Method for optimizing process conditions for multi-modal electroless plating

Abstract

The application discloses a kind of multi-modal electroless plating process condition optimization methods, the method includes: obtaining semiconductor component topography appearance size and plating specification requirement, topography characterization model is established;Establish metal electroless plating solution component, electroless plating device and process condition database, cluster data model is established;Two models are as the different branch of neural network and are fused, and multi-modal neural network model is constructed;The process condition predicted by model is input into electroless plating device and is implemented plating, and experimental data is obtained;Experimental data and predicted data are compared and discriminated;According to the discrimination result, the model is corrected, and the prediction result approaches experimental data.The method of the application realizes the collaborative optimization of multiple performance indicators of plating layer and the continuous accumulation of process knowledge, significantly improves the precision and efficiency of process condition optimization, and provides an intelligent solution for the transformation from experience-driven to data-driven for double-sided electroless plating of semiconductor components.

Description

Technical Field

[0001] This invention belongs to the field of chemical plating technology for semiconductor devices, specifically relating to a method for optimizing multimodal chemical plating process conditions. Background Technology

[0002] Chemical plating is a surface treatment technology that deposits a metal coating on a substrate surface through a self-catalytic redox reaction. This technology enables nanoscale metallization on the surface of semiconductor devices and is a key foundational process in the production of high-end electronic components such as semiconductor wafer manufacturing, 3D integrated packaging, microelectromechanical systems (MEMS), lead frames, printed circuit boards, and connectors. It plays an irreplaceable role in realizing electronic logic interconnects and the fabrication of micro- and nano-structures.

[0003] As semiconductor integration technology evolves towards higher density and higher performance, the trend towards three-dimensional and complex chip and packaging structures is becoming increasingly significant. Especially for semiconductor devices with double-sided circuitry or three-dimensional structures, the control of double-sided chemical plating processes is far more challenging than that of single-sided plating. The uniformity, density, adhesion of the plating thickness, and the morphological adaptability of different regions of the device are all affected by the coupling of multiple process parameters, including plating solution composition, temperature, pH value, stirring intensity, plating time, pump flow rate, and workpiece oscillation frequency.

[0004] Currently, the determination of chemical plating process conditions in the industry typically relies on the accumulated experience of technical personnel or on trial and error using traditional methods such as single-factor and orthogonal experiments. These methods have the following significant shortcomings: (1) Low optimization efficiency and high cost: It mainly relies on a large number of trial and error experiments to find a better process window. Not only is the R&D cycle long, but it also consumes a lot of manpower, materials and time costs, making it difficult to adapt to the current rapidly iterating semiconductor manufacturing needs.

[0005] (2) Insufficient parameter correlation analysis: Traditional methods are difficult to establish a quantitative mapping relationship between the complex morphology of semiconductor devices (such as deep holes, concave and convex structures, different aspect ratios) and multi-dimensional process parameters (such as solution flow field, temperature field, mass transfer rate), which leads to the optimized process conditions often not being the global optimal solution.

[0006] (3) Low data utilization: The massive amount of "workpiece characteristics-process parameters-performance results" data generated during the experiment could not be effectively accumulated, managed and reused, and could not form a process knowledge base that can be self-iterated and continuously optimized.

[0007] (4) Single judgment criteria: The evaluation of coating quality is often limited to thickness and appearance, lacking comprehensive consideration and joint optimization of multi-dimensional performance such as coating adhesion, corrosion resistance, wear resistance and electrical properties (such as impedance stability).

[0008] Therefore, how to overcome the shortcomings of existing technologies that rely on single experience, have low trial-and-error efficiency, and are difficult to achieve high-performance plating of complex workpieces, and how to establish a method for determining chemical plating process conditions that can integrate multi-source information, have self-learning capabilities, and can achieve rapid and accurate optimization has become an urgent technical problem to be solved in this field. Summary of the Invention

[0009] The technical problem to be solved by this invention is that existing chemical plating process optimization methods are difficult to establish the correlation between the morphology of complex semiconductor devices and multi-dimensional process parameters in an efficient and accurate manner, and to achieve synergistic optimization of multiple performance indicators of the plating layer.

[0010] To address the aforementioned problems, this invention provides a method for optimizing multimodal chemical plating process conditions, applied to the double-sided chemical plating process of semiconductor devices. By comparing experimental data from a nano-manufacturing process using a double-sided chemical plating apparatus with chemical plating process conditions predicted by a multimodal neural network model, the computational software in the neural network model is corrected to make the predicted results closer to the experimental data, thus achieving a rapid method for optimizing the double-sided chemical plating process conditions of semiconductor devices. The method includes the following steps: Obtain the morphology, appearance, dimensions, and plating specifications of the semiconductor device, and construct a morphology characterization model of the semiconductor device; Establish a database of metal electroless plating solution composition, an electroless plating device database, and an electroless plating process condition database, and build a clustering data model of electroless plating process conditions based on the databases; The morphology characterization model and the clustering data model are fused as different branches of a neural network to construct a multimodal neural network model for predicting chemical plating process conditions. The steps of constructing the multimodal neural network model include: calculating the electrochemical relationship between the standard electrode potential of the target metal and the plating time, introducing correction parameters to compensate for the influence of dynamic process factors on the plating results during the chemical plating process, and establishing an initial mapping relationship between the chemical plating process conditions and the theoretical time required to reach the target metal film thickness, as the basis for the multimodal neural network model to make predictions. From the numerical range of the chemical plating process conditions predicted by the multimodal neural network model, at least one low-end value and one high-end value are selected and input into the chemical plating device to perform plating on the semiconductor device, and the corresponding first experimental data and second experimental data are obtained. The first and second experimental data are compared and judged with the prediction data of the multimodal neural network model, respectively. Based on the discrimination results, the correction parameters are adjusted, and the adjusted model is used for the next prediction of process conditions to achieve iterative optimization of chemical plating process conditions.

[0011] Optionally, the morphological dimensions include length, width, thickness, and three-dimensional shape; the plating specifications include the type of metal, the thickness of the plating metal film, and the plating area.

[0012] Optionally, the chemical plating process conditions include oscillation frequency, plating time, solution temperature, and pump flow rate.

