# Robust Physics-Sparse Neural Dynamics

- ID: 1623
- Canonical URL: https://synthcore.org/idea/1623/robust-physics-sparse-neural-dynamics
- API JSON: https://synthcore.org/api/idea/1623.json
- API Markdown: https://synthcore.org/api/idea/1623.md
- Verification status: failed_benchmark
- Source: [arXiv:2608.00213](https://arxiv.org/abs/2608.00213)
- Category: dynamics
- Solves: stability, accuracy, generalization
- ML areas: world-model, rnn, ssm, training-dynamics
- Math tags: dynamical-systems, control-theory, optimization, linear-algebra, statistics
- Ratings: usefulness 7/10; difficulty 6/10; novelty 7/10

## Idea description

Replace an unconstrained neural transition model with a hybrid sparse dynamics model: retain analytically known first-order relaxation or control terms and learn only a sparse set of candidate interactions from a physics-guided library. Fit the library coefficients using a robust TLS-plus-RANSAC procedure, then use the identified model as the transition function or as a residual correction to a neural state-space model.

## Mathematical statement

The paper's droop mechanism contains a stable first-order relaxation term, \(\dot{\omega}_i=\omega_i^c(\omega_i^s-\omega_i)+\omega_i^c\sigma_i(p_i^s-p_i)\), where \(\omega_i\) is frequency, \(\omega_i^s\) its setpoint, \(p_i\) output active power, \(p_i^s\) its setpoint, \(\sigma_i\) droop slope, and \(\omega_i^c>0\) filter cutoff frequency. Generalize this to a learned state \(x\) and input \(u\) as \(\dot{x}=f_{\mathrm{known}}(x,u)+\Theta(x,u)\xi\), where \(\Theta\) is a physics-guided library and \(\xi\) is a sparse coefficient matrix. Estimate \(\xi\) by minimizing a total-least-squares residual, which accounts for noise in both derivative targets and library regressors, followed by RANSAC to reject corrupted samples. Discretize with step \(h\) as \(x_{k+1}=x_k+h[f_{\mathrm{known}}(x_k,u_k)+\Theta(x_k,u_k)\xi]\). For local rollout stability, require the spectral radius of the transition Jacobian \(J_k=\partial x_{k+1}/\partial x_k\) to satisfy \(\rho(J_k)<1\); a practical training penalty is \(\max(0,\rho(J_k)-(1-\epsilon))^2\), with margin \(\epsilon>0\).

## Key formulas

- $$\dot{\omega}_i=\omega_i^c(\omega_i^s-\omega_i)+\omega_i^c\sigma_i(p_i^s-p_i),\qquad \omega_i^c>0.$$
- $$\dot{x}=f_{\mathrm{known}}(x,u)+\Theta(x,u)\xi,\qquad \xi^*=\arg\min_{\xi}\|\dot{X}-F_{\mathrm{known}}-\Theta\xi\|_2^2+\lambda\|\xi\|_1.$$
- $$\hat{\xi}_{\mathrm{TLS}}=\arg\min_{\xi,\,\Delta A,\Delta b}\|[\Delta A\ \Delta b]\|_F^2\ \mathrm{s.t.}\ (A+\Delta A)\xi=b+\Delta b,$$
- $$\mathcal{L}_{\mathrm{stab}}=\frac{1}{B}\sum_{k=1}^{B}\left[\max\left(0,\rho\left(I+h\,\partial_x[f_{\mathrm{known}}+\Theta\xi](x_k,u_k)\right)-(1-\epsilon)\right)\right]^2.$$

