number of samples in the same batch is less than the minimum sample size, it rolls back to the most recent valid batch of the same formula and saves the rollback flag.

[0120] Lifespan range by and Generation, particle risk range according to , The same clipping rules apply to the size risk range; all lower bounds must not exceed the upper bound, and the range width... The risk interval width is written into the calibration results table;

[0121] The module updates the batch residual distribution with the difference between the actual failure time of the witness and the estimated predicted point, and uses order-preserving calibration to ensure that the absolute value of the residual does not change due to sample sorting; the calibration results must pass the coverage check, which checks the proportion of the actual value falling into the interval on the reserved witnesses. If the proportion is lower than the target coverage, the current interval is stopped from being published and marked as needing high-resolution retesting.

[0122] When the lower limit of lifespan Less than or equal to the replacement threshold Timely generation of replacement decisions; The maintenance strategy table is provided by component type and recipe version, with units and Consistent, with lead time and safety margin; lifespan retest threshold Similarly, the maintenance strategy table provides values ​​by component type, formula version, and lifespan unit, taking non-negative values ​​with corresponding units and interval widths. If the width is consistent, a miss occurs, or the units are inconsistent, a width comparison is not performed, and a check-up flag is generated; when the lifespan range width... Greater than the retest threshold When the width of the risk interval for particles or size exceeds its respective threshold, a high-resolution retest decision is generated. The two decisions can coexist, but the triggering conditions are recorded separately.

[0123] If a new retest result has been generated for the same batch of witness documents, the module will combine the decision result written back by the maintenance system with the retest result to form a new sample, update the channel drift correction amount on the edge side, and summarize the inconsistency score and update the quantile value in the cloud by batch. The updated correction amount and quantile value are only allowed to be used for prediction in the next time window after the witness document coverage verification is passed again. If the verification fails, the previous passed version will continue to be used.

[0124] The module writes the lifetime upper and lower bounds, cycle and hour measurement units, particle and size risk range, target coverage, three class quantiles, sample batch identifier, replacement decision, high-resolution retest decision, calibration version and coverage verification status into the maintenance system; the maintenance system generates work orders or retest tasks based on this, and sends the execution results back as witness samples for the next window, forming a closed loop from observation, status update, lifetime prediction to maintenance decision-making;

[0125] The interval calibration sample table simultaneously stores the observation window start point, observation window end point, actual failure time, whether right censoring is used, the model version used during prediction, and the calibration version. When the witness has not failed by the end of the observation window, the end of the window is used as the right censoring time, and this time is only used for the upper inconsistency score and does not participate in the actual failure sample set of the lower lifetime.

[0126] For cases where the actual failure time of the witness test is unavailable but there are inconsistencies in the most recent valid batch of the same formula, the module only calls the quantile value of the batch that passed the coverage verification, and does not call the quantile value that failed the verification; if the batch failed the verification, the interval calibration is stopped directly and a high-resolution retest decision is generated, and the maintenance system records the reason for the stop.

[0127] Coverage verification is performed separately by component type, formulation version, and unit of measurement. The lifetime cycle count range must not be verified by mixing samples from hourly intervals. The verification results include target coverage, actual coverage, number of samples, number of right censored samples, quantiles used, and confidence flags. If the actual coverage is insufficient, the upper and lower bounds of the previous version are retained and new quantiles are prohibited from taking effect.

[0128] After receiving the write-back result, the maintenance system will replace the decision, high-resolution retest decision, execution time and retest data feedback interval calibration module. The module will use the feedback record as a new sample for the next window, and update the channel drift correction on the edge side first. After completing the batch quantile value summary in the cloud, it will perform coverage verification to ensure that the update result will not skip the unverified state and directly enter the prediction chain.

[0129] After the interval results are generated, the module writes the results back to the component life archive of the maintenance system, and uses the component number, batch identifier, window end time and calibration version as idempotent keys; when writing back repeatedly, only the verification status and execution results under the same idempotent key are updated, and replacement work orders are not created repeatedly. The maintenance system saves the upper and lower bounds, risk intervals and trigger thresholds at the same time.

[0130] When the retest is completed and the coverage verification is passed, the module adds the retest sample to the set of inconsistencies in the next batch. If the retest sample is inconsistent with the original observation in terms of unit, timestamp or component number, the original interval is retained and the sample is set to be reviewed, and it is prohibited from participating in the edge drift correction amount and cloud quantile value update.

[0131] The maintenance system's write-back interface returns the receiving time, file version, and work order status. The interval calibration module associates and saves this receipt with the current upper and lower bounds, risk interval, and coverage verification status. Only when a successful receipt is received and the calibration status is passed is the current decision marked as effective; otherwise, the previous version result is retained and a new pending task is generated.

[0132] At the start of the next time window, the maintenance system returns the work order execution status, whether the retest is completed, and whether the component has been replaced to the prediction service; the new window for replaced components starts from the new initial qualified record, while the interval and correction amount of the previous pass version are used for unreplaced components to avoid mixing the observation jumps caused by maintenance actions into the same life trajectory.

