Add files using upload-large-folder tool

configs/combustion/cno.yaml

@@ -0,0 +1,50 @@
+exp_name: "cno_combustion"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "combustion"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 12
+normalizer: "gaussian" # gaussian, range, none
+
+# data parameters for training
+mask_prob: 0.1
+noise_scale: 0.1 # only applicable for numerical data
+
+# model
+model_name: "cno"
+checkpoint_path: ./results/cno/cno_combustion_numerical_False/2025-09-16_05-34-53/model_3000.pth # for resume training
+N_layers: 3 # Number of (D) or (U) blocks in the network
+N_res: 1,                 # Number of (R) blocks per level (except the neck)
+N_res_neck: 8,            # Number of (R) blocks in the neck
+channel_multiplier: 32,   # How the number of channels evolve?
+conv_kernel: 3,             # Size of all the kernels
+cutoff_den: 2.0001,       # Filter property 1.
+filter_size: 6,             # Filter property 2.
+lrelu_upsampling: 2,      # Filter property 3.
+half_width_mult: 0.8,    # Filter property 4.
+radial: False,            # Filter property 5. Is filter radial?
+batch_norm: True,         # Add BN? We do not add BN in lifting/projection layer
+out_dim: 1,               # Target dimension
+out_size: 1,              # If out_size is 1, Then out_size = in_size. Else must be int
+expand_input: False,      # Start with original in_size, or expand it (pad zeros in the spectrum)
+latent_lift_proj_dim: 64, # Intermediate latent dimension in the lifting/projection layer
+add_inv: True,            # Add invariant block (I) after the intermediate connections?
+activation: 'LeakyReLU'   # Activation function can be 'LeakyReLU' or 'lrelu'
+
+
+# training
+is_use_tb: True
+scheduler: cosine # step, cosine
+step_size: 1000 # only applicable for step scheduler
+num_update: 5000
+train_batch_size: 16
+test_batch_size: 64
+lr: 0.0003
+clip_grad_norm: 0.
+
+# evaluation
+N_autoregressive: 3
+N_plot: 0

configs/combustion/deeponet.yaml

@@ -0,0 +1,34 @@
+exp_name: "deeponet_combustion"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "combustion"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 12
+normalizer: "gaussian" # none, gaussian, range
+
+# data parameters for training
+mask_prob: 0.5
+noise_scale: 0.0 # only applicable for numerical data
+
+# model
+model_name: "deeponet"
+checkpoint_path: ./results/deeponet/deeponet_combustion_numerical_False/2025-09-16_22-50-38/model_2220.pth # for resume training and evaluation
+p: 128
+dropout_rate: 0.1
+
+# training
+is_use_tb: True
+scheduler: cosine # step, cosine
+step_size: 1000 # only applicable for step scheduler
+num_update: 3000
+train_batch_size: 64
+test_batch_size: 64
+lr: 0.0005
+clip_grad_norm: 0.
+
+# evaluation
+N_autoregressive: 1
+N_plot: 1

configs/combustion/dmd.yaml

@@ -0,0 +1,25 @@
+exp_name: "dmd_combustion"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "combustion"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 12
+normalizer: "gaussian"
+
+# model
+model_name: "dmd"
+is_resume: False
+n_modes: 10
+n_predict: 20
+mask_prob: 0.5
+noise_scale: 0.1 # only applicable for numerical data
+input_feature: 2
+checkpoint_path: 'no model needed'
+test_batch_size: 12
+
+# evaluation
+N_plot: 1
+N_autoregressive: 1

configs/combustion/dpot_l.yaml

@@ -0,0 +1,54 @@
+exp_name: "dpot_l_combustion"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "combustion"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 16
+normalizer: "gaussian" # gaussian, range, none
+
+# data parameters for training
+mask_prob: 0.1
+noise_scale: 0.5 # only applicable for numerical data
+
+# model
+model_name: "dpot"
+checkpoint_path: "./dpot_ckpts/model_L.pth"  # Large DPOT model
+model_type: "dpot"  # "dpot" or "dpot3d"
+
+# DPOT specific parameters (Large model configuration - from checkpoint)
+img_size: 128       # Pre-trained model resolution (with auto-resize for combustion)
+patch_size: 8
+in_channels: 16     
+out_channels: 16     
+embed_dim: 1536     
+depth: 24            
+n_blocks: 16         
+modes: 32           # Keep default modes for AFNO
+mlp_ratio: 4        
+out_layer_dim: 1536   # Match checkpoint out layer dim
+normalize: False
+act: "gelu"
+time_agg: "exp_mlp"
+n_cls: 12            # Match checkpoint pre-training (12 datasets)
+
+# dpot multi-step specific parameters (from checkpoint)
+in_timesteps: 20
+out_timesteps: 20
+
+# training
+is_use_tb: True
+scheduler: "cosine" # step, cosine
+step_size: 1000 # only applicable for step scheduler
+num_update: 10000         
+train_batch_size: 16
+test_batch_size: 64
+lr: 0.0001
+clip_grad_norm: 1
+
+# evaluation
+N_autoregressive: 1
+N_plot: 5
+

configs/combustion/dpot_s.yaml

@@ -0,0 +1,54 @@
+exp_name: "dpot_s_combustion"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "combustion"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 16
+normalizer: "gaussian" # gaussian, range, none
+
+# data parameters for training
+mask_prob: 0.1
+noise_scale: 0.5 # only applicable for numerical data
+
+# model
+model_name: "dpot"
+checkpoint_path: "./dpot_ckpts/model_S.pth"  # Small DPOT model
+model_type: "dpot"  # "dpot" or "dpot3d"
+
+# DPOT specific parameters (Small model configuration - from checkpoint)
+img_size: 128       # Pre-trained model resolution (with auto-resize for combustion)
+patch_size: 8
+in_channels: 16     
+out_channels: 16     
+embed_dim: 1024     
+depth: 6            
+n_blocks: 8         
+modes: 32           # Keep default modes for AFNO
+mlp_ratio: 1        
+out_layer_dim: 32   # Keep default out layer dim
+normalize: False
+act: "gelu"
+time_agg: "exp_mlp"
+n_cls: 12            # Match checkpoint pre-training (12 datasets)
+
+# dpot multi-step specific parameters
+in_timesteps: 20
+out_timesteps: 20
+
+# training
+is_use_tb: True
+scheduler: "cosine" # step, cosine
+step_size: 1000 # only applicable for step scheduler
+num_update: 10000         
+train_batch_size: 16
+test_batch_size: 64
+lr: 0.001
+clip_grad_norm: 1
+
+# evaluation
+N_autoregressive: 1
+N_plot: 5
+

configs/combustion/fno.yaml

@@ -0,0 +1,37 @@
+exp_name: "fno_combustion"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "combustion"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 12
+normalizer: "gaussian" # none, gaussian, range
+
+# data parameters for training
+mask_prob: 0.5
+noise_scale: 0.0 # only applicable for numerical data
+
+# model
+model_name: "fno" 
+checkpoint_path: ./results/fno/fno_combustion_numerical_False/2025-09-14_08-01-31/model_0720.pth
+modes1: 4
+modes2: 16
+modes3: 16
+n_layers: 4
+width: 64
+
+# training
+is_use_tb: True
+scheduler: cosine # step, cosine
+step_size: 1000 # only applicable for step scheduler
+num_update: 2000
+train_batch_size: 64
+test_batch_size: 64
+lr: 0.01
+clip_grad_norm: 0.
+
+# evaluation
+N_autoregressive: 1
+N_plot: 0

