Characterization of Low-Temperature SU-8 Photoresist Processing for MEMS Applications

Transcription

1 Characterization of Low-Temperature SU-8 Photoresist Processing for MEMS Appications Sang Jeen Hong 1, Seungkeun Choi 2, Yoonsu Choi 3, Mark Aen 4*, and Gary S. May 5 Schoo of Eectrica and Computer Engineering Georgia Institute of Technoogy Atanta, GA gte625q@mai.gatech.edu 1, gtg737d@mai.gatech.edu 2, gte810q@prism.gatech.edu 3, mark.aen@ece.gatech.edu 4, and gary.may@ece.gatech.edu 5 * Office: (404) , FAX: (404) Abstract In this paper, negative SU-8 photoresist processed at owtemperature has been characterized in terms of deamination. Based on a 3 3 factoria designed experiment, 27 sampes are fabricated and the degree of deamination is measured. In addition, nine sampes are fabricated for the purpose of verification. Empoying a neura network modeing technique, a process mode is estabished, and response surfaces are generated to investigate the degree of deamination associated with three process parameters: post exposure bake (PEB) temperature, PEB time, and exposure energy. From the response surfaces generated, two significant parameters associated with deamination are identified, and their effects on deamination are anayzed. The higher the post exposure bake (PEB) temperature at a fixed PEB time, the more deamination occurred. In addition, the higher the dose of exposure energy, the ower the temperature at which the deamination begins and the arger the degree of deamination. The resuts identify acceptabe ranges of the three process variabes to avoid the deamination of SU-8 fim, which in turn might ead to potentia defects in MEMS device fabrication. INTRODUCTION Among numerous poymers being used in the deveopment and fabrication of MEMS devices, the popuarity of SU-8 has increased because of its mechanica stabiity, biocompatibiity, and suitabiity for fabricating high aspect ratio features [1]-[3]. SU-8 is a negative near UV photoresist designed to produce uniform thick fims in a singe spin-coating step. Vertica sidewas and high aspect ratio features resut from the product of photochemica and therma cationic processes. The exposed and subsequenty cross-inked portions of the fim are rendered insoube to iquid deveopers. SU-8 has ow optica absorption, thus aowing the patterning of very thick fims. However, standard recipes suggested for SU-8 processing have proven in practice to be very sensitive to process conditions, and the parameter vaues described in the iterature have varied over a wide range [3][4]. For these reasons, previous efforts at characterization and optimization of SU-8 process have empoyed statistica designed experiment and Taguchi method [5][6], and the resuts suggested optima processing parameters for various thicknesses of SU-8 fim to acquire better resoution of the patterned image. Despite the advantages of SU-8, previous studies reported deamination of the SU-8 microstructures and fims. The faiure of microposts in [1] was reported mainy due to the interfacia fracture at the base, as no faiure occurred in the micropost bodies. Mechanica deamination of SU-8 was aso observed in MEMS drug deivery devices [2]. Brunet et a. aso reported the deamination of SU-8 microstructures during deveoping in the deveopment of high aspect ratio magnetic cois. Thick ayers of SU-8 experience more stress, and the structures tended to deaminate more quicky than the thin ayers [3]. For muti-ayer MEMS fabrication, which is currenty under investigation at the Georgia Institute of Technoogy, deamination associated with stress has ed to concerns about defects in SU-8 fabricated microstructures. In an effort to reduce the amount of stress on SU-8 microstructures, a ow-temperature process with proonged bake time was investigated. By tria and error, deamination was reduced. However, it is necessary to perform a more systematic characterization experiment to carify the reationship between process parameters and identify suitabe ranges for process variabes to ensure fabrication without deamination. Therefore, this paper investigates the variation of ow-temperature SU-8 processing with the utimate goa of minimizing deamination, using response surfaces generated from neura network modes. The paper is organized as foows: Section 2 describes Tabe 1. Process parameters and their ranges Step Parameters Abbrev. Ranges Units Exposure Energy ENERGY mj/cm 2 PEB Temp. PEB TMP C Time PEB TIME min. (Note: Parameters in bod are three parameters corresponding to 3 3 factoria design in this study.)

