acceptodds
Under review as a conference paper at ICLR 2027

From Prediction to Compensation: Neural Operators for Event-Triggered Sliding-Mode PDE Control

Abstract

In recent years, neural operators have made strides in predicting solutions to partial differential equations; however, their reliable application in scenarios with communication-constrained feedback remains under-researched. This paper proposes a neural-operator-assisted dynamic event-triggered integral sliding mode control scheme for reaction-diffusion neural systems subject to sojourn-time-dependent semi-Markovian switching. A Fourier neural operator is employed to estimate matched uncertainties, while a robust term handles prediction residuals and state-preservation errors. Under conditions such as residual boundedness, it is proven that the sliding mode surface is reachable in finite time, the sliding mode dynamics are mean-square exponentially stable, and Zeno behavior is avoided. Independent testing and closed-loop simulations demonstrate that the proposed method enhances estimation accuracy while reducing control energy consumption and input fluctuations; comparisons of triggering mechanisms reveal that the method requires fewer communication transmissions than static strategies.

Primary area:
learning on time series and dynamical systems
Keywords:
Fourier neural operator, uncertainty compensation, dynamic event-triggered mechanism, integral sliding mode control, reaction–diffusion systems, semi-Markovian switching

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_QWH3OpenReview ↗
Rating
2
Confidence
2(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

onsiders sliding-mode control of uncertain reaction–diffusion systems with semi-Markov switching. An FNO is used to estimate matched uncertainties and is combined with a dynamic event-triggered controller. The authors establish stability results under a bounded prediction residual assumption and use numerical simulations to evaluate prediction accuracy and control effort.

Critical strengths

The idea of using an FNO to estimate matched uncertainties, rather than directly learning controller gains, is reasonable. It provides a useful way to combine learned compensation with robust PDE control. The stability analysis is also fairly detailed, and the reported reductions in control effort and input fluctuations are encouraging.

Critical weaknesses
  1. My main concern is the novelty. Learning unknown disturbances and compensating for them is a well-established idea in robust control. The paper applies this idea to PDE systems using a largely standard FNO, together with existing sliding-mode control and event-triggered techniques. While the combination is useful, I am not fully convinced that it offers a sufficiently new methodological contribution. In particular, the connection between prediction accuracy and closed-loop performance could be better established.

  2. I also have concerns about the theoretical assumption and experimental comparisons. Assumption 2 requires a uniform bound on the prediction residual for all time, whereas the bound used in the simulations is obtained from validation data with an empirical factor of 1.25. The authors acknowledge that this does not certify the assumption. In addition, the FNO performs only slightly better than U-Net in Table 1 (0.003245 vs. 0.003351). Although Table 2 reports substantial improvements, the controller without NO uses an analytical bound on the full uncertainty, while the proposed controller uses an empirically calibrated residual bound. It is unclear whether the reported improvements would remain as substantial under comparable robustness guarantees.

Questions for authors
  1. Since the residual bound is calibrated from validation data, how can Assumption 2 be verified or enforced during deployment? What happens to the stability guarantees when the actual prediction residual exceeds this bound?

  2. Why does the controller without NO use an analytical bound on the full uncertainty, while the proposed method uses an empirically calibrated residual bound? Would the improvements in control energy and total variation remain under comparable calibration procedures or robustness guarantees?

  3. Could the authors clarify the main methodological contribution relative to existing neural-operator-based PDE control and learning-assisted sliding-mode control? Beyond using an FNO for uncertainty estimation, what aspects of the approach can be generalized to other learning-based control problems?

4 comments
AuthorsOpenReview ↗

Response to Reviewer QWH3

Thank you for your careful assessment. We have clarified the guarantees and added the following comparisons.

1. Assumption 2 and exceedance

Each mode's margin is 1.25 times its maximum validation residual ∥ri−si∥L2\lVert r_i-s_i\rVert_{L_2}, fixed before testing.

In simulation, calibration uses the simulator's known uncertainty. For physical deployment, commissioning data would be needed to measure the unknown input independently or reconstruct it from measured state evolution, the nominal dynamics, and applied control. We would compare this with the applied FNO compensation, add any known reconstruction-error bound to the margin, and reassess it using new measurements when conditions change. See Appendix D.1 for calibration details.

Margin (50 paired 30-s paths)Sample coverageControl energySampled TVTransmissions
ϵ^ival\widehat\epsilon_i^{\mathrm{val}}99.9603%295.29171.9746.62
2ϵ^ival2\widehat\epsilon_i^{\mathrm{val}}99.9623%298.67520.25521.60
3ϵ^ival3\widehat\epsilon_i^{\mathrm{val}}99.9923%304.862,368.761,550.84
Full bound rˉi\bar r_i, No-NO100%369.1316,044.303,995.20

If the margin is exceeded, its stability guarantee no longer applies; this does not itself imply instability: all 200 runs meet the terminal regulation criterion. The threefold margin retains 17.4% lower energy and 85.2% lower control total variation (TV) than full-bound No-NO. See Appendix D.5 for the complete margin-sweep results.

