PL-KKT-hPINN: Enforcing Nonlinear Equality Constraints on Neural Networks via Piecewise-Linear Projection

Fateme Mohammad Mohammadi; Hector Budman; Joshua L. Pulsipher

REVIEW 1 major objections 1 minor 26 references

PL-KKT-hPINN: Enforcing Nonlinear Equality Constraints on Neural Networks via Piecewise-Linear Projection

T0 review · 1 major / 1 minor · reviewed 2026-06-27 · grok-4.3

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A piecewise-linear projection method lets neural networks satisfy nonlinear equality constraints exactly at inference.

desk verdict PL-KKT-hPINN extends KKT-hPINN to nonlinear constraints via piecewise-linear projection but only exactly satisfies the approximation, not the original equalities. read the letter →

arxiv 2606.10682 v1 pith:NQEGNBP3 submitted 2026-06-09 cs.LG

Fateme Mohammad Mohammadi , Hector Budman , Joshua L. Pulsipher This is my paper · ORCID

classification cs.LG

keywords physics-informedneuralnetworkshardconstraintspiecewise-linearprojectionKKTconditionssurrogatemodelingnonlinearequalitychemicalprocess

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper develops PL-KKT-hPINN to address the issue that physics-informed neural networks only enforce physical equations softly during training. The framework approximates the feasible region of nonlinear equality constraints with piecewise-linear segments and uses KKT conditions to project network outputs exactly onto that region. Demonstrated on a continuous stirred-tank reactor with one and two inputs, the model matches standard neural network accuracy but shows much lower constraint violations. It also yields lower error than unconstrained networks when training data is scarce. The result is a framework for building surrogate models that remain physically consistent.

What carries the argument

Piecewise-linear Karush-Kuhn-Tucker hard-constrained projection that approximates the nonlinear equality constraint feasible region to enable exact satisfaction after network prediction.

What would settle it

Measuring the value of the nonlinear equality constraints on held-out test data after applying the piecewise-linear projection; persistent nonzero violations on the original nonlinear equations would indicate the approximation does not guarantee exact satisfaction.

Watch

Extended reading notes

Core claim

The PL-KKT-hPINN framework strictly enforces nonlinear equality constraints through piecewise-linear projection. It extends the KKT-hPINN framework for linear equalities by using the Karush-Kuhn-Tucker conditions to orthogonally project neural network outputs onto an approximated feasible region. In the continuous stirred-tank reactor case study for both one and two inputs, this preserves predictive accuracy comparable to a standard neural network while achieving substantially lower constraint violations and improved robustness in low-data regimes.

Load-bearing premise

The piecewise-linear approximation of the nonlinear constraint feasible region is accurate enough that the resulting projection step exactly satisfies the original nonlinear equalities at inference time.

Editorial extensions

If this is right

Physically consistent surrogate models can be created for nonlinear chemical engineering systems without post-hoc correction.
Constraint violations are substantially reduced compared to standard neural networks or soft-constrained PINNs.
Performance improves in low-data regimes for process modeling tasks.
The approach is computationally efficient for real-time surrogate modeling applications.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Adapting the piecewise-linear projection to other nonlinear problems in optimization or control could yield similar benefits.
The method may generalize beyond CSTR to other reactor or process models with different constraint structures.
Combining this hard constraint enforcement with uncertainty quantification techniques could enhance reliability in engineering predictions.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

read the letter

The main point is that this paper adapts the KKT-hPINN projection method, which works exactly for linear equalities, to nonlinear ones by swapping in a piecewise-linear surrogate of the feasible set.

What is new is the piecewise-linear projection step itself. The CSTR example with one and two inputs shows the model keeps predictive accuracy close to a plain neural network, cuts constraint violations, and improves RMSE when training data is limited. The low-data robustness stands out as the most practical result.

