TWI221435B - Method for optimizing timing control process parameters in chemical mechanical polishing - Google Patents

Abstract

A method for optimizing timing control process parameters in chemical mechanical polishing mainly comprises: using a neural-network Taguchi method to develop a low cost but efficient technique in optimizing chemical mechanical polishing (CMP) process parameters including time control. The new technique uses a neural network to simulate a CMP process and further obtain a set of optimum process parameters including an optimum polishing time. When the process parameters are set at the optimum parameter set, a maximum removal rate and a minimum non-homogeneity can be achieved in the shortest time.

Description

1221435 _____________ A7 ----——-— B7 _ V. Description of the invention (/) [Technical field to which the invention belongs] The present invention is a method for optimizing control process parameters during chemical mechanical polishing. The combination of neural network is used to simulate and obtain the best lingling array in the system of CMP. It can provide better grinding effect under the shortening of grinding time. [Previous Technology] The chemical mechanical polishing (CMP) process is to use a chemical polishing liquid containing fine grinding and polishing powder particles, and perform mechanical engraving and mechanical polishing with the mechanical movement of the grinding wheel to achieve grinding. In recent years, IBM has taken the lead in improving this technology and applied it to the semiconductor manufacturing process, breaking through the bottleneck of wafer manufacturing in the semiconductor industry and obtaining a fairly good flat effect. First, in the semiconductor manufacturing process, the most important parts are the three major technologies of deposition, lithography, and etching. The CMP process technology is the main auxiliary technology for wafer planarization, and the resolution of a deposition layer and the lithography process. Whether the pattern transfer in China is 70% is directly or indirectly affected by wafer planarization, especially the multi-metal wire process of integrated circuits and the wafer planarization involved. This is a major challenge for VLSI process technology to advance. One. When the wire width of ultra-large integrated circuits reaches less than 0.25 micrometers, CMP is also the only technology that can be used today, Global pianarization. Therefore, the global semiconductor industry competes with research institutions for the global semiconductor industry. Develop and commit to research and development technology. The CMP process includes the effects of chemical reactions and mechanical movements. There are many influencing factors in the process; the speed of the wafer carrier and the pressure of the wafer • — — III — 1 ---- (Please read the precautions on the back first (Fill in this page again) Revision · · 3 V. Description of the invention (y) The size, supply rate of research application and PH value of research failure, etc., all have a great impact on the results, for PUD & In order to accurately control the clothing, the process can be properly ground to avoid too much or too little removal. It is also an important issue in the CMP process technology. It is a problem of End Point DeteetiQn. The traditional method of U-end point detection is performed offline. The & point of offline measurement is highly accurate. The disadvantage of '㈣its offline method' also causes excessive or insufficient removal. The direction of measurement is' and because CMP is a very dynamic process, the use of online real-time measurement will be more in line with demand. The methods include: 1. Motor current method; 2. Pad thermal method (pad thermal); 3. Optical method (Optic interfer ⑽ eter method); 4. Sonic method (Acoustic method); 5. · Small vibration method. Although the above endpoint detection technologies are based on different principles to achieve the detection endpoint, they must all be combined with accurate closed loop (close 100p) process control technology. However, until now, true real-time closed loop process control technology It is rare to see how to combine accurate on-line monitoring and endpoint detection technology to develop real-time closed-loop process control technology to increase production capacity and reduce costs, which is the basic spirit of the present invention. [Summary of the Invention] The main purpose of the present invention is to provide a method for optimizing control process parameters during chemical mechanical polishing, taking into account the time factor for its removal rate and flatness. 4 1221435

Fifth, the invention, explain the impact of (^, without the need to change the head of the organic σ # structure and add other measurement equipment

Tl ^ 〇VER ^ Γ is made as the mode of grinding end point, which can achieve the highest removal rate and the purpose of grinding with small unevenness. [Embodiment] The technical method used in the optimization method of the control process parameters in the chemical mechanical polishing of the present invention is to construct the experimental design weighting and process mode by using the neural-like Taguchi experiment method. For details, please refer to the chart and refer to The following description. The Takaguchi test method is a valid system simulation and parameter setting tool that combines neural-like network and Taguchi method. The optimization steps of CMP process parameters in parameter design are as follows: : 1. Design and guide experiments by Taguchi method: The materials used for the experiments are mainly oxide layers that are easier to obtain. Four more important and adjustable process parameters are selected as control variables. They are: A_ Wafer pressure (Down Force) B. Platen Speed C. Grinding Slurry Composition D. Back Pressure (E. Polishing Time) and Removal Rate The relationship with the grinding pressure and grinding speed is shown in the preston equation: -------- Order -------- # (Please read the precautions on the back before filling this page)