[0013] Optionally, the comparison and discrimination step further includes: pre-setting a comprehensive judgment standard, which includes six judgment items: coating adhesion judgment, morphology and appearance judgment, gas corrosion rate judgment, resistance and impedance judgment before and after gas corrosion test, resistance and impedance judgment of insertion and extraction test, and quality difference rate judgment before and after friction and wear test; only when the first experimental data and the second experimental data are evaluated according to the comprehensive judgment standard, and the judgment results of the six judgment items are all excellent, will the corresponding chemical plating process conditions be established as the preferred process conditions.

[0014] Optionally, it also includes: storing the original data of the chemical plating process conditions corresponding to the comprehensive judgment result of "excellent" and its related morphology characterization model and clustering data model into an effective database for iterative updates of the model; storing the original data of the chemical plating process conditions corresponding to the comprehensive judgment result of "unqualified" and its related data into a failure database to enrich the analogy data source.

[0015] Optionally, the metal electroless plating solution is a tungsten-nickel alloy solution, the components of which include: 280 mmol / L sodium tungstate dihydrate, 13.2 mmol / L triammonium citrate, 14.8 mmol / L triethylenetetraminehexaacetic acid, 120 mmol / L nickel aminosulfonate hydrate, 184 mmol / L sodium hypophosphite, 0.31 mmol / L polyethylene glycol octadecyl and 0.49 mmol / L 1,2,4-triazole-3-carboxamide.

[0016] Optionally, the neural network is a recurrent neural network.

[0017] Optionally, the swing frequency in the chemical plating process is 21 to 39 times per minute, the plating time is 22.6 to 24.5 minutes, the solution temperature is 77°C to 89°C, and the pump flow rate is 270 to 390 liters per minute.

[0018] Compared with existing technologies, the method for optimizing multimodal chemical plating process conditions provided by this invention has the following beneficial effects: First, this invention introduces an electrochemical principle based on the relationship between standard electrode potential and plating time, and sets correction parameters to compensate for the influence of dynamic process factors, thus constructing an accurate initial mapping foundation for a multimodal neural network model. This technique combines electrochemical mechanisms with data-driven models, solving the problem of existing technologies' difficulty in efficiently and accurately establishing quantitative correlations between complex semiconductor device morphologies and multidimensional process parameters, laying a solid algorithmic foundation for the efficient and accurate prediction of subsequent process conditions.

[0019] Second, this invention constructs a multimodal neural network model by fusing morphological characterization models and clustering data models as different branches of a neural network. This model can simultaneously process multi-source heterogeneous data such as the three-dimensional morphology, size specifications, plating solution composition, device parameters, and process conditions of semiconductor components, achieving deep coupling and quantitative mapping between workpiece features and multi-dimensional process parameters, thereby significantly improving the accuracy and efficiency of process condition prediction.

[0020] Third, this invention verifies the coating process by selecting low-end and high-end values ​​from the model's predicted numerical range and comparing the obtained experimental data with the predicted data. This "high-low end verification" strategy can quickly and effectively test the accuracy of the model's predictions and intuitively reveal the deviation between the predicted and actual values ​​through comparison results, providing crucial and quantitative evidence for subsequent model correction.

[0021] Fourth, this invention establishes a comprehensive evaluation standard encompassing six aspects: coating adhesion, morphology, gas corrosion rate, electrical resistance, and friction and wear. It stipulates that only when all evaluation results are excellent can the preferred process conditions be established. This technique directly embeds the synergistic optimization objective of multiple performance indicators into the optimization process, overcoming the limitations of traditional methods that focus only on a single or a few performance indicators. This ensures that the final process conditions fully meet the comprehensive requirements of semiconductor devices for the mechanical, corrosion-resistant, electrical, and wear-resistant properties of the coating.

[0022] Fifth, this invention establishes a closed-loop iterative optimization mechanism of "prediction-verification-discrimination-correction" by adjusting the correction parameters in the model based on the discrimination results and using the corrected model for the next process prediction. Simultaneously, excellent and unqualified experimental data, along with their associated raw data, are stored in valid and failed databases for categorized management. This mechanism enables the continuous accumulation of process knowledge and the self-evolution of the model, allowing prediction results to infinitely approach experimental data. It completes the intelligent transformation from relying on experience-based trial and error to data-driven precise decision-making, effectively improving the efficiency of process optimization and the yield of coated products.

[0023] Sixth, through verification by embodiments, this invention achieves excellent overall performance of the coating within the preferred process parameter range: oscillation frequency of 21 to 39 times / min, plating time of 22.6 to 24.5 minutes, solution temperature of 77°C to 89°C, and pump flow rate of 270 to 390 liters / min. This provides a reliable, precise, and reproducible technical solution for the industrial application of double-sided chemical plating technology for semiconductor devices.

[0024] In summary, this invention effectively solves the technical problems of existing technologies that are difficult to optimize chemical plating process conditions efficiently and accurately and cannot achieve synergistic optimization of multiple performance indicators by constructing a multimodal neural network based on electrochemical principles, establishing stringent comprehensive judgment criteria, and forming a closed-loop data iterative optimization system. This significantly improves the efficiency, accuracy, and overall quality of the plating layer. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0026] Figure 1 This is a schematic diagram of the chemical plating apparatus used in the method for optimizing the multimodal chemical plating process conditions provided by the present invention.

[0027] Figure 2 This is a schematic diagram of the semiconductor device morphology structure involved in the optimization method of multimodal chemical plating process conditions provided by the present invention.

[0028] Figure 3 A flowchart illustrating the optimization method for multimodal chemical plating process conditions provided by the present invention.

[0029] Figure 4 This is a schematic diagram of the control module of the chemical plating apparatus involved in the optimization method for multimodal chemical plating process conditions provided by the present invention.

[0030] Figure 5 This is a functional schematic diagram of the method for optimizing multimodal chemical plating process conditions provided by the present invention.

