acceptodds
Under review as a conference paper at ICLR 2027

RecastNet: Adapting Neural Solvers to Network Topology Changes Without Retraining

Abstract

Many scientific and engineering applications require repeatedly solving large-scale networked systems under varying operating conditions and topologies, with local component equations coupled through the network. Neural solvers can accelerate these computations, but training them to cover the combinatorially many possible topologies is costly and often impractical. We present Neural Network Recasting (RecastNet), which shifts topology adaptation from training to inference: a neural solver is trained only for solution prediction on a single base topology and analytically recast under unseen topology changes. Our key finding is that the input–output Jacobian of the base-topology neural solver can recover the underlying network Laplacian inverse, allowing low-rank topology changes to be incorporated analytically. This yields zero-shot adaptation to unseen topology changes with only lightweight test-time computation. Experiments on synthetic networks and real-world electric grids show that RecastNet consistently outperforms graph-learning and topology-adaptation baselines under topology changes of increasing severity. Overall, RecastNet offers a new paradigm for topology adaptation through analytical recasting rather than cross-topology training.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
neural solver, topology adaptation, graph learning, scientific machine learning, low-rank matrix updates.

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_6dhiOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Neural solvers for networked systems usually require retraining when topology changes, yet real changes typically affect only a few links. RecastNet exploits this by treating such changes as low-rank updates to the network Laplacian and handling them analytically rather than learning them. It first recovers the base-topology Laplacian inverse from the input–output Jacobian of a frozen neural solver, then combines it with the low-rank change to produce solutions for unseen topologies without retraining or fine-tuning. For linear systems, only directional inverse actions along the changed links are needed, keeping computation low-dimensional. For nonlinear systems, an auxiliary solver trained on Newton-step pairs enables zero-shot nonlinear recasting. The authors provide error bounds and recastability conditions, and validate on synthetic networks and real power grids, reporting accurate zero-shot adaptation and 18–150× inference speedups over numerical solvers.

Critical strengths
  1. The paper addresses a real and widely encountered problem. In networked infrastructure, topology changes are often localized, while training neural solvers to cover many possible configurations can be expensive and may generalize poorly to unseen topologies. Moving topology adaptation from training to an analytical correction is an appealing and well-motivated perspective.

  2. The solution-to-inverse insight in Theorem 1 is particularly elegant. It is architecture-agnostic and shows how a trained neural solver can provide useful structural information beyond its solution predictions. Combining this insight with low-rank topology updates and the SMW formula leads to a coherent method with substantial theoretical support.

  3. The experiments cover a broad range of settings, including different network sizes, three synthetic graph families, four power-grid benchmarks, several nodal relations, and topology changes of varying severity. This breadth makes the empirical evaluation convincing within the tested regimes.

Critical weaknesses
  1. The paper lacks discussion of how the method degrades under high-rank or large-scale topology changes.
Questions for authors
  1. As the rank r of the topology change grows (e.g., network partitioning, simultaneous failure of many links), how do RecastNet's error and computational cost change? Is there a characterizable "recastable rank limit" or break-even point? Could the authors provide experimental or theoretical analysis across different r/n ratios?

  2. If the base-topology neural solver itself has limited accuracy, how does recasting performance degrade?

6 comments
AuthorsOpenReview ↗

Part[1/2] Response to Reviewer 6dhi 1.1 & 1.2

We thank the reviewer for the constructive feedback and positive remarks on "a real and widely encountered problem," the "particularly elegant" solution-to-inverse insight, "substantial theoretical support," and the "breadth" of our experiments. We address each comment below and have highlighted the manuscript revisions in blue. We hope these revisions address your concerns.

Weakness and Question 1 W: The paper lacks discussion of how the method degrades under high-rank or large-scale topology changes. Q1: As the rank r of the topology change grows (e.g., network partitioning, simultaneous failure of many links), how do RecastNet's error and computational cost change? Is there a characterizable "recastable rank limit" or break-even point? Could the authors provide experimental or theoretical analysis across different r/n ratios?

We thank the reviewer for this important question. The original paper reports low recasting errors on ER, SBM, and BA networks with up to 2020\\% of the edges removed (see Figure 3, Page 7{\color{#003A8C}\text{Figure 3, Page 7}} and Table 1, Page 8{\color{#003A8C}\text{Table 1, Page 8}}). Below, we extend this study in three aspects: a rank-dependent error bound that characterizes conditions for a "recastable rank limit," a rank-dependent computational complexity analysis, and experiments evaluating prediction error and runtime with up to 3030\\% of the edges removed.

(I) Effect of topology-change rank on recasting error.

Eq. (8), Corollary 3, Page 5{\color{#003A8C}\text{Eq. (8), Corollary 3, Page 5}}, bounds the recast-solution error by two terms: the contribution from the base prediction error ε0\varepsilon_0 and the contribution from the directional inverse-action recovery error EH,ΔE_{H,\Delta}. The latter measures how accurately the neural solver recovers the base-network inverse along the topology-change directions.

We extend this bound to make the dependence on the topology-change rank rr explicit. If each recovered inverse-action column has error at most δH\delta_H, then: ∥EH,Δ∥2≤∥EH,Δ∥F≤rδH\lVert E_{H,\Delta}\rVert_{2} \leq \lVert E_{H,\Delta}\rVert_{F} \leq \sqrt{r}\delta_{H}. Substituting this relationship into Eq.(8) gives: ∥x^G−xG∥2≤c0,Δ∥ε0∥2+cH,Δr,δH\lVert \widehat{x}_{G}-x_{G}\rVert_{2}\leq c_{0,\Delta}\lVert\varepsilon_{0}\rVert_{2}+c_{H,\Delta}\sqrt{r}\\,\delta_{H} The bound identifies a square-root dependence on the topology-change rank: the inverse-action error contribution grows as r\sqrt{r}, scaled by the topology-dependent amplification coefficient cH,Δc_{H,\Delta}. It therefore characterizes how increasing rank can degrade recasting accuracy, while accounting for differences among topology changes. We provide the derivation in Appendix C.2.3, Page 22{\color{#003A8C}\text{Appendix C.2.3, Page 22}} of the revised manuscript.

