TWI264043B - Method and system for analyzing data from a plasma process - Google Patents

Abstract

A method and system for analyzing multivariate data of plasma processes in which response surface and neural networks are utilized to improve or find optimal process settings of the plasma process such that performance measurements are compared against model measurements to modify a current plasma process to achieve optimized processing performance.

Description

1264043 V. INSTRUCTIONS (1) 1. TECHNICAL FIELD OF THE INVENTION The present invention is generally related to the field of plasma processing. Especially regarding the monitoring and analysis of process parameters in plasma processing equipment. 2. [Prior Art] Den electrical processes, such as semiconductor or display parameters, can change significantly. Process conditions can change critical process parameters with micro-grounds. Minor changes can easily occur at the etch gas or wafer temperature. Therefore, the plasma process design and monitoring of these process parameters can be used to adjust the process parameters. However, in many cases, changes in the process data cannot be simply measured. In order to detect initial anomalies and processes, it is often necessary to perform various intricate, monitoring, and analytical manufacturing processes through re-entering wafer processes and care in advanced process control plants. However, it cannot be applied to various plasma processes and there is no final process relationship. Looking at the various stages of manufacturing, etc., the key processes change over time, even the lightest produces undesired results. The composition or pressure and process chambers need to be constantly monitored. Analyze valuable data at any given time. Or, it is difficult to determine the deterioration of the characteristics of a process-like material that reflects the deterioration of process characteristics. (APC) predicts and recognizes the pattern. Because of the process steps in the semiconductor manufacturing process, the computer is generally used to control the data collection system software; the data accumulated and analyzed by the law

1264043

III. SUMMARY OF THE INVENTION In one embodiment of the invention, systems and methods are used to determine the interaction of process harmonics with other measurements to achieve desired process results. It is widely used in the relationship between parameters and characteristic measurement of various plasma processes. Decision and process variables combined with existing plasma system IV. [Implementation]

The detailed description of the embodiments with reference to the present invention is intended to be <RTI ID=0.0>> Therefore, the following detailed amounts f, θ θ limit the present invention. The scope of the invention is defined by the accompanying application; BRIEF DESCRIPTION OF THE DRAWINGS Figure 1 is a diagram of a system 100 in accordance with the present invention. The plasma processing system 100 includes a process chamber: a device 104 for measuring and adjusting at least one process parameter, two to four J, a device for measuring process characteristics, and a controller 108. The control 4 is connected to the process chamber 102, the device 104, and the device 106.

The plasma processing system 1 Q of Figure 1 uses, for example, plasma treatment 4k t\\ ^ u n ZL

Wei Zhi. Alternatively, the plasma processing system 100 includes a photoresist coating such as a 弁P day '^ to ^ ^ foot-turn coating system. The system also includes a photoresist pattern transfer chamber such as a UV (ϋν) lithography system or a dielectric coating 2 that will be changed by a SOG or a spin-coated dielectric (SOD). System / ^ Er Zhu. In addition, the electro-hydraulic I-pass system 1 〇 〇 includes: deposition chamber, such as Huayu Weng 4# (CVh \ 予礼目 sink w) or physical vapor deposition (PVD) system, rapid heat treatment