[0031] The labels in the attached diagram are explained as follows: 100-Semiconductor element; 200-Plastering apparatus solution tank; 300-Heater; 400-Semiconductor element fixing and swaying mechanism; 500-Control module of the chemical plating apparatus; 510-Control display; 520-Memory; 530-Morphological characterization and plating process condition variables; 531-Semiconductor element morphology; 532-Clustering parameters of chemical plating process conditions; 540-Morphological characterization model; 550-Clustering data model; 560-Neural network model; 561-Judgment and evaluation criteria; 600-Chemical plating apparatus. Plating apparatus; 700-Optimization process flow for multimodal chemical plating; 800-Process conditions for multimodal chemical plating; 810-Morphological characterization and prediction of plating process conditions; 811-Morphological prediction parameters for semiconductor devices; 812-Clustering prediction parameters for plating process conditions; 820-Multimodal construction; 830-Multimodal evaluation computing system; 840-Multimodal evaluation training system; 850-Morphological characterization construction; 860-Clustering data construction; 870-Neural network construction; 871-Multimodal judgment criteria; 880-Variable parameters of chemical plating equipment. Detailed Implementation

[0032] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] Example 1 This embodiment provides an optimization method for multimodal chemical plating process conditions, which is applied to the double-sided chemical plating process of semiconductor devices.

[0034] like Figure 1 As shown, the chemical plating apparatus 600 used in this invention includes a plating apparatus solution tank 200, a heater 300, a semiconductor component fixing and swaying mechanism 400, and a control module 500 for the chemical plating apparatus. This apparatus is used to conduct subsequent plating experiments and obtain experimental data.

[0035] The semiconductor device 100 to be processed has a specific three-dimensional morphological structure, such as Figure 2 As shown. In this embodiment, the semiconductor element 100 is made of copper alloy, with dimensions of 60 mm × 10³ mm and a thickness of 0.127 mm. The element has 10 basic units, each with a plating area of ​​16 mm × 23 mm and a single-sided area of ​​368 mm². The total double-sided area of ​​one semiconductor electronic element is 10 mm × 368 mm × 2 = 7360 mm². The plating specifications are: a tungsten-nickel alloy is plated on the element surface, with a film thickness ≥ 2.5 micrometers (μm).

[0036] The optimized process flow of the multimodal chemical plating process of this invention is as follows: Figure 3 As shown, the complete multimodal chemical plating process optimization flow 700, from data modeling to model iteration optimization, specifically includes the following steps: First, the morphological dimensions and plating specifications of the semiconductor device are obtained. The morphological dimensions include length, width, thickness, and three-dimensional shape; the plating specifications include the metal type (tungsten-nickel alloy in this embodiment), the thickness of the plating metal film (≥2.5 μm), and the plating area (10 units on both sides). A basic database for the morphological characterization of the semiconductor device before chemical plating is established to facilitate the subsequent construction of a morphological characterization model. This step solves the problem of workpiece morphological data acquisition and standardization, providing an accurate input basis for subsequent modeling.

[0037] Secondly, a database of required metal electroless plating solution components, an electroless plating equipment database, and an electroless plating process condition database are established, and a clustering data model of the electroless plating process conditions is constructed. The metal electroless plating solution component database can contain formulation data for various solutions; the electroless plating equipment database records the structural parameters and operating characteristics of different equipment; and the process condition database collects various combinations of process parameters and their corresponding plating results. In this embodiment, the specific process of constructing this clustering data model includes: first, data cleaning and standardization preprocessing of the raw data in the process condition database to eliminate the influence of different dimensions; then, using the K-Means clustering algorithm to perform cluster analysis on the preprocessed data. The advantage of the K-Means algorithm lies in its high computational efficiency and easy interpretation of results when processing datasets with known or predictable cluster numbers. Through this cluster analysis, the potential correlation patterns between different combinations of process conditions and plating effects can be effectively mined, thereby forming a clustering data model. This step systematizes the scattered process data, providing data support for multimodal fusion.

[0038] Cluster analysis can uncover potential correlations between process conditions and coating effects, forming a clustered data model. This step systematizes scattered process data, providing data support for multimodal fusion.

[0039] This embodiment preferably uses a self-developed tungsten-nickel alloy solution with the following composition to establish the required metal electroless plating solution database: 280 mmol / L sodium tungstate dihydrate, 13.2 mmol / L triammonium citrate, 14.8 mmol / L triethylenetetraminehexaacetic acid, 120 mmol / L nickel aminosulfonate hydrate, 184 mmol / L sodium hypophosphite, 0.31 mmol / L polyethylene glycol octadecyl, and 0.49 mmol / L 1,2,4-triazole-3-carboxamide. Specific solution preparation methods can be found in Example 13 of CN117779011B. After starting the plating tank, the electroless plating process is carried out according to the solution, with 800 L of tungsten-nickel electroless plating solution in the plating tank. The management standard of the metal electroless plating solution is maintained and controlled in real time by a clustering data model. The plating device is equipped with an automatic solution component analysis and automatic replenishment system to ensure that the actual operation of the electroless plating device is carried out under the stated stable conditions. This automatic control system can control the fluctuation of solution concentration within a small range, thereby ensuring the stability and consistency of the plating quality.

[0040] Next, a morphology characterization database for semiconductor components is established to construct morphology characterization models. This database stores data on semiconductor components with different morphological features. Through feature extraction and modeling, a digital model capable of characterizing the workpiece morphology is formed. For example... Figure 4 In the control module 500 of the chemical plating apparatus shown, the memory 520 stores the aforementioned database, the control display 510 is used to display relevant parameters, and the morphology characterization and plating process condition variables 530 include semiconductor element morphology 531 and chemical plating process condition clustering parameters 532. These variables serve as inputs to the morphology characterization model 540 and the clustering data model 550.

[0041] A clustering data model based on a metal electroless plating solution composition database and an electroless plating process condition database, along with a semiconductor device morphology characterization model, is used as different branches of a neural network to form a comprehensive process condition database for semiconductor device electroless plating, which is then used to construct the neural network model. For example... Figure 4 As shown, the neural network model 560 is constructed based on the morphological representation model 540 and the clustering data model 550, and outputs the judgment evaluation criteria 561.

[0042] In this embodiment, the neural network model 560 employs a recurrent neural network. The construction and training of this multimodal neural network model includes the following steps: (1) Network structure parameter optimization: Randomly generate a set of variables ( ) values, where n1 and n2 represent the number of neurons in the first and second layers of the network, respectively; by generating multiple sets ( The values ​​of n1 and n2 are compared to determine the values ​​of n1 and n2, thereby optimizing the network structure.