Is there a recastable rank limit? Our analysis suggests that there may not be a universal recastable rank limit. For a prescribed tolerance ϵ\epsilon, requiring the right-hand side of the bound to be at most ϵ∥xG∥2\epsilon\|x_G\|_2 gives a sufficient condition for ϵ\epsilon-recastability. However, equal-rank topology changes can produce substantially different errors because they affect different links and amplify approximation errors differently. In an extreme case, even a rank-one change that removes a bridge edge can disconnect the network and make LGL_G singular. Thus, a rank threshold alone may not guarantee that recasting meets a prescribed accuracy tolerance.

(II) Effect of topology-change rank on computational cost.

For a rank-rr topology change, Proposition 2 reduces inverse-action recovery and topology correction to linear solves of dimensions at most q:=psh+rq:=p_{\mathrm{sh}}+r and rr, respectively. The inference cost of RecastNet is thus: Csingle(n,r)=Cfwd+Csens+O(rCJVP+nq2+q3+r3),C_{\mathrm{single}}(n,r) = C_{\mathrm{fwd}}+C_{\mathrm{sens}}+ \mathcal O\left(rC_{\mathrm{JVP}}+nq^2+q^3+r^3\right),

where CfwdC_{\mathrm{fwd}}, CsensC_{\mathrm{sens}}, and CJVPC_{\mathrm{JVP}} denote the costs of the base prediction, sensitivity evaluation, and one neural Jacobian-vector product (JVP), respectively. The nq2+q3nq^2+q^3 terms account for inverse-action recovery, and r3r^3 accounts for the topology-correction solve. We add this complexity analysis to Appendix C.2.1, Page 21{\color{#003A8C}\text{Appendix C.2.1, Page 21}}.

AuthorsOpenReview ↗

Part[1/2] Response to Reviewer 6dhi 1.3

(III) Experimental evaluation under increasing topology-change ranks.

Next, we present new experimental results to complement the above rank-dependent error bound and complexity analysis. We evaluate accuracy and runtime on the 250-node ER network as the topology-change rank increases from r/N=0.4r/N=0.4\\% to 3030\\%, where rr is the rank of the topology update and N=250N=250. All cases use the same frozen base-topology neural solver. Prediction accuracy is evaluated against the true target-topology solution.

Accuracy. The recast error grows gradually as the topology-change rank increases. As shown in the table below, raising r/Nr/N from 0.40.4\\% to 3030\\%--a fifteenfold increase in rank--raises the median error from 1.351.35\\% to 2.552.55\\% and the P95 error from 5.725.72\\% to 10.3810.38\\%, corresponding to roughly a 1.8×1.8\times error increase in both cases. The aforementioned theoretical error bound characterizes this dependence through rank-dependent error accumulation and topology-dependent amplification. In these experiments, their combined effect leads to a modest increase in error, with the median error remaining below 2.62.6\\% even at r/N=30r/N=30\\%.

rrr/Nr/NRecastNet error (%) MedianRecastNet error (%) P95Base solver error (%) MedianBase solver error (%) P95RecastNet time (ms/case)Numerical solver time (ms/case)Speedup
10.40.4\\%1.355.721.7213.640.0451.73938.64×38.64\times
52.02.0\\%1.395.781.817.490.2211.7507.92×7.92\times
104.04.0\\%1.486.213.6416.650.3091.7505.66×5.66\times
208.08.0\\%1.606.513.9715.060.6001.7522.92×2.92\times
3012.012.0\\%1.797.244.9518.460.8921.7441.96×1.96\times
4016.016.0\\%1.887.6414.6744.781.2031.6831.40×1.40\times
5020.020.0\\%1.877.7914.5132.631.6071.7111.06×1.06\times
6224.824.8\\%2.369.7115.9044.372.1791.7180.79×0.79\times
7530.030.0\\%2.5510.3819.7044.682.7021.6630.62×0.62\times

Computational cost. We use a batch size of 512512, consistent with the original paper, and report average inference time per case, including data transfers. As r/Nr/N increases from 0.40.4\\% to 3030\\%, the average runtime of RecastNet increases from 0.0450.045 to 2.7022.702ms. RecastNet remains comparable to the numerical solver at 2020\\% (1.6071.607 versus 1.7111.711 ms/case), while the numerical solver is faster at 24.824.8\\% and 3030\\%. The measured runtime crossover therefore lies between the sampled 2020\\% and 24.824.8\\% levels under this implementation and hardware setup.

This behavior is consistent with the complexity analysis: when considering fixed pshp_{\mathrm{sh}}, the matrix-operation cost is O(Nr2+r3)\mathcal O(Nr^2+r^3). Increasing r/Nr/N therefore gradually increases the cost to O(N3)\mathcal O(N^3), and the correction systems are no longer much smaller than the original system. The computational benefit consequently diminishes as the update rank grows, as is also the case for classical SMW updates. The original runtime experiments focus on the low-rank regime motivated by localized outages and network reconfiguration. The additional results clarify that, although recasting maintains a median error below 2.62.6\\% through r/N=30r/N=30\\%, its runtime advantage does not extend across this entire range. We thank the reviewer for prompting this broader evaluation and have clarified this distinction in the revised manuscript.