Removal rate

AH Δ /

kp * P

He

As △ t 5 V. Description of the invention (V) where ΔΗ: surface height change △ t: polishing time Kp: polishing parameter P: polishing pressure AS / Δ t The linear velocity of the M pad relative to the wafer and therefore the removal rate and polishing pressure It is directly proportional to the polishing speed, and the calculation of the polishing pressure in equat coffee is the wafer pressing force divided by the contact area of the polishing pad. Therefore, as the polishing time is different, the average surface hook of the wafer will be different. Will change, so that the pressure under the wafer changes accordingly. Therefore, it can be inferred that the material removal rate (MateriaI rem〇val rate, _ and 卩 不 surface uneven self-degree ㈣ hin Wafer N_lf_lty, W surface) is a function of polishing time, Therefore, time was selected as the fifth control variable to build a more complete _ process model (e

Polishing time). In the experimental design, the L25 (55) orthogonal table as shown in the table is selected, and the 5 rows represent the 5 control variables of A to E, and each of the 25 columns represents the 5 control variables of the experimental study. The level of the value, the observation required in the experiment still has the bigger the better, and WIWNU has the smaller the better. The calculation method of the two is as follows: (1) Use an optical film thickness gauge to measure the thickness of the wafer before grinding and measure 9 points, and make the value t !; (2) After grinding for T minutes, measure the grinding The thickness of the rear wafer "; (3) the polishing rate, the average value of the polishing rate at each point is 6 1221435

V. Description of the invention (f) MRR of this experiment; (4) WIWNU = (standard deviation of t2 σ) / (· ^ mean mean) * 100%, and because WIWNU must be compared when the thickness is similar after grinding Only makes sense, so further replace WIW with the standard deviation of t2, find the best parameters to make WIWNU meet the requirements. 2 · Signal-to-noise ratio analysis: The main factors that affect process performance are the control factor (factor) and noise fact (factor). The experimental design of Taguchi's formula is based on the signal-to-noise ratio (S / N) ratio. As a performance indicator to measure the value of a product or process, the higher the S / N ratio, the more stable the product f, and the closer the system's performance is to the ideal target. After experimenting on the process technology and the planning of the parent table, first calculate the corresponding ratio (S / N) to 0 standard deviation for each experiment. The values are shown in Table 2. Then calculate the effect of each factor to make a response table (Table 3 and Table 4) and a response chart (the first chart and the second chart), and finally perform the analysis of variation number (Analysis Wi bribe, shed M). The function of the response table is to explain the effectiveness of Xunzi at different levels.

The graph shows the trend of the effect of each factor at different levels, and the variable I analysis is to find the contribution of each factor. Therefore, from the response chart of the s / N, it can be clearly seen that each factor is In the range of the experimental process, the level of the maximum yield ^ / N ratio, that is, the optimal level of this factor, and the combination of each wIWNr ^ under the optimal level is the relative parameter combination of Zhan and IWNU . From the figure, you can find that A5B5CiD5Ei is the local best for mrr. 1221435 V. Description of the invention (t) = number of degrees ° and 疋 f: From the second figure, the secret can be found for t2. Local optimum parameter setting value. Let's use the table to analyze the variation number of _ and we delete. Using the present invention-"analysis can find out that each factor within the set range contributes to the improvement of '= low: standard deviation (can be regarded as how much WI_ has contributed ( The reference of E = f. Table 5 shows the backlash) and the effect of time = R is relatively smaller than other factors. As for the thickness after removal, the standard deviation is shown. The 5 control variables all have a certain degree of shadowing. 3.CMP The establishment of the process model (Model) ... Said that the two-pass technology is a very complicated system, so the present invention is intended to be based on the system. The simpler orthogonal tables are often beneficial: The neural network-like sample is sufficient. In the present invention, the experimental data of Table 2 is used as a training sample. 'Γ :? This is set between positive and negative 0 · 8 to build a neural network-like two = informal network such as In the third picture, the learning method used is-I have a manual method to adjust the learning parameters and the inertia factor = the two-tune learning rule. During network training, use the control variables shown in Table 1: Input, shift The removal rate and the standard deviation of the thickness after grinding are two output terms. Do n’t have the opposite nature and 1 nature, so define the objective function: a ^ MRR (S / N) + STDEV (S / N) to summarize the standard deviation of thickness and thickness (s bribe), and establish a face and a picture The process model, in which the sunny (s / N) is as large as possible when obtaining the optimal parameters (S / N), so it is out of order (please read the precautions on the back before filling this page). Order · 8 V. Description of the invention (maximum number), the maximization is the direction of parameter adjustment, and because the main goal in the CMP process is to quickly remove and ensure that WIWNU is within the required range, a is set to 0, 6, b Set to 0.4. 4 Optimization of process parameters: Straight = The levels in the table are not continuous. The best parameters obtained by s / N analysis are still rough estimates, as shown in the third figure. The CMp process is to incorporate Shun and STDEV into the temple. Therefore, the optimal parameters cannot be obtained from the S / N ratio analysis. If a more accurate optimal parameter combination is required, a neural network such as the one established must be used. MRR_S is used as a guard, and the fine-tuning function of parameter optimization is performed by using the best parameter design method. The steps are as follows: (1) Input the local optimal MRR and STDEV parameter values obtained from the S / N ratio analysis into the human path, and calculate the network output value. The network output value is (J) max ° (2) at (J ) Calculate 3⑺ / touch value around max, and change Xi within the allowable effective range to increase δ⑺ ore. If a⑺ / fat is positive, increase xi. If 3 //) is considered negative, decrease xi. When xi If it cannot be changed, skip to step ⑷ 'where Xi represents the input factor of the network. (3) Calculate j using neural network. When J > Jmax, let Jmax = j, and return to step (2). (4) The required parameter is the optimal parameter value for MRR and STDEV. According to the above steps, the optimal parameter can be set as:

