Using Nonuniform Nodal Space in Meshless Groundwater Modeling Kuo-Chin Hsu, Wen-Han Tsai, and Der-Liang Young Department of Resources Engineering, National Cheng Kung University, Tainan 70101,Taiwan, R.O.C. Department of Civil Engineering, National Taiwan University,Taipei, 106,Taiwan 1 Background and Motivation Mesh-dependent numerical methods are commonly used. However, meshing and remeshing in traditional numerical methods is a tedious work. Can numerical methods be done without meshing? Meshless method! How does the meshless method apply to groundwater modeling? 2 Types of meshless method Common used meshless methods Node-distribution for RBFCM (1) Domain-type (Hu et al. 2007) - Global Radial basis function collocation method (RBFCM) - Localized radial basis function collocation method (LRBFCM) (2) Boundary-type -Method of fundamental solution (MFS) -Method of particular solution (MPS) -Trefftz method (TM) Node distribution for MFS (Young et al. 2006) 3 R11 R12 Radial basis functions (RBF) 𝜑𝑖 = 𝑁 𝑗=1 𝛼𝑗 𝑅𝑖𝑗 (𝑟𝑖𝑗 ) R13 R14 ? R15 𝜑𝑖 : unknown , 𝛼𝑗 : undetermined coefficient , 𝑅𝑖𝑗 : radial basis function (RBF) 𝑟𝑖𝑗 : distance between ith point of interest and jth node, 𝑁 : number of source nodes 𝑟𝑖𝑗 = 𝑥𝑖 − 𝑥𝑗 (1D, Cartesian) 𝑟𝑖𝑗 = (𝑥𝑖 , 𝑦𝑖 ) − (𝑥𝑗 , 𝑦𝑗 ) (2D, Cartesian) Types of common used RBF - 𝑟𝑖𝑗2 + 𝒄2 - 𝑟𝑖𝑗2 + - 𝑒 2 −𝑐𝑟𝑖𝑗 1 2 , Multiquardrics (MQ) 1 − 𝑐2 2 , Inverse Multiquardrics , Gaussian (EXP) - 𝑟𝑖𝑗2 log(𝑟𝑖𝑗 ) , Thin plate splines (TPS) 4 Algorithm of RBFCM 𝜕2𝜑 𝜕𝑥 2 Let 𝜑𝑖 = 𝑟13 = 0, 𝜑0 = 𝑏0 , 𝜑𝐿 = 𝑏𝐿 𝑁=3 𝑗=1 𝛼𝑗 𝑅𝑖𝑗 (𝑟𝑖𝑗 ), 𝑖 𝑟12 = 1, 2, 3 For G.E. 𝑅21 (𝑟21 )′′𝛼1 + 𝑅22 (𝑟22 )′′𝛼2 + 𝑅23 (𝑟23 )′′𝛼3 = 0 For B.C.s 𝑅11 (𝑟11 )𝛼1 + 𝑅12 (𝑟12 )𝛼2 + 𝑅13 (𝑟13 )𝛼3 = 𝑏0 𝑅31 (𝑟31 )𝛼1 + 𝑅32 (𝑟32 )𝛼2 + 𝑅33 (𝑟33 )𝛼3 = 𝑏𝐿 2 3 x=0 x=L 𝑅11 𝛼1 + 𝑅12 𝛼2 + 𝑅13 𝛼3 = 𝑏0 𝑅21 𝛼1 + 𝑅22 𝛼2 + 𝑅23 𝛼3 = 0 𝑅31 𝛼1 + 𝑅32 𝛼2 + 𝑅33 𝛼3 = 𝑏𝐿 𝑅11 𝑅21 𝑅31 5 1 𝑅12 𝑅22 𝑅32 𝑅13 𝑅23 𝑅33 𝛼1 𝑏0 𝛼2 = 0 𝛼3 𝑏𝐿 𝛼1 𝜶 = 𝑨−𝟏 𝒃 = 𝛼2 𝛼3 For any 𝝋 ∗= 𝜶R= = 𝛼1 𝛼2 𝛼3 𝑅∗1 𝑅∗2 𝑅∗3 Governing equations Assume isotropic and homogenous confined aquifer 1D steady-state flow in Cartesian coordinate : 𝜕2𝜑 =0 2 𝜕𝑥 Qusi-3D steady-state radial flow in cylindrical coordinate : 𝜕 2 𝜑 1 𝜕𝜑 + =0 2 𝜕𝑟 𝑟 𝜕𝑟 Qusi-3D transient radial flow in cylindrical coordinate 𝜕 2 𝜑 1 𝜕𝜑 𝑆𝑠 𝜕𝜑 + = 2 𝜕𝑟 𝑟 𝜕𝑟 𝐾 𝜕𝑡 6 Boundary conditions (1) 2 Dirichlet boundaries 𝜑 = 𝜑1 𝜑 = 𝜑2 Unconfined aquifer Aquitard Confined aquifer (2) 1 Dirichlet + 1 Neumann boundary Unconfined aquifer Aquitard 𝑞 7 Confined aquifer 𝜑 = 𝜑3 Nodal configuration (1) Uniformly-distributed nodes 𝑑𝑐 8 𝑑𝑐(𝑎𝑣𝑔) = 𝐿 𝑁−1 𝑁 ∶ Number of total used nodes Nodal configuration (2) Non-uniformly distributed nodes Linear type-node dense at partial boundary 𝑑𝑐 𝑚𝑖𝑛 1 𝑑𝑐 𝑚𝑖𝑛 2 𝑑𝑐 𝑚𝑎𝑥 1 𝑑𝑐 𝑚𝑎𝑥 2 𝑅𝑖𝑐𝑟 ∶ distance Increasing rate 9 Nodal configuration (3) Non-uniformly distributed nodes Linear type-node dense at all boundary 10 Steady-state flow in Cartesian coord. with 2 Dirichlet BCs (Case1) Uniformly-distributed nodes (Configuration 1) A n a l y t ic a l s o l u ti o n M Q _ 1 1 p ts _ d c ( a v g ) = 1 0 0 (m ) _ c = 1 0 0 0 10 2 0 0 -h (m ) 8 6 4 2 0 11 0 2 00 4 00 60 0 x (m ) 80 0 1 00 0 Steady-state flow in Cartesian coor. with 1 Dirichlet & 1 Neumann BCs (Case 2) Uniformly-distributed nodes (Configuration 1) A n a ly tic a l so lu ti o n M Q _ 1 1 p t s _ d c ( av g ) = 1 0 0 m _ c = 1 0 0 0 M Q _ 2 1 p t s _ d c ( av g ) = 5 0 m _ c = 4 7 0 M Q _ 5 1 p t s _ d c ( av g ) = 2 0 m _ c = 1 9 0 M Q _ 1 0 1 p t s _ d c(a v g)= 1 0 m _ c = 9 5 50 2 0 0 -h ( m ) 40 30 20 10 0 12 0 2 00 4 00 60 0 x (m ) 80 0 1 00 0 Steady-state radial flow with 2 Dirichlet BCs (Case 3) Uniformly-distributed nodes (Configuration 1) 10 A n a l y t ic a l s o l u t i o n M Q _ 1 1 p t s _ d c ( a v g ) = 1 0 0 m _ c = 1 1 2 5 .8 M Q _ 1 0 1 p t s _ d c ( a v g ) = 1 0 m _ c = 9 5 .6 8 M Q _ 1 0 0 1 p t s _ d c ( a v g ) = 1 m _ c = 1 0 .3 2 0 0 -h (m ) M Q _ 2 0 0 1 p t s _ d c ( a v g ) = 0 .5 m _ c = 5 . 4 M Q _ 5 0 0 1 p t s _ d c ( a v g ) = 0 .2 m _ c = 2 . 1 9 6 M Q _ 1 0 0 0 1 p t s _ d c ( a v g ) = 0 . 1 m _ c = 1 .0 4 M Q _ 2 0 0 0 1 p t s _ d c ( a v g ) = 0 . 0 5 m _ c = 0 .5 2 4 2 0 2 00 4 00 60 0 r (m ) 13 80 0 1 00 0 Steady-state radial flow in with 2 Dirichlet BCs (Case 3) Cases be tested in linear-type non-uniformly-distributed nodes dense at partial boundary (Configuration 2) 14 N 𝑑𝑐(min) (m) 𝑑𝑐(max) (m) 𝑅𝑖𝑐𝑟 (m) 𝑐 𝛿(%) MAE(m) 498 0.1 2.7 0.01 1.398 11.96 1.30 × 10−1 614 0.1 1.9 0.01 1.395 5.67 6.18 × 10−2 956 0.1 1.1 0.01 1.451 0.48 5.26 × 10−3 1149 0.1 0.9 0.01 1.392 0.28 3.08 × 10−3 1189 0.1 0.9 0.005 1.26 0.35 3.82 × 10−3 1469 0.1 0.9 0.001 1.261 0.22 2.45 × 10−3 1688 0.1 0.7 0.001 1.241 0.24 2.59 × 10−3 2163 0.1 0.5 0.001 1.216 0.21 2.26 × 10−3 2405 0.05 0.45 0.001 0.663 0.48 5.18 × 10−3 2988 0.05 0.35 0.001 0.696 0.18 2.01 × 10−3 4083 0.05 0.25 0.001 0.682 0.084 9.16 × 10−4 9146 0.01 0.11 0.001 0.141 0.85 9.27 × 10−3 11149 0.01 0.09 0.001 0.202 0.064 6.97 × 10−4 14318 0.01 0.07 0.001 0.156 0.052 𝟓. 