AuthorsOpenReview ↗

Part[2/2] Response to Reviewer 6dhi 2.1 & 2.2

Question 2 If the base-topology neural solver itself has limited accuracy, how does recasting performance degrade?

We thank the reviewer for this question. Corollary 3, Page 5{\color{#003A8C}\text{Corollary 3, Page 5}} in the original submission characterizes how the base-topology neural solver's approximation errors propagate through recasting. Below, we add additional experiments to study how performance degrades as the base predictions become less accurate.

(I) Error propagation from the base neural solver.

Corollary 3 bounds the recast-solution error as ∥x^G−xG∥2≤c0,Δ∥ε0∥2+cH,Δ∥EH,Δ∥2,\lVert \widehat{x}_{G}-x_{G}\rVert_{2}\leq c_{0,\Delta}\lVert\varepsilon_{0}\rVert_{2}+c_{H,\Delta}\lVert E_{H,\Delta}\rVert_{2}, where ε0\varepsilon_0 is the base prediction error and EH,ΔE_{H,\Delta} is the directional inverse-action recovery error. The first term captures the propagation of the base prediction error, while the second accounts for inaccuracies in the inverse information recovered from the neural solver. The coefficients c0,Δc_{0,\Delta} and cH,Δc_{H,\Delta} characterize how these errors are amplified under a particular topology change. We examine the resulting effect on recasting accuracy experimentally below.

(II) Recasting with increasingly inaccurate base predictions.

We use the IEEE 118-bus test system and evaluate DC power flow as the power injections progressively move beyond the training range of the frozen base-topology neural solver. This increases the base prediction error from 0.070.07\\% to 10.7310.73\\%. At each input, we evaluate the same ten randomly sampled N−1N-1 outages in each of the medium- and high-impact groups.

The results show a gradual loss of recasting accuracy as the base prediction becomes less accurate. In these tested cases, median recast errors remain close to the base error even when it exceeds 1010\\%, while high-impact outages produce greater additional error, particularly in P95. Thus, recasting can accommodate topology changes despite substantial base prediction errors, but its accuracy still depends on the quality of the base approximation.

Base-topology errorRecast error (medium-impact outages) MedianRecast error (medium-impact outages) P95Recast error (high-impact outages) MedianRecast error (high-impact outages) P95
0.0710.071\\%0.1180.118\\%0.3430.343\\%0.3290.329\\%1.3851.385\\%
0.6020.602\\%0.6270.627\\%0.8100.810\\%1.1521.152\\%1.4991.499\\%
2.2842.284\\%2.2952.295\\%2.4802.480\\%2.9712.971\\%3.4533.453\\%
5.7505.750\\%5.7655.765\\%6.2546.254\\%6.2766.276\\%7.7097.709\\%
7.1417.141\\%7.1697.169\\%7.7837.783\\%7.5927.592\\%9.2979.297\\%
9.2659.265\\%9.3119.311\\%10.12810.128\\%9.6149.614\\%11.60411.604\\%
10.73110.731\\%10.79010.790\\%11.75311.753\\%11.01611.016\\%13.11213.112\\%
Reviewer_Kc6eOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The authors introduce RecastNet for adapting neural solvers to unseen graph topology changes without retraining. It recovers Laplacian inverse information from the neural solver's Jacobian and uses low-rank updates to correct predictions. Experiments on synthetic networks and real-world power grids show improved accuracy over learning-based methods and faster inference than numerical solvers.

Critical strengths
  1. The authors solve the adaptation of neural solvers to graph topology changes, which is practically relevant to applications such as power systems that require repeated numerical solutions.

  2. The paper shows a connection between the input–output Jacobian of a neural solver and the system sensitivity matrix, giving a method to recover the inverse of the graph Laplacian.

  3. The method only requires the inverse action along the topology-change directions, which avoids the explicit construction of the full inverse matrix.

Critical weaknesses
  1. Eqs. (1)–(2) rely on a specific Laplacian structure and node-local nonlinearities, but the paper does not provide a concrete derivation to show how realistic AC power flow equations satisfy these assumptions.

  2. Since the base Laplacian is known, the authors should compare against direct sparse matrix factorization combined with SMW updates to demonstrate the advantages of approximating the inverse operator using a neural network.

  3. Theorem 1 mainly follows from implicit differentiation of the network equations, while Proposition 2 relies on the classical SMW formula. However, it remains unclear what distinctive theoretical or computational contribution arises from combining these two techniques.

Questions for authors
  1. In Eqs. (4)–(6), the authors recover LG0−1BΔL_{G_0}^{-1}B_\Delta through the Jacobian of the neural solver. However, since the base Laplacian LG0L_{G_0} is already known, why not compute this quantity directly using sparse matrix factorization or the SMW formula?

  2. Eqs. (1)–(2) assume that the system can be expressed as LG=BWB⊤L_G = BWB^\top, with node-local nonlinearities bi=fi(xi,ξ)b_i = f_i(x_i, \xi). However, Section 4.2 directly applies the method to AC power flow. I am curious how the authors extend this formulation to the AC power flow equations.

  3. Eq. (10) uses the base model's prediction as initialization, followed by iterative residual corrections. Why is there no comparison with Newton–Raphson initialized with the same prediction? It is currently difficult to distinguish whether the reported improvements come from the learned inverse operator or simply from better initialization.

6 comments
AuthorsOpenReview ↗

Part[1/4] Response to Reviewer Kc6e

We thank the reviewer for the constructive feedback and recognition of our method’s strengths. We address each comment below and have highlighted the manuscript revisions in blue. We hope these revisions address your concerns and would appreciate it if you could consider revisiting your rating.