The soft spot is the gap in the enforcement claim. Projection onto the piecewise-linear set satisfies those linear equalities exactly, but the original nonlinear g(x)=0 holds only up to the approximation error. The abstract reports substantially lower violations rather than machine-zero residuals on the true constraints, so the quality of the linear pieces carries the central claim. I would want to see the actual residual sizes on the nonlinear equations and how the approximation error behaves.

This is for people working on constrained surrogate models in chemical engineering or PINN applications to process systems. A reader focused on hard physical constraints in nonlinear settings would get the most from the method and the CSTR results.

It deserves a serious referee. The extension is direct and the low-data numbers are worth checking even with the approximation nuance.

Referee Report

1 major / 1 minor

Summary. The manuscript introduces PL-KKT-hPINN as an extension of KKT-hPINN that replaces the linear feasible region with a piecewise-linear surrogate and performs orthogonal projection to enforce nonlinear equality constraints in neural networks for physics-informed surrogate modeling. The approach is demonstrated on a continuous stirred-tank reactor (CSTR) example with one and two inputs, with claims of predictive accuracy comparable to unconstrained networks, substantially lower constraint violations, and improved robustness in low-data regimes.

Significance. If the piecewise-linear approximation is shown to be sufficiently accurate and the projection step is computationally efficient, the framework could provide a practical route to hard-constrained surrogates for nonlinear process systems where constraint satisfaction at inference is required. The reported low-data performance advantage would strengthen the contribution if supported by detailed experiments with error bars and multiple runs.

major comments (1)

[Abstract] Abstract: The central claim that PL-KKT-hPINN 'strictly enforces nonlinear equality constraints through piecewise-linear projection' is not supported by the reported outcomes. The construction enforces the PL surrogate exactly by construction, yet the abstract states only 'substantially lower constraint violations' rather than machine-zero residuals on the original nonlinear equalities g(x)=0. The approximation error between the PL set and the true nonlinear manifold is therefore load-bearing for the 'strictly enforces' phrasing, but no quantitative bound on |g(x)| or comparison of violation magnitudes is referenced.

minor comments (1)

[Abstract] Typo in abstract: 'framewor' should read 'framework'.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their constructive feedback. We address the concern regarding the abstract phrasing below.

read point-by-point responses

Referee: [Abstract] Abstract: The central claim that PL-KKT-hPINN 'strictly enforces nonlinear equality constraints through piecewise-linear projection' is not supported by the reported outcomes. The construction enforces the PL surrogate exactly by construction, yet the abstract states only 'substantially lower constraint violations' rather than machine-zero residuals on the original nonlinear equalities g(x)=0. The approximation error between the PL set and the true nonlinear manifold is therefore load-bearing for the 'strictly enforces' phrasing, but no quantitative bound on |g(x)| or comparison of violation magnitudes is referenced.

Authors: We agree that the framework exactly enforces the piecewise-linear surrogate of the nonlinear constraints g(x)=0 by construction via the orthogonal projection, and that residuals on the original nonlinear equalities are governed by the quality of the PL approximation. The abstract phrasing is therefore imprecise. We will revise the abstract to state that PL-KKT-hPINN strictly enforces the piecewise-linear surrogate of the nonlinear equality constraints, resulting in substantially lower (but not necessarily machine-zero) violations of the original g(x)=0. We will also add or explicitly reference quantitative comparisons of |g(x)| violation magnitudes between methods to support the claims. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity in the derivation chain.

full rationale

The paper presents PL-KKT-hPINN as a new extension of the prior KKT-hPINN framework, using piecewise-linear projection to handle nonlinear equality constraints. The abstract and description introduce the method, demonstrate it on the CSTR example, and report empirical results on accuracy and constraint violations without any equations or claims reducing to fitted parameters or self-citations by construction. The central construction (orthogonal projection onto the PL surrogate) is independent of the target nonlinear equalities and does not rely on load-bearing self-citations for its validity. This is a standard case of an independent methodological extension with external demonstration.

