HNO-Hammerstein

Hammerstein Neural Operator with Latent Dynamics for Magnetic Hysteresis

This package implements a Neural Operator based on the classic Hammerstein structure, designed to model non-linear dynamic systems with memory, such as magnetic hysteresis.

Unlike traditional neural networks (DeepONet, FNO), the HNO architecture explicitly separates the static non-linearity (input) from the latent dynamics (memory), ensuring physical interpretability and generalization across different temporal resolutions.

Arquitetura / Architecture

O fluxo de dados é:

1. Não Linearidade Estática (Static Non-linearity): z(t) = N_theta(H(t))
2. Contexto Global (Global Context): c = E_phi(z(t))
3. Dinâmica Latente (Latent Dynamics): s(t) = Integrate(f_psi(s, z, c))
4. Saída (Output): B(t) = G_omega(s(t), z(t))

Instalação / Installation

pip install hno-hammerstein

#Uso Rápido / Quick Start
import torch
from hno_hammerstein import HNOHammerstein

# Instancia a rede / Instantiate the network
model = HNOHammerstein(
    nonlinear_dim=32,
    context_dim=32,
    state_dim=64,
    hidden_dim=64,
    dropout=0.05
)

# Cria uma entrada fictícia / Create a dummy input: (batch=1, Nt=198, 1)
h = torch.randn(1, 198, 1)

# Predição / Prediction
B_pred = model(h)
print(B_pred.shape)  # torch.Size([1, 198, 1])


#Licença / License
Este projeto está sob a licença MIT. Veja o arquivo LICENSE para mais detalhes.
This project is licensed under the MIT License. See the LICENSE file for details.

#Citação / Citation
bibtex
@mastersthesis{freitas2026hno,
  title={Hammerstein Neural Operator (HNO): Um Novo Operador Neural para Modelagem de Histerese Magnética},
  authors={Freitas, Daniel,  Almeida, Luiz}
  school={Universidade Federal do ABC (UFABC)},
  year={2026}
}