Weakness 1 and Question 2 W1: Eqs. (1)–(2) rely on a specific Laplacian structure and node-local nonlinearities, but the paper does not provide a concrete derivation to show how realistic AC power flow equations satisfy these assumptions. Q2: Eqs. (1)-(2) assume that the system can be expressed as LG=BWB⊤L_G = BWB^\top, with node-local nonlinearities bi=fi(xi,ξ)b_i = f_i(x_i, \xi). However, Section 4.2 directly applies the method to AC power flow. I am curious how the authors extend this formulation to the AC power flow equations.

Below, we clarify the AC formulation and its correspondence to Eqs. (1)-(2).

AC nodal variables and network equations. At bus ii, let ViV_i, IiI_i, and Si=Pi+jQiS_i=P_i+\mathrm{j}Q_i denote the complex voltage, current injection, and power injection, respectively. By the definition of complex power: Si=ViIi∗,Ii=Pi−jQiVi∗S_i=V_iI_i^{\ast}, I_i=\frac{P_i-\mathrm{j}Q_i}{V_i^{\ast}}, where ∗{}^{\ast}denotes complex conjugation. The network coupling follows from Kirchhoff's current law and the branch voltage-current relationships: I=Y~GV,Y~G=B~diag⁡(y)B~⊤I=\widetilde{ Y}_G V, \widetilde{Y}_G = \widetilde{B}\operatorname{diag}(y)\widetilde{B}^{\top}. Here, Y~G\widetilde{ Y}_G is the full bus-admittance matrix, yy contains the branch admittances, and B~\widetilde{B} encodes the edge connection (e.g., branches, transformer ratios, shunts).

AC formulation in the form of Eq.(2b). In AC power flow, the slack bus has a fixed voltage magnitude and phase angle. Its contribution can thus be moved to the right-hand side. Substituting the current-injection relation and stacking the real and imaginary parts then gives: [GG−BGBGGG]⏟LG[VRVI]⏟x=[Re⁡(ϕG)Im⁡(ϕG)]⏟fAC,G,ϕG=(P−jQ)⊘V∗−bG\underbrace{\begin{bmatrix}G_G & -B_G \\\\ B_G & G_G\end{bmatrix}}_{\displaystyle {\mathcal L}_G}\underbrace{\begin{bmatrix}V_{\mathrm R} \\\\ V_{\mathrm I}\end{bmatrix}}_{\displaystyle x}=\underbrace{\begin{bmatrix}\operatorname{Re}(\phi_G) \\\\ \operatorname{Im}(\phi_G)\end{bmatrix}}_{\displaystyle f_{\mathrm{AC},G}},\qquad \phi_G=(P-\mathrm{j}Q)\oslash V^{\ast}-b_G Here, $ V= V_{\mathrm R}+\mathrm{j} V_{\mathrm I}$ contains only non-slack voltages, $ G_G+\mathrm{j} B_G$ is the reduced admittance matrix, $ b_G$ is the contribution from the known slack-bus voltage. PQ buses have fixed active and reactive power injections. The unknown voltage states VR,iV_{\mathrm R,i} and VI,iV_{\mathrm I,i} are already included in xx, while the specified injections PispP_i^{\mathrm{sp}} and QispQ_i^{\mathrm{sp}} enter ϕG\phi_G. Thus, the PQ equations directly take the form shown above. PV buses have fixed active power injections and voltage magnitudes. The voltage states thus remain in xx, while the unknown reactive injection QiQ_i is also included in the state and the specified active injection PispP_i^{\mathrm{sp}} enters ϕG\phi_G. To enforce the voltage-magnitude setpoint VispV_i^{\mathrm{sp}}, we append the local equation Qi=Qi+(Visp)2−∣Vi∣2Q_i=Q_i+(V_i^{\mathrm{sp}})^2-|V_i|^2, which is equivalent to ∣Vi∣2=(Visp)2|V_i|^2=(V_i^{\mathrm{sp}})^2. This adds an identity block to the linear operator LG{\mathcal L}_G and the corresponding local terms to the nonlinear map ϕG\phi_G, without changing the network-admittance block. The complete AC system therefore retains the form of Eq.(2b), which is equivalent to standard rectangular AC power flow with the network Laplacian written explicitly.

Network Laplacian in AC power flow and its use in RecastNet. In Eq.(1), the network Laplacian is given in the form of $ L_G(\xi_{\mathrm L}) = B W(\xi_{\mathrm L}) B^\top$ where $ W=\operatorname{diag}(w_1,\ldots,w_m)$ is diagonal. In the AC formulation, each bus has two unknown voltage components, so LG{\mathcal L}_G defined above is the corresponding block matrix with 2×22\times2 blocks: LG=(B⊗I2),W,(B⊤⊗I2),W=blkdiag⁡(w1,…,wm),we=[ge−bebege]{\mathcal L}_G=(B\otimes I_2)\\,{\mathcal W}\\,(B^{\top}\otimes I_2),\qquad {\mathcal W}=\operatorname{blkdiag}(w_1,\ldots,w_m),\qquad w_e=\begin{bmatrix}g_e & -b_e \\\\ b_e & g_e\end{bmatrix} where ge+jbeg_e+\mathrm{j}b_e is the edge admittance and ⊗\otimes denotes the tendor product. Thus, scalar 1×11\times1 weights in the original Laplacian become 2×22\times2 blocks; however, the incidence structure is preserved.

The RecastNet workflow remains unchanged: only the local scalar inversions become independent 2×22\times2 block inversions, each of negligible cost, while topology updates remain low rank. No additional large-matrix inversion is required. This block-valued extension was originally noted in Footnote 2, Page 3{\color{#003A8C}\text{Footnote 2, Page 3}}; now Appendix C.4, Page 24{\color{#003A8C}\text{Appendix C.4, Page 24}} provides the complete details.

