Datasets:
File size: 37,863 Bytes
79067a6 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 | {
"paper_id": "wdno",
"paper_title": "Wavelet Diffusion Neural Operator (WDNO)",
"D1": [
{
"id": "wdno-D1-001",
"claim": "WDNO BRM/SRM training config (all experiments): Adam optimizer, lr=1e-4, cosine annealing scheduler, 190000 training steps, batch_size=16. 1D experiments on 1 A100 GPU (2.4-2.5h), 2D on 2 A100 GPUs (7.8-7.9h).",
"source": "Table 18, Table 19, Table 20, Table 12"
},
{
"id": "wdno-D1-002",
"claim": "WDNO 1D U-Net architecture hyperparameters (for Burgers/Advection/Navier-Stokes): init_dim=128, down_up_layers=4, conv_kernel=3, dim_multiplier=[1,2,4,8], resnet_block_groups=8, attn_hidden_dim=32, attn_heads=4.",
"source": "Table 18, Table 19"
},
{
"id": "wdno-D1-003",
"claim": "1D DDIM sampling config: DDIM_sampling_steps=50, DDIM_eta=1 (equivalent to DDPM stochastic sampling) for Burgers, Advection, and Navier-Stokes inference.",
"source": "Table 18, Table 19"
},
{
"id": "wdno-D1-004",
"claim": "1D Burgers control guidance config: guidance_weight=120000 with cosine scheduler for guidance intensity during control optimization.",
"source": "Table 18, Section C.4 Table 8"
},
{
"id": "wdno-D1-005",
"claim": "1D Burgers PDE and dataset config: T=8, D=[0,1], nu=0.01 Dirichlet BC; 40000 train trajectories; test: 50 (control), 2000 (super-res 0x), 100 (super-res 1x/2x/3x); 81 stored timesteps; state shape [81,120], force shape [80,120]; solver internal grid 120x16 spatial, 4800x16 temporal, downsample factor 16.",
"source": "Appendix F.1, Appendix F.2"
},
{
"id": "wdno-D1-006",
"claim": "1D Burgers initial condition generation: sum of 2 Gaussians; amplitudes a1~U(0,2), a2~U(-2,0); positions b1~U(0.2,0.4), b2~U(0.6,0.8); widths sigma~U(0.05,0.15).",
"source": "Appendix F.2"
},
{
"id": "wdno-D1-007",
"claim": "1D Burgers control force generation: sum of 8 Gaussians; spatial/temporal positions b1,i~U(0,1), b2,i~U(0,1); widths sigma1,i~U(0.1,0.4), sigma2,i~U(0.1,0.4); amplitude a1~U(-1.5,1.5); for i>=2, ai~U(-1.5,1.5) or 0 with 50% probability.",
"source": "Appendix F.2"
},
{
"id": "wdno-D1-008",
"claim": "1D wavelet transform config (Burgers/Advection/Navier-Stokes): 2D DWT using bior2.4 wavelet, periodization padding mode, pytorch_wavelets (Cotter 2019). Output shape [4,41,60] from 81x120 input. Applies to 81 timesteps (Burgers) and 80 timesteps (Advection, from PDEBench).",
"source": "Appendix F.3, Appendix G.2"
},
{
"id": "wdno-D1-009",
"claim": "1D Navier-Stokes dataset config: shear and bulk viscosity eta=zeta=1e-8, heat capacity ratio Gamma=5/3, data resolution [81,120], 9000 training samples, PDEBench CFD shock dataset ('1D_CFD_Shock_Eta1.e-8_Zeta1.e-8_trans_Train.hdf5').",
"source": "Section 4.3, Appendix G.1"
},
{
"id": "wdno-D1-010",
"claim": "2D wavelet transform config (Fluid/ERA5): 3D DWT using bior1.3 wavelet, zero padding mode, implemented via ptwt (Wolter 2024). Input shape [32,64,64], output wavelet coefficient shape [8,18,34,34] (1 coarse + 7 detail channels).",
"source": "Section 4.4, Appendix H.2"
},
{
"id": "wdno-D1-011",
"claim": "2D incompressible fluid dataset config: 32 time steps, 64x64 spatial grid, 3584 peripheral control variables per time step, fluid-solid coupling with no-slip boundary conditions at obstacles.",
"source": "Section 4.4"
},
{
"id": "wdno-D1-012",
"claim": "WDNO 3D U-Net hyperparameters (2D experiments): Conv3D kernel=(3,3,3), padding=(1,1,1), stride=(1,1,1); spatial-only downsampling kernel=(1,4,4), stride=(0,1,1); spatial-only upsampling kernel=(1,4,4); attn_heads=4. Inspired by Video Diffusion Models (Ho et al., 2022).",
"source": "Table 20"
},
{
"id": "wdno-D1-013",
"claim": "2D DDIM sampling config: DDIM_sampling_steps_2D=100, DDIM_eta_2D=1, used for 2D fluid and ERA5 inference.",
"source": "Table 20"
},
{
"id": "wdno-D1-014",
"claim": "2D fluid control guidance config: guidance_weight_2D=100 for indirect smoke navigation control task.",
"source": "Table 20"
},
{
"id": "wdno-D1-015",
"claim": "Zero-shot super-resolution config: 1D base [80,120], targets [160,240] (1x), [320,480] (2x), [640,960] (3x); 2D base [32,64,64], target [32,128,128]. Multi-res wavelet coeff shapes: 1D [4x21x30, 4x11x15, 4x6x8]; 2D [8x18x18x18, 8x18x10x10].",
"source": "Section 4.6, Appendix F.3, Appendix H.2"
},
{
"id": "wdno-D1-016",
"claim": "ERA5 dataset config: 0.25-degree latitude-longitude resolution, surface to 100km altitude, variable=temperature. Prediction task: 12h input history to forecast 20h ahead.",