[0133] To prevent batch sample drift, the cloud performs monthly group statistics on the actual coverage, average interval width, and retest trigger rate of different formulations; only the quantile version with actual coverage not lower than the target coverage and interval width not continuously increasing remains valid, while other versions are rolled back to the most recently verified quantile value, and the edge side is notified to freeze the corresponding drift correction amount.

[0134] The final write-back record also includes the decision reason code, the difference between the lower bound and the replacement threshold, the difference between the interval width and the retest threshold, and the upper bound of three types of risks. Maintenance personnel can review the trigger direction based on the same record. The system uses the interpretation that the smaller the lower bound of lifespan, the more urgent the risk, and the larger the upper bound of risk, the higher the risk, and does not use the opposite sorting rules for the same field.

[0135] The replacement threshold table and retest threshold table are managed by the maintenance strategy version. Fields include component type, formula identifier, life measurement unit, threshold, lead time, risk level, effective time and approval version. The interval calibration module queries by component type, formula and unit. If no match is found or the units are inconsistent, automatic comparison is not performed. Instead, a decision to be reviewed is generated and the missing threshold record is specified to prevent the hourly threshold and the cycle threshold from being mixed.

[0136] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0137] This invention combines multi-source quartz component observation with residuals from witness components in the same batch, enabling calibration results to participate in damage status updates, remaining life predictions, and interval decisions, forming a closed-loop data chain for a write-back maintenance system.

[0138] This invention adds channel reliability, monotonicity constraints, and order-preserving distribution calibration between feature calibration, degradation status update, and maintenance decision-making. It also adapts to batch and formulation changes through edge-side drift correction and cloud quantile updates, thereby more fully realizing the consistency of lifespan early warning and maintenance decisions.

Claims

1. A method for predicting the lifetime of a semiconductor quartz component, characterized in that, include: S1. Collect multi-source observations of a single semiconductor quartz component, process context, and observations of witness components and zero-point wafers in the same batch. Use the residuals of the witness components to calibrate the observation bias, drift, and missing measurement status, and output a calibration observation sequence with channel quality weights. S2. Extract thermal, chemical, ionic, particle, geometric and crack features from the calibration observation sequence, establish five types of monotonic sub-states, and output the multi-state damage vector as the degradation update input; S3. Input the multi-state damage vector and process context into the multi-state degradation equation constrained by monotonicity, update the thermal damage state with temperature integral or Arrhenius constraint, update the structural damage state with crack growth rate upper bound constraint, process the missing channel according to the channel quality weight, and output the damage state trajectory and state uncertainty. S4. Input the damage state trajectory, state uncertainty and process context into the time-series remaining lifetime prediction header, and output the point estimate of the remaining cycle number or remaining hours, as well as the risk of particle, size and component failure. S5. Utilize the residual prediction of witness parts from the same batch to perform order-preserving interval calibration based on point estimation and risk distribution. Output the upper and lower bounds of remaining lifetime, particle and size risks with target coverage. Generate a replacement decision when the lower bound of remaining lifetime reaches the replacement threshold. Generate a high-resolution retest decision when the interval width reaches the retest threshold. Write the results back to the maintenance system.

2. The method for predicting the lifetime of a semiconductor quartz component according to claim 1, characterized in that, S1 includes: The batch residual and zero-point residual are obtained by subtracting the observations of the witness pieces from the zero-point reference pieces in the same process batch, the same formula and the same spatial location. The median of the residuals within a preset sliding window is used as the bias correction amount, the time difference of the median of the residuals is used as the drift correction amount, and the quality weight of the channels with no missing measurements is set to one, and the quality weight of the channels with missing measurements is set to a preset value between zero and one, and the calibration observation sequence is output.

3. The semiconductor quartz component lifetime prediction method according to claim 1, characterized in that, S2 include: The cumulative thermal cycling amount is obtained by integrating the temperature time series; the effective cleaning dose is obtained by combining the pickling solution concentration, temperature and action time; the adsorption cumulative amount is obtained by integrating the ion or metal signal over time; the particle count per unit time and the change of the hydroxyl peak relative to the zero point plate are used as particle and hydroxyl features, respectively; the signed difference of the size measurement value relative to the initial size is used as the size drift feature; and the slope of the adjacent window of the microcrack proxy signal is used as the crack growth feature. The cumulative amount of thermal cycling is classified into the thermal damage sub-state, the effective cleaning dose and hydroxyl characteristics are classified into the chemical damage sub-state, the cumulative amount of adsorption is classified into the ionic damage sub-state, the particle characteristics are classified into the particle damage sub-state, and the size drift characteristics and crack growth characteristics are classified into the structural damage sub-state. The five monotonic states are updated based on the non-negative increments of each feature and the channel quality weights.