On 50 rotating-direction out-of-distribution (OOD) paths, a causal monitor checks completed-interval residual anchors. An alert disables compensation and selects full-bound No-NO using rˉi\bar r_i. Mean exceedance time with compensation active falls 0.2603→0.0282 s (89.2%), with 17.0% less energy than always-full-bound operation. This mitigates violations but does not enforce Assumption 2. FNO uses smaller residual margins to reduce control costs; full-bound No-NO remains the conservative alternative. See Section 5.5 and Appendix D.3 for monitoring and fallback details.

2. Methodological contribution and transferability

Relative to gain learning [1], learned PDE control [4], and finite-dimensional SMC [2,3], we estimate spatial uncertainty in switching PDEs from intermittent states. See Appendix E for the comparison with related work.

Between transmissions, the predictor uses model-propagated states, past interval residual estimates, and mode/sojourn-time information. Deployment gates regulate when compensation is applied. The gain hi(t)=ϵˉi+χi(t)+δih_i(t)=\bar\epsilon_i+\chi_i(t)+\delta_i separately covers the remaining prediction residual, state-holding error, and reaching margin.

To quantify these design choices, all variants use the same 287,362-parameter FNO and three training seeds. Validation-selected checkpoints are evaluated on 50 paired rotating-direction OOD trajectories. The main ablations are reported in Section 5.5 (Table 5); see Appendix D.7 for the complete results.

VariantMean residualMean calibrated margin across modesControl energyTransmissions
Full causal design0.0208570.3231293.2347.10
No residual anchors0.0213510.3274293.4746.62
Held state instead of propagation0.0357400.2974293.2848.74
No mode/sojourn-time inputs0.0243352.4115398.024,052.86
No equilibrium/deployment gates0.0303330.3108292.9664.86

Removing switching inputs raises the calibrated margin to 7.46 times its original value and control energy by 35.7%; replacing propagation with held states raises residuals by 71.4%; removing gates raises transmissions by 37.7%. The margin sweep in Question 1 keeps the trained FNO unchanged and shows that larger margins increase control and communication costs.

The generalizable idea is to predict and compensate the unknown term using only available information, then design the robust term to cover the remaining prediction error and state-holding error. Other estimators can replace FNO if they satisfy the same information and residual-bound conditions, with an appropriate stability analysis for the target system.

AuthorsOpenReview ↗

Response to Reviewer QWH3

3. Comparable calibration, analytic protection, and estimator comparisons

We added matched-calibration and analytic-bound comparisons to test whether the benefits depend on how the margin is chosen.

  • Same calibration rule. Both controllers use 1.25 times their mode-wise maximum validation residual: ri−sir_i-s_i for FNO and rir_i for Zero (no compensation). Margins are fixed before testing on 50 paired trajectories per condition.

    Test conditionFNO control energyZero control energyFNO sampled TVZero sampled TV
    Original setup295.29350.18171.9713,485.81
    Combined changes575.91634.41270.4912,748.28

    Combined changes comprise a 1.5× multi-frequency initial field, doubled switching rates, and rotating uncertainty directions. With the same calibration rule, FNO retains 15.7% and 9.2% energy savings, respectively, and over 97% lower TV. These comparisons are reported in Section 5.5 (Table 4); see Appendix D.5 for complete protocols and results.

  • Analytic uncertainty protection. We also retain FNO compensation while covering its residual by ∥ri−si∥L2≤rˉi+∥si∥L2\lVert r_i-s_i\rVert_{L_2}\leq\bar r_i+\lVert s_i\rVert_{L_2}. Full-bound No-NO disables compensation and uses rˉi\bar r_i.

    Test conditionFNO control energyFull-bound No-NO energyFNO sampled TVFull-bound No-NO TV
    Original setup382.96369.1316,217.7616,044.30
    Combined changes671.92653.5015,398.6015,207.59

    The matched-calibration results demonstrate the value of FNO: compensating uncertainty enables a smaller residual margin and lower control costs. The full analytic bound does not exploit this smaller residual, so the same energy and TV savings are not retained. See Appendix D.5 for the analytic-bound comparison details.

  • Broader estimator comparison. We compare six methods on the same 50 in-distribution and 50 rotating-direction OOD trajectories under the same information constraints. All 600 runs meet the terminal criterion; the table reports OOD results.