configs/combustion/galerkin_transformer.yaml

@@ -0,0 +1,72 @@
+exp_name: "galerkin_transformer_combustion"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "combustion"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 12
+normalizer: "gaussian"
+
+# data parameters for training
+mask_prob: 0.5
+noise_scale: 0.1 # only applicable for numerical data
+
+# model
+model_name: "galerkin_transformer"
+checkpoint_path: ./results/galerkin_transformer/galerkin_transformer_combustion_real_True/2025-09-16_23-02-38/model_4900.pth
+
+pos_dim: 1 # pos dim
+n_hidden: 256
+num_feat_layers: 0
+num_encoder_layers: 1
+n_head: 4
+dim_feedforward: 256
+feat_extract_type: null
+attention_type: galerkin
+xavier_init: 0.01
+diagonal_weight: 0.01
+symmetric_init: False
+layer_norm: False
+attn_norm: True
+norm_eps: 0.0000001
+batch_norm: False
+return_attn_weight: False
+return_latent: False
+decoder_type: ifft2
+spacial_dim: 3
+spacial_fc: True
+upsample_mode: interp
+downsample_mode: interp
+freq_dim: 128
+boundary_condition: None
+num_regressor_layers: 1
+fourier_modes_x: 16
+fourier_modes_y: 16
+fourier_modes_t: 4
+regressor_activation: silu
+downscaler_activation: relu
+upscaler_activation: silu
+last_activation: True
+dropout: 0.0
+downscaler_dropout: 0.05
+upscaler_dropout: 0.0
+ffn_dropout: 0.05
+encoder_dropout: 0.05
+decoder_dropout: 0
+debug: False
+
+# training
+is_use_tb: True
+scheduler: cosine # step, cosine
+step_size: 1000 # only applicable for step scheduler
+num_update: 5000
+train_batch_size: 16
+test_batch_size: 16
+lr: 0.01
+clip_grad_norm: 0.
+
+# evaluation
+N_autoregressive: 1
+N_plot: 1

configs/combustion/mwt.yaml

@@ -0,0 +1,43 @@
+
+
+exp_name: "mwt_combustion"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "combustion"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 12
+normalizer: "gaussian" # gaussian, range, none
+
+# data parameters for training
+mask_prob: 0.5
+noise_scale: 0.0 # only applicable for numerical data
+
+# model
+model_name: "mwt"
+checkpoint_path: "./results/mwt/mwt_combustion_numerical_False/2025-09-18_14-45-34/model_3800.pth"
+
+k: 3            # wavelet filter size
+alpha: 5        # fourier filter parameter, should be less or equal to T/2
+c: 4            # increase factor
+nCZ: 4         # number of CZ blocks
+L: 0
+base: "legendre" # chebyshev
+
+
+# training
+is_use_tb: True
+scheduler: cosine # step, cosine
+step_size: 1000 # only applicable for step scheduler
+num_update: 5000
+test_interval: 200
+train_batch_size: 32
+test_batch_size: 64
+lr: 0.001
+clip_grad_norm: 0.
+
+# evaluation
+N_autoregressive: 3
+N_plot: 5

configs/combustion/surrogate_model/fno.yaml

@@ -0,0 +1,28 @@
+exp_name: "surrogate_fno_combustion"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "combustion"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 10
+normalizer: "gaussian"
+
+# model
+model_name: "fno"
+modes1: 4
+modes2: 16
+modes3: 16
+n_layers: 4
+width: 64
+
+# training
+is_use_tb: True
+scheduler: cosine # step, cosine
+step_size: 1000 # only applicable for step scheduler
+num_update: 1000
+train_batch_size: 16
+test_batch_size: 16
+lr: 0.0001
+clip_grad_norm: 0.

configs/combustion/surrogate_model/unet.yaml

@@ -0,0 +1,25 @@
+exp_name: "surrogate_unet_combustion"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "combustion"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 10
+normalizer: "gaussian"
+
+# model
+model_name: "unet"
+dim_mults: [1, 2]
+
+
+# training
+is_use_tb: True
+scheduler: cosine # step, cosine
+step_size: 1000 # only applicable for step scheduler
+num_update: 1000
+train_batch_size: 2
+test_batch_size: 2
+lr: 0.0001
+clip_grad_norm: 0.

configs/combustion/trainsolver.yaml

@@ -0,0 +1,46 @@
+exp_name: "transolver_combustion"
+gpu: 0
+seed: 0
+results_path: "./results/"
+# data
+dataset_name: "combustion"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 12
+normalizer: "gaussian" # none, gaussian, range
+
+# data parameters for training
+mask_prob: 0.1
+noise_scale: 0.1 # only applicable for numerical data
+
+# model
+model_name: "transolver"
+space_dim: 16
+n_layers: 1
+n_hidden: 256
+n_head: 8
+H: 64
+W: 64
+D: 20
+fun_dim: 0 # 1
+out_dim: 16
+ref: 4
+dropout: 0.1
+act: "gelu"
+mlp_ratio: 4
+slice_num: 16
+
+checkpoint_path: "./results/transolver/transolver_combustion_numerical_False/2025-09-15_13-23-53/model_3900.pth"
+# training
+is_use_tb: True
+scheduler: cosine # step, cosine
+step_size: 1000 # only applicable for step scheduler
+num_update: 5000
+test_interval: 200
+train_batch_size: 16
+test_batch_size: 16
+lr: 0.0007
+clip_grad_norm: 0.
+
+# evaluation
+N_autoregressive: 3
+N_plot: 0

configs/combustion/unet.yaml

@@ -0,0 +1,34 @@
+exp_name: "unet_combustion"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "combustion"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 12
+normalizer: "gaussian"
+
+# model
+model_name: "unet"
+is_resume: False
+checkpoint_path: './results/unet/unet_combustion_numerical_False/2025-09-17_00-56-36/model_7600.pth' # for resume training
+dim_mults: [1, 2, 4]
+mask_prob: 0.5
+noise_scale: 0.1 # only applicable for numerical data
+
+# training
+is_use_tb: True
+epochs: 200
+step_size: 50
+gamma: 0.5
+train_batch_size: 12
+test_batch_size: 12
+lr: 0.0001
+clip_grad_norm: 0.
+scheduler: cosine # step, cosine
+num_update: 10000
+
+# evaluation
+N_plot: 1
+N_autoregressive: 1

configs/combustion/wdno.yaml

@@ -0,0 +1,41 @@
+exp_name: "wdno_combustion"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "combustion"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 12
+normalizer: none # none, gaussian, range
+
+# data parameters for training
+mask_prob: 0.5
+noise_scale: 0.0 # only applicable for numerical data
+
+# model
+model_name: "wdno"
+checkpoint_path: ./results/wdno/wdno_combustion_numerical_False/2025-09-16_12-34-36/model_64000.pth
+dim: 256 
+dim_mults: [1, 2]
+wave_type: "bior1.3"
+pad_mode: "zero"
+beta_schedule: "sigmoid" 
+
+# training
+is_use_tb: True
+scheduler: cosine # step, cosine
+step_size: 0 # only applicable for step scheduler
+num_update: 100000
+train_batch_size: 16
+test_batch_size: 64
+lr: 0.0001
+clip_grad_norm: 1.0 # 0 for no clip
+
+# sampling
+sampling_timesteps: 15 # number of sampling timesteps (ddim for faster inference)
+ddim_sampling_eta: 1.
+
+# evaluation
+N_autoregressive: 1
+N_plot: 0

