Gaussian-mixture kinetic neural solver
Implementation & benchmark of arXiv:2608.14477 — Linear response across interaction regimes in two-dimensional ferromagnets
Source paper: Linear response across interaction regimes in two-dimensional ferromagnets arXiv:2608.14477 ⓘ · analyzed Sep 1, 2026
AI-generated research hypothesis, automatically tested. Not peer-reviewed.
Idea description
Make a neural network predict a positive Gaussian-mixture representation of the distribution function rather than independent values on a momentum grid. Use the mixture parameters inside a differentiable Boltzmann collision operator, so training directly enforces the interaction mechanism and exposes the relaxation spectrum responsible for ballistic-to-hydrodynamic crossover.
Formulas
Mathematical statement
The paper starts from interacting bosons with Hamiltonian $H_b=\sum_{\bm p}\epsilon_{\bm p}a^\dagger_{\bm p}a_{\bm p}+(2N)^{-1}\sum \mathcal T(\bm k,\bm p;\bm k',\bm p')\delta_{\bm k'+\bm p',\bm k+\bm p}a^\dagger_{\bm k'}a^\dagger_{\bm p'}a_{\bm k}a_{\bm p}$, where $\epsilon_{\bm p}=p^2/(2m)+\Delta$, $N$ is the site count, and $\mathcal T$ is the two-body scattering matrix. Represent the occupation as $f_\theta(\bm p)=\sum_{j=1}^{K}w_j(2\pi\sigma_j^2)^{-1}\exp[-\|\bm p-\bm\mu_j\|^2/(2\sigma_j^2)]$, with $w_j\geq0$, $\sum_jw_j=1$, means $\bm\mu_j$, and widths $\sigma_j>0$. The bosonic collision operator is $Q[f](\bm p)=\int W_{\bm p\bm p_1\to\bm p'\bm p_1'}[f'f_1'(1+f)(1+f_1)-ff_1(1+f')(1+f_1')]\delta^{(2)}(\bm p+\bm p_1-\bm p'-\bm p_1')\delta(\epsilon+\epsilon_1-\epsilon'-\epsilon_1')\,d\bm p_1d\bm p'd\bm p_1'$, where $W\propto|\mathcal T|^2$ and primed quantities are evaluated at outgoing momenta. Linearizing $f=f_0-f_0(1+f_0)\phi$ gives $\partial_t\phi=-\mathcal L\phi+s$; the eigenvalues $\gamma_r$ of $\mathcal L$ are relaxation rates. Gaussian products make the required momentum integrals low-dimensional and numerically stable.
Implementation notes
1. Exact integration point: use this architecture for a neural surrogate of a 2D kinetic equation, not for ordinary image or language prediction. Feed $(t,\bm x)$, external field, temperature, and model parameters to an MLP. Its final layer emits logits $u_j$, unconstrained means $\bm\mu_j$, and log-widths $s_j$ for $K=8$--$32$ Gaussian components. Evaluate the mixture on quadrature points in momentum space. 2. Pseudocode: emit $(u,\mu,s)$; form $w=softmax(u)$ and $\sigma=softplus(s)+10^{-4}$; evaluate $f_\theta$; sample or precompute momentum-conserving collision tuples; compute $Q[f_\theta]$ using the displayed Bose gain-minus-loss expression; minimize $L=\|\partial_t f_\theta+v\cdot\nabla_xf_\theta-Q[f_\theta]\|_2^2+\lambda_NL_N+\lambda_EL_E$, where $L_N$ and $L_E$ penalize drift of particle number and energy. 3. Compute from the paper: the collision structure, conservation delta functions, and linearized response operator. Estimate empirically: mixture parameters, quadrature error, and eigenvalues of the discretized Jacobian $\mathcal L_\theta=-\partial Q/\partial f$. 4. First cheap experiment: use a spatially homogeneous 2D quadratic-dispersion Boltzmann toy problem with a constant scattering amplitude, compare this model against a grid-output MLP and a conventional finite-volume solver. Apply a sinusoidal perturbation of wavevector $k$ and frequency $\omega$. The quantitative prediction is that the measured response changes character near $|k|v_{\mathrm{th}}\approx\gamma_1$, where $v_{\mathrm{th}}=\sqrt{T/m}$ and $\gamma_1$ is the smallest nonzero collision eigenvalue; fitted response-pole locations and widths should agree with the reference solver within 20%, while number and energy drift should remain below 1% over 100 relaxation times.
Verification
Stage 1 · Toy mechanism gate: Failed ✗
Stage 2 · Mechanism transferred to benchmark: Not tested
Stage 2 · Practical benchmark result: Not run
Methodology: Toy-system gate first; the benchmark stage runs only after a pass. How verification works
Stage 1 — Mechanism check agent confidence 9/10
Built a deterministic Gaussian-mixture kinetic toy in `kinetic_mvp.py` with positive softmax/softplus parameterization, a number/energy-preserving projected collision operator, relaxation-spectrum sweeps, response sweeps, and a coarse-grid comparison. Positivity, two conservation zero modes, and linear eigenvalue scaling with collision rate were confirmed, but the measured response FWHM did not show a crossover near k v_th/gamma = 1. The full Bose collision integral with momentum/energy delta functions and neural PDE training were not implemented, so the headline mechanism was not established.
- Agent confidence
- 9/10
- Baseline
- Coarse 7x7 grid representation RMSE = 2.19e-2 on the synthetic mixture target; response crossover not reliably detected.
- Idea
- Positive fitted 3-component Gaussian mixture RMSE = 1.35e-11; minimum density = 1.06e-20; collision operator had exactly 2 zero modes and median nonzero eigenvalues 0.25, 0.50, 1.00, 2.00 for gamma = 0.25, 0.5, 1, 2, respectively; crossover sweep FWHM values were 0.5, 4, 6, 6, 2 at ratios 0.25, 0.5, 1, 2, 4.
Limitations: This is a small projected linear toy rather than the full differentiable Bose gain-minus-loss collision operator with conservation delta functions. No MLP training, spatial PDE residual optimization, CUDA experiment, response-pole fitting, 100-relaxation-time drift test, or full finite-volume solver comparison was performed. The quadrature box truncates the mixture tails, giving an integrated mass of 0.8882 on the finite grid.
How to run: python3 kinetic_mvp.py
Verdict computed by deterministic test code from paired-seed statistics — not by the language model.
Artifacts
- REPORT.md 2.3 KB View
- kinetic_mvp.py 2.8 KB View Raw
- report.md 1.8 KB View
- results.json 1.1 KB View Raw