𝟕𝟐 × 𝟏𝟎−𝟒 Steady-state radial flow with 2 Dirichlet BCs (Case 3) Cases be tested in linear-type non-uniformly-distributed nodes dense at all boundary (Configuration 3) N 𝑑𝑐(min) (m) 𝑑𝑐(max) (m 𝑅𝑖𝑐𝑟 (m) 𝑐 𝛿(%) MAE(m) ) 15 621 0.1 2.9 0.01 1.48 9.8 1.07 × 10−1 651 0.1 2.3 0.01 1.46 7.43 8.1 × 10−2 803 0.1 1.5 0.01 1.41 2.23 2.42 × 10−2 1193 0.1 0.9 0.01 1.14 0.65 7.08 × 10−3 1257 0.1 0.9 0.005 1.14 0.58 6.27 × 10−3 1833 0.1 0.9 0.001 1.14 0.28 3.10 × 10−3 1955 0.1 0.7 0.001 1.07 0.24 2.65 × 10−3 2325 0.1 0.5 0.001 1.01 0.22 2.43 × 10−3 2585 0.05 0.45 0.001 0.61 0.215 2.35 × 10−3 3123 0.05 0.35 0.001 0.59 0.058 6.33 × 10−4 4165 0.05 0.25 0.001 0.56 0.027 2.99 × 10−4 9191 0.01 0.11 0.001 0.17 0.136 1.49 × 10−3 11193 0.01 0.09 0.001 0.15 0.067 7.27 × 10−4 14339 0.01 0.07 0.001 0.16 0.002 𝟐. 𝟐𝟑 × 𝟏𝟎−𝟓 Steady-state radial flow with 1 Dirichlet & 1 Neumann BCs (Case 4) Cases be tested in linear-type non-uniformly-distributed nodes dense at all boundary (Configuration 3) 16 N 𝑑𝑐(min) (m) 𝑑𝑐(max) (m) 𝑅𝑖𝑐𝑟 (m) 𝑐 𝛿(%) MAE(m) 2325 0.1 0.5 0.001 1.23 9.69 7.74 × 10−2 10001 0.1 0.1 0 1.08 10.74 8.57 × 10−2 4165 0.05 0.25 0.001 0.66 1.76 1.4 × 10−2 20001 0.05 0.05 0 0.601 1.865 1.49 × 10−2 14339 0.01 0.07 0.001 0.16 0.027 𝟐. 𝟏𝟑 × 𝟏𝟎−𝟒 10001 and 20001 points of nodes are uniformly-distributed for the purpose of comparison with non-uniformly-distributed nodes. Transient radial flow with 1 Dirichlet & 1 Neumann BCs (Case 5) Cases be tested in linear-type non-uniformly-distributed nodes dense at all boundary (Configuration3) 17 Shape parameter c and the nodal interval Cases for uniformly-distributed nodes (Configuration 1) c 11.273 d c ( avg ) 2.8578 1200 y = 11.273x - 2.8578 R² = 0.9998 1000 800 c 600 400 200 0 0 20 40 60 dc_(avg) 18 80 100 120 Shape parameter c and the nodal interval Cases for linearly increasing nodal space (Configuration 2) c 1.45 [1 exp(2.3* d c ( avg ) )] 1.6 1.4 1.2 c 1 0.8 c=1.45*(1-exp(2.3*d_(avg))) 0.6 0.4 0.2 0 0 0.5 1 1.5 dc_(avg) 19 2 2.5 Shape parameter c and