[0043] (2) Preliminary training: Using n1 and n2 obtained in step (1), construct a neural network and train the neural network using training samples to obtain a preliminarily trained neural network.

[0044] (3) Network performance test: Input the test set data into the neural network trained in step (2), compare the difference between the output value and the true value, calculate the average percentage error value, and evaluate the prediction accuracy of the network.

[0045] (4) Optimal parameter selection and update: When the average percentage error value obtained in step (3) is less than the best average percentage error value in the historical record, then update ( The value of ) is taken and used as the new historical best average percentage error value.

[0046] (5) Iterative loop: Set the loop condition, repeat steps (1) to (4) until the termination condition is met, and output the final determined optimal ( )value.

[0047] (6) Train the optimal network: using the optimal network obtained in step (5) The optimal neural network is constructed by taking the values ​​of the complete training set samples and then training the optimal neural network to obtain the trained optimal neural network model.

[0048] (7) Application and verification: Input the morphology data of the semiconductor device to be predicted and the relevant data of the process conditions into the trained optimal neural network to obtain the predicted chemical plating process conditions (such as plating time). The accuracy of the predicted value can be further verified through actual plating experiments, and the percentage error value can be calculated.

[0049] Through the above steps, deep fusion of multi-source information and accurate model construction were achieved, enabling the neural network to simultaneously learn the complex mapping relationship between workpiece morphology and process conditions, laying the foundation for accurate prediction of process conditions. By introducing multimodal fusion technology, the shortcomings of traditional single models in handling multidimensional heterogeneous data were overcome.

[0050] Specifically, in constructing a clustering data model, simulation software relating the standard electrode potential of a chemical redox reaction to the mass of the deposited metal film can be used to predict the thickness of the film that can be deposited within a given chemical plating time. The smaller the difference between the predicted data and the experimental data from the chemical plating apparatus—that is, the closer the predicted data is to the measured data—the higher the probability that it can be used to construct a clustering data model. The simulation software is based on the standard electrode potential of metal M. The reaction equation can be expressed as follows:

[0051] in, Represents metal atoms; Represents metal ions, that is, the positively charged form of metal dissolved in solution; Representing two electrons, it is a microscopic particle that participates in charge transfer during a reaction; "Indicates the direction of the reaction, that is, the metal ion gains electrons and is reduced to a metal atom; The standard electrode potential is a physical constant that measures the tendency or ease with which a metal ion gains electrons and is reduced. represent The numerical unit is millivolt.

[0052] In this embodiment of the invention, the standard electrode potentials of the metals involved are common knowledge data, which can be obtained by consulting standard electrochemical data sheets. Specifically, for the tungsten-nickel alloy plating solution used in this embodiment, the standard electrode potentials of its main metal ions are as follows: For tungsten (W): Its standard electrode potential -0.2 V (relative to the standard hydrogen electrode).

[0053] For nickel (Ni): Its standard electrode potential It is -0.23 V (relative to the standard hydrogen electrode).

[0054] By substituting the specific electrode potential values ​​mentioned above into the simulation software for calculation, and combining this with subsequent experimental verification, the relationship between coating time and film thickness can be determined more accurately, thereby improving the model's prediction accuracy.

[0055] like Figure 5 The functional diagram of the multimodal chemical plating process conditions 800 shown illustrates the functional module composition of the entire optimization method.

[0056] Among them, morphology characterization and plating process condition prediction 810 includes two types of core prediction parameters: The first category is semiconductor device morphology prediction parameters 811, which serve as input features for the model and specifically include the device's geometric dimensions (such as length, width, and thickness) and three-dimensional shape features.

[0057] The second category is the clustering prediction parameter 812 for plating process conditions, which is the main output of the model. It includes not only the specific process conditions predicted by the multimodal neural network model (such as swing frequency, plating time, solution temperature, pump flow rate, etc.) and their corresponding prediction confidence intervals; more importantly, this module also includes predictions of the final performance indicators of the plating layer, such as gas corrosion area, gas corrosion rate, resistance changes before and after insertion / removal tests, and mass differences before and after friction and wear tests. These performance prediction data will serve as a benchmark for subsequent comparison and judgment with measured experimental data.

[0058] These prediction parameters are input into the multimodal evaluation computation system 830 and the multimodal evaluation training system 840 through multimodal construction 820 (which integrates morphological representation construction 850, clustering data construction 860 and neural network construction 870).

[0059] The multimodal evaluation computing system 830 is the core computational engine of this optimization method. It contains a pre-built computational model based on electrochemical principles, which includes a formula relating the standard electrode potential of the plated metal to the plating time, used to initially calculate the theoretical time required to reach the target plated metal film thickness. Simultaneously, since dynamic process factors such as pump flow rate and oscillation frequency can affect the plating results, the algorithm formula of this computing system also incorporates correction parameters for these factors. It should be noted that this correction parameter is a key coefficient in this model, comprehensively reflecting the correction of the metal deposition rate by reaction kinetic factors (such as solution temperature, plating solution formulation, pump flow rate, oscillation method, plating tank size, solution volume, etc.) and the mixed potential effect (i.e., the combined effect of the potential fusion of all components in the plating solution formulation) during the actual chemical plating process. For a specific metal type, this coefficient has a theoretical value at a reference temperature (e.g., 25°C). However, due to the complexity of the actual chemical plating process, there is a deviation between the theoretical value and the actual optimal value. Therefore, the core optimization logic of this invention is to obtain measured data through subsequent actual plating verification, and to continuously correct the calibration parameter (i.e., the key coefficient) using this measured data. The more experimental data accumulated, the more accurate the corrected key coefficient becomes, and the more closely the model's prediction results can approach the true value, thereby achieving continuous iterative optimization of the model and refined evaluation and correction of process conditions.

[0060] The multimodal evaluation and training system 840 is responsible for continuously iteratively training and optimizing the aforementioned computational model and correction parameters using feedback data obtained from actual plating experiments. In a preferred embodiment of the present invention, the multimodal evaluation and training system 840 employs a Bayesian optimization algorithm to iteratively optimize the correction parameters. Bayesian optimization is a highly efficient global optimization algorithm based on a surrogate model, particularly suitable for optimization problems with unknown objective functions and high evaluation costs, such as the optimization of chemical plating process conditions in this invention. The initial values ​​of the correction parameters can be initially calibrated by screening historical experimental data (including successful and failed cases), such as selecting a set of relatively reasonable experimental data obtained from previous attempts as the starting point for Bayesian optimization.

