• Tidak ada hasil yang ditemukan

Discussion and Conclusion

4.4 Discussion and Conclusion

BCP films but also on the lamella-forming BCP films. In the case of lamella-forming BCPs, a more accurate renormalization is possible using the Morse Calibration, and the simulation is easier to perform because its morphology is simpler than that of the cylindrical phase. Using the (χN)ODT equation of symmetric BCPs with respect to ¯N proposed by Morse and Matsen, [27, 28,100] it is possible to accurately compare the results with the theoretical prediction in an experimentally relevant regime, without using the TIM.

In chapter 3, the effect of the wetting condition on the ODT behavior of thin films was investigated using the SCFT adopting finite-ranged interaction, but in chapter 4, the effect of fluctuation on the ODT of thin film was investigated using the discrete chain model without the finite-ranged interaction. In order to simulate a polymeric system with a more realistic polymer model, it is necessary to incorporate the finite-ranged interaction with FTS. In chapter 2, the field transform using an arbitrary pair interaction potential u(r) has already been intro- duced, and it seems possible to develop an FTS adopting finite-ranged interaction. Since the renormalization methods known so far have been developed on the assumption of a continuous Gaussian chain model, it is also necessary to modify the renormalization method. One possible suggestion is to perform the RPA theory adopting finite-ranged interaction and try an RPA renormalization.

Unlike the SCFT which is a mean field theory, the FTS can account for the fluctuation effect, but it is practically difficult to perform FTS in a large polymeric system because the simulation time increases by a factor of thousands to tens of thousands of times compared to the SCFT calculation. In this study, the simulation was performed efficiently by parallelizing the calculation using GPUs, but further improvements are required in order to perform L- FTS of various systems which had been mostly studied by SCFT. One possible suggestion is to accelerate L-FTS using the machine learning method, which is rapidly evolving in recent years. In L-FTS, it is necessary to find the partial saddle point at every Langevin step. Since at least tens of thousands Langevin steps are performed in one L-FTS execution, data on the corresponding pressure field in a given exchange field can be collected as the training data.

Using this data and online learning, L-FTS could be accelerated tens of times by generating the pressure field directly through machine learning method during each simulation. This method is very versatile and could be applied to a variety of systems without prior training of a neutral network model.

Appendix A

Gaussian Integral

The one-dimensional Gaussian integral of a quadratic equation is evaluated as Z

−∞

dxexp −ax2/2 +Jx

= 2π

a 1/2

exp

J2/(2a)

, (A.1)

where a constanta >0 andJ are real. This Gaussian integral can be expanded to multivariable integral ofx= (x1, x2, . . . , xN)T

Z

dxexp −xTAx/2 +JTx

= (2π)N/2

(detA)1/2 exp JTA1J/2

, (A.2)

whereAis a symmetric and positive definiteN×N matrix, and a vector J= (J1, J2, . . . , JN)T is a real-valued vector. The functional integral version of Gaussian integral is

R Dfexp

−(1/2)R dxR

dx0f(x)A(x, x0)f(x0) +R

dxJ(x)f(x) R Dfexp

−(1/2)R dxR

dx0f(x)A(x, x0)f(x0)

= exp 1

2 Z

dx Z

dx.J(x)A1(x, x0)J(x0)

. (A.3)

The function A(x, x0) is also symmetric and positive definite, and A1(x, x0) is a functional inverse which satisfies

Z

dx00A(x, x00)A1(x00, x0) =δ(x−x0), (A.4) whereδ(x) is the Dirac delta distribution. [61]

Appendix B

GPU Optimization

Currently, the main trend of general-purpose computing on graphics processing units (GPGPU) is to use GPU hardware manufactured by NVIDIA and to develop programs on its platform, compute unified device architecture (CUDA). In our L-FTS study, GPU parallelization was implemented using CUDA, and this appendix is devoted to describe the optimization of GPU parallelization using CUDA.

