APPLIED IN PHYSICAL SYSTEMS
By
Fabio Milentiansen Sim 11601062
BACHELOR’S DEGREE in
MECHANICAL ENGINEERING - MECHATRONICS CONCENTRATION FACULTY OF ENGINEERING AND INFORMATION TECHNOLOGY
SWISS GERMAN UNIVERSITY The Prominence Tower
Jalan Jalur Sutera Barat No. 15, Alam Sutera Tangerang, Banten 15143 - Indonesia
Revision after Thesis Defence on 07 July 2020 July 2020
STATEMENT BY THE AUTHOR
I hereby declare that this submission is my own work and to the best of my knowledge, it contains no material previously published or written by another person, nor material which to a substantial extent has been accepted for the award of any other degree or diploma at any educational institution, except where due acknowledgement is made in the thesis.
Fabio Milentiansen Sim
Student Date
Approved by:
Dr Eka Budiarto, S.T., M.Sc.
Thesis Advisor Date
Dr Rusman Rusyadi
Thesis Co-Advisor Date
Dr Maulahikmah Galinium, S.Kom, M.Sc
Dean Date
ABSTRACT
NEURAL SOLVERS FOR DIFFERENTIAL EQUATIONS APPLIED IN PHYSICAL SYSTEMS
By
Fabio Milentiansen Sim
Dr Eka Budiarto, S.T., M.Sc., Advisor Dr Rusman Rusyadi, Co-Advisor
SWISS GERMAN UNIVERSITY
Differential equations are ubiquitous in many fields of study, yet analytical solutions to such equations, whether ordinary or partial, have long eluded discovery. With the advent of modern deep learning, neural networks have become a viable alternative to numerical methods. By reformulating the problem as an optimisation task, neural networks can be trained to approximate nonlinear solutions. In this paper, neural solvers are implemented in TensorFlow for a variety of equations. Their overall performance is analysed and even found to surpass traditional schemes in certain cases. Experimental data is also used to validate the neural solutions. Stiff and nonlinear systems of equations are attempted as well, and the stability of the method is investigated. A normalisation technique, akin to feature scaling for supervised learning, is proposed and shown to achieve superior convergence. Although neural solvers will not replace the computational speed offered by traditional schemes in the near future, they remain a feasible, easy-to-implement substitute when all else fails.
Keywords: Differential Equations, Deep Learning, Neural Networks, Numerical Analysis, Computer Simulation
© Copyright 2020 by Fabio Milentiansen Sim
All rights reserved
DEDICATION
For the pursuit of knowledge in not only mathematics, but also every other aspect of the universe.
ACKNOWLEDGEMENTS
I would like to extend my gratitude to my advisors, Dr Eka Budiarto, for guiding me through the intricacies of applied mathematics, and Dr Rusman Rusyadi, for igniting my passion for deep learning.
I thank my family and friends for supporting me throughout my academic journey, as well as in my personal life.
TABLE OF CONTENTS
STATEMENT BY THE AUTHOR 2
ABSTRACT 3
DEDICATION 5
ACKNOWLEDGEMENTS 6
TABLE OF CONTENTS 7
LIST OF FIGURES 10
LIST OF TABLES 13
CHAPTER 1 - INTRODUCTION 14
1.1 Background 14
1.2 Research Problems 15
1.3 Research Objectives 15
1.4 Significance of the Study 16
1.5 Research Questions 16
1.6 Hypotheses 16
CHAPTER 2 - LITERATURE REVIEW 17
CHAPTER 3 - RESEARCH METHODS 20
3.1 Theoretical Approach 20
3.1.1 Differential Equations Preliminaries 21
3.1.2 Neural Network Preliminaries 22
3.1.3 Combination of Neural Networks and ODEs 25
3.2 Implementation Considerations 26
3.3 Model Validity 30
3.4 Model Performance 30
3.5 Model Robustness 30
3.6 Timeline 31
CHAPTER 4 - RESULTS AND ANALYSIS 32
4.1 Environment Setup 32
4.2 Ordinary Differential Equations (ODEs) 33
4.2.1 First Order Semilinear ODE 33
4.2.1.1 Ansatz Variant 38
4.2.1.2 Convolutional Variant 39
4.2.1.3 Recurrent Variant 40
4.2.2 First Order Linear ODE 41
4.2.3 First Order Nonlinear ODE 44
4.2.4 Second Order Linear ODE 46
4.2.5 Second Order Quasilinear ODE 50
4.2.6 Second Order Nonlinear ODE 52
4.3 Partial Differential Equations (PDEs) 53
4.3.1 Poisson’s Equation for the Dirichlet Problem 53 4.3.2 Heat Equation for the Dirichlet Problem 59 4.3.3 Heat Equation for the Neumann Problem 62
4.3.4 Wave Equation 64
4.3.5 Inviscid Burgers’ Equation 66
4.4 Systems of Differential Equations and Additional Examples 70
4.4.1 Linear System of ODEs 72