[0061] The core process of Bayesian optimization includes: Step A: Construct a probabilistic surrogate model based on existing experimental data (including historical data from initial calibration and newly acquired measured data each time), typically a Gaussian process regression model.

[0062] Step B: Select the next most promising evaluation point, i.e., a set of candidate correction parameter values, by maximizing the acquisition function (e.g., the desired improvement in EI); this process can automatically balance the "utilization" of known superior areas and the "exploration" of unexplored areas with high uncertainty.

[0063] Step C: Substitute the selected correction parameters into the multimodal neural network model to predict process conditions and perform actual plating verification to obtain real coating performance data.

[0064] Step D: Update the proxy model using the newly acquired real data to make it more accurately reflect the real objective function.

[0065] Step E: Repeat steps B to D until the correction parameters converge to the optimal value or the preset number of iterations is reached.

[0066] Bayesian optimization can efficiently approximate the optimal calibration parameters with fewer experiments, thereby achieving precise optimization of complex process conditions. This process feeds measured data back into the model, replacing theoretical values ​​with key coefficients that more closely resemble reality, allowing the model's prediction accuracy to continuously improve through successive "prediction-verification-correction" loops.

[0067] The system comprehensively evaluates the model's prediction and calculation results using the multimodal decision criterion 871, and finally outputs optimized process conditions.

[0068] The variable parameters 880 of the chemical plating equipment identify the set of key operating parameters of the chemical plating apparatus. The equipment variables included in this module (such as oscillation frequency, pump flow rate, solution temperature, etc.) are the basic data source for constructing the clustering data model 860, and also the dynamic process factors that the multimodal evaluation calculation system 830 needs to consider when processing predictions. By inputting the equipment variable parameters and the semiconductor component morphology prediction parameters 811 into the system together, deep coupling between workpiece characteristics and equipment operating conditions is achieved, providing a complete input feature space for the multimodal neural network model.

[0069] The oscillation frequency of the chemical plating apparatus has a unique correlation with the performance test results of the plated semiconductor components. Optimizing the oscillation frequency can improve the flow and mass transfer of the plating solution, thereby enhancing the uniformity of the plating layer. Controlling the pump flow rate of the chemical plating apparatus, through the control of the chemical plating solution flow rate, can regulate the plating solution renewal rate and stirring intensity, thus obtaining the optimal performance test results for the plated products. Optimal control of the solution temperature in the chemical plating apparatus can regulate the reaction rate and plating structure, achieving the best performance test results for the plated products. The synergistic optimization of these process parameters is the core of this invention.

[0070] Next, based on the inspection results of the semiconductor morphology appearance in the morphology characterization model and the performance test data of the semiconductor component plating products in the constructed neural network model, a judgment standard for the neural network model is established. For example... Figure 4 As shown, the evaluation criteria 561 output by the neural network model 560 are used to evaluate the coating quality obtained under the process conditions predicted by the model, providing a basis for subsequent comparison and judgment. The morphological appearance inspection of the coated product includes the size of the coated area, its morphological appearance, the thickness of the coated film, and the adhesion between the coated film and the substrate. The performance test data of the coated product includes the circuit resistance and impedance before and after the sulfur dioxide gas corrosion test, as well as the sulfur dioxide gas corrosion rate, wear and tear test, and the circuit resistance and impedance before and after the insertion and removal test.

[0071] A comprehensive evaluation criterion for the neural network model is constructed, as shown in Table 1. This comprehensive criterion includes multiple evaluation criteria, and the overall evaluation result is considered excellent only when all evaluation criteria are rated as excellent. This rigorous multi-index evaluation system ensures that the optimized process conditions can simultaneously meet the comprehensive requirements of the coating's mechanical properties, corrosion resistance, electrical properties, and wear resistance, avoiding the one-sidedness of traditional methods that only focus on a single performance index.

[0072] Table 1 Comprehensive Judgment Criteria

[0073] The specific criteria for judgment in Table 1 include: (1) The adhesion of the metal coating and the appearance of the coating were tested according to national standards. Test method: The surface of the tungsten-nickel alloy coating was bent 180° and the condition around the bend was observed under magnification at 150x. The coating adhesion judgment standard: The coating adhesion was judged according to the ratio of the area of ​​peeling and detachment (coating adhesion judgment): =0 indicates excellent; 0 < ≤1.0% is considered good; <1.0% is considered good. The following are the criteria for judging the appearance of the coating: a smooth and flat surface is excellent; a surface with pits is unacceptable; a surface with pinholes is unacceptable.

[0074] (2) Conduct a sulfur dioxide gas corrosion test according to national standards, and test the resistance of the semiconductor electronic circuit board before and after the test. Corrosion rate judgment standard: The corrosion rate is judged based on the ratio of the corroded area to the total area of ​​the coating (gas corrosion rate). =0 indicates excellent; 0 < ≤0.5% is considered good; <0.6% is considered good. This is considered unqualified. Resistance / Impedance Judgment Criteria (Resistance / Impedance Judgment Before and After Gas Corrosion Test): Based on the difference in resistance / impedance before and after the mixed gas corrosion test. , compared with the resistance and impedance before the experiment Determine based on proportion: Excellent; Good; It is unqualified.

[0075] (3) Before and after the insertion and removal test of the semiconductor electronic circuit board according to national standards, test the resistance of the semiconductor electronic circuit board. The judgment criteria for the insertion and removal test of the semiconductor electronic circuit board (insertion and removal test / resistance judgment): After 1000 insertion and removal tests, the resistance difference before and after the test is used to determine the resistance. , compared with the resistance and impedance before the experiment Determine based on proportion: Excellent; Good; It is unqualified.

[0076] (4) The CSM ball friction and wear testing equipment was used to conduct sliding friction and wear tests on the test samples according to national standards. The sliding friction and wear rate determination standard (mass difference rate before and after the friction and wear test): based on the mass difference of the test piece before and after the sliding friction and wear test. The quality of the experimental piece before the experiment The ratio is used to determine the difference: Excellent; Good; It is unqualified.