    MethodMean residualControl energySampled TVTransmissionsEstimator time (ms)
    FNO0.020826293.22171.9846.845.089
    U-Net0.021775292.73173.4150.764.081
    SD-DOB0.021934401.8219,567.034,062.720.0868
    Adaptive RLS0.030270349.5413,520.683,906.880.1196
    Zero predictor0.028934348.4013,480.853,886.880.0129
    Full-bound No-NO—367.3916,022.943,978.18—

    FNO and U-Net have comparable in-distribution residuals; on rotating-direction OOD paths, FNO achieves 4.4% lower mean residuals and 7.7% fewer transmissions. Relative to SD-DOB and RLS, FNO reduces control energy by 27.0% and 16.1%, respectively. We select FNO for its balance of prediction accuracy, control smoothness, and communication efficiency under intermittent state transmission. See Section 5.3 (Table 2) and Appendix D.4 for the comparison and implementation details.

In summary, we added matched-calibration and analytic-bound comparisons and causal-design ablations, and clarified deployment calibration and fallback. These revisions address comparison fairness, identify the source of the control-cost reductions, and show what the causal design contributes.

References

[1] Bhan, Shi, and Krstic. Neural Operators for Bypassing Gain and Control Computations in PDE Backstepping. IEEE Transactions on Automatic Control, 2024.

[2] Li et al. A novel learning-based asynchronous sliding mode control for discrete-time semi-Markov jump systems. Automatica, 2022.

[3] Cheng et al. Neural Network-Based Sliding Mode Control for Semi-Markov Jumping Systems With Singular Perturbation. IEEE Transactions on Cybernetics, 2025.

[4] Holl et al. Learning to Control PDEs with Differentiable Physics. ICLR, 2020.

I apologize for the confusion. Due to an error in my visibility settings, you were unable to see my previous comment. Thank you for your efforts in preparing the rebuttal. You have addressed most of my concerns. After reading the responses to the other reviewers, I am inclined to raise my score from 2 to 3 based on the quality of your rebuttal. However, the major remaining issue is that the presentation style and clarity of the figures, tables, and visualizations still leave considerable room for improvement. Therefore, I will maintain my current score for now.

The overall quality of the manuscript has improved substantially. I encourage you to consider submitting it to another journal or conference.

Thank you very much for your time and thoughtful feedback. If you feel that our responses and revisions have addressed your concerns, we would be sincerely grateful if you could consider updating your assessment and rating accordingly. We would be happy to clarify any remaining questions and greatly appreciate your consideration.

Reviewer_V69LOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a control method for reaction-diffusion neural systems with uncertain dynamics and switching behavior. The authors use a Fourier neural operator to estimate unknown disturbances and compensate for their effects in the control design. A dynamic event-triggered mechanism reduces the number of state transmissions, while a sliding mode controller handles the remaining errors. The paper provides theoretical results concerning stability, reaching the sliding surface, and avoiding excessively frequent triggering. Numerical experiments compare the proposed method with alternative prediction approaches and control strategies. The results show improved uncertainty estimation, lower control energy, smoother control inputs, and fewer transmissions than the tested baselines.

Critical strengths

The first strength is the clean separation between learning and robustness. The neural operator only reduces the uncertainty, and the stability proofs only need a bound on what is left, so a poor prediction does not break the guarantees as long as the bound holds. I think this is a useful and honest way to bring learned components into control with guarantees. The second strength is the complete theory, which covers the reaching phase, the sliding phase, and the exclusion of Zeno behavior, together with LMI conditions for gain design. These two points are the main reasons I see value in the paper.

Critical weaknesses

I think that the main weakness is the very limited validation. There is only one small synthetic system, and the experimental uncertainty is a fixed matrix times a simple sine wave, which is an easy target for any learner. The U-Net is almost as good as the FNO, and the only closed-loop baseline is sliding mode control designed with a worst-case bound, which is expected to chatter much more. I recommend comparing with classical tools such as a disturbance observer or an adaptive estimator.

Questions for authors

Could the authors clarify how they would obtain and maintain the required prediction error bounds in a real application? What happens to the controller if the actual prediction error exceeds the assumed bound? Could the authors also provide results for a more complex system or different operating conditions to demonstrate the method's reliability?

3 comments
AuthorsOpenReview ↗

Response to Reviewer V69L

Thank you for the constructive and thoughtful feedback. We hope that the following responses address your concerns.

1. Classical baselines and FNO

We added SD-DOB and adaptive RLS as you suggested. Tests use 50 ID and 50 rotating-direction OOD paired paths (in-/out-of-distribution). Estimators share causal information and calibration rules. The expanded comparison is reported in Section 5.3 (Table 2); see Appendix D.4 for complete results and implementation details.