Assumptions & free parameters 0 free parameters · 0 assumptions · 0 invented entities

Abstract provides no explicit free parameters, axioms, or invented entities; the method description implies standard neural-network training plus a projection operator but does not introduce new fitted constants or postulated objects.

how reviews work

0 comments

Cite this review

Pith. "Pith review of PL-KKT-hPINN: Enforcing Nonlinear Equality Constraints on Neural Networks via Piecewise-Linear Projection." pith.science (2026). https://pith.science/paper/NQEGNBP3

@misc{pith2026260610682,
  author       = {Pith},
  title        = {Pith review of: PL-KKT-hPINN: Enforcing Nonlinear Equality Constraints on Neural Networks via Piecewise-Linear Projection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NQEGNBP3}},
  note         = {Machine review of arXiv:2606.10682}
}

read the original abstract

While physics-informed neural networks (PINNs) have shown strong potential for process modeling, physical equations are only enforced as soft constraints during training, and thus, they do not guarantee constraint satisfaction at inference. We propose a framework, called piecewise-linear Karush--Kuhn--Tucker hard-constrained PINNs (PL-KKT-hPINNs), that strictly enforces nonlinear equality constraints through piecewise-linear projection. This extends the KKT-hPINN framewor, which exactly enforces linear equalities through the Karush--Kuhn--Tucker (KKT) conditions associated with orthogonally projecting neural network outputs onto the constraint feasible region. The method is demonstrated on a continuous stirred-tank reactor (CSTR) case study for both one and two inputs. Results show that PL-KKT-hPINN preserves predictive accuracy comparable to that of a standard neural network while achieving substantially lower constraint violations. In addition, the proposed model shows improved robustness in low-data regimes, yielding lower RMSE than the unconstrained neural network for limited training sample sizes. These results demonstrate that PL-KKT-hPINN provides a computationally efficient and physically consistent framework for surrogate modeling of nonlinear chemical engineering systems.

Figures

Figures reproduced from arXiv: 2606.10682 by the authors.

**Figure 1.** Schematic of the KKT-hPINN architecture for enforcing linear equality constraints. For a given input x, the neural network first produces an unconstrained prediction yˆ. This prediction is then projected onto the feasible hyperplane of the linear constraint Ax ` By “ b to obtain the corrected prediction y˜. The resulting projection matrices are used in two non-trainable layers with fixed parameters A˚ , B˚ , and b ˚… view at source ↗

**Figure 2.** Schematic of the proposed PL-KKT-hPINN architecture for enforcing nonlinear equality constraints using piecewise-linear KKT projections. For a given input x, the neural network first produces an unconstrained prediction yˆ. In each linearization region Rj , this prediction is projected onto the local feasible hyperplane defined by the linearized constraint Ajx ` Bjy “ bj to obtain the regional corrected prediction y… view at source ↗

**Figure 3.** Training behavior of the standard neural network (NN) and the proposed PL-KKT-hPINN for the 1D CSTR case study. The panels show: (a) training and validation RMSE over epochs, and (b) training and validation nonlinear constraint violation over epochs. Figures 3 and 4 show the loss function root-mean-square-error (RMSE) and mean constraint violation values during each epoch of training for the 1D and 2D cases, respec… view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: Training behavior of the standard neural network (NN) and the proposed PL-KKT-hPINN for the twodimensional CSTR case, where T and CA0 are the inputs. The panels show: (a) training and validation RMSE over epochs, and (b) training and validation nonlinear constraint vi…

**Figure 5.** Figure 5: Effect of the number of linearization regions on the performance of PL-KKT-hPINN for the one-dimensional CSTR case. The panels show: (a) test RMSE, (b) nonlinear constraint violation, and (c) experiment time. The standard NN is shown as a baseline because its architect…

**Figure 6.** Figure 6: Effect of the number of linearization regions on the performance of PL-KKT-hPINN for the two-dimensional CSTR case. The panels show: (a) test RMSE, (b) nonlinear constraint violation, and (c) experiment time. The standard NN is shown as a baseline because its architect…

