acceptodds
Under review as a conference paper at ICLR 2027

NOVA: Nullspace Dynamics in Neural Basis Solvers for Moving Boundary PDEs, Transport, Rank Transitions, and Reprojection

Abstract

Multiphysics coupling with moving boundaries poses a major challenge to neural solvers, as evolving interfaces induce time dependent constraints. Existing boundary treatments face an inherent trade-off between constraint accuracy and update cost. To address this challenge, we propose NOVA (**N**ullspace Dynamics in Neural Basis S**O**lvers for Mo**V**ing Bound**A**ry PDEs, Transport, Rank Transitions, and Reprojection). The method explicitly decomposes neural basis coefficients into a boundary lift and a component parameterized by instantaneous nullspace coordinates, then reprojects the previous physical state onto the updated nullspace. This construction avoids boundary parameter sweeps by eliminating penalty weights and stabilization parameters. To avoid costly singular value decompositions (SVDs) at every time step, we derive horizontal transport equations for the nullspace basis on the Stiefel manifold and develop an accompanying algorithm. Furthermore, spectral analysis establishes a rank saturation criterion for boundary collocation. Across eight multiphysics benchmarks covering phase change and flow, multiphase flow, and free boundary problems, NOVA achieves best or competitive field accuracy with substantially lower boundary errors than baselines, providing a new paradigm for physics-informed neural networks based on features and mesh-free representations to efficiently handle complex moving geometries. Our code is available [here](https://anonymous.4open.science/r/NOVA-3C44).

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
neural PDE solvers, moving boundary, multiphysics coupling problems, nullspace methods, manifold transport, rank saturation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_hQ3pOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces NOVA, a neural basis solver designed for moving boundary PDEs, null space transport, rank transition, and state reprojection. By using null-space coordinate parametrization and boundary lifting, NOVA avoids manual tuning for boundary parameters and penalty terms. It demonstrates superior performance and accuracy on moving boundaries without sacrificing field accuracy across eight PDE benchmarks.

Critical strengths
  1. NOVA combines boundary-compatible coefficients with time marching, allowing the admissible solution space to change as the geometry evolves. This represents a useful extension of fixed-domain null-space methods.
  2. The benchmarking is thorough across different scenarios and criteria. The paper evaluates 8 different physics problems, comparing useful metrics such as residual sum of squares (RSS) and computation time.
  3. The authors extensively tune penalty, Baumgarte, and Nitsche treatments, reporting their best-performing configurations. Combined with boundary-count and refresh-interval studies, this helps isolate the effects of boundary enforcement and computational choices.
Critical weaknesses
  1. Limited comparison with specialised moving-boundary solvers. Although the Frozen-PINN boundary-treatment comparisons and refresh-interval ablations are useful, the paper does not directly compare NOVA with specialised moving-boundary neural methods, such as Wang and Perdikaris (2021) and Ren et al.’s Stefan PIELM (2025). Including these baselines on a shared Stefan problem—with matched reporting of field error, interface-position error, and computational cost—would better establish NOVA's advantages over existing approaches beyond simply achieving smaller boundary-condition residuals.

  2. Multiphysics framing. The main contribution is handling moving boundaries rather than multiphysics. The phrasing could be adjusted to help readers better understand the core framing. For example, line 25 mentions "Across eight multiphysics benchmarks." While I acknowledge that some benchmarks are strong multiphysics cases (e.g., Two-phase Stefan, Spherical Scriven), others are less so in a strict sense for a broader audience (e.g., the Burgers example solves a single PDE on a fixed domain, and the moving cylinder problem is generally not considered multiphysics). Consider refining this wording to be more rigorous.

Questions for authors

Could the authors provide a direct comparison between NOVA and specialized moving-boundary neural methods on a shared benchmark like the Stefan problem? Specifically, how do field errors, interface-position errors, and computational costs compare?

4 comments
AuthorsOpenReview ↗

Response to Reviewer hQ3p

We sincerely appreciate the reviewer's constructive comments and careful assessment.

W1 & Q1 — Stefan comparison and computational cost

We compare NOVA with PINN and PIELM on the same two-phase Stefan benchmark. Field and interface errors are relative L2L^2 errors, and EBE_B is the boundary error.