configs/controlled_cylinder/cno.yaml

@@ -0,0 +1,54 @@
+
+
+exp_name: "cno_control"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "controlled_cylinder"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 12
+normalizer: "gaussian" # gaussian, range, none
+
+# data parameters for training
+mask_prob: 0.1
+noise_scale: 0.1 # only applicable for numerical data
+
+# model
+model_name: "cno"
+checkpoint_path: ./results/cno/cno_control_numerical_False/2025-09-07_00-40-24/model_4400.pth # for resume training
+
+N_layers: 3 # Number of (D) or (U) blocks in the network                 
+N_res: 1,                 # Number of (R) blocks per level (except the neck)
+N_res_neck: 8,            # Number of (R) blocks in the neck
+channel_multiplier: 32,   # How the number of channels evolve?
+conv_kernel: 3,             # Size of all the kernels
+cutoff_den: 2.0001,       # Filter property 1.
+filter_size: 6,             # Filter property 2.
+
+lrelu_upsampling: 2,      # Filter property 3.
+half_width_mult: 0.8,    # Filter property 4.
+radial: False,            # Filter property 5. Is filter radial?
+batch_norm: True,         # Add BN? We do not add BN in lifting/projection layer
+out_dim: 1,               # Target dimension
+out_size: 1,              # If out_size is 1, Then out_size = in_size. Else must be int
+expand_input: False,      # Start with original in_size, or expand it (pad zeros in the spectrum)
+latent_lift_proj_dim: 64, # Intermediate latent dimension in the lifting/projection layer
+add_inv: True,            # Add invariant block (I) after the intermediate connections?
+activation: 'LeakyReLU'   # Activation function can be 'LeakyReLU' or 'lrelu'
+
+
+# training
+is_use_tb: True
+scheduler: cosine # step, cosine
+step_size: 1000 # only applicable for step scheduler
+num_update: 5000
+train_batch_size: 32
+test_batch_size: 64
+lr: 0.0003
+clip_grad_norm: 0.
+
+# evaluation
+N_autoregressive: 3
+N_plot: 0

configs/controlled_cylinder/deeponet.yaml

@@ -0,0 +1,34 @@
+exp_name: deeponet_control
+gpu: 0
+seed: 0
+results_path: ./results/
+
+# data
+dataset_name: controlled_cylinder
+dataset_root: /wutailin/real_benchmark/
+num_workers: 12
+normalizer: gaussian   # none, gaussian, range
+
+# data parameters for training
+mask_prob: 0.1
+noise_scale: 0.1 # only applicable for numerical data
+
+# model
+model_name: deeponet
+checkpoint_path: ./results/deeponet/deeponet_control_numerical_False/2025-09-08_14-27-56/model_0600.pth
+p: 256
+dropout_rate: 0.1
+
+# training
+is_use_tb: true
+scheduler: cosine # step, cosine
+step_size: 1000 # only applicable for step scheduler
+num_update: 5000
+train_batch_size: 64
+test_batch_size: 64
+lr: 0.00005
+clip_grad_norm: 0.
+
+# evaluation
+N_autoregressive: 1
+N_plot: 1

configs/controlled_cylinder/dmd.yaml

@@ -0,0 +1,26 @@
+exp_name: "dmd_control"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "controlled_cylinder"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 12
+normalizer: "gaussian"
+
+# model
+model_name: "dmd"
+is_resume: False
+n_modes: 10
+n_predict: 10
+mask_prob: 0.5
+noise_scale: 0.1 # only applicable for numerical data
+input_feature: 2
+checkpoint_path: 'no model needed'
+test_batch_size: 12
+
+
+# evaluation
+N_plot: 1
+N_autoregressive: 1

configs/controlled_cylinder/dpot_l.yaml

@@ -0,0 +1,54 @@
+exp_name: "dpot_l_control"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "controlled_cylinder"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 16
+normalizer: "gaussian" # gaussian, range, none
+
+# data parameters for training
+mask_prob: 0.1
+noise_scale: 0.5 # only applicable for numerical data
+
+# model
+model_name: "dpot"
+checkpoint_path: "./dpot_ckpts/model_L.pth"  # Large DPOT model
+model_type: "dpot"  # "dpot" or "dpot3d"
+
+# DPOT specific parameters (Large model configuration - from checkpoint)
+img_size: 128       # Pre-trained model resolution (with auto-resize for combustion)
+patch_size: 8
+in_channels: 5      
+out_channels: 4     # We follow the original DPOT workflow to pad the channels to 4, if channels is less than 4
+embed_dim: 1536     
+depth: 24            
+n_blocks: 16         
+modes: 32           # Keep default modes for AFNO
+mlp_ratio: 4        
+out_layer_dim: 1536   # Match checkpoint out layer dim
+normalize: False
+act: "gelu"
+time_agg: "exp_mlp"
+n_cls: 12            # Match checkpoint pre-training (12 datasets)
+
+# dpot multi-step specific parameters (from checkpoint)
+in_timesteps: 10
+out_timesteps: 10
+
+# training
+is_use_tb: True
+scheduler: "cosine" # step, cosine
+step_size: 1000 # only applicable for step scheduler
+num_update: 50000         
+train_batch_size: 16
+test_batch_size: 64
+lr: 0.00005
+clip_grad_norm: 1
+
+# evaluation
+N_autoregressive: 1
+N_plot: 5
+

configs/controlled_cylinder/dpot_s.yaml

@@ -0,0 +1,54 @@
+exp_name: "dpot_s_control"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "controlled_cylinder"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 16
+normalizer: "gaussian" # gaussian, range, none
+
+# data parameters for training
+mask_prob: 0.1
+noise_scale: 0.5 # only applicable for numerical data
+
+# model
+model_name: "dpot"
+checkpoint_path: "./dpot_ckpts/model_S.pth"  # Small DPOT model
+model_type: "dpot"  # "dpot" or "dpot3d"
+
+# DPOT specific parameters (Small model configuration - from checkpoint)
+img_size: 128       # Pre-trained model resolution (with auto-resize for combustion)
+patch_size: 8
+in_channels: 5      
+out_channels: 4     # We follow the original DPOT workflow to pad the channels to 4, if channels is less than 4
+embed_dim: 1024     
+depth: 6            
+n_blocks: 8         
+modes: 32           # Keep default modes for AFNO
+mlp_ratio: 1        
+out_layer_dim: 32   # Keep default out layer dim
+normalize: False
+act: "gelu"
+time_agg: "exp_mlp"
+n_cls: 12            # Match checkpoint pre-training (12 datasets)
+
+# dpot multi-step specific parameters
+in_timesteps: 10
+out_timesteps: 10
+
+# training
+is_use_tb: True
+scheduler: "cosine" # step, cosine
+step_size: 1000 # only applicable for step scheduler
+num_update: 50000         
+train_batch_size: 16
+test_batch_size: 64
+lr: 0.0005
+clip_grad_norm: 1
+
+# evaluation
+N_autoregressive: 1
+N_plot: 5
+

configs/controlled_cylinder/fno.yaml

@@ -0,0 +1,37 @@
+exp_name: "fno_control"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "controlled_cylinder"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 12
+normalizer: "gaussian" # none, gaussian, range
+
+# data parameters for training
+mask_prob: 0.1
+noise_scale: 0.1 # only applicable for numerical data
+
+# model
+model_name: "fno"
+checkpoint_path: ./results/fno/fno_control_numerical_False/2025-09-02_18-38-24/model_3760.pth
+modes1: 4
+modes2: 12
+modes3: 16
+n_layers: 4
+width: 64
+
+# training
+is_use_tb: True
+scheduler: cosine # step, cosine
+step_size: 1000 # only applicable for step scheduler
+num_update: 4000
+train_batch_size: 32
+test_batch_size: 64
+lr: 0.0001
+clip_grad_norm: 0.
+
+# evaluation
+N_autoregressive: 1
+N_plot: 0