MethodOOD residualControl energySampled TVTransmissions
FNO0.020826293.22171.9846.84
U-Net0.021775292.73173.4150.76
SD-DOB0.021934401.8219,567.034,062.72
Adaptive RLS0.030270349.5413,520.683,906.88
Zero predictor0.028934348.4013,480.853,886.88
Full-bound No-NO—367.3916,022.943,978.18

All 600 runs meet the terminal regulation criterion. FNO and U-Net have comparable ID residuals. On the 50 paired rotating-direction OOD trajectories, FNO achieves 4.4% lower mean residuals and 7.7% fewer transmissions than U-Net. For this same OOD test, the mean residual difference (U-Net minus FNO) is 0.000948, with a 95% confidence interval of [0.000503, 0.001417]. FNO also uses 21% fewer parameters than U-Net. Under the same OOD condition, FNO reduces control energy by 27.0% relative to SD-DOB and 16.1% relative to adaptive RLS. We therefore select FNO for its balance of prediction accuracy, control smoothness, and communication efficiency under intermittent state transmission.

2. Broader validation

Even with a sinusoidal coefficient, ri(t,x)=ΔBi(t)tanh⁡(Υ(t,x))r_i(t,x)=\Delta B_i(t)\tanh(\Upsilon(t,x)) is a state-dependent spatial field. We add three OOD families without retraining or recalibration. These additional tests are summarized in Section 5.5; see Appendix D.2 for detailed protocols and results.

UncertaintySample-time coverageRegulated paths
ID single sinusoid99.9517%100/100
OOD dual sinusoid99.8864%50/50
OOD piecewise linear99.8847%50/50
OOD rotating direction99.1325%50/50

All 250 unseen 30-s paths (150 OOD) remain finite, with terminal L2L_2 norm at most 5% of its initial value.

Models and parameters stay frozen; both controllers regulate 50/50 paired paths per condition.

ConditionFNO control energyNo-NO control energyFNO transmissionsNo-NO transmissions
Original295.29369.1346.623,995.20
Initial amplitude ×1.5618.85697.0665.983,844.18
Multi-frequency initial field273.19347.9149.944,093.26
Fast switching (×2)296.93370.1046.483,967.92
Slow switching (×0.5)294.25368.6946.683,997.42
Combined changes575.91653.5071.083,914.74

Combined changes: a 1.5× multi-frequency initial field, fast switching, and rotating directions. See Appendix D.6 for the operating-condition protocols and complete results. These tests confirm reliable regulation beyond the original setting, with lower control and communication costs under the tested changes, without retraining or retuning.

3. Obtaining and maintaining the margin

Each mode's allowance is 1.25 times its maximum validation residual ∥ri−si∥L2\lVert r_i-s_i\rVert_{L_2}, using the deployed compensation, and is fixed before testing. Physical calibration requires measured or independently reconstructed uncertainty, including reconstruction error. Changed conditions require revalidation or full-bound operation; finite calibration data do not certify an all-time bound. See Section 5.1 and Appendix D.1 for calibration details.

The original-condition margin sweep uses 50 paired paths:

AllowanceSample coverageControl energySampled TVTransmissions
ϵ^ival\widehat\epsilon_i^{\mathrm{val}}99.9603%295.29171.9746.62
2ϵ^ival2\widehat\epsilon_i^{\mathrm{val}}99.9623%298.67520.25521.60
3ϵ^ival3\widehat\epsilon_i^{\mathrm{val}}99.9923%304.862,368.761,550.84
Full bound rˉi\bar r_i, No-NO100%369.1316,044.303,995.20

The threefold allowance retains 17.4% energy savings and 85.2% lower TV, quantifying the coverage–control-burden trade-off. See Appendix D.5 for the complete margin-sweep results.

4. Residual exceedance and fallback

Exceedance removes the bound-dependent guarantee during that interval, but does not itself imply instability. On 50 paired rotating-direction OOD paths, the monitor uses transmitted interval anchors to trigger full-bound fallback. Exposure measures time with compensation active while its residual exceeds the margin.

ControllerUnprotected exposure (s)Control energy
Frozen FNO0.2603293.22
Causal fallback0.0282305.04
Always full-bound No-NO0367.39

Exposure falls by 89.2%, with 17.0% less energy than always-full-bound operation. True residuals are used only for offline evaluation. The monitor is an empirical safeguard; when prediction-independent protection is required from the outset, compensation is disabled and the known full bound rˉi\bar r_i is used under the remaining theoretical conditions. See Section 5.5 and Appendix D.3 for monitoring and fallback details.

In summary, we added SD-DOB and adaptive RLS, broader uncertainty and operating-condition tests, and margin-sensitivity and fallback experiments. These additions address the limited validation and missing classical baselines, and explain residual calibration and the handling of bound exceedances.

Thank you very much for your time and thoughtful feedback. If you feel that our responses and revisions have addressed your concerns, we would be sincerely grateful if you could consider updating your assessment and rating accordingly. We would be happy to clarify any remaining questions and greatly appreciate your consideration.

Reviewer_QxLqOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper integrates three components: causal Fourier Neural Operator (FNO) into an integral sliding mode control framework, reaction–diffusion systems with semi-Markovian switching, and event-triggered control. The three components have independently been studied and existing results are plenty. The motivation, challenges, and benefits of integrating them is not clear. Therefore, the contributions seems to be incremental.