MethodField errorInterface errorEBE_BTime (s)
PINN, Wang & Perdikaris, Figure 6 of [1]5.83e-042.81e-04Not reportedapprox. 300
PINN reimplementation (Wang & Perdikaris [1])1.12e-033.48e-041.32e-033743.19
PIELM, Ren et al. [2], Table 21.04e-074.26e-08Not reported16
PIELM reimplementation [2] (10-seed median)1.41e-062.55e-071.74e-061.15
NOVA, 80 steps3.19e-052.08e-053.74e-161.60
NOVA, 320 steps2.06e-061.35e-064.52e-162.87
NOVA, 640 steps5.15e-073.40e-073.92e-165.12
NOVA, 1280 steps1.29e-078.60e-083.74e-169.74
NOVA, 2560 steps3.25e-082.17e-084.39e-1619.36

Ren et al. [2] use MATLAB, and our PIELM reimplementation uses Python.

At the Table 1 time step, NOVA reduces field and interface errors by factors of approximately 35 and 17 compared with the reproduced PINN. Refining NOVA's time step yields approximately second-order convergence in both metrics. At 640 steps, its field error is 5.15e-7, below the PIELM reimplementation's 1.41e-6, and its interface error is 3.40e-7, close to 2.55e-7. At 2560 steps, these errors reach 3.25e-8 and 2.17e-8, with boundary errors near machine precision.

The PIELM reimplementation has lower field and interface errors and a shorter solve time than NOVA at 80 and 320 steps. PIELM fits the solution jointly over the full space–time interval, using PDE constraints at all sampled times. NOVA advances sequentially using the current equations and previous states. This difference helps explain the short-horizon accuracy and cost tradeoff.

Long-horizon cost. We also compare the methods on a periodic Stefan problem with a manufactured solution. As the number of periods increases, the global PIELM configuration uses more features and collocation points. Added memory is the increase in peak process memory from the start of the solve.

PeriodsMethodField errorSolve time (s)Added memory (MiB)
1NOVA7.64e-057.29.0
1PIELM3.67e-051.125.4
2NOVA7.57e-0514.610.1
2PIELM7.37e-057.286.4
4NOVA7.66e-0529.112.7
4PIELM1.27e-0454.5330.3
8NOVA7.79e-0558.717.2
8PIELM1.66e-04471.41303.0

NOVA maintains a field error near 7.6e-5–7.8e-5 with approximately linear runtime growth. At eight periods, it is approximately eight times faster and uses approximately one seventy-sixth of the additional memory used by this dense global PIELM configuration.

We will include these comparisons in the revision.

References

[1] Wang, S., & Perdikaris, P. (2021). Deep learning of free boundary and Stefan problems. Journal of Computational Physics, 428, 109914.

[2] Ren, F., Zhuang, P. Z., Chen, X., Yu, H. S., & Yang, H. (2025). Physics-informed extreme learning machine (PIELM) for Stefan problems. Computer Methods in Applied Mechanics and Engineering, 441, 118015.

AuthorsOpenReview ↗

Response to Reviewer hQ3p

W2 — Moving and free boundaries

Thank you for this suggestion. We will describe the scope as PDEs with moving and free boundaries and revise the following text.

  • Abstract. The opening will read “Moving boundaries and free interfaces pose a major challenge to neural PDE solvers, because the evolving geometry induces time-dependent constraints on the solution representation.” The benchmark sentence will distinguish one fixed-domain problem with time-dependent boundary data, four prescribed moving-domain problems, and three coupled free-boundary problems.
  • Terminology and contributions. We will describe the two Navier–Stokes benchmarks as single-phase flows past moving bodies, use “for PDEs on moving domains and with free boundaries” in the contribution statement, and title the evaluation contribution “Evaluation on moving and free boundaries.” Long-time claims will follow the measured accuracy and cost in the horizon study.
  • Introduction and conclusion. Multiphysics coupling will remain a motivation and a description of the Scriven cases, where heat transfer, phase change, and modal flow couple through the interface. General descriptions of NOVA will use “moving-boundary” or “free-boundary.”
  • Table 1 and experiment overview. We will state the three benchmark groups in the table caption and the opening paragraph of the experiments, making the coverage of prescribed geometry and coupled interface dynamics explicit.

Thank you for the detailed response and the additional Stefan comparisons, which clarify the accuracy and computational tradeoffs. I also appreciate the proposed revisions to the terminology and scope. These responses address my main concerns, and I will maintain my score of 3 (Weak acceptance).

Thank you very much for your time, thoughtful engagement, and positive assessment of our work. We truly appreciate your constructive suggestions.

