README.md

metadata

tags:
  - fourier-neural-operator
  - seismic
  - structural-engineering
  - time-series
  - regression
  - pytorch
library_name: pytorch
pipeline_tag: other

SeFNO — Seismic Floor Acceleration Response Prediction (FNO v1.0+)

Pre-trained Fourier Neural Operator (FNO) models for predicting multi-floor acceleration response time histories of MDOF shear buildings subjected to seismic ground motions.

Code: github.com/HKUJasonJiang/Seismic-FNO
Dataset: []

---

Task

Given a scaled ground motion acceleration time series (3 000 time steps, 50 Hz), the model predicts the roof-floor acceleration response of a target building — a pure regression task over 1-D signals.

Input	Shape	Description
Ground motion	(1, 3000)	Scaled accelerogram (m/s²)

Output	Shape	Description
Floor acceleration	(1, 3000)	Roof acceleration response (m/s²)

---

Available Models

Baseline Models

Three FNO configurations trained for 50 epochs on the full KNET dataset (3 474 GMs × 57 amplitude scale factors, 250 building configurations):

Folder	Hidden (h)	Modes (m)	Layers (l)	Parameters
Base-FNO_v1.0+_h64_m64_l4_e50_*/	64	64	4	~3 M
Large-FNO_v1.0+_h64_m512_l8_e50_*/	64	512	8	~12 M
Huge-FNO_v1.0+_h128_m1024_l12_e50_*/	128	1024	12	~48 M

Experimental Series

Folder	Runs	Purpose
Test-Series (Test-1~10)/	10	Hyper-parameter sweep (modes, layers, hidden channels)
Efficiency-Series (E-Base, E-Test-1~14)/	15	Dataset-size ablation (varying number of GMs and scale factors)

Each model folder contains:

<model_folder>/
├── model/
│   └── fno_best.pth          # Best checkpoint (lowest validation loss)
└── details/
    ├── training_log.csv       # Epoch-by-epoch MSE / RMSE / MAE / R²
    ├── training_config.txt    # Full hyperparameter configuration
    ├── dataset_indices.pkl    # Reproducible train / val / test split indices
    └── test_results.txt       # Final test-set metrics

---

Usage

Install dependencies

git clone https://github.com/HKUJasonJiang/Seismic-FNO.git
cd Seismic-FNO
pip install -r requirement.txt   # Windows
# pip install -r requirement_linux.txt  # Linux

Load a checkpoint and run inference

import torch
from neuralop.models import FNO

def load_fno(checkpoint_path, n_modes, hidden_channels, n_layers, device="cuda"):
    model = FNO(
        n_modes                  = (n_modes,),
        hidden_channels          = hidden_channels,
        in_channels              = 1,
        out_channels             = 1,
        n_layers                 = n_layers,
        projection_channel_ratio = 2,
        domain_padding           = 0.1,
    )
    ckpt  = torch.load(checkpoint_path, map_location=device, weights_only=False)
    state = ckpt.get("model_state_dict", ckpt)
    # strip torch.compile prefix if present
    state = {(k[10:] if k.startswith("_orig_mod.") else k): v for k, v in state.items()}
    model.load_state_dict(state)
    return model.to(device).eval()


device = "cuda" if torch.cuda.is_available() else "cpu"

# --- Base model ---
model = load_fno(
    checkpoint_path = "output/Base-FNO_v1.0+_h64_m64_l4_e50_.../model/fno_best.pth",
    n_modes         = 64,
    hidden_channels = 64,
    n_layers        = 4,
    device          = device,
)

# --- Inference on a single ground motion ---
import numpy as np

gm_array = np.load("my_ground_motion.npy")   # shape (3000,), unit m/s²
x = torch.from_numpy(gm_array).float().unsqueeze(0).unsqueeze(0).to(device)  # (1,1,3000)

with torch.no_grad():
    pred = model(x).cpu().numpy().squeeze()   # shape (3000,)

print("Predicted roof acceleration shape:", pred.shape)

Quick review notebook

Open quick_inference.ipynb in the cloned repository to run inference on the held-out test set and visualise time-history and Fourier amplitude spectrum comparisons interactively.

---

Training Details

Property	Value
Framework	PyTorch ≥ 2.0 + neuralop
Loss	MSE
Optimizer	AdamW
LR schedule	StepLR (step=20, γ=0.5)
Batch size	2 560
Epochs	50 (baseline models)
Train / Val / Test split	70 % / 20 % / 10 % (grouped by GM to prevent leakage)
Dataset	KNET — 3 474 GMs × 57 amplitude scale factors × 250 buildings

---

Citation

If you use these models, please cite (forthcoming):

@misc{jiang2025sefno,
  title  = {SeismicFNO: Fourier Neural Operators for Seismic Structural Response Prediction},
  author = {Jason Jiang},
  year   = {2025},
  url    = {https://github.com/HKUJasonJiang/Seismic-FNO}
}

License

Creative Commons Attribution 4.0 (CC BY 4.0)