Further, the role of the neural operator in the theoretical development appears limited. In particular, the stability analysis assumes bounded estimation residuals without establishing whether or under what conditions the proposed FNO satisfies this assumption. Consequently, the theoretical guarantees appear to be independent of the specific learning architecture and potentially applicable to any uncertainty estimator with a bounded approximation error.

While the proposed event-triggering mechanism and stability analysis may constitute contributions to control theory, they are already well-researched in control theoretic domains. This paper severely lacks literature review and acknowledges limited works from control literature that have address and developed some of the results presented here. The paper does not appear to introduce a new neural operator architecture, training methodology, or learning-specific theoretical guarantee. As a result, the contribution to the ICLR community remains unclear.

Critical strengths

None

Critical weaknesses

The theoretical stability, reachability, and Zeno-exclusion analyses rely on an assumed bounded estimation residual rather than deriving this bound from the Fourier Neural Operator's approximation or generalization properties. This limits the significance of the theoretical results as a contribution to learning-enabled control.

The use of a causal FNO for uncertainty estimation appears primarily to adapt an existing neural operator architecture to a control application.

Control theoretic contributions are primary merits of the paper but the paper does not do a good job in presenting a good review of relevant literature and does not acknowledge works relevant to the proposed approach.

Because the theoretical analysis depends only on bounded estimation error, the manuscript does not establish whether the FNO provides meaningful advantages over alternative estimators.

Questions for authors
  1. The theoretical guarantees rely on Assumption 2. Can the authors provide theoretical or empirical justification for this assumption, particularly for unseen closed-loop trajectories? How is the residual bound determined in practice?

  2. What are the specific methodological contributions of the proposed causal FNO beyond adapting an existing neural operator architecture for uncertainty estimation? How does the proposed formulation differ from existing FNO-based approaches?

  3. Can the authors clarify how the proposed error-modulated dynamic event-triggering mechanism and sojourn-time partitioning differ from closely related approaches in the control literature? What are the key technical advances over existing methods?

  4. Given that the theoretical guarantees depend only on bounded estimation error, what motivates the choice of FNO over alternative uncertainty estimators? Can the authors provide comparative experiments or ablation studies demonstrating the specific benefits of FNO in terms of estimation accuracy, closed-loop performance, robustness, and computational efficiency?

4 comments
AuthorsOpenReview ↗

Response to Reviewer QxLq

Thank you for your careful review and helpful suggestions. We hope our responses address your concerns.

1. Assumption 2 and unseen trajectories

Each mode's allowance is fixed before testing at 1.25 times its maximum validation residual ∥ri−si∥L2\lVert r_i-s_i\rVert_{L_2}, using the deployed compensation sis_i. Simulation uses known uncertainty offline. Physical calibration requires independent measurements or reconstruction from measured state evolution, nominal dynamics and applied inputs, including reconstruction-error bounds; changed operating conditions require reassessment. See Appendix D.1 for calibration details.

We froze the predictor, controller and allowances and tested 250 unseen 30-s trajectories, without retraining or recalibration. ID/OOD denote in-/out-of-distribution testing. All uncertainties retain the original norm constraint.

UncertaintySample-time coverageRegulated paths
ID single sinusoid99.9517%100/100
OOD dual sinusoid99.8864%50/50
OOD piecewise linear99.8847%50/50
OOD rotating direction99.1325%50/50

Regulation requires finite records and

∥Υ(30)∥L2≤0.05 ∥Υ(0)∥L2.\lVert \Upsilon(30) \rVert_{L_2} \leq 0.05\,\lVert \Upsilon(0) \rVert_{L_2}.

Exceedance does not itself imply instability, but the theorem remains conditional on a valid residual bound. These additional tests are summarized in Section 5.5; see Appendix D.2 for details.

A causal monitor uses completed-interval residual anchors formed locally and transmitted with state packets (Appendix A.1), rather than the instantaneous true uncertainty. An alert disables compensation and uses the full bound rˉi\bar r_i.

Strategy: 50 paired rotating-direction OOD pathsMean exceedance time with compensation active (s)Control energy
Frozen FNO0.2603293.22
Monitor + fallback0.0282305.04
Always full-bound No-NO0367.39

Fallback reduces exposure by 89.2%, with 17.0% less energy than always-full-bound operation. This is empirical mitigation, not enforcement of Assumption 2. With compensation disabled, the known rˉi\bar r_i supplies the residual bound under the remaining design conditions. See Appendix D.3 for the monitoring and fallback protocol.

2. Contribution of the causal FNO

Unlike learning PDE solutions [1] or gain/kernel mappings [2], we predict the distributed matched uncertainty from intermittently available information. Model-propagated states, completed-interval residual anchors, mode/sojourn-time inputs and deployment gates make the estimator causal and compatible with the feedback law.