Reviewer_fJWaOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

NOVA extends frozen neural-basis solvers to moving-boundary PDEs through boundary lifts, nullspace representations, and compatible time stepping. It proposes horizontal frame transport to reduce repeated decompositions and a rank-saturation criterion for boundary sampling. Eight benchmarks demonstrate low boundary residuals and competitive field accuracy, with runtime savings from less frequent rebuilds.

Critical strengths

The framework provides a useful way to enforce evolving boundary constraints without tuning penalty or stabilization parameters.

Eight test cases, extensive baseline parameter sweeps, and refresh-interval ablations provide substantial evidence for boundary accuracy and computational savings.

Critical weaknesses

The distinct contribution of horizontal transport is unclear: the stated construction satisfies PBP=PPBP=P, so the subsequent correction annihilates the Euler predictor term in exact arithmetic. A comparison against projection-only frame updates is missing.

The experiments insufficiently validate coupled interface dynamics. Most cases use prescribed geometry, while the modal Scriven field error measures only the spherical temperature component; interface, higher-mode, and conservation errors are not quantified.

Questions for authors

Given (I−PB)P=0(I-PB)P=0, how does the corrected update in Appendix B.5.1 differ from directly projecting the previous frame? Can you compare their accuracy and runtime?

Can you report interface-position errors, conservation errors, and higher-mode accuracy against an independent reference for the free-boundary cases, particularly modal Scriven?

2 comments
AuthorsOpenReview ↗

Response to Reviewer fJWa

We sincerely appreciate the reviewer's constructive comments and careful assessment.

W1 & Q1 — Projection derived from the Euler step

Substituting the endpoint secant B˙=(Bn+1−Bn)/Δt\dot B=(B_{n+1}-B_n)/\Delta t used in the released implementation into the Euler step in line 9 gives

V_n-\Delta t\\,P_{n+1}\dot B V_n &=V_n-P_{n+1}(B_{n+1}-B_n)V_n\\\\ &=(I-P_{n+1}B_{n+1})V_n+P_{n+1}B_nV_n. \end{aligned}$$ Since the previous frame satisfies $B_nV_n=0$ in exact arithmetic, the Euler step yields the projection $(I-P_{n+1}B_{n+1})V_n$. In finite precision, $B_nV_n\approx0$, and the difference is the residual term $P_{n+1}B_nV_n$. Thus, the Euler step followed by two corrections acts approximately as three successive $P$ projections. In exact arithmetic, $PBP=P$ implies $(I-PB)^2=I-PB$, so repeated projections reduce to one. In finite precision, repeated corrections can improve accuracy for ill-conditioned boundary matrices. The following ablation uses the Table 1 settings. The tables report boundary and field errors for all eight benchmarks. No update keeps the frame fixed between scheduled rebuilds. **Boundary error $E_B$** | Case | No update | Euler step | One projection | Euler + one correction | Two projections | Algorithm 1 | Three projections | |---|---:|---:|---:|---:|---:|---:|---:| | Burgers | 2.47e-16 | 2.47e-16 | 2.33e-16 | 2.33e-16 | 2.30e-16 | 2.30e-16 | 2.29e-16 | | Dynamic disk | 1.88e+00 | 2.51e-07 | 6.43e-08 | 2.09e-10 | 2.09e-10 | 2.09e-10 | 2.09e-10 | | Moving cylinder | 5.71e-02 | 6.01e-08 | 2.02e-08 | 4.34e-14 | 4.44e-14 | 1.14e-15 | 1.14e-15 | | Navier–Stokes | 4.40e-06 | 5.82e-08 | 3.57e-08 | 3.57e-08 | 3.57e-08 | 3.57e-08 | 3.57e-08 | | Topology change | 7.97e-01 | 6.52e-05 | 2.65e-05 | 2.71e-06 | 2.43e-06 | 3.57e-08 | 3.24e-08 | | Stefan | 2.49e-02 | 4.83e-16 | 4.14e-16 | 4.06e-16 | 4.22e-16 | 4.11e-16 | 4.25e-16 | | Scriven bubble | 6.29e-02 | 2.12e-12 | 7.74e-13 | 8.43e-13 | 9.37e-13 | 8.79e-13 | 8.59e-13 | | Modal Scriven | 2.26e-07 | 8.28e-15 | 4.09e-15 | 4.09e-15 | 4.15e-15 | 4.15e-15 | 4.08e-15 | **Field error $E_{\mathrm{field}}$** | Case | No update | Euler step | One projection | Euler + one correction | Two projections | Algorithm 1 | Three projections | |---|---:|---:|---:|---:|---:|---:|---:| | Burgers | 3.02e-06 | 3.02e-06 | 3.02e-06 | 3.02e-06 | 3.02e-06 | 3.02e-06 | 3.02e-06 | | Dynamic disk | 2.45e+00 | 8.53e-07 | 8.24e-07 | 8.24e-07 | 8.23e-07 | 8.23e-07 | 8.24e-07 | | Moving cylinder | 1.26e-01 | 7.89e-02 | 7.89e-02 | 7.89e-02 | 7.89e-02 | 7.89e-02 | 7.89e-02 | | Navier–Stokes | 4.45e-05 | 8.26e-06 | 7.93e-06 | 7.92e-06 | 7.93e-06 | 7.93e-06 | 7.93e-06 | | Topology change | 1.22e+00 | 6.57e-05 | 2.39e-05 | 3.49e-06 | 3.27e-06 | 2.36e-06 | 2.35e-06 | | Stefan | 2.54e-02 | 3.22e-05 | 3.22e-05 | 3.22e-05 | 3.22e-05 | 3.22e-05 | 3.22e-05 | | Scriven bubble | 1.98e-02 | 6.37e-06 | 6.37e-06 | 6.37e-06 | 6.37e-06 | 6.37e-06 | 6.37e-06 | | Modal Scriven | 2.54e-07 | 2.54e-07 | 2.54e-07 | 2.54e-07 | 2.54e-07 | 2.54e-07 | 2.54e-07 | Updating the frame reduces field error by several orders of magnitude on four benchmarks compared with keeping it fixed between rebuilds. Repeated corrections further improve boundary accuracy. Algorithm 1 reduces boundary error by approximately two to seven orders of magnitude relative to one projection on three benchmarks. On Topology change, field error also decreases from 2.39e-5 to 2.36e-6. Across all eight benchmarks, Euler with one correction and two projections give comparable accuracy, as do Algorithm 1 and three projections. This supports the interpretation that the accuracy gain comes from successive projections in finite precision.
AuthorsOpenReview ↗

