Online reduced-order modeling from primal solver trajectories for PDE-constrained optimization
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.
Kai: We're looking at the paper: "Online reduced-order modeling from primal solver trajectories for PDE-constrained optimization".
Paper summary: Kai: So, this paper, "Online reduced-order modeling from primal solver trajectories for PDE-constrained optimization," it’s about how to handle the cost of running high-fidelity simulations repeatedly when you're trying to optimize something constrained by a partial differential equation.
Mira: Right, and the core idea here is moving away from just using those final converged solutions for building your reduced model, because that can make the model inaccurate if you move too far from where you got your data one <ref:2610.19821#pg1>.
Kai: They propose an online framework called RETO which builds these local reduced-order models directly from unconverged primal solver trajectories instead of waiting for a full convergence before they start modeling.
Mira: That means they are constructing the model locally as the design point moves, and they use those intermediate states to create a snapshot matrix for the reduced space, which is then rebuilt whenever you need to update it ten <ref:2610.19821#pg2>.
Lev: From an engineering standpoint, that’s smart because it lets you do multiple low-cost state evaluations and adjoint gradient evaluations without needing that full-order convergence at every single design update two <ref:2610.19821#pg2>.
Kai: They claim this approach supports those low-cost state evaluations and gradient calculations while also allowing for design updates without needing a converged high-fidelity state for model construction one <ref:2610.19821#pg1>.
Mira: It’s about reducing the cost of those repeated high dimensional computations by using the intermediate states generated during the primal solver advancement as a continuous source of information three <ref:2610.19821#pg2>.
Lev: So, what they’re saying is that instead of stopping and restarting the whole process every time you move a step, you keep reusing what you already computed to maintain a valid model for that local area two <ref:2610.19821#pg2>.
Kai: It matters because in large scale problems, those repeated full-order convergences can be really expensive overall two <ref:2610.19821#pg2>.
Mira: And they show that this online construction method helps maintain the validity of the reduced model in the current design region more directly than building one big global model beforehand three <ref:2610.19821#pg2>.
Conclusion: Kai: Thinking about the title, "Online reduced-order modeling from primal solver trajectories for PDE-constrained optimization," it really highlights that this method isn't just about creating a model once and then using it; it’s an online process.
Mira: Exactly, the focus is on constructing these local reduced spaces right as you are optimizing along the design trajectory, which is what makes it different from traditional approaches that wait for convergence ten <ref:2610.19821#pg2>.
Lev: And what this implies for us in terms of real hardware, it means we can keep sampling states and calculating gradients without needing a perfect solution at every single step two <ref:2610.19821#pg2>.
Kai: The authors show that this method successfully maintains an effective reduced representation even as the design changes, specifically capturing evolving shock-related flow features in transonic cases two <ref:2610.19821#pg2>.
Mira: They tested this across three different problems, including Poisson parameter field inversion and logistic resource allocation, and they showed solution quality comparable to full order optimization five.
Lev: The computational cost analysis shows that in the subsonic case, high fidelity snapshot acquisition takes about twenty-two percent of the total time, while Jacobian assembly takes thirty-six point one six percent twenty-four.
Kai: But in the transonic case involving moving shocks, that snapshot acquisition fraction jumps up to fifty-seven point four six percent and becomes the main computational cost twenty-five.
Mira: And they found that in the logistic resource allocation case, they achieved a time-to-threshold speedup of twenty-two point three eight times sixteen <ref:2610.19821#pg3>.
Zhe Lu, Wenbo Caod, Weiwei Zhanga
School of Aeronautics, Northwestern Polytechnical University · International Joint Institute of Artificial Intelligence on Fluid Mechanics, Northwestern Polytechnical University · National Key Laboratory of Aircraft Configuration Design, Northwestern Polytechnical University · Institute of AI for Industries, Chinese Academy of Sciences · Institute of Computing Technology, Chinese Academy of Sciences
physics.comp-ph, math.OC
Submitted: 2026-10-14
Updated: 2026-11-17
Comments: 24 pages, 10 figures, 3 tables. Substantially revised version with updated numerical results and figures
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 82/100
The gist: Online reduced-order modeling from primal solver trajectories for PDE-constrained optimization presents an online framework that constructs local reduced-order models directly from unconverged primal…
Key concepts
- Reduced-order Trajectory-based Optimization (RETO)
- RETO is an online method that constructs local reduced models by using a sequence of intermediate, unconverged solutions from the primal solver. Instead of waiting for a full solution, it uses these partial trajectories to build and update the model incrementally during optimization.