1264043

The (RTP) chamber, such as the RTp system or batch device 104 for thermal annealing, can adjust process parameters such as electrode spacing back maintaining ae) air pressure, total flow rate, flow ratio, lower electrode RF power, upper electrode rating power, Wafer holder temperature, or other. 2 is a more detailed illustration of a plasma processing system 20 in accordance with the present invention. The plasma manufacturing system 2 〇 〇 includes, as shown in FIG. 1, a process chamber i 〇 2, a substrate holder 212, and a substrate 2〇4 to be processed is fixed on the holder, the gas injection system 206, and the vacuum pump system 208. Substrate 204 can be, for example, a semiconductor substrate, a wafer, or a liquid crystal display (LCD). The process chamber 1〇2 can be configured, for example, to facilitate the generation of plasma adjacent the surface of the substrate 204 within the process region 202. The plasma can be formed by collision between hot electrons and ionizable gas in the process chamber 1〇2. The ionizable gas or mixed gas 21〇 can be introduced into the process area 2〇2 via the gas injection unit 206. Control mechanisms not shown can be used with the final pump system 2〇8. The plasma in the process zone 2〇2 can be used to create materials that are different from the predetermined material process and to help deposit material into the substrate 204 or to remove material that is exposed on the surface of the substrate 2〇4. Substrate 204, for example, can enter or exit process chamber 102 via a slot valve (not shown) and a chamber feeder (not shown) of the automated substrate transport system. The substrate 20 can be obtained by a substrate plug (not shown) fixed to the substrate holder 212, and the substrate can be mechanically transferred by means of a device fixed thereto. The substrate 205 can be fixed to the substrate holder H2 via, for example, an electrostatic tong system 2丨4. The substrate holder 2 12 includes a cooling system having a recyclable circulating coolant that absorbs heat from the substrate holder 21 and transports the tropics to the heat exchanger (not shown in Figure 12 (4). The cooling system also includes means 2 1 6 configured to monitor the temperature of the substrate 2 〇 4 and/or the substrate holder 2 1 2 . The device 216 can be, for example, a thermoelectric device such as a kappa type thermoelectric device. In addition, gas can be transferred to the back side of the substrate 248 via the back gas system 218 to improve the inter-gas thermal conductivity between the substrate 604 and the substrate holder 212. The backside gas system 2 18 can be utilized when the substrate 250 needs to be warmed or cooled. For example, balancing the heat flux transferred from the plasma to the substrate 2 及4 and the heat flux removed from the substrate 204 is conducted to the substrate holder 212 causing the temperature to exceed the steady state temperature, and the temperature control of the substrate 204 is useful. . The substrate holder 212, for example, can further function as an electrode from which rf power can be transferred to the plasma within the process chamber 202. For example, the substrate holder 2丨2 can be biased to an RF voltage, and RF power can be transmitted from the RF generator 22 through the impedance matching network 22 2 to the substrate holder. The RF bias can act as a hot electron, thus forming and maintaining Plasma. With this setting, the system operates as a reactive ion etching (RIE) reactor in which the chamber and the upper gas injection electrode act as a ground plane. The frequency of the general RF bias ranges from "" to "10" and can be 13.56 MHz. The plasma process system is well known to those skilled in the art. Or 'a plurality of frequencies of RF power can be supplied to the substrate holder electrode. In addition, the 'impedance matching network 2 2 2 maximizes the R ρ power delivered to the plasma in the process chamber 2 0 2 by reducing the reflected power. Matching the network architecture (eg [type, π, T, etc. And the automatic control method is well known in the art. n The process gas 2 1 图 of FIG. 2 can be introduced into the process area 20 2 via the gas injection system 2 〇 6. For example, the process gas 210, for example, includes a mixed gas for etching. Argon gas, CF4 and % or argon, qF8 and 〇2.

Page 8 1264043 V. INSTRUCTIONS (5) The body injection system 2 0 6 includes a gas jet head, wherein the process gas 2 1 〇 can pass through a gas injection system (not shown) through a gas injection system (not shown), a series of buffers A plate (not shown) and a porous jet gas injection plate (not shown) are supplied to the process area 202. The gas injection system is well known to those skilled in the art.

The vacuum pump system 208 includes, for example, a turbomolecular vacuum pump (TMP) with a pumping rate of 50 〇 liters per second (and larger) and a gate valve that regulates chamber pressure. In the conventional plasma process, a dry plasma etching apparatus is used, which requires 1 to 3 liters per second. ^1? is usually useful in low presses below 5〇111. At higher pressures, the TMP pumping rate drops dramatically. For high pressure processes, such as above 100 mTorr, mechanical booster pumps and dry pumping pumps are available. Additionally, the monitoring chamber pressure passage 224 can be coupled to the process chamber 102. The pressure measuring device 224, for example, can be a heart-shaped Baratr〇n absolute capacitance manometer sold by MKS Instruments (Andover, MA). The electric table process system 2000 includes the calving tool 2 3 0 set to measure the characteristic measurement. For example, in the insect engraving system, the residual ratio and the residual selection ratio (also "one material versus another material" Etching rate ratio), etching uniformity, pattern angle, and critical dimensions. The metrology tool 23 can be set to internal or external f. The 'Wall Gallery Internal Unit' metrology tool 230 can be, for example, a scatterometer that combines the ^^ wheel and the beam profile reflectometer. The scatterometer can be in the A 1 偏 4 4 疋 疋 疋 疋 ( ( ( ( ( ( ( ( ( ( ( ( ( ( ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' ' 2 04. If the mirror (SEM), which cuts the substrate, the illumination map is fortunately

1264043

The controller 1 0 8 includes a microprocessor, a memory and a digital port capable of generating a control voltage sufficient to communicate and initiate input to the plasma processing system, and an output of the meter system. In addition, the controller 1 〇8 is connected to and coupled to the RF generator 220, the impedance matching network 2 22, the gas injection system 2〇6, the vacuum pump system 208, the pressure measuring device 24, the back gas transmission system 218_, and the substrate. The substrate holder temperature measurement system 216, the electrostatic clamp system 214, and the metrology tool 230 exchange information. The program stored in the memory can be used to initiate the input to the plasma processing system 2 根据 根据 according to the stored process. 3 illustrates a plasma processing system 300 in accordance with another embodiment of the present invention. As shown in FIG. 3, in addition to the parts described with reference to FIG. 2 and FIG. 2, the plasma processing system 2 shown in FIG. 2, for example, further includes a mechanical or electronic dc rotating magnetic field system 3 〇 4, Thereby increasing the plasma density and/or improving the uniformity of the electric paddle process. Further, the controller 108 can be connected to the rotating magnetic field system 3 〇 4 to adjust the rotational speed and the magnetic field strength. Designing and implementing a rotating magnetic field is well known to those skilled in the art.