Response to Reviewer fJWa

W2 & Q2 — Coupled interface dynamics

We evaluate the coupled dynamics through interface trajectories, global energy balances, and higher-order modes. The following results use the Table 1 configurations, including each baseline's selected parameters.

Interface accuracy. Relative L2L^2-in-time errors against the analytic interface trajectories are

CaseNOVABaumgarteNitschePenalty
Stefan2.12e-052.06e-061.00e-023.37e-02
Scriven bubble2.07e-056.15e-061.02e-041.02e-04
Modal Scriven3.86e-072.46e-077.16e-062.93e+01

These results directly assess interface accuracy alongside the boundary and field errors.

Global conservation. We evaluate the energy balance using the reconstructed thermal field, interface motion, and boundary fluxes. The maximum balance defects, normalized by the enthalpy change for Stefan and the latent-heat change for the Scriven cases, are

CaseNOVA normalized energy-balance defect
Stefan5.62e-5
Scriven bubble7.77e-5
Modal Scriven, base state8.69e-7

These diagnostics directly assess consistency between thermal evolution and interface motion. For the Scriven models, mass conservation is built into the radial velocity prescribed by the interface speed.

Higher-order modes. We compare the modal solution with an independent Chebyshev-collocation reference using adaptive BDF time integration. For the flow-feedback benchmark at T=5T=5 ms, the relative errors in shape a2a_2, temperature T2T_2, streamfunction ψ2\psi_2, and vorticity g2g_2 are

Methoda2a_2T2T_2ψ2\psi_2g2g_2
NOVA2.03e-042.90e-043.81e-031.76e-03
Baumgarte1.62e+029.86e+012.31e+031.50e+05
Nitsche2.85e-022.59e-021.03e-011.54e-01
Penalty3.73e+025.37e+008.04e+011.35e+00

NOVA tracks the shape and temperature modes to approximately 10−410^{-4} relative error and the flow modes to 10−310^{-3} against this reference. The Baumgarte configuration selected by the spherical field metric develops a large shape-mode error: its final a2/R0a_2/R_0 is -6.76, compared with the reference value 0.01185. The mode-resolved comparison therefore reveals a substantial accuracy advantage for NOVA in the coupled perturbation dynamics.

