Unverified Re-invented 2026

Entropy-safe non-diagonal neural mobility

Implementation & benchmark of arXiv:2608.26297 — Generalised dissipative solutions for a non-isothermal phase-field system: existence, weak-strong uniqueness, and long-time behaviour

Usefulness6/10

Difficulty6/10

Novelty7/10

Dynamics Stability Accuracy Generalization World model Graph nn Training Dynamics Pde Dynamical systems Linear algebra Optimization

Source paper: Generalised dissipative solutions for a non-isothermal phase-field system: existence, weak-strong uniqueness, and long-time behaviour arXiv:2608.26297 ⓘ · analyzed Aug 29, 2026

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 Not peer-reviewed.

Idea description

Replace an unconstrained neural PDE right-hand side with a thermodynamic operator acting on two order parameters and inverse temperature. The network predicts a non-diagonal mobility factor, but the actual mobility is constructed as M = B B^T, so cross-coupling remains expressive while the dissipative quadratic form is nonnegative by construction. Train the model on trajectories while monitoring free-energy change, entropy production, conservation, and temperature positivity.

Formulas

$$\Psi(z,\theta,\nabla z)=\psi(z,\theta)+\frac{\gamma_{\rho}}{2}|\nabla\rho|^{2}+\frac{\gamma_{\eta}}{2}|\nabla\eta|^{2},\qquad z=(\rho,\eta),\quad \gamma_{\rho},\gamma_{\eta}>0.$$

$$e(z,\theta)=\partial_{\theta}\psi(z,\theta),\qquad s_{\rm bulk}(z,\theta)=\theta e(z,\theta)-\psi(z,\theta).$$

$$\mathscr{S}(z,\theta)=\int_{\Omega}s_{\rm bulk}(z,\theta)\,\mathrm{d}x-\frac{\gamma_{\rho}}{2}\left\lVert\nabla\rho\right\rVert_{L^{2}}^{2}-\frac{\gamma_{\eta}}{2}\left\lVert\nabla\eta\right\rVert_{L^{2}}^{2}.$$

$$M(u,x)=B_{\phi}(u,x)B_{\phi}(u,x)^{\mathsf T}\succeq0,\qquad J=-M q,\qquad q^{\mathsf T}Mq=\left\|B_{\phi}(u,x)^{\mathsf T}q\right\|_{2}^{2}\ge0.$$

Mathematical statement

The paper defines z = (rho, eta), where rho and eta are the two order parameters, theta = T^{-1} > 0 is inverse temperature, psi(z, theta) is the reduced bulk Helmholtz free-energy density, and gamma_rho, gamma_eta > 0 are interfacial coefficients. The reduced free-energy density is Psi(z, theta, grad z) = psi(z, theta) + gamma_rho |grad rho|^2 / 2 + gamma_eta |grad eta|^2 / 2. Internal energy and bulk entropy are e(z, theta) = partial_theta psi(z, theta) and s_bulk(z, theta) = theta e(z, theta) - psi(z, theta). The total entropy is S = integral of s_bulk over the domain minus the two gradient-energy terms. We adapt the Onsager principle by letting q = delta A / delta u be the generalized force for state u = (rho, eta, theta), and defining the dissipative flux J = -M q with M = B B^T. Here B is a learned square matrix containing off-diagonal entries. Since q^T M q = ||B^T q||^2 >= 0 for every q, the learned operator cannot create negative quadratic dissipation even when it couples all state variables. Temperature is parameterized as theta = softplus(a) + epsilon, where a is unconstrained and epsilon > 0.

Implementation notes

Integrate the construction into the state-update layer of a neural PDE simulator or neural ODE. Represent the state on a periodic grid as rho, eta, and an unconstrained thermal field a, then compute theta = softplus(a) + epsilon with epsilon = 1e-4. A CNN or graph neural network receives local fields and finite-difference gradients and outputs the six entries of a lower-triangular 3 by 3 matrix B_phi at every grid cell; retaining off-diagonal entries is essential because they model cross-coupling. Compute e = d psi(z, theta) / d theta, s_bulk = theta * e - psi, and the discrete free energy containing the gradient terms. Use automatic differentiation to compute the generalized force q = delta A / delta u. Form M = B_phi @ transpose(B_phi), compute J = -M @ q, and convert fluxes for rho and eta into conservative updates using a discrete divergence. For nonconserved relaxation use the corresponding local component of J. After each update, remove the spatial mean of the rho and eta increments to enforce mass conservation. If a known internal-energy invariant is available, apply a scalar correction to the thermal increment so that its discrete internal energy is unchanged. Train with trajectory loss plus soft penalties for entropy decrease, energy drift, and conservation drift, while keeping PSD mobility and positive temperature as hard constraints. The exact structural quantities are M, q^T M q, e, entropy, and free energy; psi, coefficients, and B_phi are learned or fitted. First compare this model against an unconstrained CNN vector field and a diagonal-mobility model on synthetic 2-D periodic thermal Cahn-Hilliard trajectories generated with a known non-diagonal PSD mobility. Measure one-step and long-rollout error, entropy violations, mass drift, temperature failures, and stability at timesteps larger than the training timestep. Success means comparable one-step error, lower long-horizon error, zero negative-temperature states, and zero negative quadratic-dissipation events.

Verification

This idea has not been verified yet.

Verification happens in two stages: Stage 1 — a mechanism check on a toy system confirms the claimed mathematical phenomenon reproduces; Stage 2 — a benchmark implements the idea on a real (small) neural network task and compares it against a tuned baseline over 8 paired seeds with a permutation test.

Artifacts

Artifacts unavailable.