"source": "Section 4.5"
},
{
"id": "wdno-D1-017",
"claim": "Ablation study configs: measurement noise scales [0.01, 0.001, 0.0001] times data_std on 1D Burgers; training sample fractions [0.2, 0.4, 0.6, 0.8] of 9000 samples on 1D Navier-Stokes; control sequence noise prob=0.1 in robustness test.",
"source": "Section 4.7, Table 9, Section C.5"
},
{
"id": "wdno-D1-018",
"claim": "SAC baseline hyperparameters (1D Burgers control): discount=0.5, target_smoothing=0.05, critic_lr=3e-4, entropy_lr=3e-3, policy_lr=3e-3, batch=8192, episodes=1500, model_updates_per_step=50, target_updates=15, replay_buffer=1M, energy_weight=2e-5.",
"source": "Table 22"
},
{
"id": "wdno-D1-019",
"claim": "BPPO baseline hyperparameters (1D Burgers control): state value net (lr=1e-4, 2M steps, 3 layers, hidden=512); Q net (lr=1e-4, 2M steps, 2 layers, hidden=1024, discount=0.99, target_update=2, soft_update=0.005); BC pretrain (lr=1e-4, 500k episodes); main (100k episodes, 2 layers, hidden=1024, lr=1e-5, clip=0.25, ReLU, batch=512).",
"source": "Table 24"
},
{
"id": "wdno-D1-020",
"claim": "BC baseline hyperparameters (1D Burgers control): lr=1e-4, batch=512, episodes=500k, replay_buffer=2M, 2 layers, hidden=1024, ReLU activation.",
"source": "Table 25"
},
{
"id": "wdno-D1-021",
"claim": "SL (Supervised Learning) baseline hyperparameters (1D Burgers control): LBFGS optimizer, lr=0.1, epochs=100, objective loss weight=1, reconstruction loss weight=0.01, tolerance=1e-5.",
"source": "Table 23"
},
{
"id": "wdno-D1-022",
"claim": "ANN PID baseline hyperparameters (1D Burgers control): 1D conv kernel=3, padding=1, stride=1, Softsign activation, batch=16, lr=1e-4, MAE loss.",
"source": "Table 21"
},
{
"id": "wdno-D1-023",
"claim": "FNO baseline hyperparameters (1D and 2D): 1D: modes=16, width=64, in_ch=3, out_ch=1, 4 Fourier layers, lift/proj hidden=256, GeLU activation, lr=1e-4; 2D: modes=16, width=64, in_ch=6, out_ch=4, 4 Fourier layers, lr=1e-4, epochs=1000, cosine scheduler.",
"source": "Table 26, Table 32"
},
{
"id": "wdno-D1-024",
"claim": "WNO baseline hyperparameters (1D and 2D): 1D: sym4 wavelet, 5 levels, uplift_dim=40, 4 layers, batch=100, lr=1e-3, epochs=1000, StepLR scheduler; 2D: db4 wavelet, 2 levels, uplift_dim=8, 3 layers, batch=50, lr=0.05, epochs=500, StepLR scheduler.",
"source": "Table 27, Table 33"
},
{
"id": "wdno-D1-025",
"claim": "MWT baseline hyperparameters (1D and 2D): 1D: Legendre wavelet, 10 Fourier modes, kernel=4, batch=256, epochs=300, MultiStepLR scheduler; 2D: Legendre wavelet, 12 Fourier modes, kernel=3, batch=200, epochs=300, MultiStepLR scheduler.",
"source": "Table 29, Table 34"
},
{
"id": "wdno-D1-026",
"claim": "OFormer 1D baseline: SpatialEncoder2D (in_ch=3, embed_dim=96, encoded_dim=256, heads=4, depth=6, res=120, dropout=0.05), PointWiseDecoder2DSimple (latent=256, out_ch=1), batch=32, 50000 iterations, lr=1e-4.",
"source": "Table 30"
},
{
"id": "wdno-D1-027",
"claim": "OFormer 2D baseline: SpatialTemporalEncoder2D (in_ch=3, embed_dim=496, encoded_dim=192, heads=1, depth=5), PointWiseDecoder2D (embed_dim=96, out_ch=1, out_seq_len=32, propagate_forward=1, curriculum_ratio=0.1, curriculum_steps=6), batch=8, 100000 iterations, lr=1e-4.",
"source": "Table 35"
},
{
"id": "wdno-D1-028",
"claim": "MS-L-NODE (MSVI) baseline hyperparameters (1D simulation): encoder CNN channels=128, latent_dim=8, aggregation heads=1, 6 static + 4 dynamic layers, batch=64, 37500 iterations, lr=1e-3.",
"source": "Table 31"
},
{
"id": "wdno-D1-029",
"claim": "CNN 1D baseline: conv kernel=5, padding=2, ELU activation, latent=256, batch=100, lr=1e-3, epochs=500, cosine scheduler.",
"source": "Table 28"
},
{
"id": "wdno-D1-030",
"claim": "Sensitivity analysis configs: DDIM steps grid [20,40,50,100,200]; DDIM eta grid [0.2,0.5,0.8,1.0] for 1D Burgers control; guidance weight grid [9,10,11.5,12.5,13] x1e4 for 2D control.",
"source": "Table 7, Table 8"
}
],
"D2": [
{
"id": "wdno-D2-001",
"claim": "DDPM forward process (noise addition): x_{k+1} = sqrt(alpha_k) * x_k + sqrt(1 - alpha_k) * epsilon, where epsilon ~ N(0, I) and {alpha_k}_{k=1..K} is the variance schedule. Implemented via: x_k = sqrt(alpha_bar_k) * x_0 + sqrt(1 - alpha_bar_k) * epsilon, where alpha_bar_k = prod_{i=1..k} alpha_i.",
"source": "Section 2.2"
},
{
"id": "wdno-D2-002",