The errors in the table use a reference with a smooth far-field taper at ξ=0.9\xi=0.9 for the initial vorticity's 1/r1/r tail. The measured sensitivity of the final shape amplitude to far-field treatment is 0.4–2%. In the compatible flow-feedback-off variant, the independent reference is converged to 10−910^{-9}. As the time step decreases from 10−610^{-6} to 5×10−75\times10^{-7} and 2.5×10−72.5\times10^{-7} s, NOVA's shape-mode error decreases from 2.43e-3 to 8.41e-4 and 4.41e-4, with comparable reductions in temperature and flow errors. This supplies additional convergence evidence for the higher-order dynamics.

Reviewer_ZyqoOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces NOVA, a causal, time-marching mesh-free framework for solving moving-boundary PDEs. It addresses the high cost of frequent SVDs and the difficulty of enforcing boundary conditions without penalty weights. By decomposing coefficients into a boundary lift and a nullspace component, and updating the nullspace frame via horizontal transport on the Stiefel manifold, the method avoids full rebuilds at every step. State reprojection maintains physical continuity. Experiments across eight benchmarks demonstrate that NOVA achieves machine-precision-like boundary residuals while remaining competitive in field accuracy and avoiding parameter tuning.

Critical strengths

The primary strength is the structural enforcement of boundary conditions. By mapping the solution to a boundary-compatible affine space via nullspace decomposition, the method eliminates the need for tuning penalty parameters.

Another critical strength is the mathematical formulation of the horizontal transport. Updating the nullspace frame between full SVDs drastically reduces computational overhead, and defining the velocity field to minimize the Frobenius norm on the Stiefel manifold ensures that the basis updates remain physically meaningful and stable.

Critical weaknesses

The experimental setup regarding hyperparameter scanning is somewhat unbalanced. The paper emphasizes that NOVA requires no boundary parameter tuning. However, reporting baselines at their absolute best over a massive grid search while NOVA uses a single fixed configuration makes the field error comparisons slightly biased in favor of the baselines. A more indicative comparison might involve a fixed computational budget for tuning.

Additionally, while the state reprojection ensures continuity across basis refreshes, the paper does not thoroughly discuss the accumulation of projection error over long time horizons, especially during complex topological changes. It is unclear if the least-squares reprojection introduces numerical dissipation that affects long-term dynamics.

Questions for authors
  1. Regarding the state reprojection, how does the method handle the accumulation of least-squares projection errors over extended time integrations? Does the projection error scale with the frequency of basis refreshes, and is there a theoretical bound on this error?

  2. For the topology change benchmark, the interface splits. How does the algorithm dynamically detect the need for a full rebuild versus relying on horizontal transport when the boundary matrix undergoes such a drastic structural change?

3 comments
AuthorsOpenReview ↗

Response to Reviewer Zyqo

We sincerely appreciate the reviewer's constructive comments and careful assessment.

W1 — Experimental setup

We admit that the extensive parameter sweeps favor the baselines. This intentional design tests whether NOVA can still maintain its advantage against well-tuned soft-constraint methods. NOVA achieves competitive field accuracy and boundary errors several orders of magnitude lower on multiple benchmarks. Its hard-constraint formulation eliminates the need to tune additional boundary penalty or stabilization parameters. These results demonstrate its structural advantage and robustness to boundary-parameter tuning. We will make this experimental motivation explicit.

W2 & Q1 — Error evolution

Equation (5) fits the initial state. Subsequent steps, including basis-refresh steps, solve Eq. (6) in the current compatible space using BDF2, with BE for startup and restart. For the linear update, Comment 2 bounds the field error in terms of the initial error, errors from the discrete equations, and approximation errors from the finite basis and truncation. The retained singular values control error amplification in each solve, and the update matrices describe how these errors propagate over time.

We complement this analysis with time-resolved field errors across all eight benchmarks.

We evaluate all eight benchmarks against their analytical references at 4096 independent spatial points at every time step, using the Table 1 settings and each benchmark's original time interval. The maximum errors include initialization.