- Online Model Construction
- This means creating or updating the simplified mathematical representation of the complex PDE system in real-time as the optimization progresses. The model is built from snapshots of solver trajectories rather than relying on a single, final, converged state.
- Trajectory Reuse
- The method reuses information from partially converged primal solutions across multiple subsequent evaluations (like state and gradient calculations). This reuse allows the system to support several low-cost evaluations before needing to rebuild the reduced model, saving time.
Terminology
Summary
Online reduced-order modeling from primal solver trajectories for PDE-constrained optimization presents an online framework that constructs local reduced-order models directly from unconverged primal solver trajectories to reduce the computational cost of repeated high-fidelity computations in large-scale PDE-constrained optimization
The Core Problem Addressed
Repeated high-fidelity state and gradient evaluations make large-scale PDE-constrained optimization computationally expensive <ref:2610.19821#pg3>. Existing reduced-order modeling approaches often rely on converged high-fidelity states for model construction or updating, which may cause the reduced spaces to lose accuracy as the design moves beyond the region represented by available data <ref:2610. p. 019821#pg3>. The authors address this by presenting Reduced-order Trajectory-based Optimization (RETO), an online framework that constructs local reduced-order models (ROMs) directly from unconverged primal solver trajectories <ref:2610. p. 019821#pg2>. This new method supports multiple low-cost state evaluations, adjoint-based gradient evaluations, and design updates without requiring full-order convergence for model construction <ref:2610. p. 019821#pg2>.
Framework for Online Model Construction
RETO operates by advancing the primal solver for a finite number of iterations at the current design point to generate unconverged states <ref:2610.19821#pg4>. These unconverged states are then combined with previously retained states to form a snapshot matrix, which is used to construct the local reduced space via proper orthogonal decomposition <ref:2610. p. 719821#pg10>. The resulting reference state and reduced basis remain fixed throughout the subsequent local optimization stage <ref:2610. p. 819821#pg10>. This process alternates between trajectory acquisition and reduced-order optimization along the design trajectory <ref:2610. p. 819821#pg11>.
Benefits of Trajectory Reuse
The primary advantage of RETO is that it does not require a newly converged high-fidelity state at every design update <ref:2610.19821#pg5>. Instead, the method reuses information from partially converged primal solutions to support several consecutive state evaluations, gradient evaluations, and design updates before the local model is rebuilt <ref:2610. p. 319821#pg11>. This reuse reduces the need for repeated full-order convergence throughout the optimization process <ref:2610. p. 819821#pg11>.
Performance Across Diverse Problems
RETO was assessed across three distinct PDE-constrained optimization problems: Poisson parameter field inversion, logistic resource allocation, and subsonic and transonic RANS airfoil pressure inverse design <ref:2610.19821#pg5>. The method achieves solution quality comparable to full-order optimization, with speedups of approximately 8–22× in cases without discontinuities in the state fields <ref:2610. p. 019821#pg2>. In the transonic inverse-design case involving moving shocks, RETO retains an approximately threefold speedup <ref:2610. p. 019821#pg2>. The results demonstrate that the online local ROM progressively captures evolving shock-related flow features as new primal solver trajectories are incorporated <ref:2610. p. 019821#pg2>.
Computational Cost Analysis
The computational cost analysis shows that in the subsonic case, high-fidelity snapshot acquisition accounts for 22.00% of the total time, while Jacobian assembly and reduced-order adjoint evaluation account for 36. p. 516% and 32.65% respectively <ref:2610. p. 919821#pg24>. In the transonic case, the fraction associated with snapshot acquisition increases to 57. p. 046%, becoming the dominant computational cost <ref:2610. p. 119821#pg25>. The time-to-threshold speedup achieved in the logistic resource allocation case was 22. p. 938× <ref:2610.19821#pg16>.