All five variants use the same 287,362-parameter FNO, three training seeds and validation-selected checkpoints, evaluated on 50 paired rotating-direction OOD paths. The main ablations are reported in Section 5.5 (Table 5); see Appendix D.7 for the complete results.

VariantMean residualControl energySampled TVTransmissions
Full causal design0.020857293.23172.0347.10
No residual anchors0.021351293.47172.2546.62
Held state instead of propagation0.035740293.28177.3148.74
No mode/sojourn-time inputs0.024335398.0219,147.024,052.86
No equilibrium/deployment gates0.030333292.96173.6864.86

Removing switching inputs raises the mean calibrated margin 7.46-fold and energy 35.7%; held-state inputs raise residuals 71.4%; removing gates raises transmissions 37.7%. These experiments identify the value of causal information and deployment design at fixed model capacity. The transferable design is to compensate the unknown term from available information and retain robust control for prediction and state-holding errors.

3. Related work

Thank you for highlighting this point. We have strengthened the closest-work comparisons in the Introduction and Appendix E, covering neural-operator PDE control, learning-assisted SMC, dynamic event-triggering, and semi-Markov controller synthesis. The added comparison table in Appendix E (Table 23) explicitly summarizes the existing results and the distinctions addressed here. These additions clarify the established foundations and our specific contribution: causal spatial uncertainty estimation under intermittent transmission, with prediction residuals, holding errors, and residence-stage-dependent switching jointly incorporated into the PDE control and stability analysis.

AuthorsOpenReview ↗

Response to Reviewer QxLq

4. Triggering and residence-time-dependent switching

Triggering. Girard [3] introduces an internal dynamic variable; Cheng et al. [4] use historical transmissions and two auxiliary variables in finite-dimensional singular semi-Markov systems. Our mode-dependent PDE trigger uses Fi(t)=ξie−νiEi(t)F_i(t)=\xi_i e^{-\nu_i E_i(t)}: for νi>0\nu_i>0, larger holding-error energy tightens the threshold. Its bound ∥Kim∥L2≤χi(t)\lVert K_i m\rVert_{L_2}\leq\chi_i(t) enters hi(t)=ϵˉi+χi(t)+δih_i(t)=\bar\epsilon_i+\chi_i(t)+\delta_i, separating prediction residual, holding error and reaching margin.

Switching. Tian et al. [5] study discrete-time systems whose semi-Markov kernels are regulated by a higher-level Markov chain. We consider continuous-time reaction–diffusion PDEs with residence-stage-dependent transition rates. Within the same mode, successor-mode probabilities differ between ℓ<ϑ⋆\ell<\vartheta^\star and ℓ≥ϑ⋆\ell\geq\vartheta^\star, while the total departure rate remains fixed. This residence-stage dependence is retained explicitly in the PDE stability conditions.

The technical contribution is the joint treatment of spatial diffusion/reaction dynamics, residence-time-dependent switching, prediction residuals and transmission-induced holding errors across the reaching and sliding analyses. The finite-dimensional methods [3–5] do not address this joint PDE learning/control setting. See Appendix E for the detailed comparison with related work.

On 100 paired 30-s paths, we isolate the dynamic threshold by removing only η(t)/θi\eta(t)/\theta_i:

TriggerMean transmissions
Proposed DETM44.41
Same trigger without η/θi\eta/\theta_i573.17

The mean pathwise reduction is 92.12%. This comparison measures the dynamic threshold's contribution; the error-modulation term is unchanged. See Section 5.4 (Table 3) for this comparison.

5. Why FNO rather than another estimator?

The robust analysis accepts other estimators satisfying the information and residual-bound conditions. We therefore justify FNO empirically, comparing six methods on the same 50 ID and 50 rotating-direction OOD paths. Predictors use the same information constraint and validation-calibration rule; full-bound No-NO uses the analytic uncertainty bound. The expanded comparison is reported in Section 5.3 (Table 2); see Appendix D.4 for complete results and implementation details.

MethodMean residual: ID / OODOOD energyOOD TVOOD transmissionsTime/call (ms)
FNO0.006637 / 0.020826293.22171.9846.845.089
U-Net0.006563 / 0.021775292.73173.4150.764.081
SD-DOB0.026374 / 0.021934401.8219,567.034,062.720.0868
Adaptive RLS0.036248 / 0.030270349.5413,520.683,906.880.1196
Zero0.034574 / 0.028934348.4013,480.853,886.880.0129
Full-bound No-NO—367.3916,022.943,978.18—

All 600 method runs meet the terminal criterion. FNO/U-Net have comparable ID residuals; on rotating-direction OOD paths, FNO has 4.4% lower residuals, 7.7% fewer transmissions and 21% fewer parameters. OOD energy is 27.0%/16.1% lower than SD-DOB/RLS. Although FNO is slower per call, we select it for its balance of prediction accuracy, control smoothness, communication efficiency and model size.