AuthorsOpenReview ↗

Part[2/4] Response to Reviewer Kc6e

Weakness 2 and Question 1 W2: Since the base Laplacian is known, the authors should compare against direct sparse matrix factorization combined with SMW updates to demonstrate the advantages of approximating the inverse operator using a neural network. Q1: In Eqs. (4)-(6), the authors recover LG0−1BΔL_{G_0}^{-1}B_{\Delta} through the Jacobian of the neural solver. However, since the base Laplacian LG0L_{G_0} is already known, why not compute this quantity directly using sparse matrix factorization or the SMW formula?

We thank the reviewer for suggesting this comparison. We would like to clarify the distinction between topology and link weights in our formulation. A fixed base topology specifies which nodes are connected, while the weights of those connections may vary with physical operating conditions and control settings.

Specifically, the governing equation is (Eq.(1-2), Page 3{\color{#003A8C}\text{Eq.(1-2), Page 3}}):$ L_{\mathcal G_0}(\xi_{\mathrm L}) x = b(\xi)$ with LG0(ξL)=B0W(ξL)B0⊤L_{\mathcal G_0}( \xi_{\rm L}) = B_0 W(\xi_{\mathrm L}) B_0^\top where the base topology G0\mathcal G_0 determines the incidence matrix $ B_0$, while ξL\xi_{\mathrm L} determines the link weights that can be impacted by physical operating conditions and control settings. Thus, a fixed base topology does not imply a fixed network matrix: even with unchanged topology, different operating conditions can yield different matrices through changes in the link weights.

This dependence on link parameters is common in physical infrastructure networks:

  • In electrical networks, conductor temperature and operating frequency change line resistance and hence the AC admittance matrix, while transformer tap settings and series compensation change the network matrices used in AC and DC power flow. These changes preserve network connectivity. Thus, evaluating the same grid under different environmental and operational conditions requires a family of network matrices.
  • In thermal networks, cooling airflow changes heat-transfer coefficients, so different fan settings yield different thermal conductance for the same network.
  • In fluid networks, viscosity and valve openings change hydraulic resistance while pipe connectivity remains fixed.

The reviewer's proposed factorization reuse is effective when link parameters remain fixed: one base-topology factorization for a fixed base-topology matrix can support multiple topology updates. In our setting, however, varying link parameters also changes the base-topology matrix, so a single cached inverse is insufficient and storing a separate inverse for every parameter value is infeasible. Our learned model instead recovers the required inverse actions at the requested condition from the same frozen base-topology neural solver, then applies the low-rank topology update.

Following the reviewer's suggestion, we compare RecastNet with a numerical baseline that uses SuperLU to factorize the condition-dependent base-topology matrix and applies an analytical low-rank update (SMW formula) for topology changes. We evaluate N-1 outages on all four test grids, with randomly varying link parameters and batch sizes of 32, 128, and 512. Speedups are computed from mean runtimes over ten independent runs. The table below shows speedups of 3.443.44-12.23×12.23\times at B=32B=32 and 29.6529.65-72.47×72.47\times at B=512B=512. Relative to Figure 5, this comparison changes both the baseline and the test networks: the baseline now exploits low-rank updates, and the power grid matrices yield sparser L and U factors than the synthetic matrices, so the SuperLU itself also becomes less costly. Even with both changes, RecastNet remains faster.

SystemBatch size 32Batch size 128Batch size 512
IEEE 57-bus3.44×3.44\times13.23×13.23\times29.65×29.65\times
IEEE 118-bus3.70×3.70\times15.24×15.24\times38.22×38.22\times
Carolina 500-bus5.57×5.57\times20.95×20.95\times71.05×71.05\times
European 1354-Bus12.23×12.23\times45.21×45.21\times72.47×72.47\times

In the revised paper, we have clarified condition-dependent inverse recovery in Remark 3, Page 4{\color{#003A8C}\text{Remark 3, Page 4}} and added the DC runtime comparison to Section 4.3, Page 9{\color{#003A8C}\text{Section 4.3, Page 9}} of the revised manuscript. As a supplementary check, we also explicitly implement the reviewer's proposed reuse strategy. For topology changes at a fixed operating condition, the base-network LU factorization is shared across outages. The numerical baseline is faster at B=32B=32, while RecastNet retains speedups of 1.821.82-4.12×4.12\times at B=512B=512. However, as explained above, these LU factors cannot be reused when changes in operating conditions alter the link parameters and hence the base-network matrix; the corresponding factorization must then be recomputed.

AuthorsOpenReview ↗

Part[3/4] Response to Reviewer Kc6e

Weakness 3 Theorem 1 mainly follows from implicit differentiation of the network equations, while Proposition 2 relies on the classical SMW formula. However, it remains unclear what distinctive theoretical or computational contribution arises from combining these two techniques.

We thank the reviewer for the opportunity to clarify our contribution. Our contribution lies in establishing how a neural solver trained only for base-topology solution prediction can be repurposed for analytical adaptation to unseen topology changes, together with an efficient computation procedure and recastability analysis. Implicit differentiation and the Sherman–Morrison–Woodbury (SMW) formula serve as mathematical tools in this derivation.

(I) Exposing inverse information beyond the neural solver's original prediction task.