"claim": "DDPM reverse process (denoising): x_{k-1} = 1/sqrt(alpha_k) * (x_k - (1-alpha_k)/sqrt(1-alpha_bar_k) * epsilon_theta(x_k, k)) + sigma_k * z, where z ~ N(0, I) if k > 1 else z = 0. epsilon_theta is the learned denoising U-Net.",
"source": "Section 2.2"
},
{
"id": "wdno-D2-003",
"claim": "DDPM training loss (simplified variational bound): L = E_{k~U(1,K), x_0~p(x), epsilon~N(0,I)} [|| epsilon - epsilon_theta(sqrt(alpha_bar_k) * x_0 + sqrt(1 - alpha_bar_k) * epsilon, k) ||_2^2]",
"source": "Section 2.2, Eq. 3"
},
{
"id": "wdno-D2-004",
"claim": "Classifier-free guidance sampling: epsilon_theta_guided(x, y) = epsilon_theta(x, empty) + omega * (epsilon_theta(x, y) - epsilon_theta(x, empty)), where omega in [0, 1] is the guidance weight and 'empty' is the null-condition identifier.",
"source": "Section 2.2"
},
{
"id": "wdno-D2-005",
"claim": "Discrete wavelet transform (1D signal decomposition): phi_{l,m}(x) = 2^{l/2} * phi(2^l * x - m); psi_{l,m}(x) = 2^{l/2} * psi(2^l * x - m). With l_0 = L (maximum level), the decomposition is: u(x) = sum_m c_L(m) * phi_{L,m}(x) + sum_m d_L(m) * psi_{L,m}(x), where c_L are coarse coefficients and d_L are detail coefficients. In code: apply 2D DWT via pytorch_wavelets (1D) or 3D DWT via ptwt (2D) to get coefficient arrays.",
"source": "Section 3.1"
},
{
"id": "wdno-D2-006",
"claim": "WDNO simulation: denoising in wavelet domain: Initialize W_{u,[0,T]}^{(K)} ~ N(0, I). For k = K down to 1: W_{u,[0,T]}^{(k-1)} = W_{u,[0,T]}^{(k)} - eta * epsilon_theta(W_{u,[0,T]}^{(k)}, W_a, k) + xi, where xi ~ N(0, sigma_k^2 * I) if k > 1 else xi = 0. W_a is the wavelet-transformed equation parameter (e.g., initial condition). Final output is obtained by inverse DWT on W_{u,[0,T]}^{(0)}. DDIM acceleration is used during inference.",
"source": "Section 3.1, Eq. 3"
},
{
"id": "wdno-D2-007",
"claim": "WDNO control: guided denoising with objective gradient: Initialize W_{f,[0,T]}^{(K)} ~ N(0, I). For k = K down to 1: W_{f,[0,T]}^{(k-1)} = W_{f,[0,T]}^{(k)} - eta * (epsilon_theta(W_{f,[0,T]}^{(k)}, W_a, k) + lambda * grad_{W_f} J(W_hat_{f,[0,T]}^{(k)})) + xi, where xi ~ N(0, sigma_k^2 * I) if k > 1 else xi = 0. W_hat_{f,[0,T]}^{(k)} is the estimated noise-free W_f^{(0)} extracted via: W_hat_{f,[0,T]}^{(k)} = (W_{f,[0,T]}^{(k)} - sqrt(1 - alpha_bar_k) * epsilon_theta(W_{f,[0,T]}^{(k)}, W_a, k)) / sqrt(alpha_bar_k). lambda is the guidance weight. J is the control objective.",
"source": "Section 3.1, Eq. 4, Eq. 5"
},
{
"id": "wdno-D2-008",
"claim": "WDNO noise-free estimate from noisy sample (one-step prediction): W_hat_f^{(k)} = (W_f^{(k)} - sqrt(1 - alpha_bar_k) * epsilon_theta(W_f^{(k)}, W_a, k)) / sqrt(alpha_bar_k). This is used to compute the objective J on the estimated clean data rather than noisy data, avoiding noise-induced errors in the guidance gradient.",
"source": "Section 3.1, Eq. 5"
},
{
"id": "wdno-D2-009",
"claim": "Multi-resolution training: dataset preparation by downsampling: Given original resolution N x M (time x space), create data pairs by downsampling: (N, M) and (N/2, M/2); (N/2, M/2) and (N/4, M/4); (N/4, M/4) and (N/8, M/8); etc. Apply wavelet transform to each resolution level. For 1D Burgers with original [80, 120]: downsampled to [40, 60], [20, 30], [10, 15]. Wavelet coefficients: original 81x120 -> 4x41x60; downsampled 41x60 -> 4x21x30; 21x31 -> 4x11x15; 11x15 -> 4x6x8.",
"source": "Section 3.2, Appendix F.3"
},
{
"id": "wdno-D2-010",
"claim": "Super-Resolution Model (SRM) training: Train conditional diffusion model p(W_h | W_l, W_{a_h}) using paired high-resolution W_h and low-resolution W_l data. To align sizes, duplicate low-resolution coefficients to match high-resolution dimensions (with boundary duplication for odd numbers). Each training batch randomly selects data pairs from a given resolution level. Training loss is the standard DDPM MSE loss between predicted and true noise.",
"source": "Section 3.2"
},
{
"id": "wdno-D2-011",
"claim": "WDNO zero-shot super-resolution inference: 1) Downsample a to base resolution NxM, wavelet transform to get W_a. 2) Use BRM to generate wavelet coefficients W_base at resolution NxM. 3) Use SRM iteratively: condition on W_current (low-res) and W_{a_next} (high-res a at next level) to generate W_next at 2Nx2M. 4) Repeat SRM until target resolution reached. 5) Apply inverse wavelet transform to get final trajectory.",
"source": "Section 3.2"
},
{
"id": "wdno-D2-012",
