🔢 mathematics

Time-Dependent PDE-Constrained Optimization via Weak-Form Latent Dynamics

This paper introduces a weak-form latent-space reduced-order modeling framework (WLaSDI) that accelerates gradient-based optimization of high-dimensional, time-dependent PDE-constrained problems by compressing solution trajectories into robust, noise-resilient latent dynamics, achieving speedups of up to five orders of magnitude while maintaining accuracy across diverse benchmark applications.

Original authors: April Tran, Terry Haut, David Bortz, Youngsoo Choi

Published 2026-05-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: April Tran, Terry Haut, David Bortz, Youngsoo Choi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to design the perfect shape for a rocket nozzle, or the ideal temperature settings for a nuclear fusion reactor. To do this, you need to run a computer simulation that solves complex physics equations (Partial Differential Equations, or PDEs) over and over again. You tweak a knob, run the simulation, see the result, tweak the knob again, and repeat.

The problem is that these simulations are like trying to calculate the trajectory of every single raindrop in a hurricane. They are incredibly accurate, but they take so much computing power and time that running them thousands of times to find the "best" design is practically impossible. It's like trying to find the perfect recipe by baking a full-sized, five-course banquet every time you want to test adding a pinch more salt.

The Solution: A "Smart Sketch" Instead of the Full Painting

This paper introduces a new method called WLaSDI (Weak-form Latent Space Dynamics Identification) to solve this problem. Think of it as a way to create a highly accurate, low-resolution "sketch" of the physics that runs lightning-fast, allowing you to test thousands of designs in the time it used to take to bake one cake.

Here is how the paper explains it, broken down into simple concepts:

1. The "Compression" Trick (The Suitcase)

The full simulation involves millions of data points (like a massive, heavy suitcase full of clothes). WLaSDI first uses a "compressor" (an encoder) to fold all that data down into a tiny, manageable suitcase (a low-dimensional "latent space").

Analogy: Instead of carrying a whole library of books to find a specific fact, you carry a single index card that summarizes the whole library.
The Catch: Usually, when you compress data, you lose some detail, or the "summary" breaks if the data is a little bit messy (noisy).

2. The "Weak-Form" Secret Sauce (The Noise Filter)

This is the paper's biggest innovation. Traditional methods try to figure out how the system changes by looking at the exact speed of change at every single moment. If your data has a little bit of static or noise (like a scratch on a record), these methods get confused and the simulation goes off the rails.

WLaSDI uses a "Weak-Form" approach.

Analogy: Imagine you are trying to guess the speed of a car by looking at its position every second. If your stopwatch is slightly off (noise), your speed calculation is wrong.
The WLaSDI Way: Instead of looking at the exact speed at every split second, WLaSDI looks at the average behavior over a short period, smoothing out the bumps. It's like asking, "Did the car generally move forward?" rather than "Exactly how fast was it at 12:00:01?"
Result: This makes the method incredibly robust. Even if the training data is "dirty" or noisy, the "sketch" remains accurate.

3. The Optimization Loop (The Chef's Shortcut)

Once WLaSDI has learned this "sketch" (the latent dynamics), it can predict the outcome of a new design almost instantly.

The Old Way: To find the best design, you had to bake the full banquet 1,000 times.
The WLaSDI Way: You bake the full banquet once to learn the recipe. Then, you use the "sketch" to test 1,000 variations in seconds. When you find the winner, you can verify it with the full simulation if needed.

What Did They Test It On?

The authors tested this "smart sketch" method on three very different, complex physics problems to prove it works:

Burgers' Equation (Traffic Jams): They simulated how waves crash into each other. Even with 40% "noise" (like static on a radio), WLaSDI found the correct design parameters, while other methods failed or gave wrong answers. It was 20 times faster than the full simulation.
Vlasov-Poisson (Plasma Physics): They simulated how charged particles move in a plasma (like in a star or fusion reactor). The full simulation takes 72 seconds on a supercomputer. WLaSDI took 0.1 seconds on a single laptop core. That is a speedup of 50,000 times.
Thermal Radiative Transfer (Nuclear Fusion): They tried to design a "hohlraum" (a gold cylinder used in fusion experiments) to heat a fuel capsule evenly. The full simulation takes 30 minutes on a massive supercomputer. WLaSDI did the job in 2.7 seconds. This allowed them to find the optimal design in a fraction of the time, even when the training data was noisy.

The Bottom Line

The paper claims that WLaSDI is a game-changer for engineering and science because:

It's Fast: It speeds up optimization by up to 100,000 times (five orders of magnitude) compared to traditional methods.
It's Tough: It doesn't break when the data is messy or noisy, unlike other "smart" shortcuts.
It's Accurate: It finds the correct "best design" just as well as the slow, expensive method, but without the wait.

In short, WLaSDI lets scientists and engineers skip the long, expensive calculations and go straight to the best solution, even when the data they start with isn't perfect.

Technical Summary: Time-Dependent PDE-Constrained Optimization via Weak-Form Latent Dynamics

Problem Statement

High-dimensional, time-dependent partial differential equation (PDE) constrained optimization problems are computationally prohibitive in many-query design and control settings. These problems require repeated forward and sensitivity solves of high-fidelity physical models within iterative optimization loops. The computational cost of invoking high-fidelity PDE solvers for every gradient evaluation or function query renders such workflows infeasible for complex systems. While surrogate-based strategies exist, many rely on interpolating specific quantities of interest (QoIs), requiring retraining when objectives change, or use reduced-order models (ROMs) that may lack robustness to noisy training data or fail to provide scalable gradient information.