Solid content: 20% 9 1221435 V. Description of the invention (

Down force: 8.Opsi

Back pressure: 3.2psi

Plsten speed.41 rpm

Polishing time: 42sec Among them, the physical meaning of 42Sec is that it can stop grinding when it is sent, and at this time, it can produce the maximum;, sed is 5. Verification experiment: Take half of the small uneven hook.焉 The double experiment is to use the aforementioned neural Taguchi e experiment and compare the t tea number into the parameters of the real nanometer component of the country. As shown in Table 7, the original value used for M ^ eye, T Larrier speed and two This parameter sets the original data in the combination of MRR and WIWNU's shadow change. U This parameter can be found from the experimental results in Table 7. 'In the case where the crystal (coffee) meets the requirements, the needle found in the present invention: = the optimal parameter set of the material, can indeed shorten the grinding time / 3 Below, ^ to the purpose of rapid grinding. For other materials that are different from the oxide layer, the same process can be used to find the optimal parameter combination and research room for the shape. In other words, without the need to update the machine equipment, in addition to the optimization history and number combination of the CMP process can be selected In addition, it can also be used as an effective tool for end-point detection. Regardless of its cost considerations or effectiveness, 'all have high feasibility of its development: economic benefits' actually-excellent inventions'. Therefore, the invention patent is requested in the text. . [Brief description of the diagram] Response diagram The first diagram ... The method of the present invention is directed to MRR 10 1221435 A7 B7 5. Description of the invention (I) The second diagram: The response diagram of the method of the present invention directed to STDEV. Third diagram: A simple schematic diagram of a neural network architecture such as a CMP process in the method of the present invention. [Description of main component symbols] (None) --- 1 ------ AWII .— (Please read the precautions on the back before filling this page) Order