"claim": "1D Burgers control objective function J: J = integral_D |u(T,x) - u*(x)|^2 dx + alpha * integral_{[0,T]xD} |f(t,x)|^2 dt dx. The first term penalizes deviation from target state at final time. The second term penalizes control energy. In code: MSE between u_pred(T) and u_target, plus alpha * MSE of control force f over all time steps.",
"source": "Section 4.1, Eq. 6, Appendix F.1"
},
{
"id": "wdno-D2-013",
"claim": "1D Burgers initial condition generation: u(0,x) = sum_{i=1}^{2} a_i * exp(-(x - b_i)^2 / (2 * sigma_i^2)). Parameters sampled: a_1 ~ U(0,2), a_2 ~ U(-2,0), b_1 ~ U(0.2, 0.4), b_2 ~ U(0.6, 0.8), sigma_i ~ U(0.05, 0.15).",
"source": "Appendix F.2"
},
{
"id": "wdno-D2-014",
"claim": "1D Burgers control force generation: f(t,x) = sum_{i=1}^{8} a_i * exp(-(x - b_{1,i})^2 / (2 * sigma_{1,i}^2)) * exp(-(t - b_{2,i})^2 / (2 * sigma_{2,i}^2)). Parameters: b_{1,i} ~ U(0,1), b_{2,i} ~ U(0,1), sigma_{1,i} ~ U(0.1, 0.4), sigma_{2,i} ~ U(0.1, 0.4), a_1 ~ U(-1.5, 1.5), and for i >= 2: a_i ~ U(-1.5, 1.5) or 0 with equal probability.",
"source": "Appendix F.2"
},
{
"id": "wdno-D2-015",
"claim": "1D Burgers finite difference solver: Finite difference method solving: du/dt = -u * du/dx + nu * d^2u/dx^2 + f(t,x), with Dirichlet boundary u=0, initial u(0,x)=u_0(x), nu=0.01. Internal grid: 120x16 spatial, 4800x16 temporal steps. Control f is kept constant between two stored time stamps. After simulation, downsample by factor 16 spatially and temporally before saving.",
"source": "Appendix F.1, F.2"
},
{
"id": "wdno-D2-016",
"claim": "2D wavelet data preparation for WDNO (1D Burgers/Advection/Navier-Stokes): 1) Apply 2D DWT to trajectory data (81x120) using bior2.4 wavelet, periodization mode -> outputs 4 coeff sets each (41x60): 1 coarse + 3 detail. 2) Apply 1D DWT to initial condition (1D), repeat coeffs, concatenate with 2D coeffs. 3) For multi-resolution data: downsample data to 41x60, 21x31, 11x15; apply 2D DWT -> 4x21x30, 4x11x15, 4x6x8. 4) Align low-res coeffs to high-res by duplication, with boundary duplication for odd dimensions.",
"source": "Appendix F.3"
},
{
"id": "wdno-D2-017",
"claim": "3D wavelet data preparation for WDNO (2D incompressible fluid): 1) Apply 3D DWT to trajectory data (32x64x64) using bior1.3 wavelet, 'zero' padding mode -> outputs 8 coeff sets each (18x34x34): 1 coarse + 7 detail. 2) Apply 2D DWT to initial condition (2D) and 1D DWT to smoke percentage, repeat coeffs and concatenate. 3) For multi-resolution: downsample to 32x32x32, 32x16x16; apply 3D DWT -> 8x18x18x18, 8x18x10x10.",
"source": "Appendix H.2"
},
{
"id": "wdno-D2-018",
"claim": "WDNO training pseudocode: Training: 1) Apply DWT to u_{[0,T]}^{(0)} ~ q(u) to get W_u^{(0)}. 2) Sample k ~ Uniform(1,...,K). 3) Sample epsilon ~ N(0,I). 4) Take gradient descent step on ||epsilon - epsilon_theta(sqrt(alpha_bar_k) * W_u^{(0)} + sqrt(1-alpha_bar_k) * epsilon, k)||^2. Repeat until converged. Sampling (simulation): 1) W_u^{(K)} ~ N(0,I). 2) For k=K..1: W_u^{(k-1)} = W_u^{(k)} - eta * epsilon_theta(W_u^{(k)}, W_a, k) + xi (xi~N(0,I) if k>1 else 0). 3) Apply inverse DWT to W_u^{(0)} to get u_{[0,T]}^{(0)}. For control: same but with additional +lambda * grad J(W_hat_f^{(k)}) term in step 2.",
"source": "Algorithm 1 (Appendix E)"
},
{
"id": "wdno-D2-019",
"claim": "1D Navier-Stokes sound speed and Mach number computation: Sound speed: c_s = sqrt(Gamma * p / rho), where Gamma=5/3. Mach number: M = |v| / c_s. Initial velocity field: v(x, t=0) = sum_{i=1..n} A_i * sin(k_i * x + phi_i), where A_i = v_bar / |k|^d, d=1 for 2D case, v_bar = c_s * M. Shock-tube initialization: Q(x, t=0) = (Q_L, Q_R) with piecewise constant values generating shocks and rarefactions, where Q = (rho, v, p).",
"source": "Appendix G.1"
},
{
"id": "wdno-D2-020",
"claim": "SAC reward function for 1D control: r(t, u_t, u_d, w_t) = -integral_Omega |u_t - u_d|^2 dx - alpha * integral_Omega |w_t|^2 dx. In code: negative MSE between current state and target state minus alpha * MSE of action. Alpha is energy weight (0.00002 for Burgers).",
"source": "Appendix I.2, Eq. 13"
},
{
"id": "wdno-D2-021",
"claim": "BPPO reward function for 1D control: r(t, u_t, u_d, f_t) = -integral_Omega |u_t - u_d|^2 dx - alpha * integral_Omega |f_t|^2 dx. Same form as SAC reward with state u and control f.",
"source": "Appendix I.4, Eq. 14"
},
{
"id": "wdno-D2-022",