4 is a plasma processing system 400 according to another embodiment of the present invention. The plasma processing system 400 of FIG. 4 is similar to the system of FIGS. 1 and 2; however, the system 40 of FIG. 4 further includes an upper electrode. 402. RF power can be coupled from RF generator 404 to upper electrode 422 via impedance matching network 222. The power frequency supplied to the upper electrode 40 2 generally ranges from 0·1 MHz to 30 MHz and may be 2 MHz. Further, the controller 1 〇 8 can be connected to the RF generator 404 to control the RF power supplied to the upper electrode 402. Designing and implementing an upper electrode is familiar with the technology

Page 10 1264043 Five 'Inventions (7) Well known

Figure 5 illustrates a plasma processing system 5 in accordance with another embodiment of the present invention. As shown in FIG. 5, the system of FIG. 1 further includes an induction coil 510. The RF power is connected from the RF generator 52 to the induction coil 51 via the impedance matching network 530, and the RF power is transmitted from the induction coil 510 via a dielectric. The layer window (not shown) is inductively coupled to the plasma process area 202. The power frequency of the induction coil 5 10 is generally from 10 MHz to 100 MHz and may be 13.56 MHz. Likewise, the power frequency typically supplied to the wafer holder electrodes ranges from 丨0 MHz to 100 MHz and can be 13.56 MHz. In addition, a slotted Faraday bell (not shown) can be used to reduce the capacitive coupling between the induction coil 5 1 〇 and the plasma. Further, the controller 1 〇 8 can connect the RF generator 520 and the impedance matching network 530 to control the power supplied to the induction coil 51 。. Designing and implementing an inductively coupled plasma (ICP) source is well known to those skilled in the art. Alternatively, the plasma may be formed using electron cyclotron resonance (ECR), a transmitting spiral wave, or a propagating surface wave. Each type of plasma source is well known to those skilled in the art. In Figures 1 through 5, the substrate 204 can be processed in the process chamber 1 〇 2 and can be utilized, for example, by the metrology tool 230 for characterization. Characteristic measurement

Including, for example, etch rate, etch selectivity ratio (ie, etch rate ratio of first material to first material) and etch critical dimension (eg, length of pattern), etch pattern anisotropy (eg, etch pattern sidewalls) The porch), film properties (eg film stress, porosity, etc.), enamel paint ___ — I can be found, for example, from the Langmui r probe), ion energy (for example, from the early _, body T Xiao b 5 Jingong analyzer knows), the concentration of chemical substances (for example, from the optical radio frequency wrinkle eight (two) yJ jrfr

Page 11 1640043 Fifth invention description (8) ::, etching mask (such as photoresist), film thickness, etching mask (such as σ photoresist) pattern key size, ^ ^ ^ ^ ^ ^ ΠΓ A ^ 』 from Self-sensing measured by electricity & sister pin: J flat £VDC, peak-to-peak RF substrate bias & and voltage or current amplitude. The nonlinear nature of the plasma produces harmonics, and the harmonics can also be measured along with their 匕 characteristic parameters. t includes, for example, process pressure, gas pressure on the back side of helium, process gas (eg, cf4, c4f8, 〇2, Ar, etc.), partial pressure

1. Process gas flow rate, upper electrode "electricity, lower electrode power, substrate (or wafer holder) temperature, electrode spacing, and focus ring size. Process pressure 1 can be changed, for example, by a pressure measuring device 224 during the process. Set or total process gas flow rate adjustment and monitoring. Adjust and monitor forward and reflected by sub-command to order generator 220-, matching network 2 22, bidirectional coupler (not shown) and power meter RF power. Use flow control to adjust process gas composition flow rate to adjust and monitor process gas partial pressure: use back gas transmission system 21 8 including a pressure regulator to adjust and monitor (helium) back gas pressure. In addition, use temperature monitoring The system parameters can be monitored by the system. The process parameters include the film material viscosity, the film surface tension, the exposure time and the depth of focus. 1 Refer to the process and process parameters of the process system in more detail. Figure 3 shows A simplified diagram 600 shows the RF signal and harmonic component w factory' and the amplitude measurement Χι-& measured in the process chamber. Figure 6, base The frequency 6〇2 is indicated by the χ axis as w. The mother-harmonic frequency is marked by the X-axis as ', w2, w3, w4, w5, etc. In order to obtain the best characteristics of the process system, the reaction surface analysis can be used.