I 11 1221435 A7 B7 V. Description of the invention (/ 〇) Table 1 L25 (55) orthogonal array V ίτ) ABCDE column \ ABCDE Solid Content (wt%) Down Force (psi) Back Pressure (psi) Platen Speed (rpm) Time (sec) I 1 1 1 1 1 5% 4 0 20 40 2 1 2 2 2 2 5% 5 1 30 45 3 1 3 3 3 3 5% 6 2 40 50 4 1 4 4 4 4 5 % 7 3 50 55 5 1 5 5 5 5 5% 8 3.5 60 60 6 2 1 4 2 3 10% 4 3 30 50 7 2 2 5 3 4 10% 5 3.5 40 55 8 2 3 1 4 5 10% 6 0 50 60 9 2 4 2 5 1 10% 7 1 60 40 10 2 5 3 1 2 10% 8 2 20 45 11 3 1 2 3 5 15% 4 1 40 60 12 3 2 3 4 1 15% 5 2 50 40 13 3 3 4 5 2 15% 6 3 60 45 14 3 4 5 1 3 15% 7 3.5 20 50 15 3 5 1 2 4 15% 8 0 30 55 16 4 1 5 4 2 20% 4 3.5 50 45 17 4 2 1 5 3 20% 5 0 60 50 18 4 3 2 1 4 20% 6 1 20 55 19 4 4 3 2 5 20% 7 2 30 60 20 • 4 5 4 3 1 20% 8 3 40 40 21 5 1 3 5 4 25% 4 2 60 55 22 5 2 4 1 5 25% 5 3 20 60 23 5 3 5 2 1 25% 6 3.5 30 40 24 5 4 1 3 2 25% 7 0 40 45 25 5 5 2 4 3 25% 8 1 50 50 (Please read the notes on the back before filling out this page) Δ λ i 4 ow / r 9 1 η Y 9Q7 / entry, 1221435 A7 B7 V. Description of the invention (//) Table 2 Experimental results and signal noise ratio (S / N Ratio) L25 (55) MRR (S / N) STDEV (S / N) ABCDE (A / S) (/ 2) 1 1 1 1 1 1 9.14 19.22 137.540 17.23 2 1 2 2 2 2 14.5 23.23 109.591 19.20 3 1 3 3 3 3 16.34 24.26 99.567 20.04 4 1 4 4 4 4 20.31 26.15 128.673 17.81 5 1 5 5 5 20.55 26.26 98.534 20.13 6 2 1 4 2 3 16.09 24.13 51.999 25.68 7 2 2 5 3 4 23.45 27.4 152.866 16.31 8 2 3 1 4 5 28.92 29.22 298.643 10.50 9 2 4 2 5 1 34.64 30.79 88.683 21.04 10 2 5 3 1 2 22.51 27.05 112.108 19.01 11 3 1 2 3 5 19.44 25.77 92.238 20.70 12 3 2 3 4 1 27.68 28.84 114.126 18.85 13 3 3 4 5 2 36.89 31.34 114.363 18.83 14 3 4 5 1 3 22.11 26.89 55.401 25.13 15 3 5 1 2 4 31.84 30.06 110.232 19.15 16 4 1 5 4 2 20.44 26.21 100.058 19.99 17 4 2 1 5 3 29.22 29.31 169.638 15.41 18 4 3 2 1 4 18.37 25.28 167.419 15.52 19 4 4 3 2 5 28.72 29.16 114.870 18.80 20 4 5 4 3 1 39.28 31.88 69.075 23.21 21 5 1 3 5 4 23.25 27.33 113.494 18 .90 22 5 2 4 1 5 17.1 24.66 46.615 26.63 23 5 3 5 2 1 25.74 28.21 52.903 25.53 24 5 4 1 3 2 34.59 30.78 139.578 17.10 25 5 5 2 4: 3 44.11 32.89 206.076 13.72 1IIII — — — — I- -(Please read the precautions on the back before filling this page) · 13 1221435 A7 B7 V. Description of the invention (Table 3 MRR response form MRR response form factor level ABCDE Solid Content (wt%) Down Force (psi) Back Pressure (psi) Platen Speed (rpm) Jime (sec) 1 23.80363 24.53263 27.71867 24.62076 27.78998 2 27.71891 26.68942 27.59314 26.95966 27.72017 3 28.58194 27.66443 27.32932 28.02038 27.49836 4 28.37092 28.75611 27.63418 28.66399 26296 29296 296.296. · II. (Please read the notes on the back before filling this page) Order ·-^ 4® m (/ ^ XTC \ / ΌΐΠν 9Q7 / N ^ ^ 1221435 A7 _ B7 V. Description of the invention (丨 xi) Table 4 is for Response table for t2 standard deviation 回应 Response table for standard deviation \ ^ Factor level \\ ABCDE Solid Content (wt%) Down Force (psi) Back Pressure (psi) Platen S peed (rpm) Time (sec) 1 18.88245 20.50176 15.8791 20.7043 21.17418 2 18.50823 19.28189 18.03852 21.6729 18.82889 3 20.53432 18.08461 19.11871 19.4741 19.99524 4 18.58753 19.97647 22.43348 16.8772 17.54044 5 20.37666 19.04445 21.41936 20.6428 19.35043 Γ-load ----- Note on the back page, please fill in this page again)--/ mrc, λ / mrw 9Q7 vol. 8, 1221435 A7 B7 V. Description of the invention (丨 f) Degree A 86.2601 4 21.565 37.5417 B 77.4767 4 19.3692 33.719 C 1.83596 4 0.45899 0.79904 D 62.0033 4 15.5008 26.9848 E 2.19542 4 0.54886 0.95548 Sum 229.771 20 100 — — III — .Installation -------- Order · ---- --- # (Please read the notes on the back before filling this page) lb 1221435 A7 Β7 V. Description of the invention (^) Table 6 Analysis results of the variance of the STDEV factor squared degrees of freedom mean square and contribution A 19.8066 4 4.95164 6.93045 B 17.0717 4 4.26793 5.97351 C 138.034 4 34.5086 48.2991 D 74.4472 4 18.6118 26.0496 .E 36.4308 4 9.10771 12.7474 Total 285.791 20 100 (Please read the precautions on the back before filling out this page) '1 1221435 A7 _B7 V. Description of the invention (1 t) Table 7 Best parameter confirmation experiment Best parameter confirmation Experiment original parameter Best parameter Solid Content 15% 20% Down Force 7 psi 8 psi M Back Pressure 3 psi 3.2 psi Platen Speed 20 rpm 41 rpm Carrier Speed 25 rpm 25 rpm Oscillation Speed 1 mm / sec 2 mm / sec Time 60s A2s Average removal amount 15992 A 1897.9 4 average thickness after removal 3509.9 / 3149.9 / standard deviation of thickness after removal 112.2 / US.OA experiment MRR 値 26.65 A / s 45Λ9 A / s experiment WIWNU 値 3.1963% 3.7506% -------- --- Aw- I --- (Please read the notes on the back before filling this page). # ·