configs/controlled_cylinder/galerkin_transformer.yaml

@@ -0,0 +1,78 @@
+exp_name: "galerkin_transformer_control"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "controlled_cylinder"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 12
+normalizer: "gaussian" # none, gaussian, range
+
+# data parameters for training
+mask_prob: 0.1
+noise_scale: 0.1 # only applicable for numerical data
+
+# model
+model_name: "galerkin_transformer"
+checkpoint_path: ./results/galerkin_transformer/galerkin_transformer_control_numerical_False/2025-09-08_15-47-23/model_4500.pth # for resume training
+
+# modes1: 4
+# modes2: 12
+# modes3: 16
+# n_layers: 4
+# width: 64
+
+pos_dim: 1 # pos dim
+n_hidden: 256
+num_feat_layers: 0
+num_encoder_layers: 1
+n_head: 4
+dim_feedforward: 256
+feat_extract_type: null
+attention_type: galerkin
+xavier_init: 0.01
+diagonal_weight: 0.01
+symmetric_init: False
+layer_norm: False
+attn_norm: True
+norm_eps: 0.0000001
+batch_norm: False
+return_attn_weight: False
+return_latent: False
+decoder_type: ifft2
+spacial_dim: 3
+spacial_fc: True
+upsample_mode: interp
+downsample_mode: interp
+freq_dim: 128
+boundary_condition: None
+num_regressor_layers: 1
+fourier_modes_x: 12
+fourier_modes_y: 16
+fourier_modes_t: 4
+regressor_activation: silu
+downscaler_activation: relu
+upscaler_activation: silu
+last_activation: True
+dropout: 0.0
+downscaler_dropout: 0.05
+upscaler_dropout: 0.0
+ffn_dropout: 0.05
+encoder_dropout: 0.05
+decoder_dropout: 0
+debug: False
+
+# training
+is_use_tb: True
+scheduler: cosine # step, cosine
+step_size: 1000 # only applicable for step scheduler
+num_update: 5000
+train_batch_size: 32
+test_batch_size: 32
+lr: 0.0001
+clip_grad_norm: 0.
+
+# evaluation
+N_autoregressive: 1
+N_plot: 1

configs/controlled_cylinder/mwt.yaml

@@ -0,0 +1,44 @@
+
+
+exp_name: "fno_control"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "controlled_cylinder"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 12
+normalizer: "gaussian" # gaussian, range, none
+
+# data parameters for training
+mask_prob: 0.1
+noise_scale: 0.1 # only applicable for numerical data
+
+
+# model
+model_name: "mwt"
+checkpoint_path: "/hupeiyan/along/realpdebench/realpdebench/results/mwt/fno_control_numerical_False/2025-09-08_08-27-42/model_4900.pth"
+
+k: 3            # wavelet filter size
+alpha: 5        # fourier filter parameter, should be less or equal to T/2
+c: 4            # increase factor
+nCZ: 4         # number of CZ blocks
+L: 0
+base: "legendre" # chebyshev
+
+
+# training
+is_use_tb: True
+scheduler: cosine # step, cosine
+step_size: 1000 # only applicable for step scheduler
+num_update: 5000
+test_interval: 200
+train_batch_size: 32
+test_batch_size: 64
+lr: 0.0005
+clip_grad_norm: 0.
+
+# evaluation
+N_autoregressive: 3
+N_plot: 5

configs/controlled_cylinder/trainsolver.yaml

@@ -0,0 +1,47 @@
+exp_name: "transolver_control"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "controlled_cylinder"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 12
+normalizer: "gaussian" # none, gaussian, range
+
+# data parameters for training
+mask_prob: 0.1
+noise_scale: 0.1 # only applicable for numerical data
+
+# model
+model_name: "transolver"
+space_dim: 5
+n_layers: 1
+n_hidden: 256
+n_head: 8
+H: 64
+W: 128
+D: 10
+fun_dim: 0 # 1
+out_dim: 3
+ref: 4
+dropout: 0.1
+act: "gelu"
+mlp_ratio: 4
+slice_num: 16
+
+checkpoint_path: "./results/transolver/transolver_control_numerical_False/2025-09-09_12-46-08/model_2100.pth"
+# training
+is_use_tb: True
+scheduler: cosine # step, cosine
+step_size: 1000 # only applicable for step scheduler
+num_update: 5000
+test_interval: 200
+train_batch_size: 32
+test_batch_size: 32
+lr: 0.001
+clip_grad_norm: 0.
+
+# evaluation
+N_autoregressive: 2
+N_plot: 0

configs/controlled_cylinder/unet.yaml

@@ -0,0 +1,34 @@
+exp_name: "unet_control"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "controlled_cylinder"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 12
+normalizer: "gaussian"
+
+# model
+model_name: "unet"
+is_resume: False
+checkpoint_path: './results/unet/unet_control_numerical_False/2025-09-04_14-39-03/model_4800.pth' # for resume training
+dim_mults: [1, 2, 4]
+mask_prob: 0.5
+noise_scale: 0.1 # only applicable for numerical data
+
+# training
+is_use_tb: True
+epochs: 200
+step_size: 50
+gamma: 0.5
+train_batch_size: 12
+test_batch_size: 12
+lr: 0.0001
+clip_grad_norm: 0.
+scheduler: cosine # step, cosine
+num_update: 10000
+
+# evaluation
+N_plot: 1
+N_autoregressive: 1

configs/controlled_cylinder/wdno.yaml

@@ -0,0 +1,41 @@
+exp_name: "wdno_control"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "controlled_cylinder"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 12
+normalizer: "none" # none, gaussian, range
+
+# data parameters for training
+mask_prob: 0.1
+noise_scale: 0.1 # only applicable for numerical data
+
+# model
+model_name: "wdno"
+checkpoint_path: ./results/wdno/wdno_control_numerical_False/2025-09-05_06-25-58/model_48000.pth
+dim: 256
+dim_mults: [1, 2, 4]
+wave_type: bior1.1
+pad_mode: "zero"
+beta_schedule: "sigmoid"
+
+# training
+is_use_tb: True
+scheduler: cosine # step, cosine
+step_size: 0 # only applicable for step scheduler
+num_update: 120000
+train_batch_size: 16
+test_batch_size: 64
+lr: 0.00001
+clip_grad_norm: 1.0
+
+# sampling
+sampling_timesteps: 25 # number of sampling timesteps (ddim for faster inference)
+ddim_sampling_eta: 1.
+
+# evaluation
+N_autoregressive: 1
+N_plot: 1

configs/cylinder/cno.yaml

@@ -0,0 +1,54 @@
+
+
+exp_name: "cno_cylinder"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "cylinder"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 12
+normalizer: "gaussian" # gaussian, range, none
+
+# data parameters for training
+mask_prob: 0.1
+noise_scale: 0.1 # only applicable for numerical data
+
+# model
+model_name: "cno"
+checkpoint_path: ./results/cno/cno_cylinder_numerical_False/2025-09-16_14-34-26/model_4100.pth # for resume training
+
+N_layers: 3 # Number of (D) or (U) blocks in the network                 
+N_res: 1,                 # Number of (R) blocks per level (except the neck)
+N_res_neck: 8,            # Number of (R) blocks in the neck
+channel_multiplier: 32,   # How the number of channels evolve?
+conv_kernel: 3,             # Size of all the kernels
+cutoff_den: 2.0001,       # Filter property 1.
+filter_size: 6,             # Filter property 2.
+
+lrelu_upsampling: 2,      # Filter property 3.
+half_width_mult: 0.8,    # Filter property 4.
+radial: False,            # Filter property 5. Is filter radial?
+batch_norm: True,         # Add BN? We do not add BN in lifting/projection layer
+out_dim: 1,               # Target dimension
+out_size: 1,              # If out_size is 1, Then out_size = in_size. Else must be int
+expand_input: False,      # Start with original in_size, or expand it (pad zeros in the spectrum)
+latent_lift_proj_dim: 64, # Intermediate latent dimension in the lifting/projection layer
+add_inv: True,            # Add invariant block (I) after the intermediate connections?
+activation: 'LeakyReLU'   # Activation function can be 'LeakyReLU' or 'lrelu'
+
+
+# training
+is_use_tb: True
+scheduler: cosine # step, cosine
+step_size: 1000 # only applicable for step scheduler
+num_update: 5000
+train_batch_size: 16
+test_batch_size: 64
+lr: 0.0003
+clip_grad_norm: 0.
+
+# evaluation
+N_autoregressive: 3
+N_plot: 0