The GPU hardware has tens of streaming multiprocessors (SMs), and each SM consists of many CUDA cores. For example, the GeForce RTX 3080 we use has 68 SMs and each SM has 128 CUDA cores; there are 8704 CUDA cores. Each CUDA core has relatively limited computing power compared to a single CPU core, but it shows high parallel performance using the large number of GPU cores. In order to reduce possible bottlenecks, the GPU has many types of memories and the most important ones are global and shared memories. The main task of GPU optimization is to reduce the bottlenecks caused by data transfer between memories and to maximize usage of GPU cores.

Delaney and Fredrickson first published a paper about polymeric field-based simulations on GPUs. [98] The computation employing pseudo-spectral method was parallelized using CUDA.

The calculations using the pseudo-spectral method are largely element-wise multiplications and FFT operations in between them. In their study, element-wise multiplications were calculated using CUDA kernels, and FFT operations were performed using FFT functions provided by the CUDA library. [87] In order to reduce the time loss in data transfer between the main memory and the global memory, all partition function steps were stored in the global memory, and only the calculated segment densities were transferred to the main memory. In addition, by storing the product of the partial partition function and the complementary partition function instead of storing two partition functions separately, the memory usage could be reduced by 50 percents.

With such techniques, they achieved a speedup of 30 in a simulation of 1283 grids compared to a single CPU calculation using double precision FFTW. [101]

Recently, Cheong et al. provides an SCFT implementation utilizing GPU as an open source.

[99] In their implementation, the reduction operations are also conducted in GPU to efficiently calculate the total partition function and the inner product of the Anderson mixing. The re- duction operation on GPU using shared memory is a very non-trivial task due to issues such as bank conflict. They use the most optimized reduction method suggested by the NVIDIA.

The FFT library provided by CUDA internally determine the number of threads and blocks depending on the simulation grid number. If the simulation grid number is small, it appears that all SMs are not utilized. In Cheong et al., the FFT parallelization efficiency in the case of a small grid number is doubled by mixing calculations of both partial partition functions and by using batching system of FFT operations. Instead of using the FFT batch process, we utilize the concurrent kernel execution. In this approach, two concurrent streams are created and each is responsible for one partial partition function. Compared to the implementation of Cheong et al., the performance is the same, but it is easier to upgrade for multiple GPU implementation.

Figure B1 explains the performance difference between each implementation as a time- line. The FFT parallelization efficiency is doubled by the batch process compared to the basic implementation. With batch process, the calculation time of two FFT operations equals the calculation time of one FFT operation. Similarly, in the case of using double streams, there is a performance gain in the FFT calculations. Note that this illustration is only valid when the simulation grids number is small.

Figure B1: Comparison of GPU implementations. (left) Implementation without batch process and streams. (middle) Two FFT operations performed using the batch process. (right) Each partial partition function is performed simultaneously in each stream.

The performance of GPU parallelization increases as the simulation grid number increases.

Since all partial partition steps are stored in the global memory, the maximum limit of grid number is about 1283due to limit of GPU memory space when using a typical graphics card such as GeForce RTX 3080. Delaney and Fredrickson proposed a method that transfers the partial

partition functions to the main memory as they are computed in order to reduce the GPU memory usage when the simulation grid number is large. This method doubles the execution time in the simulation of 1283 grid number due to the cost of transferring data and loss of parallelization in density calculation.

When using large grid numbers, we implemented another method that saves GPU memory to deal with this case. The GPU supports kernel execution and data transfer overlapping, which simultaneously executes kernel and transfers data between main and global memories. For this, we first create two streams. The first stream is responsible for the kernel execution, and the second stream is responsible for the data transfer. Then, we allocate memory space that can store two steps of each partial partition function in the global memory. For the first half of the memory space, the partial partition function calculated in the previous step is stored, and the other half is used to calculate the partial partition function in the next step, using the result of the previous step. These calculations are performed on the first stream. In the meantime, the second stream transfers the calculated partial partition function to the main memory. The two streams are synchronized at the end of each ∆s step. When calculation of one ∆s step is finished, the roles of the two memory spaces are exchanged and the above process is repeated.

Figure B2:Illustration of overlapping of kernel execution and data transfer. The yellow and green bars denote data transfer and kernel execution times, respectively. The data transfer can be com- pleted during the kernel execution when the simulation grid number is large enough.

