🔢 mathematics

Real-time inverse solutions via neural matrix operators

This paper introduces NEMO, a reduced-basis neural operator that approximates the parameter-to-observable matrix for physics-constrained inverse problems, enabling real-time, closed-form solutions with over three orders of magnitude speedup compared to traditional PDE solvers while maintaining high accuracy and providing uncertainty quantification.

Original authors: Julie Pham, Thomas O'Leary-Roseberry, Omar Ghattas, Karen Willcox

Published 2026-08-26

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Julie Pham, Thomas O'Leary-Roseberry, Omar Ghattas, Karen Willcox

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 a world where complex physical systems, from the spread of a chemical plume in the air to the stress on a spacecraft flying at hypersonic speeds, could be understood and predicted instantly. This is the promise of "digital twins," virtual replicas of real-world objects that update in real time as new data arrives. For these twins to be useful for emergency response or flight control, they must solve a specific type of puzzle known as an inverse problem. Instead of asking what will happen if we change a setting, an inverse problem asks the reverse: given what we are seeing now, what caused it? If a sensor detects a spike in pollution, where did the leak start? If a plane's skin vibrates in a certain way, what pressure is hitting it? Traditionally, answering these questions requires running massive, time-consuming computer simulations of physics equations. These simulations are so slow that by the time the answer is ready, the situation has already changed, making real-time control impossible.

Researchers at The University of Texas at Austin and The Ohio State University have developed a new method to solve these puzzles in a fraction of a second, making true real-time digital twins a reality. Their approach, called the Neural Matrix Operator, or NEMO, uses a specialized type of artificial intelligence to bypass the need for slow, repeated physics calculations. Rather than training a computer to guess the answer directly, NEMO learns to predict the underlying mathematical structure that connects the unknown cause to the observed effect. By understanding this structure, the system can calculate the solution instantly, without needing to re-simulate the entire physical world every time a new sensor reading comes in. This breakthrough allows for rapid data assimilation, where a digital twin can update its understanding of a crisis or a flight condition the moment new information is available.

The core challenge the team addressed is that many physical systems are governed by partial differential equations, which are complex mathematical rules describing how things change over space and time. In these systems, the relationship between the unknown source (like a leak or a force) and the data we measure is often linear, meaning the effects add up in a predictable way, but the system itself is influenced by other changing factors, like wind speed or temperature, in a non-linear way. Traditional methods struggle here because they treat every new scenario as a fresh, expensive calculation. The researchers realized that while the full physics simulation is heavy, the mathematical map connecting the unknowns to the observations has a simpler, compressed structure that can be learned. They built a neural network, a type of machine learning model, not to output a single number or a picture, but to output a matrix—a grid of numbers that represents the entire rulebook for that specific moment.

This matrix acts as a shortcut. Once the neural network predicts this matrix for a given set of conditions, the computer can use a standard, fast mathematical formula to find the answer. It is like having a master key that fits a specific lock; instead of trying to pick the lock every time (which is slow and difficult), you simply turn the key (which is instant). The researchers tested this on two very different real-world scenarios. First, they simulated an emergency response to a contaminant release on a university campus. The goal was to identify the initial location and shape of a gas leak based on sparse sensor readings taken at different times. In this test, the NEMO system identified the source with high accuracy, matching the quality of the slow, traditional method but doing it thousands of times faster.

In a second test, the team tackled the extreme environment of hypersonic flight, where a vehicle travels at speeds greater than five times the speed of sound. At these speeds, the air heats the vehicle's surface so intensely that direct pressure sensors cannot survive. Instead, engineers must infer the pressure loads by measuring how the vehicle's structure bends and strains. This is an incredibly difficult inverse problem because the material properties of the structure change with temperature. NEMO successfully learned to map the temperature conditions to the pressure loads, allowing it to estimate the forces on the vehicle in real time. In both cases, the new method was more than a thousand times faster than the traditional approach of running the physics solver, and it was up to 45,000 times faster in the contaminant scenario.

Beyond just speed, the method offers a significant advantage in reliability and safety. Because the solution is derived from a clear mathematical structure rather than a black-box guess, the system can also calculate how uncertain it is about its answer. It can tell operators not just what the leak location is, but how confident it is in that location, which is crucial for making life-or-death decisions during an emergency or guiding a spacecraft. The researchers compared their approach to other advanced AI methods that try to learn the inverse problem directly. They found that while those other methods could sometimes produce similar results, they required solving complex optimization problems every time a new measurement came in, which slowed them down and made their performance unpredictable. NEMO, by contrast, provides a fixed, instant calculation that does not waver based on the specific data it receives.

The success of this work relies on the fact that many physical systems, despite their complexity, can be described using a relatively small number of fundamental patterns. The researchers used a technique called reduced-basis approximation to identify these patterns, effectively teaching the AI to ignore the noise and focus on the essential features of the problem. This allows the neural network to be small and efficient enough to run on standard computer hardware, yet powerful enough to handle the massive scale of real-world physics. The team demonstrated that their approach works even when the data is noisy or imperfect, a common reality in the field. By preserving the physical laws of the system within the learning process, the method ensures that the answers it generates are not just statistically likely, but physically possible.