the nodal interval Cases for linearly increasing nodal space with symmetric distribution (Configuration 3) c 1.5 [1 exp(2* d c ( avg ) )] 1.6 1.4 1.2 c 1 0.8 c=1.5*(1-exp(2*d_(avg))) 0.6 0.4 0.2 0 0 0.5 1 dc_(avg) 20 1.5 2 Conclusions 1. Meshless method can be a useful tool for groundwater modeling 2. The shape parameter c has a strong relation with nodal arrangement. 3. Numerical errors is higher for meshless modeling with Neumann boundary condition. 4. Numerical errors is higher for meshless modeling with cylindrical coordinate. 5. Apply non-uniformly-distributed nodes in MQ-RBFCM, the total number of node can be dramatically reduced. 21 Finite element method Meshless method 刪繁就簡三秋樹 領異標新二月花 鄭板橋 22 Thank you for your attention~ 23 Groundwater flow equation (1) Mass conservation : inflow mass flux − outflow mass flux = change of density with time 𝑀𝑥1 − 𝑀𝑥2 + 𝑀𝑦1 − 𝑀𝑦2 + 𝑀𝑧1 − 𝑀𝑧2 𝜕 = (𝜌𝑓 𝑛∆𝑥∆𝑦∆𝑧) 𝜕𝑡 𝑀𝐷𝑖 = 𝜌𝑓 𝑞𝐷𝑖 𝐴𝐷𝑖 Taylor series expansion : 𝑀𝐷1 − 𝑀𝐷2 = − − 24 𝜕(𝜌𝑓 𝑞𝐷 ) ∆𝑥∆𝑦∆𝑧 𝜕𝑥 REV for deriving groundwater flow equations (from Schwartz and Zhang, 2003) 𝜕(𝜌𝑓 𝑞𝑥 ) 𝜕(𝜌𝑓 𝑞𝑦 ) 𝜕(𝜌𝑓 𝑞𝑧 ) 𝜕 + + ∆𝑥∆𝑦∆𝑧 = 𝜌 𝑛∆𝑥∆𝑦∆𝑧 𝜕𝑥 𝜕𝑦 𝜕𝑧 𝜕𝑡 𝑓 𝐷 ∶ 𝑑𝑖𝑟𝑒𝑐𝑡𝑖𝑜𝑛 𝑜𝑓 𝑡ℎ𝑒 𝑓𝑙𝑜𝑤 𝐷𝑖 : unit face in REV Groundwater flow equation (2) Assume fluid density is constant in spatial 𝜕𝑞𝑥 𝜕𝑞𝑦 𝜕𝑞𝑧 1 𝜕 − + + = 𝜌 𝑛 𝜕𝑥 𝜕𝑦 𝜕𝑧 𝜌𝑓 𝜕𝑡 𝑓 Definition of Specific storage : 𝑆𝑠 = 𝑣𝑜𝑙𝑢𝑚𝑒 𝑜𝑓 𝑤𝑎𝑡𝑒𝑟 𝑢𝑛𝑖𝑡 𝑎𝑟𝑒𝑎 𝑢𝑛𝑖𝑡 𝑎𝑞𝑢𝑖𝑓𝑒𝑟 𝑡ℎ𝑖𝑐𝑘𝑛𝑒𝑠𝑠 𝑢𝑛𝑖𝑡 ℎ𝑒𝑎𝑑 𝑐ℎ𝑎𝑛𝑔𝑒 𝑣𝑜𝑙𝑢𝑚𝑒 𝑜𝑓 𝑤𝑎𝑡𝑒𝑟 = (𝑣𝑜𝑙𝑢𝑚𝑒 𝑜𝑓 𝑅𝐸𝑉)(𝑢𝑛𝑖𝑡 ℎ𝑒𝑎𝑑 𝑐ℎ𝑎𝑛𝑔𝑒) 1 𝜕 𝜕ℎ 𝜌 𝑛 = 𝑆𝑠 𝜌𝑓 𝜕𝑡 𝑓 𝜕𝑡 Darcy’s law : 𝑞𝐷 = −𝐾𝐷 𝜕ℎ 𝜕𝐷 − 𝜕𝑞𝑥 𝜕𝑞𝑦 𝜕𝑞𝑧 𝜕 𝜕ℎ 𝜕 𝜕ℎ 𝜕 𝜕ℎ + + = 𝐾𝑥 + 𝐾𝑦 + 𝐾𝑧 𝜕𝑥 𝜕𝑦 𝜕𝑧 𝜕𝑥 𝜕𝑥 𝜕𝑦 𝜕𝑦 𝜕𝑧 𝜕𝑧 𝜕 𝜕ℎ 𝜕 𝜕ℎ 𝜕 𝜕ℎ 𝜕ℎ 𝐾𝑥 + 𝐾𝑦 + 𝐾𝑧 = 𝑆𝑠 𝜕𝑥 𝜕𝑥 𝜕𝑦 𝜕𝑦 𝜕𝑧 𝜕𝑧 𝜕𝑡 25 (Storage Equation) Nodal arrangement (4) Non-uniformly distributed nodes Exponential type-node dense at partial boundary 𝑑𝑐 𝑑𝑐 𝑖 𝑖 𝑖=1 = 𝑖 ln 𝐿+1 𝑁−1 e 2≤𝑖≤𝑁−1 = −1 𝑖 ln 𝐿+1 𝑁−1 e L − 𝑖−1 ln 𝐿+1 𝑁−1 e N N 0 26 Nodal arrangement (5) Non-uniformly distributed nodes Exponential type-node dense at all boundary 𝑑𝑐 𝑑𝑐 𝑖 𝑖=1 = 𝑖 𝐿 𝑁−1 ln 2+1 e 2 𝑖 2≤𝑖≤𝑁−1 =e 𝑖 𝑁−1 2 𝐿 2 −1 ln +1 −e 𝑖−1 𝑁−1 2 ln 𝐿 +1 2 2 𝑑 27 𝑐 𝑖 𝑁+1 ≤𝑖≤𝑁−1 2 =e 𝑁−𝑖+1 𝐿 𝑁−1 ln 2+1 2 −e 𝑁−𝑖 𝐿 𝑁−1 ln 2+1 2 Sensitivity analysis (Transient problem for example) - effect of shape parameter c 28 Sensitivity analysis (Transient problem for example) - effect of θ and time step (Δt) effect of θ 29 effect of Δt