4. The method for predicting the lifetime of a semiconductor quartz component according to claim 1, characterized in that, S3 includes: The temperature integral is input into the Arrhenius function, which is determined by the activation energy and the gas constant. The activation energy is taken from the batch calibration record of the quartz part material, and the gas constant is a preset constant. The function output is used as a non-negative increment of the thermal damage state. The slope of the crack growth characteristic is limited to a preset upper bound, which is taken from the crack growth rate quantile value of the historical failure record of the same type of quartz component. The increment exceeding the upper bound is truncated to the upper bound. The five monotonic substates at each time step are subject to constraints no less than those of the previous time step, and the incremental states of each state are weighted and summed using channel quality weights to obtain a physically consistent damage state trajectory.

5. The method for predicting the lifetime of a semiconductor quartz component according to claim 1, characterized in that, S4 includes: The damage state trajectory, uncertainty and process context of multiple consecutive time windows are encoded as time-series features. The time-series remaining lifetime prediction head is pre-trained with the historical records of quartz parts with failure markers. The remaining lifetime point estimates measured in cycle number and hours are output respectively. The corresponding risk quantities are output according to the conditions of particle count exceeding particle threshold, size drift exceeding size threshold and damage state reaching failure boundary. The particle threshold is taken from the particle limit specified in the process formula, the size threshold is taken from the dimensional tolerance of the quartz part, and the failure boundary is taken from the preset quantile value of the historical failure state.

6. The semiconductor quartz component lifetime prediction method according to claim 1, characterized in that, S5 include: First, the actual failure time of the witnesses in the same batch is converted into the same cycle number or hour unit of measurement as the remaining lifetime point estimate. Then, the difference between the converted actual failure time and the point estimate is used to construct the remaining lifetime inconsistency score. The difference between the actual particle risk, size risk and predicted risk of the witnesses is used to construct particle inconsistency score and size inconsistency score respectively. The quantile values ​​of the three types of inconsistency scores are selected according to the preset target coverage. The remaining lifetime point estimate is subtracted from the remaining lifetime quantile value to form the lower bound of the remaining lifetime, and the quantile value is added to form the upper bound of the remaining lifetime. The particle risk point estimate and the size risk point estimate are subtracted from the corresponding quantile value to form the lower bound of particle risk and the lower bound of size risk respectively, and the corresponding quantile value is added to form the upper bound of particle risk and the upper bound of size risk respectively. The target coverage, the three types of quantile values, the unit of measurement and the sample batch identifier are recorded. When the witness document is not invalid within the observation window, the end point of the observation window is used as the right censoring time and is only used for the upper inconsistency score. When the witness document's invalidation time is unavailable, the inconsistency score of the most recent valid batch of the same formula is called. When the coverage verification fails, the interval calibration is stopped and a high-resolution retest decision is generated. A replacement decision is generated when the lower bound of the remaining lifetime is less than or equal to the replacement threshold, and a high-resolution retest decision is generated when the difference between the upper and lower bounds is greater than the retest threshold.

7. The method for predicting the lifetime of a semiconductor quartz component according to claim 2, characterized in that, Step S1 also includes: The batch residuals are grouped according to the formula identifier, cavity position and quartz component type. The median and interquartile range of the group residuals are used to correct the domain offset of the sensor output. The reciprocal of the interquartile range is normalized to the channel confidence. The channel confidence is multiplied by the quality weight and used for feature calculation in step S2.

8. The method for predicting the lifetime of a semiconductor quartz component according to claim 3, characterized in that, The five types of monotonic sub-states are composed of heterogeneous graph node attributes of components, process formulas and spatial locations. Edges are established between nodes based on batch usage relationship, same formula relationship and adjacent spatial relationship. The state difference between adjacent nodes is used as graph constraint residual. The graph constraint residual and the channel quality weight are jointly input into the degradation state update in step S3.

9. The method for predicting the lifetime of a semiconductor quartz component according to claim 6, characterized in that, The decision results written back by the maintenance system and the subsequent retest results are used as new samples to update the channel drift correction amount on the edge side. The inconsistency scores are summarized in batches and the quantile values ​​are updated in the cloud. The updated correction amount and quantile values ​​are only used for prediction in the next time window after the witness coverage is verified.

10. A semiconductor quartz component lifetime prediction system, used to execute the semiconductor quartz component lifetime prediction method according to any one of claims 1 to 9, characterized in that, include: The multi-source observation acquisition and witness calibration module is used to acquire observations from semiconductor quartz components, witness components, and zero-point plates and output calibration observation sequences with channel quality weights. The feature extraction and monotonic state construction module is used to generate thermal, chemical, ionic, particle, and structural features and five types of monotonic sub-states; The degradation state update module is used to execute multi-state degradation equations subject to monotonicity constraints and output damage state trajectories. The time-series remaining lifetime prediction module is used to output point estimates of the remaining number of cycles or remaining hours and the risk of failure. The Sequence Preservation Interval Calibration and Maintenance Decision Module is used to output the upper and lower limits of remaining lifetime with coverage, particle and size risks, generate replacement or high-resolution retest decisions, and write them back to the maintenance system.

Images