paper:
  title: Thompson Sampling in Function Spaces via Neural Operators
  arxiv: '2506.21894'
  target: Figure 1 and Table 1 Darcy negative total flow rates NOTS
paths:
  data_dir: data
  output_dir: .
  train_file: darcy_train_16.pt
  test_file: darcy_test_16.pt
model:
  n_modes: [16, 16]
  hidden_channels: 64
  in_channels: 1
  out_channels: 1
  n_layers: 4
  lifting_channel_ratio: 2
  projection_channel_ratio: 2
  positional_embedding: grid
  channel_mlp_expansion: 0.5
  channel_mlp_dropout: 0.0
  channel_mlp_skip: soft-gating
  fno_skip: linear
  norm: null
  domain_padding: null
  factorization: null
  fno_block_precision: full
training:
  epochs: 10
  batch_size: 32
  lr: 0.001
  lambda_reg: 0.0001
  optimizer: AdamW
  weight_decay: 0.0
  scheduler: CosineAnnealingLR
  normalization: none
  shuffle: false
  precision: float32
experiment:
  iterations: 300
  n_init: 1
  seeds: [42, 43, 44, 45, 46, 47, 48, 49, 50, 51]
  pool_size: 1000
  resolution: 16
  noise_variance_fraction: 0.01
  shrink_pool: false
  prediction_batch_size: 64
  run_random_baseline: true
  std_ddof: 0
  objective_sign: appendix_D1_negative_physical_outflow
runtime:
  require_case_slurm_job: false
  device: cuda
  num_threads: 8
reference:
  NOTS: {mean: 0.012, std: 0.001}
  RS: {mean: 0.872, std: 0.022}
  BO: {mean: 0.703, std: 0.045}
  BFO: {mean: 0.788, std: 0.066}
  GP-TS: {mean: 0.674, std: 0.050}
  STO-NTS: {mean: 0.068, std: 0.002}