2 how the experiment was performed for 100µm thick fims of SU-8. Section 3 provides background information of neura network modeing. Resuts wi be provided in Section 4, foowed by a summary and discussion of future work in the fina section. EXPERIMENT Statistica designed experiment If a process has more than a very sma number of steps whose possibe vaues have a arge range, the number of experiments needed for process characterization can be prohibitivey arge. In addition, the roe of each step in determining the fina outcome is generay not cear. The traditiona method of coecting arge quantities of data by hoding each factor constant in turn unti a possibiities have been tested is an approach that quicky becomes impossibe as the number of factors increases. Statistica experimenta design is a systematic and efficient aternative methodoogy for characterization and modeing using a reativey sma number of experiments [7]. In this study, five parameters at three eves each were initiay considered. The parameters were soft baking temperature/time, exposure energy, and post exposure baking (PEB) temperature/time. Since cross-inking takes pace after exposure, the variabes in the soft baking step were ater omitted. Instead, the soft baking step was performed in a consistent manner for a sampes. Deveop time after PEB pays an important roe in adhesion to the surface. Extended deveoping time may increase the chance of deamination of the exposed area from the substrate, but insufficient time may negativey affect on the ithographic resoution [8]. In this research, the deveoping time that aowed decent ithographic resoution was consistenty appied in order to avoid any additiona compexity in characterization. The process variabes and their ranges L =L Figure 1. Microscopic pictures of deaminated structure that show different degrees of deamination: taken with Oympus Vanox microscope. Figure 2. An exampe picture of bar patterns used for the measurements of the degree of deamination (DoD). appear in Tabe 1. A 3 3 factoria design requiring 27 experiments was conducted, and this design was further augmented with nine randomy seected experiments for mode verification purposes. Sampe fabrication and measurement SU-8 was spin coated on 4 siicon wafers to a thickness of 100 um, and the sampes were soft baked at 70 C on a hot pate to drive off sovents. Based on the designed experiment, a possibe orthogona combinations of the three parameters were appied. A sampes were deveoped for a fixed time, and the amount of deamination was measured. The degree of deamination (DoD) was quantified by the foowing expression: ( L ) DoD = 100 (%) (1) L where L is the ength of the origina bar pattern, and is the ength of bar pattern that remained on substrate (see Figure 1). To minimize the measurement error, was averaged over eight bar patterns in one ocation as shown in Figure 2. NEURAL NETWORKS Neura networks have become usefu toos in process modeing and demonstrated the capabiity of earning compex reationships between groups of reated parameters [9]. A neura network is a structured interconnection of computationa nodes caed neurons that contribute to parae computation in a manner simiar to the human brain. The interconnection of neurons estabishes knowedge that is acquired by the network through a earning process, and that knowedge is stored in the form of inter-neuron connection strengths known as weights. Each neuron contains the weighted sum of its inputs fitered by a sigmoida squashing function, providing neura networks with the abiity to generaize