Limitations and Future Work
Several limitations warrant further investigation, including the focus on surface-pressure inverse design and the need for studies on objectives involving other flow quantities such as skin friction <ref:2610.19821#pg24>. The framework currently lacks an online error estimator or a model-management mechanism to assess the reliability of local reduced-order predictions <ref:2610. p. 319821#pg24>. Developing adaptive strategies based on information available during primal iterations may improve both computational efficiency and reliability <ref:2610. p. 019821#pg24>. Further analysis of model accuracy and convergence is also needed to establish the conditions under which RETO can be applied reliably <ref:2610. p. 019821#pg24>.
Conclusion
The paper concludes that RETO reduces the frequency of high-fidelity computations and avoids requiring a converged full-order solution at every model update by constructing reduced spaces from intermediate primal states <ref:2610.19821#pg11>. The method successfully maintains an effective reduced representation along the design trajectory through repeated local reconstruction, even capturing changes in shock position and strength in transonic cases <ref:2610. p. 019821#pg22>. The final inverse-design results agree closely with those obtained using the traditional direct adjoint loop (DAL) across all tested problems <ref:2610. p. 919821#pg20>. The method achieves significant time-to-threshold speedups across all evaluated cases <ref:2610. p. 019821#pg21>.
References
[1] A. Jameson, Aerodynamic design via control theory, Journal of Scientific Computing, 3 (1988) 233-260
[2] T. Rees, H.S. Dollar, A.J. Wathen, Optimal solvers for PDE-constrained optimization, SIAM Journal on Scientific Computing, 32 (2010) 271-298 <ref:2610. p. 019821#pg6>
[3] G.K.W., C., P., J., Effective adjoint approaches for computational fluid dynamics, Progress in Aerospace Sciences, 110 (2019) 100542 <ref:2610. p. 019821#pg7>
[4] P. Benner, S. Gugercin, K. Willcox, A survey of projection-based model reduction methods for parametric dynamical systems, SIAM Review, 57 (2015) 483-531 <ref:2610. p. 619821#pg25>
[5] D.J., P.S., W.A., Reduced-order modeling: new approaches for computational physics, Progress in Aerospace Sciences, 40 (2004) 51-117 <ref:2610. p. 019821#pg6>
[6] K. Ito, S.S., A reduced-order method for simulation and control of fluid flows, Journal of Computational Physics, 143 (1998) 403-425 <ref:2610. p. 119821#pg7>
[7] P.A., J.J., Airfoil design optimization using reduced order models based on proper orthogonal decomposition, in: AIAA Fluids 2000 Conference and Exhibit, AIAA Paper 2000-2545, 2000 <ref:2610. p. 119821#pg8>
[8] S.S., A reduced‐order approach for optimal control of fluids using proper orthogonal decomposition, International Journal for Numerical Methods in Fluids, 34 (2000) 425-448 <ref:2610. p. 119821#pg9>
[9] E. Iuliano, D. Quagliarella, Proper orthogonal decomposition, surrogate modelling and evolutionary optimization in aerodynamic design, Computers & Fluids, 84 (2013) 327-350 <ref:2610. p. 119821#pg10>
[10] M.J., C., C., C., Progressive construction of a parametric reduced‐order model for PDE‐constrained optimization, International Journal for Numerical Methods in Engineering, 102 (2015) 1111-1135 <ref:2610. p. 319821#pg5>
[13] L.Z. Li, J.J., J. Zhang, K. Lu, M.N., Aerodynamic shape optimization by continually moving ROM, Aerospace Science and Technology, 99 (2020) 105729 <ref:2610. p. 219821#pg9>
[14] T. Wen, M.J., A globally convergent method to accelerate large-scale optimization using on-the-fly model hyperreduction: application to shape optimization, Journal of Computational Physics, 484 (2023) 112082 <ref:2610. p. 219821#pg9>
[15] E. Qian, M. Grepl, K. Veroy, K. Willcox, A certified trust region reduced basis approach to PDE-constrained optimization, SIAM Journal on Scientific Computing, 39 (2017) S434-S460 <ref:2610. p. 019821#pg9>
[16] H.F., P., Anderson acceleration for fixed-point iterations, SIAM Journal on Numerical Analysis, 49 (2011) 1715-1735 <ref:2610. p. 219821#pg9>
[17] W. Cao, Y. Liu, X. Shan, C. Gao, W. Zhang, A novel convergence enhancement method based on online dimension reduction optimization, Physics of Fluids, 35 (2023) 036124 <ref:2610. p. 019821#pg9>