Page 12 1264043 V. INSTRUCTIONS (9) Reaction surface analysis is the connection of important mathematical or graphical representations. Independent variables, numbers, control factors, and dependent variable factors. For example, flow rate and temperature.彳^ | = The result of a controlled or considered variable. In the second embodiment of the present invention, in order to set a single or more independent improvement or to find the best process of the plasma process, the measurement of the characteristics of the reaction surface method will be combined with the reaction surface, and the device 6 of FIG. The method of measuring the surface is designed to evaluate the two::3 type f ratio. The reaction table should have the shape of the surface. The surface vibration effect is reflected in order to understand the inverse or their quadratic term or the underlying term. The main effect and interaction are related to the curvature of the Pill 2 term. In some environments, the reaction surface may only require a secondary or tertiary term with the main effect and interaction behavior: m. However, the complete description of the process changes affects the selection of the tongue of the process. The important factor of 疋 is to make a test on the process operation capability. The analysis of this data can show the factors that adjust or improve the process characteristics. The reaction surface method of the present invention is based on an experimental design (D0E) using process parameters of a gastric, e.g., general etch process. Table 1 provides exemplary process parameters that can be measured and faded from device 104. Table 1 shows the valid range for each process parameter. For example, the process parameters can be adjusted to high (+), medium (〇), or low (-) levels, based on the Beοχ-Behnken design, with a factor of two per level. B〇x-Behnken is designed as an independent secondary design' and does not contain inline factorial or fractional matrices. In this design, the combination is at the midpoint and center of the edge of the process space. Table 1

Page 13 1640043 V'Invention Description (10) Process parameters (factor) P Pre-interval, run between electrodes +, ft ' — He, back He pressure: + ' 0, -HP Ρ, process house force ρ +, 0 , Q, total flow rate # + ' 〇 , , flow rate ratio ρ + , 0 , , lower electrode RF power p +, 0, - segment D, upper electrode RF power p + ' 0, — T, wafer holder temperature p +, 0, -

The technique is based on systematically changing the independent variable level. Careful analysis of the data provides invaluable information on how the input variables affect the response and can significantly improve the product and process. ... use a screening design to determine the number of reasonable runs in order to develop a reaction surface based on the number of relevant factors. The following is provided as an example of determining the number of reasonable operations: for a three-level, four-factor efficiency, considering the possible combinations of all factors, the number of operations required is required. Therefore, the reaction surface based on more than four process parameters determines the primary interaction and the reaction surface using a partial combined number of runs.

In the case of Bayes, it is possible to construct a reaction surface for etching uniformity, for example, collecting data on uniformity from an etching tool at intervals of 100 msec,

The power emitted by Vdc and the first three RF harmonics (Wi, f3). The l〇〇mSec interval is chosen to match the general servo response time of this system. Systems with longer response times can have longer sampling intervals; with shorter servos

1264043 V. INSTRUCTIONS (11) The system of reaction time requires a short sampling interval. Each value of the normalized data set is normalized by subtracting the standard deviation of the measured parameters by the following equation: (Vt — v ) /v , complex. The measured value of the regular t, Vm is the average value of the operation 30 seconds m ^ / is the time. For example, if the average Vdc measurement is in the total of 3 〇: quasi- and the standard deviation of the measurements is 6,5 volt m Peach - 234) / 6.5. From this data, the surface of the reaction surface can be established: - the secondary surface can produce acceptable results, such as can be connected; = the result is stored in the database for subsequent processing of the process 2 Figure 7A is a process BRIEF DESCRIPTION OF THE DRAWINGS A method of reacting a surface is illustrated in accordance with an embodiment of the present invention. In Fig. 7A, the process starts from P7〇2 and continues to p7〇4 and ^70. At Ρ70 4 and Ρ7 0 6, measure the characteristic characteristics of the tool. Characteristic measurements include, for example, critical dimension measurements (CD), peak-to-peak RF voltage (Vpp), and self-developed DC offset (VDC) and n + i RF harmonic measurements X to Xn. The process continues until Ρ708. The characteristics of Ρ708, Ρ704 and Ρ706 and the beat wave measurement are compared with the reaction surface model. This portion first completes the 'optimal reaction surface 722' by the decision of the optimal reaction surface 722, for example, the set reaction 720 of Figure 7A. Next, it was decided to measure the difference between the reaction 7 2 〇 and the optimal reaction surface 722. This gap is used to adjust the operating parameters to move the measurement reaction 72 朝向 toward the optimal reaction surface 722. For example, by comparing the surface model 7 2 2 closest to the measurement reaction surface 72 2, the measurement reaction 72 2 is compared with the reaction surface model 7 2 2 . Thereafter, the process parameters corresponding to the closest surface model 7 2 2 are used to generate the measurement reaction 7 2 0

Page 15 1264043 V. The difference between the process parameters of the invention (12) is based on the adjustment of the process parameters. Characteristics and harmonic measurement trends, for example, CD, I, I, χ, ^, &,···', can determine the complex relationship between the characteristic measurement and the process parameters in the following ways: x two f (gap, He , P, Q, %Q, RFb, RFf, τ);