CaseMaximum field error, including initializationFinal field error
Burgers4.455e-062.376e-06
Dynamic disk1.552e-061.445e-06
Moving cylinder5.842e-023.630e-02
Navier–Stokes9.064e-067.751e-06
Topology change4.503e-061.940e-07
Stefan8.793e-051.008e-05
Scriven bubble1.012e-055.440e-06
Modal Scriven5.123e-075.123e-07

Errors are relative L2L^2 field errors. Moving cylinder uses RMS velocity error normalized by the prescribed velocity scale, accommodating times of zero reference velocity. Modal Scriven evaluates the spherical temperature component T0T_0. The measured maxima quantify the combined error over each tested interval and provide empirical evidence for the accuracy of the complete evolution. The complete error trajectories are presented in Figure 6 in Appendix E of our revision.

AuthorsOpenReview ↗

Response to Reviewer Zyqo

W2 & Q1 — Error propagation bound

For PDEs that are linear in the unknown field, we derive the error bound directly from the BDF2 and truncated-SVD update in exact arithmetic.

Let h be the time step, MnM_n the interior feature matrix, JnJ_n the spatial operator matrix, and WnW_n the least-squares weights. Define

An=Mn−2h3Jn,bn=Mn(43cn−1−13cn−2)+2h3fn.A_n=M_n-\tfrac{2h}{3}J_n,\quad b_n=M_n(\tfrac43c_{n-1}-\tfrac13c_{n-2})+\tfrac{2h}{3}f_n.

Writing cn=cp,n+Vnznc_n=c_{p,n}+V_nz_n, set Aˉn=Wn1/2An\bar A_n=W_n^{1/2}A_n, Dn=AˉnVnD_n=\bar A_nV_n, and dn=Wn1/2bn−Aˉncp,nd_n=W_n^{1/2}b_n-\bar A_nc_{p,n}. The actual update is

cn=cp,n+Pndn,Pn=VnDn,τ†,c_n=c_{p,n}+P_nd_n,\qquad P_n=V_nD_{n,\tau}^{\dagger},

where Dn,τ†D_{n,\tau}^{\dagger} inverts the retained singular directions and accommodates rank-deficient matrices. PnP_n maps residuals to coefficient corrections.

For an independently chosen reference representation Φcn∗\Phi c_n^{\ast} of the exact solution, define enc=cn−cn∗e_n^c=c_n-c_n^{\ast} and

ηnc=Wn1/2[Ancn∗−Mn(43cn−1∗−13cn−2∗)−2h3fn],\eta_n^c=W_n^{1/2}\left[A_nc_n^{\ast}-M_n(\tfrac43c_{n-1}^{\ast}-\tfrac13c_{n-2}^{\ast})-\tfrac{2h}{3}f_n\right],

ξn=(I−PnAˉn)(cn∗−cp,n).\xi_n=(I-P_n\bar A_n)(c_n^{\ast}-c_{p,n}).

The term ηnc\eta_n^c is the residual obtained by inserting the reference representation into the discrete equation. The term ξn\xi_n is the part of that representation left out by the current affine projection, including the effects of compression and truncation.

In exact arithmetic, the LS residual ℓn=Dnzn−dn\ell_n=D_nz_n-d_n is orthogonal to the retained left singular space, so Pnℓn=0P_n\ell_n=0. Also, PnAˉn(cn−cp,n)=cn−cp,nP_n\bar A_n(c_n-c_{p,n})=c_n-c_{p,n}. Subtracting the reference equation therefore gives, with en=(enc,en−1c)T\mathbf e_n=(e_n^c,e_{n-1}^c)^T,

0\end{bmatrix},\quad H_n=\begin{bmatrix}\tfrac43P_nW_n^{1/2}M_n&-\tfrac13P_nW_n^{1/2}M_n\\\\ I&0\end{bmatrix}.$$ Iterating and taking Euclidean/induced matrix norms yields $$\lVert \mathbf e_n\rVert _2\le\lVert H_n\cdots H_2\rVert _2\lVert \mathbf e_1\rVert _2+ \sum_{j=2}^n\lVert H_n\cdots H_{j+1}\rVert _2 \left(\lVert P_j\rVert _2\lVert \eta_j^c\rVert _2+\lVert \xi_j\rVert _2\right).$$ Empty products are the identity. For a nonempty retained set, $$\lVert P_j\rVert _2\le\frac{\lVert V_j\rVert _2}{\sigma_{\min,\mathrm{ret}}(D_j)}.$$ The field error on $\Omega_n=\Omega(t_n)$ then satisfies $$\lVert u_h^n-u(t_n)\rVert _{L^2(\Omega_n)}\le \lVert \Phi\rVert _{2\to L^2(\Omega_n)}\lVert e_n^c\rVert _2+ \lVert \Phi c_n^{\ast}-u(t_n)\rVert _{L^2(\Omega_n)}.$$ Here $\lVert \Phi\rVert _{2\to L^2(\Omega_n)}$ converts a bound on coefficient error into a bound on field error. The final field error is bounded by the propagated initial error, errors from the discrete equations, and representation errors from the finite basis and truncation.
AuthorsOpenReview ↗

Response to Reviewer Zyqo

Q2 — Topology-event handling

We thank the reviewer for this question.