**Figure 7.** Figure 7: Effect of training sample size on test RMSE for the CSTR case study. Panel (a) corresponds to the one-dimensional input case, where inlet concentration CA0 is the input. Panel (b) corresponds to the two-dimensional input case, where temperature T and feed concentration…

**Figure 8.** Figure 8: Comparison of the final mean absolute nonlinear constraint violation during training and inference for the standard NN, PINN, and PL-KKT-hPINN. The PINN results correspond to the selected penalty weight µ “ r0.01 0.05s T . Error bars indicate variability across repeat…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 3 canonical work pages

[1]

Overview of surrogate modeling in chemical process engineering.Chemie Ingenieur T echnik, 91(3):228–239, 2019

Kevin McBride and Kai Sundmacher. Overview of surrogate modeling in chemical process engineering.Chemie Ingenieur T echnik, 91(3):228–239, 2019

2019
[2]

Formulating data-driven surrogate models for process opti- mization.Computers & Chemical Engineering, 179:108411, 2023

Ruth Misener and Lorenz Biegler. Formulating data-driven surrogate models for process opti- mization.Computers & Chemical Engineering, 179:108411, 2023

2023
[3]

Advances in surrogate based modeling, feasibility analysis, and optimization: A review.Computers & Chemical Engineering, 108:250–267, 2018

Atharv Bhosekar and Marianthi Ierapetritou. Advances in surrogate based modeling, feasibility analysis, and optimization: A review.Computers & Chemical Engineering, 108:250–267, 2018

2018
[4]

Perspectives on the integration between first-principles and data-driven modeling.Computers & Chemical Engineering, 166:107898, 2022

William Bradley, Jinhyeun Kim, Zachary Kilwein, Logan Blakely, Michael Eydenberg, Jordan Jalvin, Carl Laird, and Fani Boukouvala. Perspectives on the integration between first-principles and data-driven modeling.Computers & Chemical Engineering, 166:107898, 2022

2022
[5]

Selection of surrogate modeling techniques for surface approximation and surrogate-based optimization.Chemical Engineering Research and Design, 170:76–89, 2021

Bianca Williams and Selen Cremaschi. Selection of surrogate modeling techniques for surface approximation and surrogate-based optimization.Chemical Engineering Research and Design, 170:76–89, 2021

2021
[6]

Physics-informed neural networks with hard linear equality constraints.Computers & Chemical Engineering, 189:108764, 2024

Hao Chen, Gonzalo E Constante Flores, and Can Li. Physics-informed neural networks with hard linear equality constraints.Computers & Chemical Engineering, 189:108764, 2024

2024
[7]

Integration of planning, scheduling and control prob- lems using data-driven feasibility analysis and surrogate models.Computers & Chemical Engi- neering, 134:106714, 2020

Lisia S Dias and Marianthi G Ierapetritou. Integration of planning, scheduling and control prob- lems using data-driven feasibility analysis and surrogate models.Computers & Chemical Engi- neering, 134:106714, 2020

2020
[8]

Surrogate modeling and surrogate- based optimization with stochastic simulations

Samira Mohammadi, Bianca Williams, and Selen Cremaschi. Surrogate modeling and surrogate- based optimization with stochastic simulations. InComputer Aided Chemical Engineering, vol- ume 49, pages 31–40. Elsevier, 2022

2022

Show all 26 references

[9]

Approximation by superpositions of a sigmoidal function.Mathematics of control, signals and systems, 2(4):303–314, 1989

George Cybenko. Approximation by superpositions of a sigmoidal function.Mathematics of control, signals and systems, 2(4):303–314, 1989

1989
[10]

Multilayer feedforward networks are universal approximators.Neural networks, 2(5):359–366, 1989

Kurt Hornik, Maxwell Stinchcombe, and Halbert White. Multilayer feedforward networks are universal approximators.Neural networks, 2(5):359–366, 1989

1989
[11]