[18] X. Wang, Y. Liu, X. Yang, W. Zhang, A data-driven convergence booster for accelerating and stabilizing pseudo time-stepping, Journal of Computational Physics, 555 (2026) 114781 <ref:2610. p. 219821#pg9>
[19] Y. Bin, X.I.A., S.J., R.F., Data-enabled reduction of the time complexity of iterative solvers, Journal of Computational Physics, 529 (2025) 113859 <ref:2610. p. 219821#pg9>
[20] W. Yao, S. Marques, T. Robinson, C. Armstrong, L. Sun, A reduced-order model for gradientbased aerodynamic shape optimisation, Aerospace Science and Technology, 106 (2020) 106120 <ref:2610. p. 219821#pg9>
[21] K. Carlberg, M. Barone, H. Antil, Galerkin v. p. 3 least-squares Petrov–Galerkin projection in nonlinear model reduction, Journal of Computational Physics, 330 (2017) 693-734 <ref:2610. p. 319821#pg9>
[22] I. Mazari, G. Nadin, Y. Privat, Optimal location of resources maximizing the total population size in logistic models, Journal de Mathématiques Pures et Appliquées, 134 (2020) 1-35 <ref:2610.19821#pg15>
[23] W. Ding, H. Finotti, S. Lenhart, Y. Lou, Q. Ye, Optimal control of growth coefficient on a steady-state population model, Nonlinear Analysis: Real World Applications, 11 (2010) 688-704 <ref:2610. p. 019821#pg15>
[24] R.H., P., J., Nocedal, C., A limited memory algorithm for bound constrained optimization, SIAM Journal on Scientific Computing, 16 (1995) 1190-1208
[25] A.H., M. Yano, Registration-based nonlinear model reduction of parametrized aerodynamics problems with applications to transonic Euler and RANS flows, Journal of Computational Physics, 521 (2025) 113576 <ref:2610. p. 919821#pg25>
[26] P.R. Spalart, S.R. Allmaras, A one-equation turbulence model for aerodynamic flows, in: 30th Aerospace Sciences Meeting and Exhibit, AIAA Paper 92-0439, Reno, NV, U.S.A., 1992
[27] P.L. Roe, Approximate Riemann solvers, parameter vectors, and difference schemes, Journal of Computational Physics, 43 (1981) 357-372
[28] B.M. Kulfan, Universal parametric geometry representation method, Journal of Aircraft, 45 (2008) 142-158
[29] D.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Online local reduced-order modeling allows for construction of local reduced-order models directly from unconverged primal solver trajectories, which
does not require a newly converged high-fidelity state at every design update.
This enables the system to supportseveral consecutive state evaluations, adjoint-based gradient evaluations, and design updates before the basis is reconstructed.
-
State and gradient evaluation can be performed using a single high-fidelity advancement to support multiple subsequent evaluations. The paper shows that this approach allows
a single high-fidelity advancement to support multiple subsequent state evaluations, gradient evaluations, and design updates before the local model is rebuilt.
-
The system can achieve significant speedups in optimization for various problems. For example, in the logistic resource allocation problem, RETO showed a
time-to-threshold speedup of 22.38x,
demonstrating substantial computational efficiency gains over traditional methods like DAL. -
The reduced basis evolves with the design rather than requiring a fixed space covering the entire design domain in advance. This allows for
local reduced models to be rebuilt
as optimization progresses, meaning thereduced basis can evolve with the design.
-
In transonic inverse-design cases involving moving shocks, RETO shows that
the online local ROM progressively captures evolving shock-related flow features as new primal solver trajectories are incorporated,
leading to accurate final results comparable to full-order optimization.
Related papers
- Gradient-Based Inverse Design of Free-Energy Landscapes with Diffusion Models
- Mixture of Polyconvex Neural Potentials for Parametric Hyperelasticity: Towards Foundation Material Models
- Development and large-scale benchmarks of a protein--ligand absolute binding free energy toolkit
- Spin disorder competing with positional symmetry breaking governs the metal-insulator behavior in oxide paramagnets
- Investigation of the Asymptotic Properties of Active Impedance in Large Finite Array Antennas
- Automatic generation of exchange-correlation response kernels