By repeating this process, the calculation of the partial partial function q(r) is completed, and the complementary partial partition functionq(r) calculation begins. In this process, unlike the previous case, the result is not transferred to the main memory, butq(r) stored in the main memory is transferred to the global memory at every contour step by the second stream. At the same time, in the first stream, the calculation ofq(r) is finished and the density ofsNth segment is computed using the two partial partition functions in the global memory. We synchronize the streams at the end of each ∆s step during the computation of the complementary partial partition function. Figure B2 explains the efficiency of the computation with this overlapping method. If the simulation grid number is large enough, the data transfer can be completed during the kernel execution because the data transfer time is shorter than the kernel execution

time. This method is supposed to be somewhat slower than the previous double stream method when the grid size is small, but for a simulation in 1283 grid, this method is faster because of the reduction of the data transfer time, provided that there are no bottlenecks in the bandwidth between main and global memories.

References

[1] M. Rubinstein and R. H. Colby,Polymer Physics. Oxford University Press, Oxford, 2003.

1

[2] M. W. Matsen, The standard Gaussian model for block copolymer melts. Journal of Physics: Condensed Matter2001,14, R21. 1,61

[3] C. T. Black, K. W. Guarini, K. R. Milkove, S. M. Baker, T. P. Russell and M. T. Tuominen, Integration of self-assembled diblock copolymers for semiconductor capacitor fabrication.

Applied Physics Letters2001, 79, 409-411. 1

[4] R. A. Segalman, H. Yokoyama and E. J. Kramer, Graphoepitaxy of spherical domain block copolymer films.Advanced Materials2001, 13, 1152-1155. 1

[5] A. V. Ruzette and L. Leibler, Block copolymers in tomorrow’s plastics. Nature Materials 2005,4, 19-31. 1

[6] J. Bang, U. Jeong, D. Y. Ryu, T. P. Russell and C. J. Hawker, Block copolymer nanolithog- raphy: translation of molecular level control to nanoscale patterns. Advanced Materials 2009,21, 4769-4792. 1

[7] I. W. Hamley, Ordering in thin films of block copolymers: Fundamentals to potential ap- plications.Progress in Polymer Science2009, 34, 1161-1210. 1

[8] H. C. Kim, S. M. Park and W. D. Hinsberg, Block copolymer based nanostructures: ma- terials, processes, and applications to electronics. Chemical Reviews 2010, 110, 146-177.

1

[9] E. J. W. Crossland, P. Cunha, S. Scroggins, S. Moratti,O. Yurchenko, U. Steiner, M.

Hillmyer and S. Ludwigs, soft-etch msoporous hole-conducting block copolymer templates.

Acs Nano 2010,4, 962-966. 1

[10] S. J. Jeong, J. Y. Kim, B. H. Kim, H. S. Moon and S. O. Kim, Directed self-assembly of block copolymers for next generation nanolithography.Materials Today2013,16, 468-476.

1

REFERENCES

[11] K. Koo, H. Ahn, S. W. Kim, D. Y. Ryu and T. P. Russell, Directed self-assembly of block copolymers in the extreme: guiding microdomains from the small to the large.Soft Matter 2013,9, 9059-9071. 1

[12] D. E. Angelescu, J. H. Waller, D. H. Adamson, P. Deshpande, S. Y. Chou, R. A. Register and P. M. Chaikin, Macroscopic orientation of block copolymer cylinders in single-layer films by shearing.Advanced Materials2004,16, 1736-1740. 1

[13] I. Park, B. Lee, J. Ryu, K. Im, J. Yoon, M. Ree and T. Chang, Epitaxial phase transition of polystyrene-b-polyisoprene from hexagonally perforated layer to gyroid phase in thin film.Macromolecules 2005,38, 10532-10536. 1

[14] R. A. Segalman, Patterning with block copolymer thin films. Materials Science & Engi- neering R-Reports2005, 48, 191-226. 1

[15] K. Kim, S. Park, Y. Kim, J. Bang, C. Park and D. Y. Ryu, Optimized solvent vapor an- nealing for long-range perpendicular lamellae in PS-b-PMMA films.Macromolecules 2016, 49, 1722-1730. 1

