# Derivative-Free Dynamic-Stiffness PINN

- ID: 2949
- Canonical URL: https://synthcore.org/idea/2949/derivative-free-dynamic-stiffness-pinn
- API JSON: https://synthcore.org/api/idea/2949.json
- API Markdown: https://synthcore.org/api/idea/2949.md
- Verification status: unverified
- Source: [arXiv:2608.28683](https://arxiv.org/abs/2608.28683)
- Category: architecture
- Solves: stability, accuracy, speedup
- ML areas: loss, training-dynamics, graph-nn
- Math tags: spectral-theory, linear-algebra, pde, numerical-analysis
- Ratings: usefulness 6/10; difficulty 5/10; novelty 7/10

## Idea description

Replace pointwise high-order derivative residuals in an eigenvalue PINN by an assembled dynamic-stiffness residual \(\mathbf W(\omega)q_\theta\), where each element matrix is obtained from homogeneous PDE solutions. The network predicts nodal degrees of freedom or element boundary traces, while the exact frequency-domain operator enforces the physics without differentiating the network multiple times with respect to coordinates.

## Mathematical statement

For the Euler–Bernoulli beam, the paper defines the frequency-dependent element dynamic stiffness matrix \(\mathbf W(\omega)\) mapping nodal displacement and rotation degrees of freedom to nodal forces. With \(\alpha=L(m\omega^2/EI)^{1/4}\), \(c=\cos\alpha\), \(s=\sin\alpha\), \(C=\cosh\alpha\), \(S=\sinh\alpha\), and \(D=1-cC\), its coefficients are \(\delta=(sC+cS)\alpha^3/D\), \(\delta_l=(s+S)\alpha^3/D\), \(\theta=sS\alpha^2/D\), \(\theta_l=(C-c)\alpha^2/D\), \(\beta=(sC-cS)\alpha/D\), and \(\gamma=(S-s)\alpha/D\). The assembled matrix acts as a linear operator on the neural nodal vector \(q_\theta\). Near zero frequency, the paper gives \(\mathbf W(\omega)\approx\mathbf K-\omega^2\mathbf M\), with \(\mathbf K\) the stiffness matrix and \(\mathbf M\) the consistent mass matrix. The transferable property is that the element basis exactly satisfies the governing ODE, eliminating spatial derivatives of the neural output from the physics loss.

## Key formulas

- $$\mathbf W(\omega)=\frac{EI}{L^{2}}\begin{bmatrix}\delta/L&\theta&-\delta_l/L&\theta_l\\\theta&\beta L&-\theta_l&\gamma L\\-\delta_l/L&-\theta_l&\delta/L&-\theta\\\theta_l&\gamma L&-\theta&\beta L\end{bmatrix}$$
- $$\delta=\frac{(sC+cS)\alpha^3}{D},\quad \delta_l=\frac{(s+S)\alpha^3}{D},\quad \theta=\frac{sS\alpha^2}{D},\quad \theta_l=\frac{(C-c)\alpha^2}{D},\quad \beta=\frac{(sC-cS)\alpha}{D},\quad \gamma=\frac{(S-s)\alpha}{D},\quad D=1-cC$$
- $$\mathbf W(\omega)\approx\mathbf K-\omega^2\mathbf M$$
- $$\mathcal L(\theta,\omega)=\frac{\|\mathbf B\mathbf W(\omega)\mathbf A q_\theta\|_2^2}{\|\mathbf B\mathbf A q_\theta\|_2^2+\varepsilon}+\lambda_n\left(\|\mathbf A q_\theta\|_{\mathbf M}^2-1\right)^2+\lambda_b\|\mathbf C\mathbf A q_\theta-d\|_2^2$$

## Implementation notes

1. Integration point: use this as the physics-loss layer for a neural eigenmode solver. Let a coordinate MLP output element or global nodal degrees of freedom \(q_\theta\); \(\mathbf A\) assembles local element boundary values into global degrees of freedom, \(\mathbf C\) selects constrained boundary degrees of freedom, and \(\mathbf B\) removes constrained rows or applies boundary-condition elimination. Construct sparse global \(\mathbf W(\omega)\) by assembling the element matrices above. 2. Pseudocode: input mesh, material parameters, trial frequency \(\omega\); compute \(q=\mathbf A q_\theta\); for every element compute \(\alpha=L(m\omega^2/EI)^{1/4}\), trigonometric and hyperbolic coefficients, and \(W_e(\omega)\); assemble \(W\); compute \(r=BWq\); minimize \(\|r\|_2^2/(\|Bq\|_2^2+\varepsilon)+\lambda_n(q^TMq-1)^2+\lambda_b\|Cq-d\|^2\). Update both \(\theta\) and \(\omega\), using softplus parameterization for positive frequency. 3. The DSM entries and sparse assembly are computed analytically; only \(q_\theta\), \(\omega\), and optional normalization weights are learned. Assemble \(M\) once from the consistent mass matrix or use the low-frequency expansion. 4. First experiment: train on a 1-D Euler–Bernoulli beam with 8, 16, and 32 elements. Compare a fourth-derivative strong-form PINN, the DSM-PINN, and a conventional FEM generalized eigensolver at equal training steps, parameter count, and wall-clock budget. Report the first four frequencies, residuals, and runtime. 5. Pre-register mechanism tests: at fixed mesh and model size, DSM-PINN should reduce median physics-residual variance by at least 3x relative to the fourth-derivative PINN; frequency error should remain below 1% as element length changes; and replacing \(W(\omega)\) with \(K-\omega^2M\) should produce increasing error away from low frequency. Measure against analytic beam frequencies or a high-resolution FEM reference. Falsification is no improvement in derivative sensitivity or failure of the exact DSM and low-frequency approximation to separate as predicted.

## Disclaimer

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