Surrogate-based optimization for mixed-integer nonlinear problems.Computers & Chemical Engineering, 140:106847, 2020

Sun Hye Kim and Fani Boukouvala. Surrogate-based optimization for mixed-integer nonlinear problems.Computers & Chemical Engineering, 140:106847, 2020

2020
[12]

Physics-constrained machine learning for chemical engineering.Current Opinion in Chemical Engineering, 51:101228, 2026

Angan Mukherjee and Victor M Zavala. Physics-constrained machine learning for chemical engineering.Current Opinion in Chemical Engineering, 51:101228, 2026

2026
[13]

Maziar Raissi, Paris Perdikaris, and George E Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear par- tial differential equations.Journal of Computational physics, 378:686–707, 2019. 17 https://...

2019
[14]

Data-driven strategies for optimization of integrated chemical plants.Computers & Chemical Engineering, 166:107961, 2022

Kaiwen Ma, Nikolaos V Sahinidis, Satyajith Amaran, Rahul Bindlish, Scott J Bury, Devin Grif- fith, and Sreekanth Rajagopalan. Data-driven strategies for optimization of integrated chemical plants.Computers & Chemical Engineering, 166:107961, 2022

2022
[15]

Enforcing analytic constraints in neural networks emulating physical systems.Physical review letters, 126(9):098302, 2021

Tom Beucler, Michael Pritchard, Stephan Rasp, Jordan Ott, Pierre Baldi, and Pierre Gentine. Enforcing analytic constraints in neural networks emulating physical systems.Physical review letters, 126(9):098302, 2021

2021
[16]

Optnet: Differentiable optimization as a layer in neural net- works

Brandon Amos and J Zico Kolter. Optnet: Differentiable optimization as a layer in neural net- works. InInternational conference on machine learning, pages 136–145. PMLR, 2017

2017
[17]

Differentiable convex optimization layers.Advances in neural information processing sys- tems, 32, 2019

Akshay Agrawal, Brandon Amos, Shane Barratt, Stephen Boyd, Steven Diamond, and J Zico Kolter. Differentiable convex optimization layers.Advances in neural information processing sys- tems, 32, 2019

2019
[18]

Hard-constrained neural networks with universal approxi- mation guarantees.arXiv preprint arXiv:2410.10807, 2024

Youngjae Min and Navid Azizan. Hard-constrained neural networks with universal approxi- mation guarantees.arXiv preprint arXiv:2410.10807, 2024

2024
[19]

Angan Mukherjee and Debangsu Bhattacharyya. Development of steady-state and dynamic mass and energy constrained neural networks for distributed chemical systems using noisy transient data.Industrial & Engineering Chemistry Research, 63(32):14211–14239, 2024

2024
[20]

Picard-kkt-hpinn: Enforcing nonlinear enthalpy balances for physically consistent neural networks.arXiv preprint arXiv:2501.17782, 2025

Giacomo Lastrucci, Tanuj Karia, Zo ¨e Gromotka, and Artur M Schweidtmann. Picard-kkt-hpinn: Enforcing nonlinear enthalpy balances for physically consistent neural networks.arXiv preprint arXiv:2501.17782, 2025

2025
[21]

Enforce: Nonlinear constrained learning with adaptive-depth neural projection.arXiv preprint arXiv:2502.06774, 2025

Giacomo Lastrucci and Artur M Schweidtmann. Enforce: Nonlinear constrained learning with adaptive-depth neural projection.arXiv preprint arXiv:2502.06774, 2025

2025 arXiv
[22]

Physics-informed neural networks with hard nonlinear equality and inequality constraints.Computers & Chemical Engineering, page 109418, 2025

Ashfaq Iftakher, Rahul Golder, Bimol Nath Roy, and MM Faruque Hasan. Physics-informed neural networks with hard nonlinear equality and inequality constraints.Computers & Chemical Engineering, page 109418, 2025

2025
[23]

Linear model decision trees as surrogates in optimization of engineering applications.Computers & Chemical Engineering, 178:108347, 2023