"claim": "Approximate scale invariance for multi-resolution alignment: Given high-res data d_+ at NxM and low-res data d_- at (N/2)x(M/2), apply linear transformation t_scaled = a_1 * t_orig + b_1, x_scaled = a_2 * x_orig + b_2 to rescale coordinates. Then the stretched function satisfies: du/(a_1*dt) = F(u, du/(a_2*dx), d^2u/(a_2^2*dx^2), ...) + f(t,x). Since resolution change factor is constant, a_1 and a_2 are fixed, so the pattern between resolutions is consistent. This consistency also holds in wavelet domain due to linearity and locality of wavelet transform.",
"source": "Section 3.2, Eq. 7"
},
{
"id": "wdno-D2-023",
"claim": "WDNO U-Net architecture (2D encoder-decoder for 1D spatiotemporal data): Two separate U-Nets: epsilon_theta(f) for force pathway and epsilon_theta(u,f) for state pathway. Each U-Net: initial dim=128, 4 down/up layers with 3x3 conv kernels, channel multipliers [1,2,4,8] per stage, 8 ResNet block groups per stage, attention at lowest resolution with hidden dim=32 and 4 heads.",
"source": "Table 18, Appendix F.4"
},
{
"id": "wdno-D2-024",
"claim": "WDNO 3D U-Net architecture (for 2D spatiotemporal data): Three modules: downsampling encoder, middle module, upsampling decoder with 3D spatiotemporal convolutions. Conv3D: kernel=(3,3,3), padding=(1,1,1), stride=(1,1,1). Downsampling: kernel=(1,4,4), padding=(1,2,2), stride=(0,1,1) - only downsamples spatial dims. Upsampling: kernel=(1,4,4), padding=(1,2,2), stride=(0,1,1) - only upsamples spatial dims. 4 attention heads. Architecture inspired by Video Diffusion Models (Ho et al., 2022).",
"source": "Table 20, Appendix H.3"
},
{
"id": "wdno-D2-025",
"claim": "Guidance loss for initial condition and target state consistency: Add auxiliary loss terms to the control guidance gradient: ∇_{W_f} J_total = ∇_{W_f} J_control + λ_IC · ∇_{W_f} ||IDWT(W_f[0,:]) - u_0_true||^2 + λ_target · ∇_{W_f} ||IDWT(W_f[T,:]) - u_T_true||^2, where W_f[0,:] is the initial timestep's wavelet coefficient channel, W_f[T,:] is the target timestep's wavelet coefficient channel, IDWT(·) is the inverse discrete wavelet transform, and u_0_true, u_T_true are the ground truth initial/target states. This ensures the generated trajectory matches the given initial condition and target state. For 1D Burgers control: λ_IC applied via cosine-scheduled guidance weight (120000); for 2D fluid control: λ_target = 100.",
"source": "Appendix F.4, Appendix H.3"
},
{
"id": "wdno-D2-026",
"claim": "DDIM accelerated sampling: Denoising Diffusion Implicit Model (DDIM) (Song et al., 2020) is used during inference to speed up sampling. Instead of running all K diffusion steps, DDIM uses a subsequence of steps. WDNO uses 50 DDIM iterations for 1D and 100 for 2D inference, with eta=1 (equivalent to DDPM stochastic sampling).",
"source": "Section 3.1, Table 18, Table 20"
},
{
"id": "wdno-D2-027",
"claim": "ANN PID controller architecture (1D baseline): Two 1D convolutional layers (kernel=3, padding=1, stride=1) with Softsign activation, followed by two fully connected layers and four activation layers. Trained with MAE loss between current and target states using Adam optimizer (lr=0.0001, batch=16). The neural network outputs MIMO PID parameters based on error Err_t, which the PID controller uses to produce control f_t.",
"source": "Appendix I.1, Table 21"
},
{
"id": "wdno-D2-028",
"claim": "Supervised Learning control algorithm (SL baseline): Stage 1: Two CNN VAEs project state u and control f into latent space; a CNN learns latent transition u_t -> u_{t+1}. Stage 2: Use these three networks as surrogate models to compute gradient of objective J w.r.t. control f via backpropagation. Optimize f with LBFGS (lr=0.1, 100 epochs, tol=1e-5). Include reconstruction loss for control (weight=0.01) alongside objective loss (weight=1.0) to avoid adversarial modes.",
"source": "Appendix I.3, Table 23"
},
{
"id": "wdno-D2-029",
"claim": "FNO architecture (1D baseline): Fourier Neural Operator: Lifting block (hidden=256) -> 4 Fourier layers (modes=16, width=64) -> Projection block (hidden=256). Each Fourier layer: FFT -> linear transform in Fourier domain on kept modes -> IFFT -> add skip connection. MLP expansion=0.5, GeLU activation, rank of tensor factorization=1.0, one-sided domain padding. Input: initial state + controls (3 channels). Output: states (1 channel). Supports zero-shot super-resolution.",
"source": "Appendix J.1, Table 26"
},
{
"id": "wdno-D2-030",