Xi= g(gap, He, P, Q, %Q, RFb, RF:, T); x2 = h(gap, He, P, Q, RFfa, 仏, T); T) T) Τ) Τ ); measure the characteristic RFt ' RFt • RFt RFt xn = m (gap, He, P, Q, %q, RF VII and CD = P (gap, He, P, Q, %q, Rf ' vpp = q(gap, He, P, Q, rf VDC = r(gap, He, P, Q, %Q, RFb ' Therefore, at p 7 1 〇, it is possible to approach the surface model 722 by the above 720 The process parameters make it necessary to measure the back-off characteristics to the optimal surface. In particular, to confirm the drive or gather at P712, it is possible to determine the existing P710 = step. The required adjustment is returned to Ρ7〇 4 ^ If the process parameters meet the expected results, proceed to the next set of measurements. Similar to Figure 7, the reaction surface method, and the expected results, such as the previously predicted process table = whether the parameters meet the surname tool is based on the existing process requirements. For example, ^. Expected result of the etch tool or <1°/. Minor change damage (shading d uniformity + /-3% or 99. 7% yield parameter can not meet the expected result, the process is a ^age), etc. Etc. If the process is controlled by advanced processes (APC) system call dinner continued until P714, wherein the desired adjustment of the back 〇4 P7 manufactured by take off clothes parameters according to this process will be returned to the process P7 4 to 04 into the network by Shen and method of the present invention

1264043 V. The system of invention (13) is used together to optimize the overall process characteristics. Neural networks are related to a large number of processors operating in parallel, each with its knowledge of and access to the data in its memory. The neural network will be circulated or enter a large amount of data and rules regarding the relationship of the data. Figure 9 ^ A neural network model 900. In the Figure 9, the input layer 904 receives the value from the input 9〇2. Each neural crest of the rim 4 has its weight value. Each neuron also has a second (10) threshold. If the sum of all weighted start inputs is greater than the critical value, the filament will start. r, , and in the network, the nodes themselves are connected to each other when the node is exposed to the pattern in the confirmation material. From this exposure, the network can learn from the traditional computing program area only by pressing a fixed continuous command. As shown in Figure 9, input 902 to input layer 9〇4 includes RF input power. ΙΝ, bias voltage Vdc, including harmonic measurement 乂, χι, χ2·% characteristic parameter ^ρ5 and time interval reading, other input 902 in (ts2Q-ts28) amount = variable number of nodes set intermediate or hidden layer 9〇6. Hidden layer 9〇6 ^ : Most of the work of the road. The output layer 9〇8 has a complex node, and as shown in the figure: No, it produces such as process OK and RFapplied. Each node in the hidden layer 9〇6 The output connection of the full input layer 9 04. Based on all the input of the information learned in the hidden node, the interaction within the network learning model is allowed to perform weighted summation for each hidden node and each round of outgoing nodes. A summation value is used to convert each sum before advancing to the next hidden layer or rounding layer 9〇 8. Establishing a neural network to simulate the process, which process parameters can be confirmed

Page 17 1264043 V. Description of invention (14): Measurement of complex characteristics of temple collection. The uniformity of the sentence, from the eclipse and the rectification (4) general servo response time in the interval of 100mSec. A time interval with a longer reaction time; then, with a shorter feeding response = 糸 ^ requires a shorter sampling interval. Each data of the data set is normalized by subtracting the average value and then dividing the standard deviation of 4 numbers. The normalized value is confirmed by the following I program: (Vt_vj/Vsd, where Vt is the measurement of time t, vm is the average value of the run 30 seconds and Vsd is the standard deviation of the measurement. For example, if the average value of Vdc is measured In each time interval, the total is 3 〇 234 volts, and the standard deviation of the measured value is β·5 v 〇its, then the normalized value is (Vf 234) / 6.5. The number of processes that are statistically significant After that, for example, 10〇, the characteristic solution analysis data of the covariance matrix is calculated for the time interval of each process. Using the actual data, an example of the feature solution for calculating the covariance matrix is as follows:

Co var(5) 0.222 0.064 0.064 0.137 - 0,1 -0107 - 0.013 -0.014 ~l_128xlCT3 A 1.929x10 Counting and finishing feature values: -0.1 -0.013 - l_128xlCT3 -0.107 - 0.014 - 1.929xl0_3

0.097 0.013 1.565xlCT3 0.013 1.653x10-3 2·055χ1(Τ4 1.565χ10~3 2.055Χ10-4 2.737 χ!0~5 V = r e v e r s e ( s o r t ( e i g e n v a 1 s ( S )))

Page 18 1640043 V. Description of invention (15) '0.34' 74.461 - 〇·1Π 25.539 F = 0 Pet = 1.821x10 A 14 0 Γ·1〇〇2.759x10—15 —0 A 9.291xlCT15 In this example, It is proved that the 3rd, 4th, and 5th harmonics are independent of the important measured values, so there is no need to consider them during this time interval. Each time interval is tested and all relevant choppings are selected for the generation of the reaction surface. Further time interval analysis of this sample data shows that the 2nd and 3rd spectral waves are important, while the 4th and 5th harmonics are negligible.