configs/cylinder/deeponet.yaml

@@ -0,0 +1,38 @@
+exp_name: deeponet_cylinder
+gpu: 0
+seed: 0
+results_path: ./results/
+
+# data
+dataset_name: cylinder
+dataset_root: /wutailin/real_benchmark/
+num_workers: 12   
+normalizer: gaussian   # none, gaussian, range
+
+# data parameters for training
+mask_prob: 0.1
+noise_scale: 0.1 # only applicable for numerical data
+
+# model
+model_name: deeponet
+checkpoint_path: 
+  ./results/deeponet/deeponet_cylinder_numerical_False/2025-09-14_19-05-53/model_0200.pth
+p: 128
+dropout_rate: 0.1
+
+# training
+is_use_tb: true
+scheduler: cosine # step, cosine
+step_size: 1000 # only applicable for step scheduler
+num_update: 5000
+train_batch_size: 32
+test_batch_size: 64
+lr: 0.0001
+clip_grad_norm: 0.
+
+# evaluation
+N_autoregressive: 10
+N_plot: 1
+probe_diagnostic: True
+N_plot_probe: 12
+

configs/cylinder/dmd.yaml

@@ -0,0 +1,29 @@
+exp_name: "dmd_cylinder"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "cylinder"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 12
+normalizer: "gaussian"
+
+# model
+model_name: "dmd"
+is_resume: False
+n_modes: 10
+n_predict: 20
+mask_prob: 0.5
+noise_scale: 0.1 # only applicable for numerical data
+input_feature: 2
+checkpoint_path: 'no model needed'
+
+test_batch_size: 12
+
+
+# evaluation
+N_plot: 1
+N_autoregressive: 1
+probe_diagnostic: True
+N_plot_probe: 12

configs/cylinder/dpot_l.yaml

@@ -0,0 +1,54 @@
+exp_name: "dpot_l_cylinder"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "cylinder"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 16
+normalizer: "gaussian" # gaussian, range, none
+
+# data parameters for training
+mask_prob: 0.1
+noise_scale: 0.5 # only applicable for numerical data
+
+# model
+model_name: "dpot"
+checkpoint_path: "./dpot_ckpts/model_L.pth"  # Large DPOT model
+model_type: "dpot"  # "dpot" or "dpot3d"
+
+# DPOT specific parameters (Large model configuration - from checkpoint)
+img_size: 128       # Pre-trained model resolution (with auto-resize for combustion)
+patch_size: 8
+in_channels: 4      # We follow the original DPOT workflow to pad the channels to 4, if channels is less than 4
+out_channels: 4     # We follow the original DPOT workflow to pad the channels to 4, if channels is less than 4
+embed_dim: 1536     
+depth: 24            
+n_blocks: 16         
+modes: 32           # Keep default modes for AFNO
+mlp_ratio: 4        
+out_layer_dim: 1536   # Match checkpoint out layer dim
+normalize: False
+act: "gelu"
+time_agg: "exp_mlp"
+n_cls: 12            # Match checkpoint pre-training (12 datasets)
+
+# dpot multi-step specific parameters (from checkpoint)
+in_timesteps: 20
+out_timesteps: 20
+
+# training
+is_use_tb: True
+scheduler: "cosine" # step, cosine
+step_size: 1000 # only applicable for step scheduler
+num_update: 2000         
+train_batch_size: 16
+test_batch_size: 64
+lr: 0.0001
+clip_grad_norm: 1
+
+# evaluation
+N_autoregressive: 1
+N_plot: 5
+

configs/cylinder/dpot_s.yaml

@@ -0,0 +1,54 @@
+exp_name: "dpot_s_cylinder"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "cylinder"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 16
+normalizer: "gaussian" # gaussian, range, none
+
+# data parameters for training
+mask_prob: 0.1
+noise_scale: 0.5 # only applicable for numerical data
+
+# model
+model_name: "dpot"
+checkpoint_path: "./dpot_ckpts/model_S.pth"  # Small DPOT model
+model_type: "dpot"  # "dpot" or "dpot3d"
+
+# DPOT specific parameters (Small model configuration - from checkpoint)
+img_size: 128       # Pre-trained model resolution (with auto-resize for combustion)
+patch_size: 8
+in_channels: 4      # We follow the original DPOT workflow to pad the channels to 4, if channels is less than 4
+out_channels: 4     # We follow the original DPOT workflow to pad the channels to 4, if channels is less than 4
+embed_dim: 1024     
+depth: 6            
+n_blocks: 8         
+modes: 32           # Keep default modes for AFNO
+mlp_ratio: 1        
+out_layer_dim: 32   # Keep default out layer dim
+normalize: False
+act: "gelu"
+time_agg: "exp_mlp"
+n_cls: 12            # Match checkpoint pre-training (12 datasets)
+
+# dpot multi-step specific parameters
+in_timesteps: 20
+out_timesteps: 20
+
+# training
+is_use_tb: True
+scheduler: "cosine" # step, cosine
+step_size: 1000 # only applicable for step scheduler
+num_update: 2000     
+train_batch_size: 16
+test_batch_size: 64
+lr: 0.001
+clip_grad_norm: 1
+
+# evaluation
+N_autoregressive: 1
+N_plot: 5
+

configs/cylinder/fno.yaml

@@ -0,0 +1,39 @@
+exp_name: "fno_cylinder"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "cylinder"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 12
+normalizer: "gaussian" # none, gaussian, range
+
+# data parameters for training
+mask_prob: 0.1
+noise_scale: 0.1 # only applicable for numerical data
+
+# model
+model_name: "fno"
+checkpoint_path: ./results/fno/fno_cylinder_numerical_False/2025-09-12_11-17-53/model_3840.pth
+modes1: 4
+modes2: 12
+modes3: 16
+n_layers: 4
+width: 64
+
+# training
+is_use_tb: 
+scheduler: cosine # step, cosine
+step_size: 1000 # only applicable for step scheduler
+num_update: 4000
+train_batch_size: 32
+test_batch_size: 64
+lr: 0.0001
+clip_grad_norm: 0.
+
+# evaluation
+N_autoregressive: 10
+N_plot: 0
+probe_diagnostic: True
+N_plot_probe: 12

configs/cylinder/galerkin_transformer.yaml

@@ -0,0 +1,72 @@
+exp_name: "galerkin_transformer_cylinder"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "cylinder"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 12
+normalizer: "gaussian" # none, gaussian, range
+
+# data parameters for training
+mask_prob: 0.5
+noise_scale: 0.1 # only applicable for numerical data
+
+# model
+model_name: "galerkin_transformer"
+checkpoint_path: ./results/galerkin_transformer/galerkin_transformer_cylinder_real_False/2025-09-15_10-50-53/model_2200.pth
+
+pos_dim: 1 # pos dim
+n_hidden: 256
+num_feat_layers: 0
+num_encoder_layers: 1
+n_head: 4
+dim_feedforward: 256
+feat_extract_type: null
+attention_type: galerkin
+xavier_init: 0.01
+diagonal_weight: 0.01
+symmetric_init: False
+layer_norm: False
+attn_norm: True
+norm_eps: 0.0000001
+batch_norm: False
+return_attn_weight: False
+return_latent: False
+decoder_type: ifft2
+spacial_dim: 3
+spacial_fc: True
+upsample_mode: interp
+downsample_mode: interp
+freq_dim: 128
+boundary_condition: None
+num_regressor_layers: 1
+fourier_modes_x: 16
+fourier_modes_y: 20
+fourier_modes_t: 4
+regressor_activation: silu
+downscaler_activation: relu
+upscaler_activation: silu
+last_activation: True
+dropout: 0.0
+downscaler_dropout: 0.05
+upscaler_dropout: 0.0
+ffn_dropout: 0.05
+encoder_dropout: 0.05
+decoder_dropout: 0
+debug: False
+
+# training
+is_use_tb: True
+scheduler: cosine # step, cosine
+step_size: 1000 # only applicable for step scheduler
+num_update: 5000
+train_batch_size: 16
+test_batch_size: 16
+lr: 0.01
+clip_grad_norm: 0.
+
+# evaluation
+N_autoregressive: 1
+N_plot: 1