In summary, we added estimator comparisons, causal-design ablations, and unseen-trajectory tests, and clarified the residual-bound requirements. These revisions explain the methodological contribution, the choice of FNO, and the conditions under which the guarantees apply.

References

[1] Li et al. Fourier Neural Operator for Parametric Partial Differential Equations. ICLR, 2021.

[2] Bhan et al. Neural Operators for Bypassing Gain and Control Computations in PDE Backstepping. IEEE TAC, 2024.

[3] Girard. Dynamic Triggering Mechanisms for Event-Triggered Control. IEEE TAC, 2015.

[4] Cheng et al. Novel event-triggered protocol to sliding mode control for singular semi-Markov jump systems. Automatica, 2023.

[5] Tian et al. Time-varying gain controller synthesis of piecewise homogeneous semi-Markov jump linear systems. Automatica, 2022.

[6] Cheng et al. Neural Network-Based Sliding Mode Control for Semi-Markov Jumping Systems With Singular Perturbation. IEEE TCYB, 2025.

Thank you for the detailed response and additional experiments. The new comparisons and ablation studies help clarify the choice of FNO and the empirical performance of the proposed framework. However, my main concerns about the novelty and significance of the contributions remain.

Specifically, the theoretical guarantees still rely on an assumed bound on the estimation residual. While the additional experiments provide some empirical justification, they do not establish when this assumption holds for unseen trajectories. My understanding is that the guarantees remain conditional and do not depend on the specific neural operator architecture.

The proposed causal FNO appears to be an adaptation of an existing architecture using information relevant to the control problem. The ablation studies demonstrate the usefulness of these design choices, but it is still unclear what the methodological novelty is. The comparison with U-Net also suggests that the benefits are not necessarily specific to FNO.

While I appreciate the authors effort toward expanding the discussions on related work, I am still not convinced that combining the different control-theoretic components constitutes a substantial technical advance. The response does not clearly explain what new challenges arise from this integration or why existing techniques are insufficient.

In my opinion, the additional experiments strengthen the empirical evaluation, and I appreciate the authors' efforts. However, I remain unconvinced about the novelty of the proposed approach and its contribution to the ICLR community.

AuthorsOpenReview ↗

Replying to Response to Reviewer QxLq

Thank you for your further feedback. We address the deployment conditions, methodological contribution, and joint theoretical analysis below.

1. Residual calibration and deployment conditions

Physical deployment would require commissioning data covering the intended initial-state range, switching behavior, and uncertainty variations. The unknown input would be independently measured or reconstructed from measured state evolution, nominal dynamics, and applied control. We would compare it with the deployed FNO compensation, include a known reconstruction-error bound, and reassess the margin using new measurements when conditions change (Appendix D.1).

Assumption 2 is satisfied over an operating interval if independent measurements or error-bounded reconstruction establish that the deployed residual remains within the selected margin throughout that interval.

Across 250 unseen trajectories, the frozen configuration achieves 99.13%–99.95% sample-time coverage, and all meet the terminal regulation criterion (Section 5.5; Appendix D.2). Exceedances are explicitly reported, and monitoring is evaluated as mitigation rather than certification (Appendix D.3).

The theoretical guarantees therefore extend to unseen trajectories when this uniform residual condition and the remaining synthesis conditions hold. The tests provide empirical support; an a priori certificate for smaller learned margins remains future work.

2. What is the methodological contribution beyond FNO?

We formulate causal estimation of evolving distributed uncertainty when the networked predictor receives state profiles only intermittently. Propagated nominal states and completed-interval residual anchors provide between-transmission inputs; mode/sojourn-time information and deployment gates condition prediction and regulate its use in feedback.

The deployed residual allowance enters both the robust gain and the reaching-phase state bound through Eq. (41):

d_R=\max_i\left\\{\varepsilon_{ri}\bar{\epsilon}_i^2+3\varepsilon_{gi}(\bar{\epsilon}_i+\delta_i)^2\right\\}.

For fixed feasible design parameters, a smaller valid allowance reduces this term and the required residual-compensation margin. Thus, prediction quality relevant to feedback extends beyond average prediction error.

Removing mode/sojourn inputs increases the mean mode-wise calibrated allowance from 0.3231 to 2.4115 and energy by 35.7% (Appendix D.7). Matched-calibration comparisons show 15.7%/9.2% energy savings and over 97% lower sampled TV (Appendix D.5).

FNO and U-Net have comparable in-distribution residuals. On 50 paired rotating-direction OOD trajectories, FNO achieves 4.4% lower mean residuals and 7.7% fewer transmissions, with 21% fewer parameters (Section 5.3, Table 2; Appendix D.4).