Sensitivity analysis (Transient problem for example) -effect of domain size (R) R=100 (m) R=10 (m) 30 R=1 (m) Steady-state flow in cylindrical coor. with 2 Dirichlet BCs (Case 3) Why not kept using configuration2. ? Reason (1). MQ_1149pts_dc(max)=0.9(m);dc(min)=0.1(m);Ricr=0.01(m)_c=1.392 MQ_1189pts_dc(max)=0.9(m);dc(min)=0.1(m);Ricr=0.005(m)_c=1.26 MQ_1469pts_dc(max)=0.9(m);dc(min)=0.1(m);Ricr=0.001(m)_c=1.261 10 Relative error (%) 1 0.1 0.01 0.001 0.0001 31 200 400 600 r (m) 800 (Configuration2.) 1000 (Configuration3.) Steady-state flow in cylindrical coor. with 2 Dirichlet BCs (Case 3) Why not kept using configuration2. ? Reason (2). MQ_1469pts_dc(max)=0.9(m);dc(min)=0.1(m);Ricr=0.001(m)_c=1.261 MQ_2405pts_dc(max)=0.45(m);dc(min)=0.05(m);Ricr=0.001(m)_c=0.663 MQ_11149pts_dc(max)=0.09(m);dc(min)=0.01(m);Ricr=0.001(m)_c=0.202 1000 100 Relative error (%) 10 1 0.1 0.01 0.001 0.0001 200 32 400 600 800 r (m) (Configuration2.) 1000 (Configuration3.) Steady-state flow in cylindrical coor. with 2 Dirichlet BCs (Case 3) Cases be tested in exponential-type non-uniformly-distributed nodes 500 Average relative error (%) Average relative error (%) 500 400 300 200 400 300 200 100 100 0 10 33 100 1000 Number of total used nodes (N) (nodes dense at partial boundary) 10000 10 100 1000 Number of total used nodes (N) (nodes dense at all boundary) 10000 Literature review Hardy (1971) developed the algorithm for scattered data 34 interpolation by MQ-RBF. Tarwater (1985) found that by increasing c in MQ-RBF, the error dropped to a minimum then increase sharply. Kansa (1990) first applied MQ-RBF in solving partial differential equations. Carlson and Foley (1991) concluded that c is problemdependent. Zerroukat et al. (1998) applied Kansa’s algorithm for heat transfer problem and found the system became “illconditioned” by increasing number of nodes and c. Literature review Mai-Duy and Tran-Cong (2001) applied different density of uniform-distributed nodes in solving differential equations by MQ-RBFCM. Wang and Liu (2002) studied the optimal chose for shape parameter in different RBFs and in different distribution of nodes. (uniformly-distributed nodes and randomlydistributed node) Amaziane et al. (2004) found that Neumann boundary condition seriously affect hydraulic head . 35
© Copyright 2026 Paperzz