During the "training" of the network, the network repeatedly displays observations from available data, all of which are related to the problem to be solved. For example, a reverse-transfer (BP) network can be learned by example, that is, for example: the provided-group includes some input examples and the known correct output is not determined. The output of the white-fan case-based output is more than the way back, and the demonstration case case is repeated until the problem is solved for this network. This network (in the beginning, the output is calculated and the mean square error is calculated and each The weighting of one layer is 'calculating the weighted change to reduce the whole process, and then the error value of the returning body is lower than some pre-notes that the network will generate the machine value. There will be a small amount of error, and the first will be no longer refined: its bond. This error has changed after this round. Signal. Case and other boundary values. It is true that the difference between the present value and the existing value will be targeted at each other. At this time, the net ideal of learning is a good knowledge.

Page 19 1264043 V. Description of invention (16) Instead, it approaches the ideal function asymptotically. Figure 11 illustrates an example of modifying neural connection weighting in accordance with the present invention. In FIG. 11, (I, 12), (Heart, H2), and ((^, 02) are respectively designed as an input hidden layer output 1104 and an output layer output of a (2, 2, 2) inverted transfer network. The output of hidden nodes 1 and 2 is (1) (2) (3) (4) J-1 and Μ where the output layer output is 7S-1 and (5) O^sgmij^HX,) or, using (1) And (2) (6) and

Page 20 1264043

Based on the above equations, it is known that a specific set of input meters is possible. In this training paradigm, for known input tolerances, is the mean squared error (MSE) between the output and the expected output. MSE is the expected _ ^ actual average of the squared difference between the results. Because we are interested in the shape of the line and not the exact MSE function, there is no need to divide the output and the minimum algorithm will still find the correct minimum. Therefore, the number of errors can be formally expressed as: °, difference E-hD^〇,) 2 /, ^ (8) or, using (6) and (7), S = ΚΣ chat r qx

M-l M \ ^ J where Dk is the expected output of kth. As an example, suppose we have outputs 〇·75 and 0·05 with the expected turnout 及 9 and 〇·1. (Real) MSE is now ((〇· 9 - 0· 75)2 + (0· 1-〇· 05) · )〆2 ' is equal to 〇· 〇1 2 5 (Note: In the inverse transfer algorithm, we No need to remove the upper Ν). Obviously, for any known training paradigm, this value is only a weighting function of the network. Therefore, to reduce the error, we can try to move the lowest point on the surface. In order to find this, you must count the nose and each network.

Page 21 1264043 V. INSTRUCTIONS (18) The gradient of the weight function associated with the weighting. Then each weight is moved slightly in the opposite direction of the gradient—if the surface is skewed in a particular direction, the weighting is adjusted so that this point on the error surface moves down. Since the differentiation of the sigmoid function can be expressed by the function itself, the calculation of the gradient is quite simple. (1 + 〇 " (10) The gradient is defined as the direction of the multivariate function for the partial differentiation of each variable. Since the error is a function of the network output, therefore, for the connection-weighted = per-output node, a set of partial differentials must first be calculated. This result is household, because when calculating the partial differential, in addition to the variables of interest The second round remains unchanged. Therefore, the linear term of the only one is the partial differential appearing at the 1 different output, and the coefficient is rounded off. This equation is (11) Now the gradient of the computable error function

Note:): =~2(Α - °») isgm(S^) · dS (12) One 2(3⁄4 -〇„)((1-落桃〇^))落^^))

1264043

The above formula - - 〇 K) ((\- sgm (y)) sgm (S.)) The main public, liA ^ table is not for each. Multiply the negative gradient by the step size parameter (the quantity plus the network weighting vector attached to the current layer = the learning rate) and merge it, unless the hidden layer weighting also updates the otherwise ramp weight. However, this will tamper with the weighted update order of the hidden layer. Zhi Ai will not be born. Obviously, the error of output will also be fragrant. However, this relationship is more complex. Get the effect of layer weighting. The weighting is regarded as a constant rather than the weight of the barrier layer. However, the weighting function of the output layer is hidden in the hidden layer, and the output is only the input node and becomes the middle layer node. Here again: / There is a so-called LM , L is 』嘈即点) This relationship can be expressed as ·· as 2 ml= (0^ 5^^))^(^))Σ sXnh (13) Hidden layer weighting can use the same layer as the rounded layer Weighted update. In the neural network, for a piece of training w Μ _ Λ output ring. Τ奵仵 Training Bellow to complete the training cycle = Note that the input layer of Figure u is really just in the network, can be =:: force: the right needs to be modified. However, in more - it is like, like. - Once the calculation is modified, all the outputs must be updated. In addition, the above is a hypothetical one (2, 2, 2) inverted transmission network. From a comparison