We establish a solution-to-inverse relation that recovers the underlying Laplacian inverse from the neural solver's input-output Jacobian through the sensitivity matrix (see Theorem 1, Page 4{\color{#003A8C}\text{Theorem 1, Page 4}}): L^G0−1(ξ):=JN0(ξ)SN0(ξ)†≈LG0−1(ξ)\widehat{L}_{\mathcal G_0}^{-1}(\xi):=J_{\mathcal N_0}(\xi)S_{\mathcal N_0}(\xi)^{\dagger}\approx L_{\mathcal G_0}^{-1}(\xi)

This repurposes a pretrained neural solver beyond its original solution-prediction task: it can now supply inverse-operator information needed to characterize topology changes. The recovery is architecture-agnostic and requires no changes to the network architecture or trained parameters. Implicit differentiation supports the derivation, while the contribution is establishing this recoverable inverse capability and its use for topology adaptation.

(II) Structured analytical recasting for unseen topologies.

We develop a structured recasting method that uses the recovered inverse information to analytically compute the impact of topology changes and update the neural solution without retraining (see Proposition 2, Page 5{\color{#003A8C}\text{Proposition 2, Page 5}}). Using SMW as a derivation tool, the method requires only the directional inverse actions LG0−1BΔL_{\mathcal G_0}^{-1} B_\Delta and small matrix solves, avoiding construction of the full inverse, sensitivity pseudoinverse, or neural Jacobian (see Proposition 2 and Appendix C.2). Together with Theorem 1, this establishes a computational procedure for zero-shot topology adaptation of a frozen base-topology neural solver.

(III) Error and recastability characterization.

We bound the inverse-recovery error (see Appendix C.1.3, Page 19) and establish a sufficient condition for linear ϵ\epsilon-recastability that accounts for base prediction error, directional inverse-action error, and their topology-dependent amplification (see Corollary 3, Page 5{\color{#003A8C}\text{Corollary 3, Page 5}}). We further extend recasting to nonlinear systems through Dual-RecastNet (see Section 3.3) and characterize the nonlinear recastability condition in terms of the nonlinear Jacobian and inverse-recasting error (see Corollary 4).

Reviewer_a4fPOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

RecastNet adapts a frozen neural solver to known low-rank network changes. It estimates inverse actions from neural derivatives. A second learned map supports iterative correction for nonlinear systems. Experiments on synthetic graphs and four power grids report strong improvements over learned baselines.

Critical strengths
  1. Computing only the inverse actions needed for changed edges is sensible, and the structured sensitivity calculation keeps the matrix inversions small.
  2. Promising empirical breadth. The study covers several graph families, sizes and outage severities, with median and P95 errors.
Critical weaknesses

Reliability and reproduction are insufficiently documented. The paper does not fully specify the AC formulation, data splits/counts, training settings or numerical tolerances. It also lacks clear failure rates and near-limit tests tied to the derivative and contraction conditions. Also the practical use of the results is under question considering the possible big errors and failures, probably some uncertainty estimates would help.

Questions for authors
  1. For the linear/DC experiments, can you compare with a numerical solver that reuses the base-network factorization and applies low-rank updates? For AC power flow, can you compare with Newton–Raphson initialized from the same base-topology prediction?
  2. Please specify the AC equations and auxiliary-Jacobian evaluation point, complete the training/data protocol, and report residuals, componentwise errors, failure rates and conditioning diagnostics.
  3. In Figure 2, please specify how the examples in panel (b) relate to the distribution in panel (a), and provide the exact formula used for the entrywise relative errors. It seems that in (a) relative errors for N=1000 are larger than reported 0.21%.
  4. In Figure 4, please define the error bars and markers and explain how they relate to the medians reported in the text.
14 comments
AuthorsOpenReview ↗

Part[1/5] Response to Reviewer a4fP 1.1

We thank the reviewer for the constructive feedback and appreciate the positive remarks that our approach is "sensible" and that the experiments demonstrate "promising empirical breadth." We address each comment below and have highlighted the manuscript revisions in blue. We hope these revisions address your concerns and would appreciate it if you could consider revisiting your rating.

Weakness Reliability and reproduction are insufficiently documented. The paper does not fully specify the AC formulation, data splits/counts, training settings or numerical tolerances. It also lacks clear failure rates and near-limit tests tied to the derivative and contraction conditions.

(I) AC formulation and its connection to Eq.(2b)

The full derivation is presented in Appendix C.4, Page 24{\color{#003A8C}\text{Appendix C.4, Page 24}}.

AC nodal variables and network equations. At bus ii, let ViV_i, IiI_i, and Si=Pi+jQiS_i=P_i+\mathrm{j}Q_i denote the complex voltage, current injection, and power injection, respectively. By the definition of complex power: Si=ViIi∗,Ii=Pi−jQiVi∗S_i=V_iI_i^{\ast}, I_i=\frac{P_i-\mathrm{j}Q_i}{V_i^{\ast}} where ∗{}^* denotes complex conjugation. The network coupling follows from Kirchhoff's current law and the branch voltage-current relationships: I=Y~GV,Y~G=B~diag⁡(y)B~⊤I=\widetilde{ Y}_G V, \widetilde{Y}_G = \widetilde{B}\operatorname{diag}(y)\widetilde{ B}^{\top}. Here, Y~G\widetilde{ Y}_G is the full bus-admittance matrix, yy contains the branch admittances, and B~\widetilde{B} encodes the edge connection (e.g., branches, transformer ratios, shunts).