The methodological contribution is the causal compensation design and its residual-dependent PDE analysis. The comparisons support selecting FNO for its balance of prediction accuracy, control smoothness, communication efficiency, and model size.

3. What challenges arise from the integration, and why are existing results not directly sufficient?

In Appendix B.3, Eq. (36) retains εgihi2(t)\varepsilon_{gi}h_i^2(t). Since hih_i includes the trigger-dependent holding-error allowance, this term must be estimated jointly with the state and trigger variable. Under the synthesis conditions, including 0<μiξi<10<\mu_i\xi_i<1, Eqs. (38)–(40) bound the gain through these quantities; Eqs. (41)–(43) establish reaching-phase mean-square boundedness of the PDE state and spatial gradient. Integration by parts and GiMi=RG_iM_i=R account for spatial diffusion.

Moreover, mode changes need not coincide with transmissions: controller and trigger parameters may change while the holding error persists. The analysis handles these transitions and retains both residence regimes in the PDE generator and synthesis conditions (Appendices B.1 and B.3–B.4).

The finite-dimensional neural SMC studies in Appendix E, Table 23, do not directly supply this PDE state-and-gradient estimate. Gain/kernel-learning studies target controller-design mappings rather than the evolving uncertainty field inferred from intermittent information. Their results therefore cannot simply be invoked to obtain the coupled estimates required here.

The theoretical contribution is deriving these PDE estimates and compatible synthesis conditions, establishing reachability, reaching-phase boundedness, sliding-phase stability, and exclusion of Zeno behavior under the stated assumptions. For learning-based control, this connects what the predictor can observe, the residual bound the controller requires, and the resulting closed-loop guarantees and empirical costs.

Thank you once again for taking the time to conduct a careful assessment. If these clarifications address your concerns, we would be grateful if you could reconsider your evaluation and score. Thank you very much.

Comments on the paper

2 comments

We thank all reviewers for their constructive feedback. New material is marked in blue; revisions to existing text are marked in bright orange.

New material (blue)

  • Additional validation (Section 5.5; Appendix D): Unseen-trajectory stress tests, causal monitoring and fallback, matched-calibration and analytic-bound comparisons, residual-margin sensitivity, operating-condition tests, and causal-design ablations.

  • Broader comparisons: Section 5.3 and Table 2 compare six methods, including SD-DOB and adaptive RLS. Appendix E adds a closest-work comparison (Table 23).

  • Residual-bound explanations (Sections 3.1 and 3.3): Validation-based residual calibration and the full-bound No-NO configuration.

Revisions to existing text (bright orange)

  • Contribution and theoretical scope: The Abstract, Introduction, contribution statements, and Conclusion clarify causal uncertainty prediction, the residual-bound conditions underlying the guarantees, and distinctions from related control literature.

  • Information and evaluation protocols (Sections 2.3, 3.2, and 5.1): Clarifications of controller-available information, completed-interval residual anchors, state-holding errors, frozen-model and paired-trajectory testing, and the terminal regulation criterion.

These updates address experimental breadth, baseline coverage, comparison fairness, methodological contribution, and residual-bound handling. Reviewer-specific responses provide further details.

Dear Area Chair,

Thank you for overseeing the review of our submission. We would like to summarize the main revisions relevant to the final assessment.

1. Clarified methodological and theoretical contribution. We clarified causal estimation of distributed uncertainty under intermittent transmission and the joint PDE analysis of prediction residuals, holding errors, state/gradient energies, and residence-dependent switching. Equal-capacity causal-design ablations are added in Section 5.5 and Appendix D.7. The Introduction and Appendix E, Table 23, now explicitly compare our formulation with closely related control studies.

2. Expanded baselines and fairer comparisons. Section 5.3, Table 2, now compares six methods, including the requested SD-DOB and adaptive RLS baselines. Under the same validation-calibration rule, FNO retains 15.7%/9.2% control-energy savings and over 97% lower sampled TV in the original/combined-change conditions (Section 5.5; Appendix D.5). We also report full analytic-bound comparisons, where these control-cost savings are not retained, making the source and scope of the improvements explicit.

3. Broader validation and residual-violation handling. All 250 unseen 30-s trajectories, including three OOD uncertainty families, meet the terminal regulation criterion without retraining or recalibration. Further tests vary initial conditions and switching rates (Appendices D.2 and D.6). Causal fallback reduces exceedance exposure by 89.2%, with 17.0% less energy than always-full-bound operation (Appendix D.3). We explicitly distinguish these empirical results from the uniform residual-bound assumption required by the theory.

Reviewer QxLq acknowledged the value of the additional comparisons and ablations. Given the substantive revisions and additional results, we would be most grateful if the reviewer could reassess the revised manuscript. We would also sincerely appreciate your help in bringing these updates to their attention. Thank you again for your time and consideration.

Sincerely

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.