"claim": "WNO architecture (1D baseline): Wavelet Neural Operator: sym4 wavelet, 5 decomposition levels, uplifting dimension=40, 4 wavelet layers. Each wavelet layer: DWT -> linear transform on wavelet coefficients -> IDWT + skip connection. Trained with Adam (lr=1e-3, batch=100, epochs=1000, StepLR scheduler).",
"source": "Appendix J.2, Table 27"
},
{
"id": "wdno-D2-031",
"claim": "MWT architecture (1D baseline): Multiwavelet Neural Operator: Legendre wavelet basis, 10 Fourier modes, kernel size=4. Batch=256, epochs=300, Adam optimizer, MultiStepLR scheduler.",
"source": "Appendix J.4, Table 29"
},
{
"id": "wdno-D2-032",
"claim": "OFormer architecture (1D baseline): Operator Transformer: SpatialEncoder2D (3 input channels, embedding dim=96, encoded sequence dim=256, 4 heads, 6 depth layers, resolution=120, dropout=0.05) -> PointWiseDecoder2DSimple (latent channels=256, output channels=1, scale=0.5, resolution=120). Batch=32, 50000 iterations, lr=1e-4.",
"source": "Appendix J.5, Table 30"
}
],
"D3": [
{
"id": "wdno-D3-001",
"claim": "1D Burgers equation simulation: Learn mapping from initial condition u_0 and force f to entire trajectory u_{[0,T]}. Compare WDNO against neural operator baselines (WNO, MWT, OFormer, FNO, CNN, DDPM). Metrics: MSE on state sequences excluding initial conditions. Datasets: 40000 trajectories with random initial conditions (2 Gaussians) and control forces (8 Gaussians).",
"source": "Section 4.1, Table 1, Appendix F"
},
{
"id": "wdno-D3-002",
"claim": "1D Burgers equation control: Find optimal control force f that minimizes objective J = integral|u(T,x)-u*(x)|^2 + alpha*integral|f|^2. Compare WDNO against PID, SAC (pseudo-online/offline), BPPO (surrogate-solver/solver), BC (surrogate-solver/solver), SL, and DDPM baselines. Metrics: control objective J. Datasets: 40000 train / 50 test trajectories.",
"source": "Section 4.1, Table 2a, Appendix F"
},
{
"id": "wdno-D3-003",
"claim": "1D Advection equation simulation: Predict 80 timesteps of advection evolution from a one-step initial condition. Evaluate on smooth dynamics from PDEBench. Compare against WNO, MWT, OFormer, FNO, CNN, DDPM. Data prep: 2D wavelet transform (bior2.4, periodization). Metric: MSE excluding initial conditions.",
"source": "Section 4.2, Table 1"
},
{
"id": "wdno-D3-004",
"claim": "1D Compressible Navier-Stokes equation simulation: Evaluate on shock-tube scenarios from PDEBench with extremely small viscosity (eta=1e-8, zeta=1e-8). Simulation only (no control). Compare against Transolver, CNO, MSVI, ACDM, DiffusionPDE, WNO, MWT, OFormer, FNO, CNN, DDPM, Diffusion+FFT, FNO Denoiser. Metrics: MSE, MAE, L_inf. Datasets: 9000 training samples, 81x120 resolution.",
"source": "Section 4.3, Table 1, Table 5, Appendix G"
},
{
"id": "wdno-D3-005",
"claim": "2D Incompressible fluid simulation: Predict smoke density, velocity field, and percentage through target bucket given initial smoke density and control sequences. Fluid-solid coupling with no-slip boundaries at obstacles. Compare against WNO, MWT, OFormer, FNO, U-Net, DDPM. Data prep: 3D DWT (bior1.3, zero mode). Metric: MSE excluding initial conditions. Datasets: 32 timesteps, 64x64 spatial grid.",
"source": "Section 4.4, Table 1, Appendix H"
},
{
"id": "wdno-D3-006",
"claim": "2D Incompressible fluid control (indirect smoke navigation): Control 3584 peripheral variables over 32 timesteps to steer smoke from below central obstacle into top-center bucket. Objective J = percentage of smoke NOT passing target bucket. Compare against BC, BPPO, SAC (pseudo-online/offline), DDPM. Highly challenging due to indirect control and need for trajectory planning.",
"source": "Section 4.4, Table 2b, Appendix H"
},
{
"id": "wdno-D3-007",
"claim": "ERA5 weather simulation: Predict temperature evolution over next 20 hours given past 12 hours of atmospheric data. Real-world dataset (0.25deg, surface to 100km). Compare against MWT, OFormer, FNO, U-Net, DDPM. Data prep: 3D DWT (bior1.3). Metrics: MSE, Relative L2 error.",
"source": "Section 4.5, Table 1"
},
{
"id": "wdno-D3-008",
"claim": "Zero-shot super-resolution (1D Burgers): Train on 80x120; evaluate at 160x240 (1x), 320x480 (2x), 640x960 (3x) unseen during training. Compare WNO, FNO, WDNO with linear and nearest interpolation. Test: 2000 trajectories (0x) + 100 shared (1x/2x/3x). Metric: MSE at finest resolution.",
"source": "Section 4.6, Table 16, Figure 4"
},
{
"id": "wdno-D3-009",