We compare the results with event-triggered rebuilding enabled and disabled, evaluating errors at 3000 independent test points.

The released implementation rebuilds the basis on the scheduled refresh steps, when the constraint-matrix shape changes, or when the geometry module reports a topology change. For the Cassini benchmark, the geometry supplies the connected-component count from b(t)−ab(t)-a.

Event-triggered rebuildFull rebuilds after initializationSpace-time relative L2L^2 errorBoundary RMS error
Enabled31.920e-64.95e-8
Disabled21.919e-64.53e-8

With topology-event triggering disabled, NOVA continues through the split using scheduled refreshes, preserving field accuracy and slightly reducing boundary error. Figure 7 in Appendix E of the revised manuscript shows that the rank recomputed at every step remains 63 across the split. The corresponding compressed dimension is 128 on both sides. Thus, accuracy is maintained with event-triggered rebuilding disabled in this Cassini benchmark.

Section 2.6 explains that sufficiently rich boundary sampling captures the independent constraints of the current trace space, with numerical evidence in Section 3.6 and Figure 3. In this benchmark, the boundary samples follow the evolving geometry, and the retained rank remains 63 across the split. Together with the preserved accuracy when event-triggered rebuilding is disabled, these results support that the existing sampling and update adequately resolve the evolving boundary constraints. We will include this ablation and the rank diagnostics in the revision.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank the Area Chair and all reviewers for their careful assessment and constructive comments. Our responses provide additional theoretical analysis, experiments, and clarifications on NOVA's error propagation, discrete updates, interface dynamics, and computational cost. The main additions are summarized below, with references to the corresponding comments in our rebuttal.

In our response to Reviewer Zyqo, Comment 1, we clarify that the extensive parameter sweeps intentionally favor the baselines to test NOVA against well-tuned soft-constraint methods. NOVA maintains competitive field accuracy and substantially lower boundary errors without additional boundary penalty or stabilization parameters. We also report time-resolved field errors across all eight benchmarks, with the complete trajectories in Figure 6 of the revised manuscript. Comment 2 derives an error propagation bound directly from the linear BDF2 and truncated-SVD update. It accounts for initial error, discrete-equation errors, and finite-basis and truncation errors, and identifies how retained singular values and update matrices control their propagation. Comment 3 compares enabling and disabling topology-triggered rebuilding in the Cassini benchmark. Disabling this trigger preserves accuracy. Figure 7 shows that the retained rank remains unchanged across the split, supporting the sampling and update mechanism discussed in Sections 2.6 and 3.6.

In our response to Reviewer fJWa, Comment 1, we expand the Euler step algebraically and show that it yields a projection when the previous frame satisfies the boundary constraints. Together with the two corrections, Algorithm 1 acts approximately as three successive projections in finite precision. We provide boundary and field errors for seven update variants across all eight benchmarks. These ablations demonstrate the benefit of updating the frame and the additional boundary-accuracy gains from repeated corrections. Comment 2 adds direct interface-trajectory errors, global energy-balance diagnostics, and separate shape, temperature, and flow-mode errors against an independent numerical reference. These results assess the coupled interface dynamics beyond the field and boundary metrics.

In our response to Reviewer hQ3p, Comment 1, we add PINN and PIELM comparisons on the shared two-phase Stefan benchmark, together with NOVA's time-step refinement results. NOVA exhibits approximately second-order convergence in field and interface errors while maintaining boundary errors near machine precision. We also compare accuracy, runtime, and additional memory over increasing time horizons. At eight periods, NOVA is approximately eight times faster and uses approximately one seventy-sixth of the additional memory of the tested dense global PIELM configuration. Comment 2 specifies revisions that describe the scope as PDEs with moving and free boundaries and distinguish fixed-domain, prescribed moving-domain, and coupled free-boundary benchmarks.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.