AC formulation in the form of Eq.(2b). In AC power flow, the slack bus has a fixed voltage magnitude and phase angle. Its contribution can thus be moved to the right-hand side. Substituting the current-injection relation and stacking the real and imaginary parts then gives:

[GG−BGBGGG]⏟LG[VRVI]⏟x=[Re⁡(ϕG)Im⁡(ϕG)]⏟fAC,G,ϕG=(P−jQ)⊘V∗−bG\underbrace{\begin{bmatrix}G_G & -B_G \\\\ B_G & G_G\end{bmatrix}}_{\displaystyle {\mathcal L}_G}\underbrace{\begin{bmatrix}V_{\mathrm R} \\\\ V_{\mathrm I}\end{bmatrix}}_{\displaystyle x}=\underbrace{\begin{bmatrix}\operatorname{Re}(\phi_G) \\\\ \operatorname{Im}(\phi_G)\end{bmatrix}}_{\displaystyle f_{\mathrm{AC},G}},\qquad \phi_G=(P-\mathrm{j}Q)\oslash V^{\ast}-b_G

Here, $ V= V_{\mathrm R}+\mathrm{j} V_{\mathrm I}$ contains only non-slack voltages, $ G_G+\mathrm{j} B_G$ is the reduced admittance matrix, $ b_G$ is the contribution from the known slack-bus voltage. PQ buses have fixed active and reactive power injections. The unknown voltage states VR,iV_{\mathrm R,i} and VI,iV_{\mathrm I,i} are already included in $ x$, while the specified injections PispP_i^{\mathrm{sp}} and QispQ_i^{\mathrm{sp}} enter ϕG\phi_G. Thus, the PQ equations directly take the form shown above. PV buses have fixed active power injections and voltage magnitudes. The voltage states thus remain in $ x$, while the unknown reactive injection QiQ_i is also included in the state and the specified active injection PispP_i^{\mathrm{sp}} enters ϕG\phi_G. To enforce the voltage-magnitude setpoint VispV_i^{\mathrm{sp}}, we append the local equation Qi=Qi+(Visp)2−∣Vi∣2Q_i=Q_i+(V_i^{\mathrm{sp}})^2-|V_i|^2, which is equivalent to ∣Vi∣2=(Visp)2|V_i|^2=(V_i^{\mathrm{sp}})^2. This adds an identity block to the linear operator LG{\mathcal L}_G and the corresponding local terms to the nonlinear map ϕG\phi_G, without changing the network-admittance block. The complete AC system therefore retains the form of Eq.(2b), which is equivalent to standard rectangular AC power flow with the network Laplacian written explicitly.

Network Laplacian in AC power flow and its use in RecastNet. In Eq.(1), the network Laplacian is given in the form of $ L_G(\xi_{\mathrm L}) = B W(\xi_{\mathrm L}) B^\top$ where $ W=\operatorname{diag}(w_1,\ldots,w_m)$ is diagonal. In the AC formulation, each bus has two unknown voltage components, so LG{\mathcal L}_G defined above is the corresponding block matrix with 2×22\times2 blocks:

LG=(B⊗I2),W,(B⊤⊗I2),W=blkdiag⁡(w1,…,wm),we=[ge−bebege]{\mathcal L}_G=(B\otimes I_2)\\,{\mathcal W}\\,(B^{\top}\otimes I_2),\qquad {\mathcal W}=\operatorname{blkdiag}(w_1,\ldots,w_m),\qquad w_e=\begin{bmatrix}g_e & -b_e \\\\ b_e & g_e\end{bmatrix}

where ge+jbeg_e+\mathrm{j}b_e is the edge admittance and ⊗\otimes denotes the tendor product. Thus, scalar 1×11\times1 weights in the original Laplacian become 2×22\times2 blocks; however, the incidence structure is preserved.

The RecastNet workflow remains unchanged: only the local scalar inversions become independent 2×22\times2 block inversions, each of negligible cost, while topology updates remain low rank. No additional large-matrix inversion is required. This block-valued extension was originally noted in Footnote 2, Page 3{\color{#003A8C}\text{Footnote 2, Page 3}}; now Appendix C.4, Page 24{\color{#003A8C}\text{Appendix C.4, Page 24}} provides the complete details.

AuthorsOpenReview ↗

Part[1/5] Response to Reviewer a4fP 1.2

(II) Specifying data splits/counts, training settings, or numerical tolerances.

We appreciate the request for additional experimental details. We have added Appendix D.3, Page 26{\color{#003A8C}\text{Appendix D.3, Page 26}} to provide the following information.

Data splits and counts.

  • Synthetic systems: 20,00020{,}000 training, 2,0002{,}000 validation, and 2,0002{,}000 test samples for N=50,100,250N=50,100,250; the training count increases to 40,00040{,}000 for N=500,1000N=500,1000.
  • DC and AC power flow: 30,00030{,}000 training, 2,0002{,}000 validation, and 2,0002{,}000 test samples for each benchmark system.

Training settings. All neural solvers are fully connected networks with SiLU activations. System-specific architectures are configured as follows:

  • Synthetic systems: We use three hidden layers with width 256256 for N=50,100N=50,100, width 512512 for N=250,500N=250,500, and width 10241024 for N=1000N=1000.
  • DC power flow: The 57-bus, 118-bus, 500-bus, and 1354-bus models use three hidden layers with widths 128128, 256256, 512512, and 768768, respectively.
  • AC power flow: The 57-bus, 18-bus, 500-bus, and 1354-bus models use three hidden layers with widths 256256, 256256, 512512, and 10241024, respectively.

We train with Adam (learning rate 10−310^{-3}, weight decay 10−810^{-8}, seed 12341234), batch size 512512, and up to 800800 epochs

Numerical tolerances and sample filtering.

  • DC targets are computed using a sparse direct linear solver.
  • AC targets are generated using a Newton-based power-flow solver with absolute and relative convergence tolerances of 10−510^{-5}.
AuthorsOpenReview ↗

Part[1/5] Response to Reviewer a4fP 1.3 - Failure rates

(III) Failure rates and near-limit tests.

We thank the reviewer for this suggestion. We report failure rates and add near-limit tests for the linear and nonlinear recasting.