[0077] As shown in Table 1, only when all evaluation criteria are rated as excellent can the overall evaluation result be considered excellent. That is, only the chemical plating process conditions predicted by the multimodal neural network model that meet the criteria for an excellent overall evaluation result can be considered excellent. The chemical plating process conditions corresponding to the aforementioned embodiments can be established as their preferred range. Any other permutation or combination of the evaluation criteria will result in an unsatisfactory overall evaluation.

[0078] Then, as Figure 3 The process shown involves inputting the electroless plating process conditions predicted by the multimodal neural network model into the electroless plating apparatus 600 (see...). Figure 1 The semiconductor element 100 is subjected to actual plating to obtain experimental data, including plating thickness, morphology, and performance test results. The chemical plating process conditions include oscillation frequency, plating time, solution temperature, and pump flow rate. During the plating process, the semiconductor element fixing and oscillation mechanism 400 drives the element to move in the plating apparatus solution tank 200 according to the set oscillation frequency, the heater 300 maintains the set solution temperature, and the pump system circulates the plating solution at the set flow rate.

[0079] The acquired experimental data is compared and judged with the predicted data of the multimodal neural network model. Specifically, based on the comprehensive judgment criteria established in the previous step, the various performance characteristics of the experimental samples are evaluated to obtain a comprehensive judgment result (excellent or unqualified). This judgment result reflects the performance of the process conditions predicted by the model in actual plating.

[0080] The evaluation criteria for neural network models determine the quality of chemical plating process conditions based on a unique correspondence between the test content included in the criteria and the chemical plating process conditions in the clustering data model. A superior evaluation results in the corresponding chemical plating process conditions, along with all original data from the related morphological characterization model and clustering data model, stored in their respective valid databases for iterative updates and continuous improvement of the model. For example... Figure 4 As shown, this data is stored in memory 520 and monitored and managed via control display 510. By feeding excellent data back to the model, knowledge accumulation and continuous model evolution are achieved. For results deemed substandard or unqualified, the corresponding chemical plating process conditions, along with all original data from the related morphology characterization model and clustering data model, are stored in their respective failure databases to enrich the analogy data source. Failure data is also valuable, serving as negative examples to help the model avoid repeating errors. The model retrains or adjusts parameters using new data from the effective database, gradually bringing the prediction results closer to the experimental data, thereby optimizing the process conditions.

[0081] like Figure 5As shown, the multimodal evaluation computing system 830 and the multimodal evaluation training system 840 are responsible for the iterative optimization of the model. The morphological representation construction 850, the clustering data construction 860 and the neural network construction 870 are continuously updated through the multimodal construction 820. The multimodal judgment criterion 871 is used to comprehensively evaluate the optimization effect.

[0082] The core principle of this invention for optimizing chemical plating process conditions is as follows: A multimodal neural network integrating a semiconductor device morphology characterization model and a chemical plating process condition clustering data model is constructed. This method... Figure 3 The closed-loop process shown achieves adaptive optimization: (1) Data modeling: Establish workpiece morphology database and process condition database respectively to form independent characterization model and clustering model. (2) Multimodal fusion: Use the above two models as inputs to different branches of the neural network to construct a fusion model that can integrate workpiece characteristics, plating device and process formula. (3) Iterative optimization and knowledge accumulation: Compare the model prediction results with the experimental data of the actual double-sided chemical plating device, and store the data into the effective database and the failure database respectively according to the judgment result of superior or inferior. The effective database is used for iterative updates of the model, and the failure database is used to enrich the analog data source, so as to realize the continuous accumulation of process knowledge and the self-correction of optimization direction. (4) The process condition framework of multimodal chemical plating is a comprehensive system. The optimization and adjustment of all aspects involved are for obtaining high-quality and high-performance chemical plating. (5) This method combines multimodal fusion computational simulation with experimental science, providing an advanced path for the transformation of chemical plating process from experience-driven to model-driven.

[0083] Meanwhile, the optimization method for multimodal chemical plating process conditions provided by this invention can be applied to the chemical plating manufacturing of other electronic components such as semiconductor wafers, printed circuit boards, lead frames, microwave devices, and connectors.

[0084] Through multiple iterations, the model's predictions gradually approached the experimental data, ultimately optimizing the process conditions. In this embodiment, the preferred process conditions ranged from 21 to 39 swings / min, plating time from 22.6 to 24.5 minutes, solution temperature from 77°C to 89°C, and pump flow rate from 270 to 390 liters / min. The optimal process conditions were: 30 swings / min, plating time from 23.5 minutes, solution temperature from 83°C, and pump flow rate from 330 liters / min.

[0085] Example 2 To verify the effectiveness of the method of the present invention, a series of verification experiments were conducted using the method and apparatus described in Example 1. All experiments used the semiconductor element and tungsten-nickel alloy solution described in Example 1, with a target film thickness ≥2.5 μm. During the experiments, optimization and verification were performed on three key process parameters: swing frequency, solution temperature, and pump flow rate. Control experiments were also set up to highlight the effectiveness of the present invention.

[0086] 1. Optimization and verification of swing frequency Using the multimodal chemical plating process condition prediction method of this invention, the predicted swing frequency ranges from 15 times / min to 45 times / min, with the corresponding plating time predicted at intervals of 3 swings. Specifically, as shown... Figure 3 The multimodal electroless plating process optimization flow 700 shown inputs the morphology characterization model and clustering data model into the neural network model, and outputs the predicted process conditions. This verification experiment selected swing frequencies of 15, 18, 39, and 42 times / min as verification points, using an electroless plating apparatus 600 (see...). Figure 1 The actual operation was conducted with the solution temperature fixed at 82℃ and the pump flow rate fixed at 280 L / min. Detailed results are shown in Tables 2 to 7.

[0087] Table 2. Plating process conditions for Experiments 1 to 7

[0088] As shown in Table 2, in Experiment 1, the actual plating test was conducted using a swing frequency of 21 times / min and a predicted plating time of 24.1 minutes. The plating thickness of the tungsten-nickel alloy layer did not reach the expected 2.5 μm. When the measured plating time was adjusted to 24.5 minutes, the plating thickness of the tungsten-nickel alloy layer reached the expected specification requirement of ≥2.5 μm. Similarly, in Experiment 7, the same method was used, and the predicted plating time of 22.9 minutes was adjusted to 23.1 minutes through actual measurement, resulting in a plating thickness that met the specification requirement of ≥2.5 μm.