[16] S. Park, Y. Kim, W. Lee, S. M. Hur and D. Y. Ryu, Gyroid structures in solvent annealed PS-b-PMMA films: controlled orientation by substrate interactions.Macromolecules 2017, 50, 5033-5041. 1

[17] Y. Kim, J. Mun, G. Yu and K. Char, Phase transition of block copolymer/homopolymer binary blends under 2D confinement. Macromolecular Research2017, 25, 656-661. 1 [18] F. S. Bates and G. H. Fredrickson, Block copolymer thermodynamics: theory and experi-

ment. Annual Review of Physical Chemistry1990, 41, 525-557. 2

[19] F. S. Bates, Polymer-polymer phase behavior. Science1991, 251, 898-905. 2

[20] A. V. G. Ruzette and A. M. Mayes, Simple free energy model for weakly interacting polymer blends.Macromolecules 2001, 34, 1894-1907. 2

[21] E. Helfand, Theory of inhomogeneous polymers: fundamentals of the gaussian random-walk model.The Journal of Chemical Physics 1975,62, 999-1005. 2,10

[22] E. Helfand and Z. R. WassermanDevelopment in Block Copolymersedited by I. Goodman, Elsevier, New York 1982. 2

[23] L. Leibler, Theory of microphase separation in block copolymers. Macromolecules 1980, 13, 1602-1617. 2,51

[24] A. N. Semenov, Contribution to the theory of microphase layering in block-copolymer

REFERENCES

[25] M. W. Matsen and M. Schick, Stable and unstable phases of a diblock copolymer melt.

Physical Review Letters1994,72, 2660–2663. 2,57

[26] G. H. Fredrickson and E. Helfand, Fluctuation effects in the theory of microphase separation in block copolymers.Journal of Chemical Physics1987, 87, 697-705. 2,50

[27] J. Glaser, P. Medapuram, T. M. Beardsley, M. W. Matsen and D. C. Morse, Universality of block copolymer melts.Physical Review Letters2014, 113, 068302. 3,66

[28] T. M. Beardsley and M. W. Matsen, Universality between experiment and simulation of a diblock copolymer melt.Physical Review Letters2016, 117, 217801. 3,66

[29] M. W. Matsen, Self-Consistent field theory for melts of low-molecular-weight diblock copolymer.Macromolecules 2012, 45, 8502-8509. 3,4,10,23,33

[30] C. Sinturel, F. S. Bates and M. A. Hillmyer, High χ-low N block polymers: how far can We Go?Acs Macro Letters 2015,4, 1044-1050. 3

[31] H. Nakanishi and M. E. Fisher, Critical point shifts in films Journal of Chemical Physics 1983,78, 3279-3293. 3

[32] A. Menelle, T. P. Russell, S. H. Anastasiadis, S. K. Satija and C. F. Majkrzak, Ordering of thin diblock copolymer films. Physical Review Letters1992,68, 67-70. 3

[33] P. Mansky, O. K. C. Tsui, T. P. Russell and Y. Gallot, Phase coherence and microphase separation transitions in diblock copolymer thin films.Macromolecules1999,32, 4832-4837.

3

[34] K. A. Orso and P. F. Green, Phase behavior of thin film blends of block copolymers and homopolymers: Changes in domain dimensions.Macromolecules1999, 32, 1087-1092. 3 [35] A. B. Croll, A. C. Shi and K. Dalnoki-Veress, Ordering of a lamella-forming fluid near an

interface. Physical Review E2009, 80, 051803. 3

[36] G. H. Fredrickson, Surface ordering phenomena in block copolymer melts.Macromolecules 1987,20, 2535-2542. 3

[37] M. Kikuchi and K. Binder, Microphase separation in thin films of the symmetric diblock- copolymer melt. Journal of Chemical Physics1994, 101, 3367. 3

[38] G. Srinivas, W. C. Swope and J. W. Pitera, Interfacial fluctuations of block copolymers: A coarse-grain molecular dynamics simulation study.Journal of Physical Chemistry B 2007, 111, 13734-13742. 3

REFERENCES

[39] C. Shin, H. Ahn, E. Kim, D. Y. Ryu, J. Huh, K. W. Kim and T. P. Russell, Transition behavior of block copolymer thin films on preferential surfaces. Macromolecules 2008, 41, 9140-9145. 3