Failure rates. We define failure rate as the percentage of cases whose relative solution error exceeds a specified threshold. The table below reports AC power flow failure rates at thresholds of 22\\%, 33\\%, and 55\\% after five recasting steps, under independently perturbed operating conditions. Across all four systems, at least 97.2997.29\\% of predictions meet the 33\\% accuracy requirement, and at least 99.1899.18\\% meet the 55\\% requirement, corresponding to failure rates of at most 2.712.71\\% and 0.820.82\\%, respectively.

SystemFailure rate at 22\\%Failure rate at 33\\%Failure rate at 55\\%
IEEE 57-bus4.034.03\\%2.542.54\\%0.640.64\\%
IEEE 118-bus0.750.75\\%0.430.43\\%0.430.43\\%
Carolina 500-bus6.676.67\\%2.712.71\\%0.000.00\\%
European 1354-Bus1.361.36\\%1.361.36\\%0.820.82\\%

To further investigate which outage conditions are more likely to cause prediction failures, we analyze the Carolina 500-Bus results by outage impact. For each of the 343 outages, we compute the apparent power flow carried by the affected line before the outage across all evaluated operating conditions. We rank outages by this power flow and divide them into three approximately equal-sized groups: low-impact (≤38.73\leq38.73 MVA), medium-impact (38.7338.73-98.6398.63 MVA), and high-impact (>98.63>98.63 MVA). This grouping is independent of prediction errors. The results show that prediction failures are concentrated in high-impact outages: this group accounts for approximately 9191\\% of all failed cases at the 22\\% threshold. At the 33\\% threshold, only high-impact outages produce failures, with a failure rate of 7.267.26\\% within this group. All evaluated the Carolina 500-Bus cases meet the 55\\% accuracy requirement.

Outage groupPre-contingency flow (MVA)Failure at 22\\%Failure at 33\\%Failure at 55\\%
Low-impact≤38.73\leq38.730.000.00\\%0.000.00\\%0.000.00\\%
Medium-impact(38.73,98.63](38.73,98.63]2.542.54\\%0.000.00\\%0.000.00\\%
High-impact>98.63>98.6319.0919.09\\%7.267.26\\%0.000.00\\%

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal Discussion

Dear AC and Reviewers,

Thank you for all your thoughtful feedback and careful evaluation of our work. We appreciate the reviewers’ recognition of RecastNet’s practical relevance, methodological insights, and empirical breadth. Through this constructive discussion, we have further elevated the quality of our manuscript with additional theoretical analyses, extensive new experiments, stronger computational comparisons, and methodological clarifications. To assist the AC’s assessment, we summarize the reviewer discussions below.

ReviewerInitial scoreUpdated scoreNotes (Quotes from reviews and discussions)
a4fP23"I appreciate the great amount of work that was done and increase my grade."
Kc6e22"Several of my concerns have been addressed."
6dhi33"the solution-to-inverse insight in Theorem 1 is particularly elegant"; "a coherent method with substantial theoretical support"; "This breadth makes the empirical evaluation convincing"

Below, we summarize the main additions and clarifications made in response to the reviewers' comments:

1. Clarifying the core methodological novelty. We further clarified RecastNet's central insight: a neural solver trained solely for solution prediction on a single base topology can provide inverse-operator information that enables analytical adaptation to unseen topologies, without retraining or fine-tuning. We establish an architecture-agnostic solution-to-inverse mechanism and develop efficient directional recovery and solution recasting, providing a unified foundation for linear and nonlinear networked systems under topology changes. Reviewer 6dhi specifically recognized the solution-to-inverse insight as “particularly elegant,” and the resulting framework as "a coherent method with substantial theoretical support."

2. Strengthening theoretical foundations and recastability analysis. Building on the theoretical framework in our original submission, we further characterized how topology-change rank affects recasting errors and computational complexity, and examined error amplification and local convergence in linear and nonlinear recasting. These analyses further establish how topology-change severity and error amplification affect recastability, while characterizing the computational trade-offs of increasingly higher-rank topology changes.

3. Substantially expanding robustness and practical validation. In response to reviewers' questions about robustness and applicability, we conducted extensive new experiments under increasingly challenging conditions, including higher-rank topology changes (up to 30% of network size), inaccurate base predictions, and successive outages approaching network disconnection. Results show that RecastNet maintains median errors below 2.6% at 30% topology-change rank in synthetic networks, while achieving 1.3% errors under severe successive outages (up to N−6N-6) in nonlinear AC power flow on real-world power grids. We also clarified the AC power flow formulation and supplemented the existing results with failure rates, componentwise errors, and conditioning diagnostics across four benchmark power grids.

4. Establishing computational advantages against stronger baselines. Following the suggestions of Reviewers a4fP and Kc6e, we added comparisons against stronger numerical baselines, including sparse factorization with SMW updates and Newton–Raphson using the same neural initialization. We further implemented GPU comparisons against cuDSS-based low-rank updates and GPU neural-initialized Newton–Raphson. Across four benchmark power grids, RecastNet achieves DC speedups up to 9.87$\times$ and AC speedups up to 10.99$\times$ across the tested batch sizes, demonstrating computational advantages against these stronger baselines and confirming that the speedups stem from RecastNet rather than GPU hardware or neural initialization alone.

We sincerely thank the AC and reviewers for their time and constructive engagement. We respectfully hope the AC will consider our clarified methodological contributions, together with the substantial new experimental evidence and the concerns addressed during the discussion, in reaching the final assessment. We believe these additional analyses, experiments, and clarifications have significantly strengthened the manuscript and further demonstrated the novelty, technical soundness, and practical value of our approach. Thank you again for your time and consideration.

Best,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.