configs/cylinder/mwt.yaml

@@ -0,0 +1,44 @@
+
+
+exp_name: "mwt_cylinder"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "cylinder"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 12
+normalizer: "gaussian" # gaussian, range, none
+
+# data parameters for training
+mask_prob: 0.1
+noise_scale: 0.1 # only applicable for numerical data
+
+
+# model
+model_name: "mwt"
+checkpoint_path: "./results/mwt/mwt_cylinder_numerical_False/2025-09-16_07-49-46/model_4200.pth"
+
+k: 3            # wavelet filter size
+alpha: 5        # fourier filter parameter, should be less or equal to T/2
+c: 4            # increase factor
+nCZ: 4         # number of CZ blocks
+L: 0
+base: "legendre" # chebyshev
+
+
+# training
+is_use_tb: True
+scheduler: cosine # step, cosine
+step_size: 1000 # only applicable for step scheduler
+num_update: 5000
+test_interval: 200
+train_batch_size: 32
+test_batch_size: 64
+lr: 0.001
+clip_grad_norm: 0.
+
+# evaluation
+N_autoregressive: 3
+N_plot: 5

configs/cylinder/trainsolver.yaml

@@ -0,0 +1,47 @@
+exp_name: "transolver_cylinder"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "cylinder"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 12
+normalizer: "gaussian" # none, gaussian, range
+
+# data parameters for training
+mask_prob: 0.1
+noise_scale: 0.1 # only applicable for numerical data
+
+# model
+model_name: "transolver"
+space_dim: 3
+n_layers: 1
+n_hidden: 256
+n_head: 8
+H: 128
+W: 64
+D: 20
+fun_dim: 0 # 1
+out_dim: 3
+ref: 4
+dropout: 0.1
+act: "gelu"
+mlp_ratio: 4
+slice_num: 16
+
+checkpoint_path: "./results/transolver/transolver_cylinder_real_True/2025-09-14_11-11-47/model_0800.pth"
+# training
+is_use_tb: True
+scheduler: cosine # step, cosine
+step_size: 1000 # only applicable for step scheduler
+num_update: 5000
+test_interval: 200
+train_batch_size: 16
+test_batch_size: 16
+lr: 0.0007
+clip_grad_norm: 0.
+
+# evaluation
+N_autoregressive: 3
+N_plot: 5

configs/cylinder/unet.yaml

@@ -0,0 +1,36 @@
+exp_name: "unet_cylinder"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "cylinder"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 12
+normalizer: "gaussian"
+
+# model
+model_name: "unet"
+is_resume: False
+checkpoint_path: './results/unet/unet_cylinder_numerical_False/2025-08-27_00-44-01/model_6400.pth' # for resume training
+dim_mults: [1, 2, 4]
+mask_prob: 0.5
+noise_scale: 0.1 # only applicable for numerical data
+
+# training
+is_use_tb: True
+epochs: 200
+step_size: 50
+gamma: 0.5
+train_batch_size: 12
+test_batch_size: 12
+lr: 0.0001
+clip_grad_norm: 0.
+scheduler: cosine # step, cosine
+num_update: 10000
+
+# evaluation
+N_plot: 1
+N_autoregressive: 5
+probe_diagnostic: True
+N_plot_probe: 12

configs/cylinder/wdno.yaml

@@ -0,0 +1,43 @@
+exp_name: "wdno_cylinder"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "cylinder"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 12
+normalizer: "none" # none, gaussian, range
+
+# data parameters for training
+mask_prob: 0.1
+noise_scale: 0.1 # only applicable for numerical data
+
+# model
+model_name: "wdno"
+checkpoint_path: ./results/wdno/wdno_cylinder_numerical_False/2025-09-17_06-05-39/model_60000.pth
+dim: 256
+dim_mults: [1, 2]
+wave_type: "bior1.1"
+pad_mode: "zero"
+beta_schedule: "sigmoid"
+
+# training
+is_use_tb: True
+scheduler: cosine # step, cosine
+step_size: 0 # only applicable for step scheduler
+num_update: 120000
+train_batch_size: 16
+test_batch_size: 64
+lr: 0.00005
+clip_grad_norm: 1.0
+
+# sampling
+sampling_timesteps: 10 # number of sampling timesteps (ddim for faster inference)
+ddim_sampling_eta: 1
+
+# evaluation
+N_autoregressive: 5
+N_plot: 0
+probe_diagnostic: True
+N_plot_probe: 0

configs/foil/cno.yaml

@@ -0,0 +1,54 @@
+
+
+exp_name: "cno_foil"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "foil"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 12
+normalizer: "gaussian" # gaussian, range, none
+
+# data parameters for training
+mask_prob: 0.1
+noise_scale: 0.1 # only applicable for numerical data
+
+# model
+model_name: "cno"
+checkpoint_path: ./results/cno/cno_foil_numerical_False/2025-09-18_05-34-15/model_4900.pth # for resume training
+
+N_layers: 3 # Number of (D) or (U) blocks in the network                 
+N_res: 1,                 # Number of (R) blocks per level (except the neck)
+N_res_neck: 8,            # Number of (R) blocks in the neck
+channel_multiplier: 32,   # How the number of channels evolve?
+conv_kernel: 3,             # Size of all the kernels
+cutoff_den: 2.0001,       # Filter property 1.
+filter_size: 6,             # Filter property 2.
+
+lrelu_upsampling: 2,      # Filter property 3.
+half_width_mult: 0.8,    # Filter property 4.
+radial: False,            # Filter property 5. Is filter radial?
+batch_norm: True,         # Add BN? We do not add BN in lifting/projection layer
+out_dim: 1,               # Target dimension
+out_size: 1,              # If out_size is 1, Then out_size = in_size. Else must be int
+expand_input: False,      # Start with original in_size, or expand it (pad zeros in the spectrum)
+latent_lift_proj_dim: 64, # Intermediate latent dimension in the lifting/projection layer
+add_inv: True,            # Add invariant block (I) after the intermediate connections?
+activation: 'LeakyReLU'   # Activation function can be 'LeakyReLU' or 'lrelu'
+
+
+# training
+is_use_tb: True
+scheduler: cosine # step, cosine
+step_size: 1000 # only applicable for step scheduler
+num_update: 5000
+train_batch_size: 16
+test_batch_size: 64
+lr: 0.0003
+clip_grad_norm: 0.
+
+# evaluation
+N_autoregressive: 1
+N_plot: 1

configs/foil/deeponet.yaml

@@ -0,0 +1,35 @@
+exp_name: deeponet_foil
+gpu: 0
+seed: 1234567
+results_path: ./results/
+
+# data
+dataset_name: foil
+dataset_root: /wutailin/real_benchmark/
+num_workers: 12
+normalizer: gaussian   # none, gaussian, range
+
+# data parameters for training
+mask_prob: 0.1
+noise_scale: 0.1 # only applicable for numerical data
+
+# model
+model_name: deeponet
+checkpoint_path: ./results/deeponet/deeponet_foil_numerical_False/2025-09-19_01-39-51/model_0160.pth
+
+p: 256
+dropout_rate: 0
+
+# training
+is_use_tb: true
+scheduler: cosine # step, cosine
+step_size: 1000 # only applicable for step scheduler
+num_update: 4000
+train_batch_size: 32
+test_batch_size: 64
+lr: 0.0001
+clip_grad_norm: 0.
+
+# evaluation
+N_autoregressive: 1
+N_plot: 1

configs/foil/dmd.yaml

@@ -0,0 +1,26 @@
+exp_name: "dmd_foil"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "foil"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 12
+normalizer: "gaussian"
+
+# model
+model_name: "dmd"
+is_resume: False
+n_modes: 10
+n_predict: 20
+mask_prob: 0.5
+noise_scale: 0.1 # only applicable for numerical data
+input_feature: 2
+checkpoint_path: 'no model needed'
+test_batch_size: 12
+
+
+# evaluation
+N_plot: 1
+N_autoregressive: 1