[40] E. Kim, H. Ahn, D. Y. Ryu, J. Kim and J. Cho, Transition behavior of weakly interact- ing PS-b-PMMA films on preferential surfaces: a direct observation by GISAXS. Macro- molecules2009, 42, 8385-8391. 3,5

[41] M. R. Hammond, E. Cochran, G. H. Fredrickson and E. J. Kramer, Temperature de- pendence of order, disorder, and defects in laterally confined diblock copolymer cylinder monolayers.Macromolecules 2005, 38, 6575-6585. 3

[42] V. Mishra, G. H. Fredrickson and E. J. Kramer, Effect of film thickness and domain spacing on defect densities in directed self-assembly of cylindrical morphology block copolymers.

Acs Nano 2012,6, 2629-2641. 3

[43] S. Kim, W. Li, G. H. Fredrickson, C. J. Hawker and E. J. Kramer, Order-disorder tran- sition in thin films of horizontally-oriented cylinder-forming block copolymers: thermal fluctuations vs. preferential wetting. Soft Matter2016, 12, 5915-5925. 3,65

[44] V. Ganesan and G. H. Fredrickson, Field-theoretic polymer simulations. Europhysics Let- ters2001,55, 814-820. 4

[45] G H. Fredrickson, V. Ganesan and F. Drolet, Field-theoretic computer Simulation methods for polymers and complex fluids.Macromolecules 2002,55, 16-39. 4

[46] G H. Fredrickson, Computational field theory of polymers: opportunities and challenges Soft Matter2007, 3, 1329-1334. 4

[47] K. T. Delaney and G. H. Fredrickson, Recent developments in fully fluctuating field- theoretic simulations of polymer melts and Solutions. The Journal of Physical Chemistry B2016,120, 7615. 4,44,55

[48] Y. Kim, D. Yong, W. Lee, S. Jo, H. Ahn, J. U. Kim and D. Y. Ryu, Preferential wet- ting effects on order-to-disorder transition in polystyrene-b-poly(2-vinylpyridine) films: a reconsideration on thickness dependence.Macromolecules2018, 51, 8550-8560. 4,5,34,37 [49] Y. Kim, D. Yong, W. Lee, H. Ahn, J. H. Kim, J. U. Kim and D. Y. Ryu, Order-to-disorder transition of lamella-forming PS-b-P2VP films confined between the preferential surface and neutral substrate. Macromolecules2019, 52, 8672-8681. 4,5,6,39,41,57

[50] M. W. Matsen, J. U. Kim and A. E Likhtman, Finite-N effects for ideal polymer chains near a flat impenetrable wall.The European Physical Journal E 2009, 29, 107-115. 4,22,

REFERENCES

[51] M. W. Matsen and P. Mahmoudi, Segregation of chain ends to the surface of a polymer melt.The European Physical Journal E 2014, 37, 78. 4,23,24

[52] P. Mahmoudi and M. W. Matsen, Entropic segregation of short polymers to the surface of a polydisperse melt. The European Physical Journal E2017, 40, 85. 4,23

[53] S. J. Park, D. Yong, Y. Kim and J. U. Kim, Numerical implementation of pseudo-spectral method in self-consistent mean field theory for discrete polymer chains. The Journal of Chemical Physics 2017,150, 234901. 4

[54] M. Beardsley and M. W. Matsen, Computationally efficient field-theoretic simulations for block copolymer melts.Macromolecules 2019, 52, 8840. 4,50,51,52,55

[55] M. W. Matsen, Field theoretic approach for block polymer melts: SCFT and FTS. The Journal of Chemical Physics2020, 152,, 110901. 4,15

[56] M. O. de la Cruz, S. F. Edwards and I. C. Sanchez, Concentration fluctuations in polymer blend thermodynamics. The Journal of Chemical Physics1988, 89, 1704. 4,50