Bashar L Ammari, Emma S Johnson, Georgia Stinchfield, Taehun Kim, Michael Bynum, William E Hart, Joshua Pulsipher, and Carl D Laird. Linear model decision trees as surrogates in optimization of engineering applications.Computers & Chemical Engineering, 178:108347, 2023

2023
[24]

Piecewise-linear approximations of multidimensional func- tions.Journal of optimization theory and applications, 145(1):120–147, 2010

R Misener and CA2602908 Floudas. Piecewise-linear approximations of multidimensional func- tions.Journal of optimization theory and applications, 145(1):120–147, 2010

2010
[25]

Sensitivity analysis of composite piecewise smooth equations

Daniel Ralph and Stefan Scholtes. Sensitivity analysis of composite piecewise smooth equations. Mathematical Programming, 76(3):593–612, 1997

1997
[26]

Zavala.Statistics for Chemical Engineers

Victor M. Zavala.Statistics for Chemical Engineers. Cambridge University Press, 2025. 18

2025

Pith tools

Reviewed June 27, 2026 · model on record in the stance chip above.

[1] [1]

Overview of surrogate modeling in chemical process engineering.Chemie Ingenieur T echnik, 91(3):228–239, 2019

Kevin McBride and Kai Sundmacher. Overview of surrogate modeling in chemical process engineering.Chemie Ingenieur T echnik, 91(3):228–239, 2019

2019

[2] [2]

Formulating data-driven surrogate models for process opti- mization.Computers & Chemical Engineering, 179:108411, 2023

Ruth Misener and Lorenz Biegler. Formulating data-driven surrogate models for process opti- mization.Computers & Chemical Engineering, 179:108411, 2023

2023

[3] [3]

Advances in surrogate based modeling, feasibility analysis, and optimization: A review.Computers & Chemical Engineering, 108:250–267, 2018

Atharv Bhosekar and Marianthi Ierapetritou. Advances in surrogate based modeling, feasibility analysis, and optimization: A review.Computers & Chemical Engineering, 108:250–267, 2018

2018

[4] [4]

Perspectives on the integration between first-principles and data-driven modeling.Computers & Chemical Engineering, 166:107898, 2022

William Bradley, Jinhyeun Kim, Zachary Kilwein, Logan Blakely, Michael Eydenberg, Jordan Jalvin, Carl Laird, and Fani Boukouvala. Perspectives on the integration between first-principles and data-driven modeling.Computers & Chemical Engineering, 166:107898, 2022

2022

[5] [5]

Selection of surrogate modeling techniques for surface approximation and surrogate-based optimization.Chemical Engineering Research and Design, 170:76–89, 2021

Bianca Williams and Selen Cremaschi. Selection of surrogate modeling techniques for surface approximation and surrogate-based optimization.Chemical Engineering Research and Design, 170:76–89, 2021

2021

[6] [6]

Physics-informed neural networks with hard linear equality constraints.Computers & Chemical Engineering, 189:108764, 2024

Hao Chen, Gonzalo E Constante Flores, and Can Li. Physics-informed neural networks with hard linear equality constraints.Computers & Chemical Engineering, 189:108764, 2024

2024

[7] [7]

Integration of planning, scheduling and control prob- lems using data-driven feasibility analysis and surrogate models.Computers & Chemical Engi- neering, 134:106714, 2020

Lisia S Dias and Marianthi G Ierapetritou. Integration of planning, scheduling and control prob- lems using data-driven feasibility analysis and surrogate models.Computers & Chemical Engi- neering, 134:106714, 2020

2020

[8] [8]

Surrogate modeling and surrogate- based optimization with stochastic simulations

Samira Mohammadi, Bianca Williams, and Selen Cremaschi. Surrogate modeling and surrogate- based optimization with stochastic simulations. InComputer Aided Chemical Engineering, vol- ume 49, pages 31–40. Elsevier, 2022

2022