Physics-Informed Test-Time Training (PI-TTT)

• Physics-Informed Test-Time Training (PI-TTT) is an approach that uses label-free, self-supervised adaptation at test time to enforce physical law constraints for accurate predictions.
• Applied in power systems and imaging, PI-TTT significantly reduces residual errors (e.g., RMSE drops as low as 0.047 in IEEE-14 cases) and minimizes operational violations.
• The method employs a few gradient updates to refine a pre-trained model without ground-truth labels, offering a computationally efficient alternative to classical solvers.

Physics-Informed Test-Time Training (PI-TTT) is a paradigm for enhancing ML models in physics-constrained tasks by incorporating test-time self-supervision derived from domain-specific physical laws or constraints. PI-TTT provides a lightweight and label-free refinement of a pre-trained model’s predictions on unseen samples so that outputs remain consistent with the governing physical principles, without requiring ground-truth data at inference. This approach has demonstrated substantial improvements in physical reliability and generalization within both power systems and computational imaging, particularly under distribution shift or previously unseen operating conditions (Dogoulis et al., 27 Nov 2025, Chandler et al., 2024).

1. Mathematical Underpinnings in Physics-Constrained Domains

The core of PI-TTT is the enforcement of physical consistency via test-time loss functions constructed from governing equations and operational constraints.

For example, in AC power flow, the injected active and reactive powers at bus $i$ are given by

$$\begin{align*} P_i(V,\theta) &= \sum_{j=1}^N V_i V_j \left[ G_{ij}\cos\theta_{ij} + B_{ij} \sin\theta_{ij} \right], \ Q_i(V,\theta) &= \sum_{j=1}^N V_i V_j \left[ G_{ij}\sin\theta_{ij} - B_{ij}\cos\theta_{ij} \right], \end{align*}$$

where $V_i$, $\theta_i$ are the voltage magnitude and angle, and $Y = G + jB$ is the admittance matrix. Given specified power injections $P_i^{\textrm{spec}}$, $Q_i^{\textrm{spec}}$, the physics-informed residuals are $\Delta P_i = P_i^{\textrm{spec}} - P_i(V, \theta)$ and $\Delta Q_i = Q_i^{\textrm{spec}} - Q_i(V, \theta)$.

Operational inequalities (e.g., $$V_i^{\textrm{min}} \leq V_i \leq V_i^{\textrm{max}}$$, $|S_\ell| \leq S_\ell^{\textrm{max}}$ for line flows) are softly enforced through smooth penalty terms such as

$$\phi_{\textrm{volt}}(V_i) = \mathrm{ReLU}(V_i - V_i^{\textrm{max}})^2 + \mathrm{ReLU}(V_i^{\textrm{min}} - V_i)^2$$

$$\phi_{\textrm{flow}}(S_\ell) = \mathrm{ReLU}(|S_\ell| - S_\ell^{\textrm{max}})^2.$$

The self-supervised loss at inference for operating condition $z$ becomes

$$L_\textrm{TTT}(\varphi; z) = \|\Delta P(f_{\theta+\varphi}(z))\|_2^2 + \|\Delta Q(f_{\theta+\varphi}(z))\|_2^2 + \lambda_V \sum_i \phi_{\textrm{volt}}(V_i) + \lambda_\ell \sum_\ell \phi_{\textrm{flow}}(S_\ell)$$

where $f_{\theta}$ is the pre-trained surrogate and $\varphi$ parameterizes the test-time adaptation (Dogoulis et al., 27 Nov 2025).

In inverse imaging, PI-TTT (as in PnP-TTT) minimizes the violation of the forward measurement model at a deep-equilibrium (DEQ) fixed point, where the loss for data-consistency is $L_{ss}(\theta; y, A) = \|A x^* - y\|_2^2$ for $x^*$ the converged image estimate (Chandler et al., 2024).

2. Algorithmic Framework

The PI-TTT workflow proceeds as follows (illustrated for AC power flow):

1. Pre-trained Surrogate: A neural network $f_\theta$, e.g., a feed-forward or graph neural network, trained to map system inputs (e.g., load/generation vectors) to predicted physical variables.
2. Test-Time Inference:
   • Compute the initial prediction $\hat{x}_0 = f_\theta(z^\dagger)$ for a new sample $z^\dagger$.
   • Introduce adaptive parameters $\theta_{\mathrm{adapt}}$; initialize perturbation $\varphi_0 = 0$.
   • For $k = 0, ..., K-1$, update

   $$\varphi_{k+1} = \varphi_k - \eta \nabla_\varphi L_\textrm{TTT}(\varphi_k; z^\dagger)$$

   where $\eta$ is the learning rate and typically $K$ is small (e.g., $3$–$5$). - Output refined prediction $$\hat{x}_K = f_{\theta_{\mathrm{adapt}} + \varphi_K, \theta_{\mathrm{frozen}}}(z^\dagger)$$.