[0089] The measured plating time data of Experiments 1 and 7 are used to replace the predicted data in the control module 500 of the chemical plating device for upgrading. Furthermore, based on the upgraded plating time data of Experiments 1 and 7, the plating time range of Experiments 2 to 6 is optimized and upgraded to 23.3 minutes to 24.2 minutes through the multimodal chemical plating process conditions 800 system.

[0090] Furthermore, by optimizing and upgrading the plating time, Experiment 4 was actually operated using a chemical plating apparatus 600 to obtain semiconductor element samples plated with tungsten-nickel alloy. The various test data and the test results of Experiment 1 and Experiment 7 are shown in Table 3.

[0091] Table 3. Swing frequency, plating time, and sample test results for Tests 1, 4, and 7.

[0092] In Table 3, the gas corrosion area ratio refers to the corrosion area ratio of the mixed gas, and the calculation formula is tungsten-nickel alloy. The ratio of the area of ​​the cracked coating to the total area of ​​the coating is calculated using the following formula: .

[0093] As shown in Table 3, under the conditions of a swing frequency of 21 to 39 times per minute and an optimized plating time of 23.1 to 24.5 minutes, the overall evaluation result of all samples was excellent. Table 4 summarizes the detailed results of each evaluation item, indicating that this parameter range can meet the requirements of synergistic optimization of multiple performance indicators.

[0094] Table 4. Results of coating evaluation for samples from Tests 1, 4, and 7.

[0095] Control experiments 1 to 4 were set up, with swing frequencies of 15, 18, 42 and 45 times per minute, respectively, and other conditions were the same. The results are shown in Tables 5 to 7.

[0096] Table 5. Plating process conditions for Control Experiments 1 to 4

[0097] Although control experiments 2 and 3 could meet the film thickness requirements after actual measurement and adjustment, the sample performance test results are shown in Table 6, with many indicators failing to meet the requirements.

[0098] Table 6. Test results of samples from control experiment 2 and control experiment 3

[0099] Table 7. Results of coating determination for samples in Control Test 2 and Control Test 3

[0100] As can be seen from Table 7, when the swing frequency deviates from 21 to 39 times per minute, the overall judgment result is unqualified. Therefore, the preferred range of swing frequency established by the method of the present invention is 21 to 39 times per minute.

[0101] 2. Optimization and verification of solution temperature The difference between this experiment and Experiment 4 lies in changing the solution temperature. The predicted temperature range was 73℃ to 93℃, and actual measurements were taken at 75℃, 77℃, 89℃, and 91℃. The swing frequency was fixed at 30 times / min, and the pump flow rate was 280 liters / min. The results are shown in Tables 8 to 13.

[0102] Table 8. Plating process conditions for experiments 8 to 14

[0103] As shown in Table 8, Experiment 8 used a plating temperature of 77℃ and a predicted plating time of 24.2 minutes for the actual plating test. The plating thickness of the tungsten-nickel alloy layer did not reach the expected 2.5μm. When the measured plating time was adjusted to 24.4 minutes, the plating thickness of the tungsten-nickel alloy layer reached the expected specification requirement of ≥2.5μm. Similarly, Experiment 14 used the same method, adjusting the predicted plating time from 22.8 minutes to 22.9 minutes through actual measurement, and obtained a result that met the specification requirement of ≥2.5μm.

[0104] The measured plating time data of Experiments 8 and 14 are used to replace the predicted data in the control module 500 of the chemical plating device for upgrading. Based on the upgraded plating time data of Experiments 8 and 14, the plating time range of Experiments 9 to 13 is obtained by processing through the multimodal chemical plating process conditions 800 system, which is 23.3 minutes to 24.4 minutes.

[0105] Furthermore, by optimizing and upgrading the plating time, Experiment 11 was actually operated using the chemical plating apparatus 600 to obtain semiconductor element samples plated with tungsten-nickel alloy. The various test data and the test results of Experiment 8 and Experiment 14 are shown in Table 9.

[0106] Table 9. Solution temperature, plating time, and sample test results for Tests 8, 11, and 14.

[0107] Table 10. Results of coating evaluation for samples from Tests 8, 11, and 14.

[0108] The plating temperatures for control tests 5 to 8 were 73℃, 75℃, 91℃, and 93℃, respectively. The results are shown in Tables 11 to 13. All of them were deemed unqualified.

[0109] Table 11 Plating process conditions for Control Experiments 5 to 8

[0110] Table 12 Test results of samples from control experiment 6 and control experiment 7

[0111] Table 13 Results of coating determination for samples in Control Test 6 and Control Test 7

[0112] Therefore, the preferred range of solution temperature established by the method of the present invention is 77°C to 89°C.

[0113] 3. Optimization and verification of pump flow rate The difference between this experiment and Experiment 11 lies in changing the pump flow rate. The predicted flow rate range was 230 L / min to 430 L / min, and actual measurements were taken at flow rates of 250 L / min, 270 L / min, 290 L / min, and 310 L / min. The oscillation frequency was 30 times / min, and the temperature was 83℃. The results are shown in Tables 14 to 19.

[0114] Table 14 Plating process conditions for experiments 15 to 21

[0115] As shown in Table 14, in Experiment 15, a pump flow rate of 270 L / min and a predicted plating time of 23.7 minutes were used for the actual plating test. The plating thickness of the tungsten-nickel alloy layer did not reach the expected 2.5 μm. When the measured plating time was adjusted to 23.8 minutes, the plating thickness of the tungsten-nickel alloy layer reached the expected specification requirement of ≥2.5 μm. Similarly, in Experiment 21, the same method was used. The predicted plating time of 22.5 minutes was adjusted to 22.6 minutes through actual measurement, and the result met the specification requirement of ≥2.5 μm.

[0116] The measured plating time data of Experiments 15 and 21 are used to replace the predicted data in the control module 500 of the chemical plating device for upgrading. Based on the upgraded plating time data of Experiments 15 and 21, the multi-modal chemical plating process conditions 800 system is used to process the data, and the optimized and upgraded plating time range of Experiments 16 to 20 is obtained as 22.9 minutes to 23.7 minutes.