[57] J. Qin, P. Grzywacz and D. C. Morse, Renormalized one-loop theory of correlations in disordered diblock copolymers.Journal of Chemical Physics2011, 135, 084902. 4,51 [58] P. Stasiak and M. W. Matsen, Monte Carlo field-theoretic simulations for melts of sym-

metric diblock copolymer.Macromolecules 2013, 46, 8037. 4,51,53

[59] B. Vorselaars, P. Stasiak and M. W. Matsen, Field-theoretic simulation of block copolymers at experimentally relevant molecular weights. Macromolecules 2015, 46, 9071. 4, 50, 51, 52,55,57

[60] T. M. Beardsley and M. W. Matsen, Calibration of the Flory-Huggins interaction parameter in field-theoretic simulations.The Journal of Chemical Physics 2019, 150, 150. 4,50 [61] G. H. FredricksonThe Equilibrium Theory of Inhomogeneous PolymersOxford University

Press, New York, 2006. 11,15,20,22,67

[62] M. W. Matsen, in Soft Matter, Volume 1: Polymer Melts and Mixtures, edited by G.

Gompper and M. Schick, Wiley-VCH, Weinheim, 2006. 14

[63] S.-M. Hur, C. J. Garc´ıa-Cervera, E. J. Kramer and G. H. Fredrickson, SCFT simulations of thin film blends of block copolymer and homopolymer laterally confined in a square well.

Macromolecules2009, 42, 5861. 15

[64] Y. B. Yang, Y. J. Choi, S. O. Kim and J. U. Kim, Directed selfassembly of cylinder-forming diblock copolymers on sparse chemical patterns.Soft Matter 2015, 11.4496-4506. 23, 24, 26

REFERENCES

[65] D. Yong and J. U. Kim, Finite volume method for self-consistent field theory of polymers:

Material conservation and application.Physical Review E2017, 96, 063312. 24,48

[66] W. H. Press, B. P. Flannery, S. A. Teukolsky and W. T. Vetterling, Numerical Recipes:

The Art of Scientific ComputingCambridge University Press, Cambridge, 2007. 26 [67] P. Sandhu, J. Zong, D. Yang and Q. Wang, On the comparisons between dissipative particle

dynamics simulations and self-consistent field calculations of diblock copolymer microphase separation.The Journal of Chemical Physics 2013, 138, 194904. 25

[68] S. J. Park, D. Yong, Y. Kim and J. U. Kim, Numerical implementation of pseudo-spectral method in self-consistent mean field theory for discrete polymer chains. The Journal of Chemical Physics 2019,150, 234901. 25

[69] K. O. Rasmussen and G. Kalosakas, Improved numerical algorithm for exploring block copolymer mesophases. Journal of Polymer Science Part B: Polymer Physics 2002, 40, 1777–1783. 25

[70] J. Douglas, Jr., Alternating direction methods for three space variables.Numerische Math- ematik1962 4, 41. 26

[71] H. Zeng and J. Zhu, Proceedings of the International Conference on Parallel Processing Workshops(IEEE,2002),4, p.320. 26

[72] D. -W. Sun, Z. -Y. Sun, H. -F. Li and L.-J. An, Study of morphology and phase diagram of the H-shaped (AC)B(CA) ternary block copolymers using self-consistent field theory.

Polymer2009, 50, 4270. 26

[73] Y.-B. Yang, S. J. Park, P. Kim and J. U. Kim, Roles of chemical pattern period and film thickness in directed self-assembly of diblock copolymers.Soft Matter 2013, 9, 5624. 26 [74] D. Yong and J. U. Kim, Finite volume method for self-consistent field theory of polymers:

Material conservation and application.Physical Review E2017, 96, 063312. 26

[75] H. D. Ceniceros and G. H. Fredrickson, Numerical solution of polymer self-consistent field theory.Multiscale Modeling & Simulation2004, 2, 452. 27

[76] R. B. Thompson, K. O. Rasmussen, and T. Lookman, Improved convergence in block copolymer self-consistent field theory by Anderson mixing. The Journal of Chemical Physics2004, 120, 31. 27

[77] P. Stasiak and M. W. Matsen, Efficiency of pseudo-spectral algorithms with Anderson mixing for the SCFT of periodic block-copolymer phases.The European Physical Journal E2011, 34, 110-954. 27,48

Dokumen terkait