"claim": "Zero-shot super-resolution (2D fluid): Train on 32x64x64; evaluate at 32x128x128 unseen during training. Compare FNO and WDNO with linear and nearest interpolation. Metric: MSE at finest resolution.",
"source": "Section 4.6, Table 17, Figure 4"
},
{
"id": "wdno-D3-010",
"claim": "Ablation: wavelet transform for abrupt changes. Compare WDNO vs DDPM MAE per time step specifically at regions with abrupt spatial changes in 1D Burgers and 1D Navier-Stokes. Demonstrate wavelet benefit at shock/discontinuity regions.",
"source": "Section 4.7, Figure 6, Figure 9"
},
{
"id": "wdno-D3-011",
"claim": "Ablation: wavelet + multi-resolution training synergy. Compare WDNO (wavelet domain) vs DDPM with multi-resolution training in original space-time domain (no wavelet). Evaluate at multiple super-resolution levels. Metric: MSE at finest resolution.",
"source": "Section 4.7, Figure 4c"
},
{
"id": "wdno-D3-012",
"claim": "Ablation: wavelet vs Fourier transform. Compare WDNO (wavelet domain generation) vs Diffusion+FFT (Fourier domain, same architecture/pipeline except transform). Also test FNO as denoiser backbone. Dataset: 1D Navier-Stokes. Metric: MSE.",
"source": "Section 4.7, Figure 5c"
},
{
"id": "wdno-D3-013",
"claim": "Ablation: long-term dependency. Measure error growth over time steps for WDNO vs WNO, MWT, OFormer, FNO, U-Net, DDPM on 2D simulation. Verify slower error accumulation for WDNO. Metric: MSE per time step.",
"source": "Section 4.7, Figure 5a"
},
{
"id": "wdno-D3-014",
"claim": "Ablation: measurement noise robustness. Add Gaussian noise (scale factors 0.01, 0.001, 0.0001 * data_std) to both training and testing data of 1D Burgers. Compare WDNO vs DDPM. Metric: MSE vs noise scale.",
"source": "Section 4.7, Figure 5d"
},
{
"id": "wdno-D3-015",
"claim": "Ablation: training sample size. Reduce training set to 0.2, 0.4, 0.6, 0.8 of full 9000 samples on 1D Navier-Stokes. Measure WDNO MSE trend. Metric: MSE.",
"source": "Section 4.7, Figure 5b"
},
{
"id": "wdno-D3-016",
"claim": "Ablation: approximate scale invariance verification. Train FNO on original, once-downsampled, twice-downsampled, and mixed datasets; test at all three resolutions to demonstrate mixed training improves generalization. Metric: MSE at each resolution.",
"source": "Appendix C.3, Table 6"
},
{
"id": "wdno-D3-017",
"claim": "Sensitivity analysis: DDIM sampling steps. Test steps in {20, 40, 50, 100, 200} for 1D Burgers simulation and 2D fluid control. Metrics: MSE (1D sim), J (2D control).",
"source": "Appendix C.4, Table 7, Table 8"
},
{
"id": "wdno-D3-018",
"claim": "Sensitivity analysis: DDIM eta and guidance weight. Test eta in {0.2, 0.5, 0.8, 1.0} for 1D Burgers control; test guidance weights {9, 10, 11.5, 12.5, 13} x1e4 for 2D control. Metrics: MSE, J.",
"source": "Appendix C.4, Table 7, Table 8"
},
{
"id": "wdno-D3-019",
"claim": "Robustness test: control sequence noise. 1D Burgers control with 0.1 probability of noise in control during evaluation. Compare WDNO vs PID, SAC, BC, BPPO, SL, DDPM. Metric: J.",
"source": "Appendix C.5, Table 9"
},
{
"id": "wdno-D3-020",
"claim": "Robustness test: variance across test samples. Report mean +/- std of control objective J over 50 test samples for 1D Burgers control. Baselines: PID, SAC, BC, BPPO, SL, DDPM. Metric: J (mean +/- std).",
"source": "Appendix C.5, Table 10"
},
{
"id": "wdno-D3-021",
"claim": "Ablation: guidance importance. Compare WDNO 1D Burgers control with (lambda=120000) vs without (lambda=0) guidance term. Metric: J.",
"source": "Appendix C.7, Table 15"
},
{
"id": "wdno-D3-022",
"claim": "Wavelet reconstruction fidelity test. Measure relative L2 reconstruction error of wavelet bases (bior1.3, bior2.4, db4, sym4) on 1D Burgers and 2D fluid training data. Metric: relative L2 error.",
"source": "Appendix A, Table 3"
},
{
"id": "wdno-D3-023",
"claim": "Computational efficiency: wavelet vs Fourier transform. Measure total transform time on 1D Navier-Stokes training set (A100 GPU, batch 2000). Compare DWT (pytorch_wavelets) vs PyTorch 2D FFT. Metric: total transform time (seconds).",
"source": "Appendix A, Table 4"
},
{
"id": "wdno-D3-024",
"claim": "Computational resource comparison: Parameter counts, inference time (batch 1, A100), and training time for all 1D Burgers baselines (PID, SAC, BC, BPPO, SL, DDPM, WNO, MWT, OFormer, FNO, CNN). Metrics: params, inference time, training time.",
"source": "Appendix C.6, Table 11, Table 12, Table 13"
},
{
"id": "wdno-D3-025",