[0117] Furthermore, by optimizing and upgrading the plating time, Experiment 18 was actually operated using the chemical plating apparatus 600 to obtain semiconductor element samples plated with tungsten-nickel alloy. The various test data and the test results of Experiment 15 and Experiment 21 are shown in Table 15.

[0118] Table 15 Pump flow rate, plating time, and sample test results for Tests 15, 18, and 21

[0119] Table 16. Results of coating evaluation for samples from Tests 15, 18, and 21.

[0120] The pump flow rates of control tests 9 to 12 were 230 L / min, 250 L / min, 410 L / min, and 430 L / min, respectively. The results are shown in Tables 17 to 19. All of them were deemed unqualified based on comprehensive evaluation.

[0121] Table 17 Plating process conditions for Control Experiments 9 to 12

[0122] Table 18 Test results of samples from control experiment 10 and control experiment 11

[0123] Table 19 Results of coating determination for samples from control tests 10 and 11

[0124] Therefore, the preferred range of pump flow rate established by the method of the present invention is 270 liters / min to 390 liters / min.

[0125] Through the above verification experiments, the preferred range of chemical plating process conditions established by the method of this invention is as follows: oscillation frequency of 21 to 39 times / min; plating time of 22.6 to 24.5 minutes; solution temperature of 77°C to 89°C; and pump flow rate of 270 to 390 liters / min. The optimal process conditions are: oscillation frequency of 30 times / min, plating time of 23.5 minutes, solution temperature of 83°C, and pump flow rate of 330 liters / min.

[0126] In summary, the multimodal chemical plating process optimization method provided by this invention achieves a quantitative mapping between workpiece morphology features and multidimensional process parameters by constructing a multimodal neural network that integrates a semiconductor device morphology characterization model and a chemical plating process condition clustering data model. This method inputs the process conditions predicted by the model into the chemical plating apparatus for actual plating, obtains experimental data, compares and judges it with the predicted data, and corrects the model based on the judgment results, so that the predicted results gradually approach the experimental data. This forms a closed-loop optimization system of "prediction-verification-judgment-storage-iteration," and the optimized process conditions can simultaneously meet the comprehensive requirements of the coating's mechanical properties, corrosion resistance, electrical properties, and wear resistance.

[0127] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for optimizing multimodal chemical plating process conditions, applied to the double-sided chemical plating process of semiconductor devices, characterized in that, Includes the following steps: Obtain the morphology, appearance, dimensions, and plating specifications of the semiconductor device, and construct a morphology characterization model of the semiconductor device; Establish a database of metal electroless plating solution composition, an electroless plating device database, and an electroless plating process condition database, and build a clustering data model of electroless plating process conditions based on the databases; The morphology characterization model and the clustering data model are fused as different branches of a neural network to construct a multimodal neural network model for predicting chemical plating process conditions. The steps of constructing the multimodal neural network model include: calculating the electrochemical relationship between the standard electrode potential of the target metal and the plating time, introducing correction parameters to compensate for the influence of dynamic process factors on the plating results during the chemical plating process, and establishing an initial mapping relationship between the chemical plating process conditions and the theoretical time required to reach the target metal film thickness, as the basis for the multimodal neural network model to make predictions. From the numerical range of the chemical plating process conditions predicted by the multimodal neural network model, at least one low-end value and one high-end value are selected and input into the chemical plating device to perform plating on the semiconductor device, and the corresponding first experimental data and second experimental data are obtained. The first and second experimental data are compared and judged with the prediction data of the multimodal neural network model, respectively. Based on the discrimination results, the correction parameters are adjusted, and the adjusted model is used for the next prediction of process conditions to achieve iterative optimization of chemical plating process conditions.

2. The method for optimizing multimodal chemical plating process conditions according to claim 1, characterized in that, The morphological dimensions include length, width, thickness, and three-dimensional shape; the plating specifications include the type of metal, the thickness of the plating metal film, and the plating area.

3. The method for optimizing multimodal chemical plating process conditions according to claim 1, characterized in that, The chemical plating process conditions include oscillation frequency, plating time, solution temperature, and pump flow rate.

4. The method for optimizing multimodal chemical plating process conditions according to claim 1, characterized in that, The comparison and discrimination step further includes: pre-setting a comprehensive judgment standard, which includes six judgment items: coating adhesion judgment, morphological appearance judgment, gas corrosion rate judgment, resistance and impedance judgment before and after gas corrosion test, resistance and impedance judgment of insertion and extraction test, and quality difference rate judgment before and after friction and wear test; only when the first experimental data and the second experimental data are evaluated according to the comprehensive judgment standard, and the judgment results of the six judgment items are all excellent, will the corresponding chemical plating process conditions be established as the preferred process conditions.

5. The method for optimizing multimodal chemical plating process conditions according to claim 4, characterized in that, Also includes: The original data of the chemical plating process conditions corresponding to the comprehensive judgment result of "excellent", as well as the related morphological characterization model and clustering data model, are stored in the effective database for iterative updates of the model; the original data of the chemical plating process conditions corresponding to the comprehensive judgment result of "unqualified" are stored in the failure database to enrich the analogy data source.

6. The method for optimizing multimodal chemical plating process conditions according to claim 1, characterized in that, The metal electroless plating solution is a tungsten-nickel alloy solution, the components of which include: 280 mmol / L sodium tungstate dihydrate, 13.2 mmol / L triammonium citrate, 14.8 mmol / L triethylenetetraminehexaacetic acid, 120 mmol / L nickel aminosulfonate hydrate, 184 mmol / L sodium hypophosphite, 0.31 mmol / L polyethylene glycol octadecyl and 0.49 mmol / L 1,2,4-triazole-3-carboxamide.

7. The method for optimizing multimodal chemical plating process conditions according to claim 1, characterized in that, The neural network is a recurrent neural network.

8. The method for optimizing multimodal chemical plating process conditions according to claim 1, characterized in that, The swing frequency in the chemical plating process is 21 to 39 times per minute, the plating time is 22.6 to 24.5 minutes, the solution temperature is 77°C to 89°C, and the pump flow rate is 270 to 390 liters per minute.

Images