1264043 V. INSTRUCTIONS (20) Some::: The difference between the old u and the mathematics is only a long sum. The road model is an electricity-map, and it is clear that in the operation mode, the neural network starts from m2 and continues to be closed. The measurement quantity = 〇, 4v, ^ set ΐ characteristic measurement: This * number characteristic measurement includes special RF1 phenotype, *dan, P|* Dc. in? 806, the number of known tools on the process tool = wave 1 test, this process tool can be any semiconductor manufacturing tool, packaged. = limit, the electro-convenient button engraving tool, photoresist tool or pattern transfer machine "" These measurements can be used as input to the wheel entry layer g 〇 4. The characteristic measurement of /, H q ΛΡ8〇1 P8 04 and the 11-wave measurement of Ρ8〇δ are set in the input layer 9〇4 of the neural network model shown in Fig. 9. At 1; 810, using the neural network 900 to confirm the output 91 〇 predictive characteristic measurement 疋 no results in a successful process. This process continues until ρ8ι 2 . At P812, confirm that the predicted results of the process tool are acceptable. If the prediction is unacceptable, you will need to make changes. This process continues until P 81 4.

At P814, the process continues to P816 via an advanced process control (apc). If the predicted result of the process tool is acceptable, . This process continues to P816. Whole process parameters. No need to change

Confirm that the time window has expired at P8 1 6 '. The time window can be extended for a whole operation or for a predetermined period of time. If the time window has expired, the 2 process continues to P8 18.

1264043 Description of invention (21) On P 81 8, based on the newly collected data, the neural network model is updated and the process continues until P8〇4. The new network weighting value is calculated by multiplying the negative gradient by the step size parameter (i.e., the learning rate) and adding the resultant vector to the current network weighting vector. If %Between® does not expire, the process returns to p 8 〇 4. Figure 10 illustrates a system in accordance with the present invention. The system 1 includes a process tool 1〇〇2 connected to the Advanced Process Control (APC) Server 1 004. The characteristic measurement 1010 data of the process tool is sent directly to the material collection center 1 008 of the Apc server. The characteristic measurement from the data collection center 1〇〇8 is sent to a module 1 0 06, and the module 1 006 utilizes the reaction surface method (RSM) or neural network (NN) as shown in FIG. 8 and FIG. )law. The control computer 1〇12 collects the =^ event and controls the process parameters based on the recommendations of the APC server 1 00 4 sigh point and maintenance counter. Although the foregoing description of the embodiments of the present invention provides illustrations and illustrations, they are only in the form disclosed above. Embodiments of the invention may be described with modifications and variations. For example, the various features of the invention described herein may be combined in a single embodiment. The various features of the invention described in the single embodiments are also applicable to any suitable In the sub-combination, it is not limited by the above illustrations and descriptions. In fact, the target range for this month is limited by the scope of the patent application and its equivalent. 1264043 BRIEF DESCRIPTION OF THE DRAWINGS [Brief Description of the Drawings] Other features of the invention will be referred to the following figures and represent similar parts of the present invention via the present and several views < component numbers, wherein Figure 1 is illustrated in accordance with an embodiment of the present invention. A plasma processing system; FIG. 2 is a schematic diagram of a plasma processing system according to another embodiment of the present invention. FIG. 3 is another embodiment of the present invention. FIG. 4 is a diagram according to another embodiment of the present invention. FIG. Another embodiment illustrates a plasma processing system. A plasma processing system is described. A plasma processing system is a simplified diagram illustrating RF signals and harmonic measurements. FIG. 7A is a flow chart according to the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS A reactive surface method is illustrated; FIG. 7B illustrates an exemplary reaction surface model in accordance with an embodiment of the present invention; FIG. 8 is a flow chart illustrating a network model in accordance with an embodiment of the present invention; FIG. Inventive embodiments, a neural network model is illustrated; FIG. 10 illustrates a complete system in accordance with an embodiment of the present invention; and FIG. 11 illustrates a neural network module in accordance with an embodiment of the present invention. . Description of component symbols: 1 〇 〇, 2 0 〇, 3 0 0, 4 0 0, 5 0 0 : plasma processing system

1264043 Schematic description 102 Process chamber 104 Process parameter device 106 Characteristic measurement device 108 Controller 202 Process area 212 Substrate holder 206 Gas injection system 208 Vacuum pump system 204 Substrate 214 Electrostatic clamping system 216 Device 218 Back gas system 22 0. 40 4, 52 0 : RF generator 2 2 2, 5 3 0 : impedance matching network 210 process gas 224 chamber pressure device 230 metric tool 304 mechanical or electronic dc rotating magnetic field system 402 upper electrode 510 induction coil P7 0 2, P 8 0 2 ··Start P704, 1010, P804: Characteristic measurement P7 0 6、P 8 0 6 : Measurement η harmonic Ρ7 0 8 ··Compat the characteristics and harmonic measurement with the reaction surface

Page 27 1264043 Simple description of the diagram P710: Predict internal process parameters P712, P812: Do you need to change the process?