Methodology

The paper proposes a framework that integrates Weak-form Latent Space Dynamics Identification (WLaSDI) into gradient-based PDE-constrained optimization. The methodology proceeds through four key stages:

Data Sampling and Compression: High-fidelity solution trajectories are generated for a set of training parameters. These high-dimensional states are compressed into a low-dimensional latent space ( $z \in \mathbb{R}^{N_z}$ , where $N_z \ll N_u$ ) using an encoder-decoder pair (e.g., Proper Orthogonal Decomposition or autoencoders).
Weak-Form Dynamics Identification: Instead of identifying latent dynamics via explicit time derivatives (as in strong-form SINDy), the framework employs Weak-form Estimation of Nonlinear Dynamics (WENDy). This approach projects the latent dynamics onto smooth, compactly supported test functions and applies integration by parts. This avoids explicit numerical differentiation of potentially noisy training trajectories, significantly improving robustness. The latent evolution is modeled as a parametric ordinary differential equation (ODE):
$\frac{dz}{dt}(t, \mu) = W^T(\mu) \theta(z(t, \mu))$
where $W(\mu)$ are coefficient matrices identified via weak-form least squares, and $\theta(z)$ is a library of candidate nonlinear functions.
Optimization Reformulation: The original PDE-constrained optimization problem is reformulated by replacing the high-fidelity PDE constraints with the learned latent ODE constraints. The objective function and constraints are evaluated by decoding the latent trajectory back to the physical space.
Gradient Derivation: To enable scalable gradient-based optimization, the paper derives both direct-sensitivity and adjoint-based gradient expressions for the reduced latent dynamics. These formulations allow for the efficient computation of gradients with respect to design parameters $\mu$ by solving linear systems involving the low-dimensional Jacobians of the latent ODE, rather than the full-order PDE residuals.

Key Contributions

The paper makes three primary contributions:

Formulation of a Reduced Framework: It formulates a WLaSDI-based reduced framework for time-dependent PDE-constrained optimization, replacing repeated high-fidelity solves with the efficient integration of learned latent dynamics.
Gradient Derivation: It derives direct-sensitivity and adjoint-based gradient expressions specifically for the learned latent dynamics. This enables scalable gradient evaluation as the parameter dimension increases, a critical requirement for practical optimization.
Demonstration of Robustness and Efficiency: Through three benchmark problems, the work demonstrates that WLaSDI enables accurate and efficient optimization while exhibiting superior robustness to noisy training data compared to strong-form LaSDI and interpolation-based surrogates.

Results

The framework was evaluated on three time-dependent benchmark problems:

Inviscid Burgers' Equation: An inverse problem identifying initial condition parameters to match a target state.
- Accuracy: WLaSDI achieved relative parameter errors of $\approx 0.5\%$ even with 40% additive Gaussian noise, outperforming strong-form LaSDI and Radial Basis Function (RBF) interpolation by up to two orders of magnitude in accuracy.
- Efficiency: It provided a $\approx 20\times$ speedup over full-order model (FOM) optimization and $\approx 5\times$ speedup over LaSDI.
Vlasov–Poisson System: An inverse problem for the two-stream instability.
- Accuracy: The method successfully recovered target parameters with high fidelity under 5% noise.
- Efficiency: WLaSDI evaluations took 0.1 seconds on a single core compared to $\approx 72$ seconds for the FOM on 16 cores, representing a speedup of up to five orders of magnitude.
Thermal Radiative Transfer (TRT) for Hohlraum Design: A design optimization problem aiming to minimize temperature variation on a fusion capsule surface subject to heating constraints.
- Robustness: In the presence of 5% noise, WLaSDI consistently converged to the noise-free global optimum, whereas strong-form LaSDI exhibited significant scatter and deviation.
- Efficiency: The surrogate reduced computational costs by approximately five orders of magnitude relative to FOM optimization. The use of local submodels allowed for parallelized training and optimization.

Significance and Claims

The paper claims that weak-form latent dynamics provide an efficient and noise-robust surrogate foundation for gradient-based optimization of complex time-dependent PDE systems.

Robustness: By avoiding explicit differentiation, the weak-form approach yields more reliable surrogate dynamics, which is particularly critical when training data is noisy—a common scenario in experimental or high-fidelity simulation contexts.
Scalability: The derivation of adjoint-based gradients for the latent ODE allows the framework to scale to high-dimensional parameter spaces without the computational penalty associated with full-order adjoint solves.
Flexibility: Unlike QoI-specific surrogates (e.g., RBF interpolation) that require retraining for new objectives, WLaSDI reconstructs the full solution field. This allows for the evaluation of arbitrary QoIs and constraints without retraining, making it adaptable to evolving design requirements.
Performance: The results demonstrate that WLaSDI can achieve speedups of up to five orders of magnitude relative to full-order optimization while maintaining high accuracy, effectively enabling many-query workflows that were previously computationally prohibitive.

The authors note that while the performance is strong, the reliability of the optimization ultimately depends on the accuracy of the surrogates, highlighting the need for future work in uncertainty quantification and multi-fidelity strategies.