## Implementation notes

1) Integration point: implement this as the transition model in a small neural state-space model. Encode observations with an MLP encoder, maintain latent state \(x_k\), and predict \(x_{k+1}\) using the sparse hybrid equation rather than a fully unconstrained MLP. For a generic benchmark, use a known linear relaxation \(f_{\mathrm{known}}(x,u)=-\Gamma(x-x^*(u))\), analogous to the droop filter, where \(\Gamma\) is a positive diagonal matrix; use a library containing \(x_j\), \(u_l\), \(x_jx_m\), \(x_ju_l\), and bounded nonlinearities. 2) Pseudocode: estimate derivatives with a centered finite difference or automatic differentiation through a short encoder; form \(A=[\Theta\ \ -\dot X]\); repeatedly sample minimal subsets, solve the TLS problem by SVD, score inliers by normalized residual, retain the best consensus set, and apply iterative hard thresholding to remove coefficients below threshold \(\tau\). During neural training, alternate 5--20 sparse-regression updates with gradient updates of the encoder and add reconstruction loss, multi-step rollout loss, and \(\alpha\mathcal L_{\mathrm{stab}}\). Estimate \(\rho(J_k)\) with 10--20 power iterations rather than explicitly forming eigenvalues. 3) Computed from the paper's mechanism: the known relaxation structure, sparse library identification, TLS treatment of errors in both regressors and derivatives, and RANSAC outlier rejection. Estimated empirically: derivative noise scale, RANSAC inlier threshold, sparsity threshold \(\tau\), and the latent Jacobian spectrum. 4) First cheap experiment: train on noisy trajectories from a two-dimensional damped oscillator with 5% Gaussian observation noise and 5% randomly corrupted time points, comparing an unconstrained GRU/MLP transition, ordinary-least-squares SINDy, and the TLS-RANSAC hybrid model. Sweep integration step \(h\) and multiply the learned transition Jacobian by a scalar gain. The quantitative prediction is a sharp rollout-divergence boundary near \(\rho(J)=1\); the constrained model should remain bounded for gains with measured \(\rho(J)<1\), while divergence should occur predominantly after crossing one. On held-out trajectories with corrupted observations, TLS-RANSAC should have lower coefficient error and substantially longer stable rollout horizon than ordinary least squares. Report the measured critical gain and test whether it predicts the 100-step rollout boundary within 20%.

## Verification

- Status: failed_benchmark
- Mechanism evidence: yes
- Mechanism confirmed: no
- Practical verdict: harms
- Verdict: Built a compact hybrid physics-sparse dynamics MVP with known relaxation, sparse cross-interaction library, TLS, RANSAC consensus selection, hard thresholding, Jacobian spectral-radius checks, and rollout evaluation. The predicted stability boundary was gain*=5.8824 at h=0.08 versus 5.8837 observed (0.023% error), and the predicted 1/h scaling across four step sizes matched within 0.03%. Under 5% corrupted derivative samples, TLS-RANSAC achieved 0.0149 relative coefficient error versus 0.0422 for OLS and 0.0934 for plain TLS, so the mechanism manifested clearly in this toy setting.

### Mechanism check

- Verdict: Built a compact hybrid physics-sparse dynamics MVP with known relaxation, sparse cross-interaction library, TLS, RANSAC consensus selection, hard thresholding, Jacobian spectral-radius checks, and rollout evaluation. The predicted stability boundary was gain*=5.8824 at h=0.08 versus 5.8837 observed (0.023% error), and the predicted 1/h scaling across four step sizes matched within 0.03%. Under 5% corrupted derivative samples, TLS-RANSAC achieved 0.0149 relative coefficient error versus 0.0422 for OLS and 0.0934 for plain TLS, so the mechanism manifested clearly in this toy setting.
- Confidence: 9/10
- Limitations: This is a deterministic two-dimensional damped-oscillator numerical verification, not a full neural state-space model with an MLP encoder or GRU baseline. It uses synthetic finite-difference data, forward Euler discretization, independently applied scalar TLS fits, and does not test real microgrid PMU data, long-horizon distribution shift, GPU execution, or a FLOP-matched unconstrained neural transition.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.0081
- Paired wins: 0/8
- Benchmark verdict: idea worse (significant)

## Artifacts

- [bench_experiment.py](https://synthcore.org/code/605/bench_experiment.py)
- [bench_report.json](https://synthcore.org/code/605/bench_report.json)
- [experiment.py](https://synthcore.org/code/605/experiment.py)
- [report.md](https://synthcore.org/code/605/report.md)
- [report_bench_2026-08-31T193914.md](https://synthcore.org/code/605/report_bench_2026-08-31T193914.md)
- [results.json](https://synthcore.org/code/605/results.json)
- [Download all files as ZIP](https://synthcore.org/download/605)

## Disclaimer

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