4.4.2 Stiff System of ODEs 74
4.4.2.1 Ansatz Variant 77
4.4.3 Nonlinear System of ODEs 78
4.4.3.1 Ansatz Variant 80
4.4.4 Proportional (P) Controller Simulation 81
4.4.5 Proportional-Integral (PI) Controller Simulation 83
4.5 Experimental Validation 85
4.5.1 Spring-Mass System 85
4.5.2 Gauss’s Law - Laplace’s Equation 92
4.5.2.1 Ansatz Variant 94
4.6 Robustness Test 96
CHAPTER 5 - CONCLUSION 97
GLOSSARY 99
REFERENCES 100
APPENDIX A - ADDITIONAL MATHEMATICAL PROOFS 103 A1 Normalisation of Ordinary Differential Equations for Neural Methods 103
A2 Normalisation of Partial Differential Equations for Neural Methods 108 APPENDIX B - FULL PROGRAM CODE EXAMPLES 110
B1 First Order Semilinear ODE Ansatz Program 110
B2 Poisson’s Equation Program 113
B3 Lotka-Volterra System Ansatz Program 117
APPENDIX C - SUPPLEMENTAL FIGURES AND TABLES 121
C1 Experimental Data for Spring-Mass System 121
CURRICULUM VITAE 122
LIST OF FIGURES
Figure 3.0.1: Research Methodology Block Diagram 20
Figure 3.1.1: Illustration of a deep neural network 22
Figure 3.2.1: Main Algorithm Flowchart 28
Figure 3.2.2: Training Callback Flowchart 29
Figure 4.2.1.1: Training Loss Curve 35
Figure 4.2.1.2: True Loss Curve 36
Figure 4.2.1.3: Solution and Error Plots 36
Figure 4.2.1.4: Error Trends 37
Figure 4.2.1.5: Ansatz True Loss Curve 38
Figure 4.2.1.6: CNN True Loss Curve 39
Figure 4.2.1.7: RNN True Loss Curve 40
Figure 4.2.1.8: RNN Solution and Error Plots 41
Figure 4.2.2.1: True Error Characteristics 42
Figure 4.2.2.2: Solution Plot 42
Figure 4.2.2.3: Error Trends 43
Figure 4.2.3.1: True Error Characteristics 44
Figure 4.2.3.2: Solution Plot 45
Figure 4.2.3.3: Error Trends 45
Figure 4.2.4.1: True Error Characteristics 47
Figure 4.2.4.2: Error Trends 49
Figure 4.2.5.1: True Error Characteristics 51
Figure 4.2.5.2: Solution Plot 51
Figure 4.2.6.1: Error Characteristics and Solution Plot 52
Figure 4.3.1.1: Error Characteristics 56
Figure 4.3.1.2: Solution Plots 56
Figure 4.3.1.3: Neural Error 57
Figure 4.3.1.4: Error Trends 58
Figure 4.3.2.1: Error Characteristics 60
Figure 4.3.2.2: Solution Plots 60
Figure 4.3.2.3: Error Trends 61
Figure 4.3.3.1: Error Characteristics 63
Figure 4.3.3.2: Solution Plots 63
Figure 4.3.4.1: Error Characteristics 64
Figure 4.3.4.2: Solution Plots 65
Figure 4.3.5.1: Error Characteristics 66
Figure 4.3.5.2: Solution Plot and Characteristic Lines 67
Figure 4.3.5.3: Solution Time Frames 67
Figure 4.3.5.4: Solution Plot and Characteristic Lines 68
Figure 4.3.5.5: Solution Time Frames 68
Figure 4.3.5.6: Solution Plot, Characteristic Lines, Time Frames 69
Figure 4.4.1.1: Error Characteristics 73
Figure 4.4.1.2: Solution Plots and Errors 73
Figure 4.4.2.1: Error Characteristics (No Scaling) 75 Figure 4.4.2.2: Error Characteristics (With Scaling) 75
Figure 4.4.2.3: Solution Plots and Errors 76
Figure 4.4.2.4: Forward Euler Comparison 76
Figure 4.4.2.5: Ansatz Error Characteristics and Solution Plot 77 Figure 4.4.3.1: Error Characteristics, Half Domain (Left), Whole Domain
(Right) 79
Figure 4.4.3.2: Solution Plots & Errors, Half Domain (Top), Whole
Domain (Bottom) 79
Figure 4.4.3.3: Error Characteristics, Before Transformation (Left), After
Transformation (Right) 80
Figure 4.4.3.4: Solution Plots and Errors, Before Transformation (Top),
After Transformation (Bottom) 80
Figure 4.4.4.1: P-Controller Block Diagram 81
Figure 4.4.4.2: Solution Plot and Error 82
Figure 4.4.5.2: Solution Plot and Error 84 Figure 4.5.1.1: Position Sensor (Left), Solenoid (Right) 86
Figure 4.5.1.2: Sensor-CASSY 86
Figure 4.5.1.3: Experiment Setup 87
Figure 4.5.1.4: Error Characteristics 88
Figure 4.5.1.5: Solution Plots 88
Figure 4.5.1.6: Error Characteristics 91
Figure 4.5.1.7: Solution Plots and Errors 91
Figure 4.5.2.1: Loss Characteristics (Left), Error Contour (Right) 93 Figure 4.5.2.2: Experimental Data (Moradi and Marvasti, 2016) (Left),
Neural Method (Right) 93
Figure 4.5.2.3: Boundary Condition Network Output 94
Figure 4.5.2.4: Loss Characteristics (Left), Error Contour (Right) 95 Figure 4.5.2.5: Experimental Data (Moradi and Marvasti, 2016) (Left),
Neural Method Ansatz Variant (Right) 95
LIST OF TABLES
Table 3.6: Timeline 31
Table 4.6: Robustness Test Results 96
Table C1: Experimental Data for Spring-Mass System 121