TL; DR

INPUT ADAPTERS + <u>**SHARED</u> CORE** + DECODER

One model per problem — One model, many tasks (Theirs) — One pretrained model, fine-tuned per task / One model, many PDEs

How they justify calling it a "foundation model"

They define a FM for the domain operationally, by reuse and extension rather than corpus scale. (They don't argue from scale.)

Main Flow

Symbol Meaning Value
B batch size —
d / D raw / padded input length 84–102 / 101
S, P latent grid side, latent points 48, 2,304
W channel width 48
E, L experts per layer, layers 8, 3
m, k wavelet subband side, Fourier modes kept 24–31, 13–16
F latent features per point 3
N, M mesh nodes, decoder edges 1,733–65,341, about 20 per node
c output fields 1–3
Step Function Shapes
[1] encode_task_input (B, d) → normalise, pad_to_common_length (B, D) → to_latent_cells (B, P) → (B, 1, S, S) → lift_channels (B, W, S, S)
[2] run_core_layer (×3) router logits (B, W, E) → softmax → run_all_experts (B, E, W, S, S) → mix_experts_per_channel (B, W, S, S) → Mish(skip + mixed)
[3] decode_to_native_mesh feature_head (B, F, S, S) → latent point features (B, P, F) → edge_features (B, M, 12) → weighted_edge_messages (B, M, c) → summed per node (B, N, c)
Part What it does Params
1. Input adapters • Normalize each task's vector
• zero-pad to 101
• Linear → 48×48 latent, 1×1 conv up to 48 channels ~0.24 M
2. Shared core • 3 layers × 8 Daubechies experts (db1–db8);
• each layer combines a dense per-channel gate mix with a 1×1 skip, then Mish 187.1 M experts + 2.36 M gates
3. Decoder • 1×1 projection to 3 features per latent point
• then a 12-feature kernel MLP over neighbours averaged onto the task's native mesh ~0.02 M

F1–F6:

  1. F1, shared pretrained core.
  2. F2, structurally different physics. The training tasks come from different physics: incompressible flow, radiation transport and elastoplastic solid mechanics. They are not one equation with different coefficients.
  3. F3, different kinds of input. Each task's input is a vector of 84 to 102 numbers, and the numbers mean different things in each task: lid-velocity samples, sensor counts, or a displacement history (盖板速度样本、传感器计数或位移历史). For pretraining, each vector is normalized, zero-padded to length 101 and linearly projected onto the 48×48 latent grid.
  4. F4, native output geometries. The decoder takes the target mesh's node coordinates and outputs values directly at those nodes. It works by kernel-weighted aggregation from nearby latent points.
  5. F5, small-fraction extension.