3 Input x 1 x 2 x 3 x i () wki Hidden h 1 h 2 h 3 h 4 h k (L) w ok Output y 1 y 2 y o Figure 3. An iustration of a mutiayer perceptron neura network. with an added degree of freedom that is not avaiabe in statistica regression techniques [10]. The earning agorithm used in this study is the error back-propagation (BP) agorithm. A typica backpropagation neura network structure is depicted in Figure 3. In the BP earning agorithm, a singe iteration consists of two parts: a forward and a backward propagation. In the forward propagation, the outputs from the i th ayer are weighted and summed, and the weighted sum is fitered through a sigmoid function. The outputs of neurons in th ayer become inputs to the neurons in the next ayer k. The forward propagation is expained by the foowing equations: s n = i i= 1 w o i (2) o = 1 1+ exp( s ) (3) where s are weighted sum input to neurons in ayer, o i is output from neurons in ayer i, n i is the number of neurons in ith and w is the weights connecting. In the same manner, y k, s k, o i, and s i aso can be derived. In backward propagation, weights are updated in the direction that minimizes an error function defined by: 1 2 Ek = ( yˆ k yk ) (4) 2 where k is a target, and y k is the actua output vaue of the ast ayer k. The generaized deta-rue based on gradient descent approach is appied to minimize the error function. The weights are initiay randomized, and forward propagation is performed. Once the outputs of the ast ayer are cacuated, weights are updated by the deta for each node cacuated from the output ayer (ayer k) and backpropagated to the input ayer (ayer i). The generaized deta rue is: w ( n + 1) = ηδ ( n + 1) oi ( n + 1) (5) w ( n + ) = η w ( n + 1) + α w ( n) 1, (6) where n is the number of iteration, is the earning rate and is the momentum. The earning rate ( ) is a constant that represents the rate at which a weight wi be changed aong its sope to the minimum error. The momentum ( ) is a constant that incudes a portion of the previous weight change to the current weights. Utiizing ObOrNNs [11], a custom neura network simuation package deveoped by the Inteigent Semiconductor Manufacturing Group at Georgia Tech, neura network based response surface modes of the SU-8 fabrication process were derived. Initiay, neura networks were trained with the data generated from the 3 3 factoria designed experiments, and these modes were verified with the data set that was not previousy introduced to the networks during training. Inputs to the networks are three parameters of interest (exposure energy, post-bake temperature, and post-bake time), and output of the networks is degree of deamination. Hidden neurons Figure 4. Performance evauation of neura process mode. Straight ine represents 100% accuracy. Figure 5. Response surface pot: fixed exposure energy at 440 mj/cm 2.

4 Region2 Region1 Figure 6. Response surface pot: fixed exposure energy at 580 mj/cm 2. (neurons in the midde ayers) extract noninear features from the data, and severa networks with different numbers of hidden neurons were constructed and tested. The average RMS error in training was 2.57%, and that in testing was 4.87%. Mode performance is depicted graphicay in Figure 4. RESULT AND DISCUSSION Once the neura process mode was estabished, response surfaces were generated to iustrate the reationships between any two process parameters and degree deamination. Any two seected variabes of three were simutaneousy varied within their ranges in Tabe 1, whie the remaining parameter was set to at its mid-range eve. The predictions of the neura process mode were then graphed using 3-D contour pots (see Figures 5-8). Figure 5 iustrates the effect of PEB time and temperature on the degree of deamination when the exposure energy is fixed at 440 mj/cm 2. The energies are setup at 580 mj/cm 2 and 720 mj/cm 2 in Figures 6 and 7 respectivey. In this experiment, PEB temperature appeared to be the most critica factor affecting deamination under the condition that the dose of energy is fixed. To ensure the cross-inking of SU-8 after exposure, sufficient PEB time is Figure 7. Response surface pot: fixed exposure energy at 720 mj/cm 2. Figure 8. Response surface pot: fixed post exposure baking temperature at 70 C. required at the proper temperature. It is observed, however, that the higher the temperature, the arger the degree of deamination. In addition, PEB time aso somewhat affects the degree of deamination at a given temperature. The shorter the PEB time, the ess