P714, P814: Adjust APC 7 2 0 : Measurement reaction 722 : Optimal reaction surface P8 08 : Set measurement in the model P810 : Predict internal process success P8 16 : Is the time window expired? P818: Update model weighting 900: Neural network model 90 2: Input 9 0 4: Input layer P3-P5: Characteristic parameters X, Xi, χ2 "·χη: Harmonic measurement 9 0 6 : Hidden layer 9 0 8 : Output layer 910 : Output 1 0 0 0 : System 1 0 0 2 : Process tool 1 0 0 4 : Advanced process control server 1 0 0 8 : Data collection center 1 0 0 6 : Module

Page 28

Claims (1)

1264043 Case No. 92126522 Amendment m - Year 2006, Patent Application 1. A monitoring and analysis system for multivariable plasma process data, including: A process chamber to help generate plasma in the process area; a first mechanism for measuring at least one process parameter; a second mechanism for measuring at least one characteristic parameter; and a device for analyzing data from the first institution and the second mechanism based on a mathematical model to improve system characteristics; a device, based on the characteristic parameters applied to the mathematical model, adjusts at least one process parameter, wherein The characteristic parameters include at least the engraving rate, deposition rate, surname selectivity, etching critical dimension, etching pattern anisotropy, plasma density, ion energy, residual mask film thickness, reticle pattern critical size, self-induction DC substrate bias and peak-to-peak RF substrate biasing; the process parameters include at least process pressure, process gas flow rate, upper electrode RF power, process gas, lower electrode RF power, substrate temperature, electrode size, film material adhesion One of the properties of the film, the surface tension of the film material, the exposure time and the depth of focus; and | The mathematical model includes a reaction surface model or a neural network model. 2. A monitoring and analysis system for multivariable plasma process data as claimed in claim 1 wherein the process chamber is an etch chamber. 3. A monitoring and analysis system for multivariable plasma process data according to item 1 of the patent application, wherein the process chamber comprises: a substrate, which is placed in an appropriate position for the process; Page 29 1264043 _ Case No. 92126522_年月日日__ VI. Patent Application Scope A gas injection mechanism introduces process gas into the process area; and a vacuum pump mechanism, wherein at least one characteristic parameter includes characteristics of the process gas and pressure in the process chamber. 4. A monitoring and analysis system for multi-variable plasma process data according to item 3 of the patent application, wherein the substrate is a wafer or a liquid crystal display. 5. The monitoring and analysis system for multi-variable plasma process data of claim 3, further comprising a device for controlling the substrate temperature, wherein the at least one process parameter comprises measuring the substrate temperature. 6. The monitoring and analysis system for multivariable plasma process data according to claim 1 of the patent scope, further comprising: an RF generator for transferring energy to the process chamber to generate plasma in the process area; a gas injection mechanism , connecting the process chamber; and a pressure measuring device connecting the process chamber and measuring the pressure in the process chamber, The at least one process parameter includes characteristics of a process gas introduced into the process area via the gas injection mechanism. 7. The monitoring and analysis system for multivariable plasma process data of claim 1 includes a mechanical or an electronic rotating magnetic field mechanism disposed in the process chamber. Page 30 1264043 Bottom "彳_案号92126522_年月日日__ VI. Patent application scope 8. A method for monitoring and analyzing multivariable plasma process data, including: measuring the characteristic parameters of the plasma process; Measure some harmonic parameters of the plasma process; confirm a mathematical model of the plasma process; apply the measured characteristic parameters and harmonic parameters to the mathematical model of the plasma process; use the mathematical parameters of the characteristic parameters, harmonic parameters and applications Based on the prediction of process parameters; and The process parameters are adjusted according to a mathematical model of the application, wherein the characteristic parameters include at least an etch rate, a deposition rate, an etch selectivity, an etch key dimension, an etch pattern anisotropy, a plasma density, an ion energy, and an etch mask film thickness. The mask size, the self-inductive DC substrate bias, and the peak-to-peak RF substrate bias; the process parameters include at least process pressure, process gas flow rate, upper electrode RF power, process gas, lower electrode RF power, One of substrate temperature, electrode size, film material viscosity, film material surface tension, bare time, and depth of focus; The mathematical model includes a reactive surface model or a neural network model. 9. For the monitoring and analysis method of the multi-variable plasma process data of the application of the patent scope, item 8, in which the internal process parameters are predicted. Page 31