"claim": "Super-resolution computational cost scaling. Measure inference time and GPU memory for batch-5 generation at 0, 1, 2, 3 levels of super-resolution on 1D Burgers. Metrics: time, memory.",
"source": "Appendix C.6, Table 14"
},
{
"id": "wdno-D3-026",
"claim": "WDNO training protocol (shared across all experiments). BRM and SRM trained with DDPM loss. Adam optimizer, lr=1e-4, cosine annealing, 190000 steps, batch 16. DDIM inference: 50 steps (1D) / 100 steps (2D), eta=1. Control guidance: 120000 (1D) / 100 (2D) with cosine scheduler. Training hardware: 1 A100 (1D, 2.4-2.5h) / 2xA100 (2D, 7.8-7.9h).",
"source": "Section 3, Table 18, Table 19, Table 20, Table 12"
}
],
"D4": [
{
"id": "wdno-D4-001",
"claim": "WDNO Training Pipeline (Section 3, Figure 1): Phase 1 - Train Base-Resolution Model (BRM) on full-resolution wavelet-domain data using DDPM loss (output: trained BRM; Adam, lr=1e-4, cosine annealing, 190k steps, batch 16). Phase 2 - Prepare multi-resolution dataset by downsampling data, applying wavelet transform at each resolution level, and aligning coefficient dimensions via duplication (output: multi-resolution wavelet dataset, feeds into Phase 3). Phase 3 - Train Super-Resolution Model (SRM) on paired multi-resolution wavelet data (input: Phase 2 dataset, output: trained SRM; condition on low-res coeffs + high-res params to predict high-res coeffs via DDPM loss). At inference: Phase 1 BRM generates base-resolution coefficients that feed Phase 3 SRM for iterative upsampling.",
"source": "Section 3, Figure 1, Algorithm 1"
},
{
"id": "wdno-D4-002",
"claim": "WDNO Simulation Inference Pipeline (Section 3.1, Algorithm 1): Step 1 - Apply Discrete Wavelet Transform (DWT) to input equation parameters (IC, BC, PDE coeffs) to obtain wavelet-domain condition W_a. Step 2 - Initialize wavelet-domain trajectory as Gaussian noise W_u^{(K)}~N(0,I); iteratively denoise via DDIM for k=K..1: W_u^{(k-1)} = W_u^{(k)} - eta * epsilon_theta(W_u^{(k)}, W_a, k) + xi. Step 3 - Apply Inverse DWT to W_u^{(0)} to recover full trajectory u_{[0,T]} in original spatiotemporal domain.",
"source": "Section 3.1, Eq. 3, Algorithm 1"
},
{
"id": "wdno-D4-003",
"claim": "WDNO Control Inference Pipeline (Section 3.1, Eq. 4-5): Step 1 - Apply DWT to input params to obtain W_a; define control objective J (state deviation at final time + energy regularization). Step 2 - Initialize noise W_f^{(K)}~N(0,I); iteratively denoise with classifier-free guidance + objective gradient for k=K..1: W_f^{(k-1)} = W_f^{(k)} - eta * (epsilon_theta(W_f^{(k)}, W_a, k) + lambda * grad_{W_f} J(W_hat_f^{(k)})) + xi, where W_hat_f is noise-free estimate of W_f^{(0)}. Step 3 - Apply IDWT to W_f^{(0)} to recover controlled trajectory.",
"source": "Section 3.1, Eq. 4, Eq. 5"
},
{
"id": "wdno-D4-004",
"claim": "WDNO Zero-shot Super-Resolution Pipeline (Section 3.2): Step 1 - Downsample input params to base resolution NxM and apply DWT. Step 2 - BRM generates wavelet coefficients at base NxM via DDIM denoising. Step 3 - SRM iteratively upsamples: condition on current low-res coeffs + high-res params to generate coeffs at 2Nx2M; repeat until target resolution. Step 4 - Apply IDWT to obtain final trajectory at target (super-resolution) resolution.",
"source": "Section 3.2"
},
{
"id": "wdno-D4-005",
"claim": "1D Burgers Data Generation Pipeline (Appendix F.2): Step 1 - Generate initial condition u(0,x) as superposition of 2 Gaussians with random parameters (output: initial condition, feeds Step 3). Step 2 - Generate control force f(t,x) as superposition of 8 Gaussians (a_i for i>=2 set to 0 with 50% prob) (output: control force, feeds Step 3). Step 3 - Run finite difference solver (input: Step 1 IC + Step 2 force; du/dt = -u*du/dx + nu*d^2u/dx^2 + f, Dirichlet u=0, nu=0.01, internal 120x16 spatial x 4800x16 temporal). Step 4 - Downsample solver output by factor 16 spatially and temporally to final [81,120] state + [80,120] force (input: Step 3 raw trajectory).",
"source": "Appendix F.1, F.2"
},
{
"id": "wdno-D4-006",
"claim": "Supervised Learning (SL) Control Two-Stage Pipeline (Appendix I.3): Stage 1 - Train surrogate models (two CNN VAEs for state u and control f encoding; CNN for latent transition u_t -> u_{t+1}). Stage 2 - Optimize control f by computing gradient of objective J via backpropagation through the three surrogate networks; optimize with LBFGS (lr=0.1, 100 epochs, tol=1e-5, obj_weight=1, recon_weight=0.01).",
"source": "Appendix I.3, Table 23"
}
]
} |