configs/foil/dpot_l.yaml

@@ -0,0 +1,54 @@
+exp_name: "dpot_l_foil"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "foil"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 16
+normalizer: "gaussian" # gaussian, range, none
+
+# data parameters for training
+mask_prob: 0.1
+noise_scale: 0.5 # only applicable for numerical data
+
+# model
+model_name: "dpot"
+checkpoint_path: "./dpot_ckpts/model_L.pth"  # Large DPOT model
+model_type: "dpot"  # "dpot" or "dpot3d"
+
+# DPOT specific parameters (Large model configuration - from checkpoint)
+img_size: 128       # Pre-trained model resolution (with auto-resize for combustion)
+patch_size: 8
+in_channels: 4      # We follow the original DPOT workflow to pad the channels to 4, if channels is less than 4
+out_channels: 4     # We follow the original DPOT workflow to pad the channels to 4, if channels is less than 4
+embed_dim: 1536     
+depth: 24            
+n_blocks: 16         
+modes: 32           # Keep default modes for AFNO
+mlp_ratio: 4        
+out_layer_dim: 1536   # Match checkpoint out layer dim
+normalize: False
+act: "gelu"
+time_agg: "exp_mlp"
+n_cls: 12            # Match checkpoint pre-training (12 datasets)
+
+# dpot multi-step specific parameters (from checkpoint)
+in_timesteps: 20
+out_timesteps: 20
+
+# training
+is_use_tb: True
+scheduler: "cosine" # step, cosine
+step_size: 1000 # only applicable for step scheduler
+num_update: 10000     
+train_batch_size: 16
+test_batch_size: 64
+lr: 0.0001
+clip_grad_norm: 1
+
+# evaluation
+N_autoregressive: 1
+N_plot: 5
+

configs/foil/dpot_s.yaml

@@ -0,0 +1,54 @@
+exp_name: "dpot_s_foil"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "foil"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 16
+normalizer: "gaussian" # gaussian, range, none
+
+# data parameters for training
+mask_prob: 0.1
+noise_scale: 0.5 # only applicable for numerical data
+
+# model
+model_name: "dpot"
+checkpoint_path: "./dpot_ckpts/model_S.pth"  # Small DPOT model
+model_type: "dpot"  # "dpot" or "dpot3d"
+
+# DPOT specific parameters (Small model configuration - from checkpoint)
+img_size: 128       # Pre-trained model resolution (with auto-resize for combustion)
+patch_size: 8
+in_channels: 4      # We follow the original DPOT workflow to pad the channels to 4, if channels is less than 4
+out_channels: 4     # We follow the original DPOT workflow to pad the channels to 4, if channels is less than 4
+embed_dim: 1024     
+depth: 6            
+n_blocks: 8         
+modes: 32           # Keep default modes for AFNO
+mlp_ratio: 1        
+out_layer_dim: 32   # Keep default out layer dim
+normalize: False
+act: "gelu"
+time_agg: "exp_mlp"
+n_cls: 12            # Match checkpoint pre-training (12 datasets)
+
+# dpot multi-step specific parameters
+in_timesteps: 20
+out_timesteps: 20
+
+# training
+is_use_tb: True
+scheduler: "cosine" # step, cosine
+step_size: 1000 # only applicable for step scheduler
+num_update: 10000     
+train_batch_size: 16
+test_batch_size: 64
+lr: 0.001
+clip_grad_norm: 1
+
+# evaluation
+N_autoregressive: 1
+N_plot: 5
+

configs/foil/fno.yaml

@@ -0,0 +1,37 @@
+exp_name: "fno_foil"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "foil"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 12
+normalizer: "gaussian" # none, gaussian, range
+
+# data parameters for training
+mask_prob: 0.1
+noise_scale: 0.1 # only applicable for numerical data
+
+# model
+model_name: "fno"
+checkpoint_path: ./results/fno/fno_foil_numerical_False/2025-09-17_14-33-39/model_3760.pth
+modes1: 4
+modes2: 12
+modes3: 16
+n_layers: 4
+width: 64
+
+# training
+is_use_tb: True
+scheduler: cosine # step, cosine
+step_size: 1000 # only applicable for step scheduler
+num_update: 4000
+train_batch_size: 32
+test_batch_size: 64
+lr: 0.0001
+clip_grad_norm: 0.
+
+# evaluation
+N_autoregressive: 1
+N_plot: 0

configs/foil/galerkin_transformer.yaml

@@ -0,0 +1,78 @@
+exp_name: "galerkin_transformer_foil"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "foil"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 12
+normalizer: "gaussian" # none, gaussian, range
+
+# data parameters for training
+mask_prob: 0.1
+noise_scale: 0.1 # only applicable for numerical data
+
+# model
+model_name: "galerkin_transformer"
+checkpoint_path: ./results/galerkin_transformer/galerkin_transformer_foil_numerical_False/2025-09-17_19-23-35/model_0600.pth
+
+# modes1: 4
+# modes2: 12
+# modes3: 16
+# n_layers: 4
+# width: 64
+
+pos_dim: 1 # pos dim
+n_hidden: 256
+num_feat_layers: 0
+num_encoder_layers: 1
+n_head: 4
+dim_feedforward: 256
+feat_extract_type: null
+attention_type: galerkin
+xavier_init: 0.01
+diagonal_weight: 0.01
+symmetric_init: False
+layer_norm: False
+attn_norm: True
+norm_eps: 0.0000001
+batch_norm: False
+return_attn_weight: False
+return_latent: False
+decoder_type: ifft2
+spacial_dim: 3
+spacial_fc: True
+upsample_mode: interp
+downsample_mode: interp
+freq_dim: 128
+boundary_condition: None
+num_regressor_layers: 1
+fourier_modes_x: 12
+fourier_modes_y: 16
+fourier_modes_t: 4
+regressor_activation: silu
+downscaler_activation: relu
+upscaler_activation: silu
+last_activation: True
+dropout: 0.0
+downscaler_dropout: 0.05
+upscaler_dropout: 0.0
+ffn_dropout: 0.05
+encoder_dropout: 0.05
+decoder_dropout: 0
+debug: False
+
+# training
+is_use_tb: True
+scheduler: cosine # step, cosine
+step_size: 1000 # only applicable for step scheduler
+num_update: 5000
+train_batch_size: 16
+test_batch_size: 16
+lr: 0.0001
+clip_grad_norm: 0.
+
+# evaluation
+N_autoregressive: 1
+N_plot: 1

configs/foil/mwt.yaml

@@ -0,0 +1,44 @@
+
+
+exp_name: "mwt_foil"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "foil"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 12
+normalizer: "gaussian" # gaussian, range, none
+
+# data parameters for training
+mask_prob: 0.1
+noise_scale: 0.1 # only applicable for numerical data
+
+
+# model
+model_name: "mwt"
+checkpoint_path: ./results/mwt/mwt_foil_numerical_False/2025-09-18_08-56-12/model_2000.pth
+
+k: 3            # wavelet filter size
+alpha: 5        # fourier filter parameter, should be less or equal to T/2
+c: 4            # increase factor
+nCZ: 4         # number of CZ blocks
+L: 0
+base: "legendre" # chebyshev
+
+
+# training
+is_use_tb: True
+scheduler: cosine # step, cosine
+step_size: 1000 # only applicable for step scheduler
+num_update: 5000
+test_interval: 200
+train_batch_size: 32
+test_batch_size: 64
+lr: 0.0005
+clip_grad_norm: 0.
+
+# evaluation
+N_autoregressive: 1
+N_plot: 1