Claims (1)

1221435 method, its patent application Fan Yuant applied for the 0th G-fW No. 54 "optimization method of control process parameters during chemical mechanical polishing" patent application revision: a best method of control process parameters during chemical mechanical polishing Including: A. Designing and guiding experiments by Taguchi method: For the target material, select * more important and adjustable process parameters as control variables and use time as the 5th control variable to construct a more complete CMP process Mode, take a variety of different level tests and calculate the removal rate ⑽R) and the wafer surface unevenness (wmiu), among which the selected process parameters are: wafer pressure size ⑺Force) 'milling platform speed (piatenSpeed) · polishing Pulp content (s〇⑽ content) '· 磨 塾 Back 璧 (Back and Polishing time); Β · Λ5 Tiger miscellaneous daily ratio analysis ·· For each experiment, calculate the corresponding earning and removal The standard deviation of the back thickness (s / magic ratio, and then calculate the effect of each factor to make a response table, and finally perform a variance analysis (⑽㈣⑽ Vanance, AN0VA) to find the factors that produce the largest s / n ratio Level value, # / ° sub-optimal level value, and then set each factor at the best Jc level, and under the conditions of the 'decided' to determine the best combination of parameters relative to hidden and prettier; ^ Zhang scale Applicable to Chinese national standards " ^ I f _ 装 ---- * ---- Order: (Please read the precautions on the back before filling this page) ^ 1435 、 C.CMP process model (M〇) The establishment of del): ^ ^ Mao arsenic was used to perform CMP I and M through the network. During network training, the actual data of step A and double data were used as the sword chain. The control at different levels became the network 7 ^, The financial R and the mill factory degree ^ vehicle difference ⑽ΕΌ are two round-out terms, and an objective function is defined to summarize MRR and STDEV. A ^ MRR {S / N) ^ b ^ STDEV {S in step c / N) The function shown is: the maximization of this objective function is the direction of parameter adjustment; D. process parameter optimization: using a neural network (such as the one established by 〇 — STDEV model as a tool, where the parameter design method is used The steps of fine tuning function for parameter optimization are as follows: a. Local optimal mrr and stdev parameter values obtained by S / N ratio analysis Enter the neural network and calculate the output value of the network. The output value of the network is (J) max; printed by the Consumer Cooperative of the Intellectual Property Bureau of the Ministry of Economic Affairs. B. Calculate the value of 3 (j) / aYz · near (J) max. , And change xi within the allowable effective range to increase 3 ⑺, if 5 0 /) pass is positive, then increase xi if ⑺ / is considered negative 'then decrease h, when h cannot be changed, skip to step d' where xi represents the input factor of the network; c • Calculate j using neural-like networks. When j> jmax, let jmax = j, and return to step b; The paper size is ㈣0 ^ standard (CNS) A4 specification TT10 x 297 public love> 1221435 The number of quality research, patent application scope # is the number for Ewa The best parameter of STDfT Γ uses the best parameter design method to fine-tune the parameters to optimize the parameters. 2 · The method for optimizing the chemical process parameters as described in the "Scope of Patent Application", wherein in the step C, the objective function of grinding the oxide layer material is set to 06 and b is set to 0. (Please read the notes on the back before filling out this page} Printed by the Consumer Cooperatives of the Intellectual Property Bureau of the Ministry of Economic Affairs This paper is sized to the Chinese National Standard (CNS) A4 (210 X 297mm f)