# Port-Lifted Dynamics Network

- ID: 1433
- Canonical URL: https://synthcore.org/idea/1433/port-lifted-dynamics-network
- API JSON: https://synthcore.org/api/idea/1433.json
- API Markdown: https://synthcore.org/api/idea/1433.md
- Verification status: unverified
- Source: [arXiv:2607.26248](https://arxiv.org/abs/2607.26248)
- Category: architecture
- Solves: stability, accuracy, sample-efficiency
- ML areas: world-model, graph-nn, loss, training-dynamics
- Math tags: dynamical-systems, geometry, pde, linear-algebra
- Ratings: usefulness 6/10; difficulty 5/10; novelty 6/10

## Idea description

Represent every predicted displacement and velocity as the sum of a prescribed boundary lift and a learned residual that is identically zero on the Dirichlet boundary. Feed the boundary velocity into the model through an explicit distributed-port feature and train an energy-balance residual so that the learned interior dynamics cannot inject arbitrary energy at the constrained boundary. This should eliminate boundary drift and reduce the burden on penalties or projection layers.

## Mathematical statement

The paper's kinematic construction writes the physical field as a relative dynamic field plus a lifting field, schematically $r=r_r+r_L$ and $v=v_r+v_L$, with $r_r$ and $v_r$ vanishing on the Dirichlet boundary while $r_L$ and $v_L=\partial_t r_L$ extend prescribed boundary displacement and velocity into the interior. In the discrete port-Hamiltonian model, the boundary velocity enters through the input map, for example equation (50): $\hat{G}=\begin{bmatrix}\hat{B}_N&0\\0&-\hat{M}_{rr}^{-1}\hat{M}_{rD}\\0&I_{N_D}\end{bmatrix}$, where $\hat{M}_{rr}$ is the interior mass matrix, $\hat{M}_{rD}$ couples interior and Dirichlet degrees of freedom, $\hat{B}_N$ maps Neumann traction to momentum, and $I_{N_D}$ is the identity on the $N_D$ Dirichlet velocity degrees of freedom. The corresponding state evolution has port-Hamiltonian form $\dot z=J\nabla_zH(z)+G u_\partial$, with $J=-J^\top$ skew-symmetric, $u_\partial=[\tau_N;v_D]$ the Neumann traction and prescribed Dirichlet velocity, and $H$ the stored kinetic-plus-strain energy. The power identity is $\dot H=\int_{\partial\Omega}y_\partial^\top u_\partial\,dS$, where $y_\partial$ is the boundary effort/output and $dS$ is boundary measure. The neural adaptation uses a boundary-vanishing mask $\rho(X)$, with $\rho=0$ on $\partial\Omega_D$ and typically $0<\rho\leq1$ in the interior: $r_\theta(X,t)=r_L(X,t)+\rho(X)\tilde r_\theta(X,t)$ and $v_\theta(X,t)=v_L(X,t)+\rho(X)\tilde v_\theta(X,t)$. Here $X$ is spatial position, $t$ is time, $r_L$ and $v_L$ are prescribed lifts, and $\tilde r_\theta,\tilde v_\theta$ are neural residual fields.

## Key formulas

- $$r=r_r+r_L,\qquad v=v_r+v_L,\qquad r_r\vert_{\partial\Omega_D}=0,\quad v_r\vert_{\partial\Omega_D}=0,$$
- $$\dot z=J\nabla_zH(z)+G u_\partial,\qquad J=-J^\top,$$
- $$\hat{G}=\begin{bmatrix}\hat{B}_N&0\\0&-\hat{M}_{rr}^{-1}\hat{M}_{rD}\\0&I_{N_D}\end{bmatrix},\qquad u_\partial(t)=\begin{bmatrix}\hat\tau_N(t)\\\hat v_D(t)\end{bmatrix},$$
- $$\dot H=\int_{\partial\Omega}y_\partial(\texttt{S},t)^\top u_\partial(\texttt{S},t)\,d\texttt{S},\qquad r_\theta(X,t)=r_L(X,t)+\rho(X)\tilde r_\theta(X,t).$$

## Implementation notes

(1) Integration point: use this representation in a mesh-based graph neural network or neural operator that predicts one-step elastodynamic updates. Let node coordinates be $X_i$, identify Dirichlet nodes $\mathcal D$, and replace the network's raw displacement and velocity outputs by lifted fields. Construct a scalar mask $\rho_i$ equal to zero on $\mathcal D$ and positive elsewhere; for a mesh, use normalized distance to the Dirichlet boundary or a harmonic extension of boundary values. Construct $r_L(X,t)$ from the prescribed boundary displacement using linear finite-element interpolation, radial-basis interpolation, or a precomputed harmonic extension. Set $v_L=\partial_t r_L$ analytically when possible. (2) Pseudocode: inputs are current state $(r_t,v_t)$, boundary data $(r_D(t),v_D(t),\tau_N(t))$, coordinates $X$, and mask $\rho$. Compute $r_L=\mathrm{Lift}(r_D)$ and $v_L=\mathrm{Lift}(v_D)$; compute residual coordinates $r_r=r_t-r_L$ and $v_r=v_t-v_L$; evaluate $(\Delta\tilde r,\Delta\tilde v)=\mathrm{GNN}_\theta(r_r,v_r,X,\tau_N,v_D)$; update $\tilde r\leftarrow\tilde r+\Delta\tilde r$ and $\tilde v\leftarrow\tilde v+\Delta\tilde v$; reconstruct $r\leftarrow r_L+\rho\odot\tilde r$ and $v\leftarrow v_L+\rho\odot\tilde v$. Do not apply the mask to the lift. (3) Use the paper's structural facts—zero residual boundary trace, explicit boundary input, and skew interconnection—but estimate the energy terms empirically. For known mass matrix $M$, kinetic energy is $T=\frac12v^\top Mv$; for a known constitutive model use strain energy $U(r)$, otherwise use a learned nonnegative energy network $U_\phi$. Add the discrete power residual $L_{power}=\left|[H_{t+1}-H_t]/\Delta t-y_\partial^\top u_\partial\right|^2$, with $y_\partial$ obtained from predicted boundary traction or autodifferentiated energy. (4) First experiment: train a small GNN on 2-D linear or nonlinear elastodynamics with moving Dirichlet boundaries. Compare an unconstrained predictor trained with boundary MSE, a projection-after-prediction baseline, and the port-lifted model. Measure maximum boundary velocity error, rollout energy drift, one-step displacement error, and long-horizon stability at equal parameter count. The expected signal is machine-precision boundary satisfaction, substantially lower energy drift, and improved long-horizon error without increasing model size; a useful target is at least a 5x reduction in boundary error and a 2x longer stable rollout.

## Disclaimer

AI-generated research hypothesis, automatically tested. Not peer-reviewed.