configs/foil/trainsolver.yaml

@@ -0,0 +1,47 @@
+exp_name: "foil"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "foil"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 12
+normalizer: "gaussian" # none, gaussian, range
+
+# data parameters for training
+mask_prob: 0.1
+noise_scale: 0.1 # only applicable for numerical data
+
+# model
+model_name: "transolver"
+space_dim: 3
+n_layers: 1
+n_hidden: 256
+n_head: 8
+H: 128
+W: 64
+D: 20
+fun_dim: 0 # 1
+out_dim: 3
+ref: 4
+dropout: 0.1
+act: "gelu"
+mlp_ratio: 4
+slice_num: 16
+
+checkpoint_path: ./results/transolver/foil_numerical_False/2025-09-18_01-26-24/model_1100.pth
+# training
+is_use_tb: True
+scheduler: cosine # step, cosine
+step_size: 1000 # only applicable for step scheduler
+num_update: 5000
+test_interval: 200
+train_batch_size: 16
+test_batch_size: 16
+lr: 0.0007
+clip_grad_norm: 0.
+
+# evaluation
+N_autoregressive: 1
+N_plot: 1

configs/foil/unet.yaml

@@ -0,0 +1,34 @@
+exp_name: "unet_foil"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "foil"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 12
+normalizer: "gaussian"
+
+# model
+model_name: "unet"
+is_resume: False
+checkpoint_path: './results/unet/unet_foil_numerical_False/2025-09-19_08-30-43/model_8800.pth' # for resume training
+dim_mults: [1, 2, 4]
+mask_prob: 0.5
+noise_scale: 0.1 # only applicable for numerical data
+
+# training
+is_use_tb: True
+epochs: 200
+step_size: 50
+gamma: 0.5
+train_batch_size: 12
+test_batch_size: 12
+lr: 0.0001
+clip_grad_norm: 0.
+scheduler: cosine # step, cosine
+num_update: 10000
+
+# evaluation
+N_plot: 1
+N_autoregressive: 1

configs/foil/wdno.yaml

@@ -0,0 +1,41 @@
+exp_name: "wdno_foil"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "foil"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 12
+normalizer: "none" # none, gaussian, range
+
+# data parameters for training
+mask_prob: 0.1
+noise_scale: 0.1 # only applicable for numerical data
+
+# model
+model_name: "wdno"
+checkpoint_path: ./results/wdno/wdno_foil_numerical_False/2025-09-17_17-49-59/model_79200.pth
+dim: 256
+dim_mults: [1, 2]
+wave_type: bior1.3
+pad_mode: "zero"
+beta_schedule: "sigmoid"
+
+# training
+is_use_tb: True
+scheduler: cosine # step, cosine
+step_size: 0 # only applicable for step scheduler
+num_update: 120000
+train_batch_size: 16
+test_batch_size: 64
+lr: 0.00005
+clip_grad_norm: 1.0
+
+# sampling
+sampling_timesteps: 20 # number of sampling timesteps (ddim for faster inference)
+ddim_sampling_eta: 1.
+
+# evaluation
+N_autoregressive: 1
+N_plot: 0

configs/fsi/cno.yaml

@@ -0,0 +1,53 @@
+
+
+exp_name: "cno_fsi"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "fsi"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 12
+normalizer: "gaussian" # gaussian, range, none
+
+# data parameters for training
+mask_prob: 0.5
+noise_scale: 0.1 # only applicable for numerical data
+
+# model
+model_name: "cno"
+checkpoint_path: ./results/cno/cno_fsi_numerical_False/numerical1/model_4300.pth # for resume training
+N_layers: 3 # Number of (D) or (U) blocks in the network
+                 
+N_res: 1,                 # Number of (R) blocks per level (except the neck)
+N_res_neck: 8,            # Number of (R) blocks in the neck
+channel_multiplier: 32,   # How the number of channels evolve?
+conv_kernel: 3,             # Size of all the kernels
+cutoff_den: 2.0001,       # Filter property 1.
+filter_size: 6,             # Filter property 2.
+lrelu_upsampling: 2,      # Filter property 3.
+half_width_mult: 0.8,    # Filter property 4.
+radial: False,            # Filter property 5. Is filter radial?
+batch_norm: True,         # Add BN? We do not add BN in lifting/projection layer
+out_dim: 1,               # Target dimension
+out_size: 1,              # If out_size is 1, Then out_size = in_size. Else must be int
+expand_input: False,      # Start with original in_size, or expand it (pad zeros in the spectrum)
+latent_lift_proj_dim: 64, # Intermediate latent dimension in the lifting/projection layer
+add_inv: True,            # Add invariant block (I) after the intermediate connections?
+activation: 'LeakyReLU'   # Activation function can be 'LeakyReLU' or 'lrelu'
+
+
+# training
+is_use_tb: True
+scheduler: cosine # step, cosine
+step_size: 1000 # only applicable for step scheduler
+num_update: 5000
+train_batch_size: 32
+test_batch_size: 64
+lr: 0.001
+clip_grad_norm: 0.
+
+# evaluation
+N_autoregressive: 3
+N_plot: 1

configs/fsi/deeponet.yaml

@@ -0,0 +1,37 @@
+exp_name: "deeponet_fsi"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "fsi"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 12
+normalizer: "gaussian" # none, gaussian, range
+
+# data parameters for training
+mask_prob: 0.5
+noise_scale: 0.1 # only applicable for numerical data
+
+
+# model
+model_name: "deeponet"
+checkpoint_path: ./results/deeponet/deeponet_fsi_numerical_False/2025-08-28_04-21-47/model_3100.pth
+p : 64
+dropout_rate: 0.1
+
+
+
+# training
+is_use_tb: True
+scheduler: cosine # step, cosine
+step_size: 1000 # only applicable for step scheduler
+num_update: 5000
+train_batch_size: 64
+test_batch_size: 64
+lr: 0.001
+clip_grad_norm: 0.
+
+# evaluation
+N_autoregressive: 1
+N_plot: 1

configs/fsi/dmd.yaml

@@ -0,0 +1,25 @@
+exp_name: "dmd_fsi"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "fsi"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 10
+normalizer: "gaussian"
+
+# model
+model_name: "dmd"
+is_resume: False
+n_modes: 10
+n_predict: 20
+mask_prob: 0.5
+noise_scale: 0.1 # only applicable for numerical data
+input_feature: 2
+checkpoint_path: 'no model needed'
+test_batch_size: 12
+
+# evaluation
+N_plot: 1
+N_autoregressive: 1

configs/fsi/dpot_l.yaml

@@ -0,0 +1,54 @@
+exp_name: "dpot_l_fsi"
+gpu: 0
+seed: 0
+results_path: "./results/"
+
+# data
+dataset_name: "fsi"
+dataset_root: "/wutailin/real_benchmark/"
+num_workers: 16
+normalizer: "gaussian" # gaussian, range, none
+
+# data parameters for training
+mask_prob: 0.1
+noise_scale: 0.5 # only applicable for numerical data
+
+# model
+model_name: "dpot"
+checkpoint_path: "./dpot_ckpts/model_L.pth"  # Large DPOT model
+model_type: "dpot"  # "dpot" or "dpot3d"
+
+# DPOT specific parameters (Large model configuration - from checkpoint)
+img_size: 128       # Pre-trained model resolution (with auto-resize for combustion)
+patch_size: 8
+in_channels: 4      # We follow the original DPOT workflow to pad the channels to 4, if channels is less than 4
+out_channels: 4     # We follow the original DPOT workflow to pad the channels to 4, if channels is less than 4
+embed_dim: 1536     
+depth: 24            
+n_blocks: 16         
+modes: 32           # Keep default modes for AFNO
+mlp_ratio: 4        
+out_layer_dim: 1536   # Match checkpoint out layer dim
+normalize: False
+act: "gelu"
+time_agg: "exp_mlp"
+n_cls: 12            # Match checkpoint pre-training (12 datasets)
+
+# dpot multi-step specific parameters (from checkpoint)
+in_timesteps: 20
+out_timesteps: 20
+
+# training
+is_use_tb: True
+scheduler: "cosine" # step, cosine
+step_size: 1000 # only applicable for step scheduler
+num_update: 10000         
+train_batch_size: 16
+test_batch_size: 64
+lr: 0.0001
+clip_grad_norm: 1
+
+# evaluation
+N_autoregressive: 1
+N_plot: 5
+