Key attributes:

• Gradients are computed by backpropagation through both the surrogate and the physics penalty functions.

• No ground-truth labels are required for $z^\dagger$; all quantities are derived from the input and physical models (Dogoulis et al., 27 Nov 2025).

In PnP-TTT for imaging, DEQ fixed points are used, and adaptation employs implicit differentiation for memory-efficient gradients (Chandler et al., 2024).

3. Empirical Assessment and Benchmarking

The empirical evaluation of PI-TTT for power system analysis was conducted on the IEEE 14-, 118-, 300-bus test cases and the PEGASE 1354-bus network, using PowerFlowNet and MF-GNN surrogates. Results show:

• Power-flow residuals (RMSE, MW/MVAR) drop by 1–2 orders of magnitude relative to pre-trained surrogates:

  • IEEE-14 (PowerFlowNet): RMSE$_P$ 0.924→0.047, RMSE$_Q$ 0.375→0.026.
  • IEEE-300: RMSE$_P$ 9.39→1.08, RMSE$_Q$ 3.56→0.73.
  • PEGASE-1354: RMSE$_P$ 9.12→0.72, RMSE$_Q$ 2.90→2.25.
• Operational-constraint mean violations (in per-unit or thermal flow) drop by an order of magnitude:
  • IEEE-14 voltage mean: 0.012→0.001 pu; max: 0.045→0.006.
  • IEEE-118 flow mean: 0.033→0.006 pu; max: 0.152→0.026.
• Runtime per sample remains competitive: PowerFlowNet + PI-TTT $17$ ms versus Newton-Raphson solver $19$ ms, maintaining the ML surrogate's speed (Dogoulis et al., 27 Nov 2025).

For PnP-TTT in MRI reconstruction, test-time adaptation closes the distribution shift gap between priors trained on natural and MRI images. For radial CS-MRI with $m/n$ from $10\%$ to $50\%$, PnP-TTT achieves PSNR/SSIM values that approach or exceed those of matched MRI-trained priors as sampling increases, e.g., at $50\%$ sampling, PnP-TTT: $39.96$ dB/$0.9873$ versus MRI prior $38.57$ dB/$0.9828$ (Chandler et al., 2024).

4. Strengths, Limitations, and Open Questions

Strengths:

• Enforces strict physical consistency (e.g., power balance, operational limits) at inference, surpassing unconstrained ML models.
• Requires only a few gradient updates and operates without ground-truth labels, making it practical for real-time deployment.
• Overhead is modest, preserving computational advantages over classical numerical solvers (Dogoulis et al., 27 Nov 2025).

Limitations:

• On large-scale systems, full feasibility may not be achieved within a few adaptation steps.
• Choice of penalty weights ($\lambda_V, \lambda_\ell$) and learning rate $\eta$ can affect both convergence and quality, typically requiring empirical tuning.
• Added adaptation steps can increase inference time, which may be a constraint in very low-latency or large-scale applications (Dogoulis et al., 27 Nov 2025).

Open questions and potential extensions include use of adaptive step sizes or higher-order test-time optimizers, expanding adaptation to a larger fraction of model parameters, combining PI-TTT with warm-start techniques (e.g., Newton–Raphson warm starts), extension to scenarios involving measurement noise (state estimation), or incorporating strict barrier functions for constraint satisfaction (Dogoulis et al., 27 Nov 2025).

5. Relation to Broader Test-Time Training and Distribution Shift Correction

PI-TTT is situated within a broader class of test-time training methodologies aimed at improving model robustness under distribution shifts. In plug-and-play imaging, the PnP-TTT approach leverages deep-equilibrium fixed-point optimization, enabling robust adaptation directly on each test instance by minimizing forward model violation via self-supervised test-time losses. This performance gains are especially notable when priors are not matched to test distributions, effectively bridging the gap induced by domain shift (Chandler et al., 2024).

A plausible implication is that the core PI-TTT principle—optimization solely on physically meaningful, label-free self-supervision—can generalize to any domain where governing equations or constraints are known and differentiable. This suggests a potential for wide applicability in scientific ML problems beyond power systems and imaging, wherever accurate physical surrogates are sought.

6. Summary and Prospects

Physics-Informed Test-Time Training (PI-TTT) bridges the gap between rapid but physics-inconsistent ML models and computationally intensive physics solvers. The framework provides on-the-fly, self-supervised adaptation, enforcing physical laws at inference and yielding reliable, interpretable outputs with minimal computational overhead. The convergence of PI-TTT with test-time adaptation paradigms underlines its role as a scalable approach for ensuring physics consistency, reliability, and robustness in critical scientific and engineering applications (Dogoulis et al., 27 Nov 2025, Chandler et al., 2024). Continued research will likely focus on scaling, on extending to new classes of physical constraints, and on integrating more sophisticated optimization strategies for enhanced feasibility and efficiency.