cross-inking, increasing the degree of deamination. The high degree of deamination at temperatures above C is primariy due to the coefficient of therma expansion (CTE) mismatch between SU-8 and the siicon wafer with native oxidation. As the exposure energy increases, the degree of deamination increases, whie the temperature at which the deamination starts to occur decreases. Higher exposure energy tends to increase cross-inking of the poymer in the exposed area, and consequenty, this increases fim stress due to voume changes. The effect of exposure energy on deamination can be observed ceary in Figure 8. By setting the temperature at a certain eve, the degree of deamination caused by CTE mismatch can be negected. At a fixed PEB temperature of 70 C and a reasonabe PEB time of minutes, the degree of deamination increases with exposure energy. Region 1, where the PEB time is ess than 30 minutes and the exposure energy is ess than 520 mj/cm 2, showed some degree of deamination due to incompete cross-inking. The Region 2, where the PEB time is onger than 25 minutes and the exposure energy is arger than 650 mj/cm 2, shows approximatey 5% deamination due to the stress induced from voume changes. CONCLUSION To summarize, two significant parameters associated with SU-8 deamination were investigated, and their effects on deamination were determined from the response surfaces generated from neura network modes. Higher PEB temperatures at a fixed PEB time resut in more deamination due to CTE mismatch. In addition, a greater dose of exposure energy owers the temperature at which deamination starts to occur and increases the degree of deamination. The response surfaces generated aso identify

5 suitabe ranges of process conditions that avoid SU-8 deamination, which can utimatey cause defects in MEMS devices. ACKNOWLEDGEMENTS Authors are gratefu to MicroChem for their materia support and the staff of Microeectronics Research Center in Georgia Institute of Technoogy. REFERENCES [1] H. Khoo, K. Liu, and F. Tseng, Mechanica strength and interfacia faiure anaysis of cantievered SU-8 microposts, J. Micromech. Microeng., vo. 13, pp , [2] G. Voskerician, M. Shive, R. Shawgo, H. Recum, J. Anderson, M. Cima, and R. Langer, Biocompatibiity and biofouing of MEMS drug deivery devices, Biomaterias, vo. 24, pp , [3] M. Brunet, T. O Donne, J. O Brien, P. McCoskey, and S. Mathuna, Thick photoresist deveopment for the fabrication of high aspect ratio magnetic cois, J. Micromech. Microeng., vo. 12, pp , [4] M. Despont, H. Lorenz, N. Fahrni, J. Brugger, P. Renaud, and P. Vettiger, High-aspect ratio utrathick, negative tone near-uv photoresist for MSMS appications, in Proc. IEEE, Tenth Anuua Internationa Workshop in Micro Eetro Mechanica Systems, Nagoya, Japan, Jan. 1997, pp [5] J. Zhang, K.L. Tan, and H.Q. Gong, Characterization of the poymerization of SU-8 photoresist and its appications in micro-eectro-mechanica systems, Poymer Testing, 20, pp , [6] B. Eyre, J. Bosiu, D. Wiberg, Taguchi optimization for the processing of Epon SU-8 resist, in Proc. IEEE, The eeventh Annua Internationa Workshop on Micro Eeectro Mechanica Systems, Jan. 1998, pp [7] G. Box, W. Hunter, and J. Hunter, Statistics for Experimenters, New York: Wiey, [8] A. Wong, D. Linton, Appication of SU-8 in fip chip bump micromachining for miimeter wave appications, in Proceedings of 3 rd Eectronics Packaging Technoogy Conference (EPTC 2000), Dec. 2000, pp [9] S. Hong, G. May and D. Park, Neura Network Modeing of Reactive Ion Etch Using Optica Emission Spectroscopy Data, IEEE Trans. Semi. Manufac. vo. 16, no. 4, pp. 1-11, Nov [10] C. Himme and G. May, Advantages of pasma etch modeing using neura networks over statistica techniques, IEEE Trans. Semi. Manufac., vo. 6, pp , May, [11] C. Davis, S. Hong, R. Setia, R. Pratap, T. Brown, B. Ku, G. Tripett, and G. May, A Java_Based Obected-Oriented Neura Network Simuator for Semiconductor Manufacturing Appications, submitted to the 8 th Word Muti-Conference on Systemics, Cybernetics and Informatics, Orando, FL, 2004.