This research marks a significant step forward in the deployment of digital twins for critical applications. It moves the field from theoretical possibility to practical utility, showing that it is possible to solve complex physics problems in real time without sacrificing accuracy. The ability to instantly update a model of a city's air quality or a spacecraft's structural integrity opens the door to new levels of automation and safety. As the authors note, the training of these models happens offline, meaning the heavy lifting is done once, and the resulting system is ready for immediate use in the field. This separation of training and operation is key to making the technology viable for real-time control systems where every millisecond counts. The work suggests that with the right mathematical framing, artificial intelligence can become a powerful tool for understanding the physical world, not just by recognizing patterns, but by solving the fundamental equations that govern it.

Technical Summary: Real-time Inverse Solutions via Neural Matrix Operators (NEMO)

Problem Statement

The paper addresses the challenge of rapid, real-time data assimilation for physical systems governed by partial differential equations (PDEs) with unknown sources (e.g., forcing terms, initial conditions, or boundary conditions). These systems are critical for "digital twin" applications in science, engineering, and medicine, where real-time prediction and control are required.

The core difficulty lies in the computational intractability of solving the associated inverse problems in real-time using traditional physics solvers. Specifically, the paper targets a class of problems where the system dynamics are:

Nonlinearly parameterized with respect to model parameters $m$ (e.g., wind velocity, material properties).
Linearly parameterized with respect to inversion parameters $q$ (the unknown sources).

The relationship between observations $d$ and parameters is governed by a parameter-to-observable (p2o) operator $F(m)$ . Traditional approaches require solving expensive PDEs repeatedly to evaluate $F(m)$ or its adjoint for every new data instance, making real-time inversion impossible. Furthermore, existing neural operator methods often treat inverse problems as generic nonlinear optimization tasks, lacking the physical structure required for closed-form solutions and efficient uncertainty quantification.

Methodology: Neural Matrix Operator (NEMO)

The authors propose NEMO (NEural Matrix Operator), a reduced-basis neural operator approach designed to learn the map from model parameters $m$ directly to the p2o operator $F(m)$ , rather than learning the action of the operator on specific parameters.

1. Reduced Basis Formulation

The method utilizes reduced basis representations for both the inversion parameters $q$ and model parameters $m$ . By projecting these infinite-dimensional functions onto low-dimensional subspaces (using orthonormal bases $\Phi_r$ and $\Psi_r$ ), the problem is reformulated in terms of coefficient vectors $q_r$ and $m_r$ .

The operator $F(m)$ is approximated as a matrix in the reduced subspace.
The inverse problem is transformed into a subspace least-squares problem, which admits a closed-form solution.

2. Neural Operator Learning

Instead of learning a function-to-function map (typical in neural operators like DeepONet), NEMO learns a map from the reduced model parameters to the reduced operator matrix:
$\mathcal{F}_\theta: \mathbb{R}^{r_m} \to \mathbb{R}^{d \times r_q}$
where $r_m$ and $r_q$ are the dimensions of the reduced bases for $m$ and $q$ , and $d$ is the number of observations. The neural network $\mathcal{F}_\theta$ is trained to minimize the Frobenius norm error between the predicted reduced operator and the true reduced operator computed via PDE solvers.

3. Real-Time Inversion and Uncertainty Quantification

Once trained, for a new model parameter $m$ , the NEMO network outputs the approximate reduced operator $\mathcal{F}_\theta(m_r)$ . The inverse solution $q^\ddagger_r$ is computed via a closed-form least-squares equation:
$q^\ddagger_r = [\mathcal{F}_\theta(m_r)^* \Gamma^{-1} \mathcal{F}_\theta(m_r) + \gamma I_r]^{-1} \mathcal{F}_\theta(m_r)^* \Gamma^{-1} d$
This formulation offers two key advantages:

Speed: The computational complexity is $O(d r_q^2 + r_q^3)$ , independent of the PDE discretization dimension.
Uncertainty Quantification (UQ): Under a Gaussian noise model, the posterior covariance of the solution has a closed-form expression (the inverse Hessian of the objective), enabling real-time probabilistic analysis without iterative sampling (e.g., MCMC).

Key Contributions

Real-Time Methodology: A novel framework for solving inverse problems for PDE-governed systems with unknown sources via nonlinearly parameterized least-squares solutions that are computable in real-time.
Neural Matrix Operator: A specific neural operator architecture that outputs the reduced parameter-to-observable operator itself (a matrix), preserving the linear structure of the problem with respect to the inversion parameters.
Theoretical Guarantees: Supporting theory establishing the consistency of the NEMO inversion error with the operator approximation error and demonstrating dimension independence for smooth maps.

Experimental Results

The authors validated NEMO on two real-world applications:

1. Contaminant Transport Initial Condition Identification

Task: Identifying the initial condition of an airborne contaminant on a university campus using sparse sensor data, with unknown wind velocity fields ( $m$ ).
Setup: Advection-diffusion equation with 1.08 million degrees of freedom (DOF). Reduced basis dimensions: $r_m=3$ (wind), $r_q=50$ (initial condition), $d=150$ (observations).
Performance: NEMO achieved a speedup of >45,000x compared to computing the reduced operator via the PDE solver (0.012s vs. 541s). The inverse solution quality was nearly identical to the true reduced operator solution, with errors decreasing as training data increased.

2. Hypersonic Aerodynamic Pressure Estimation

Task: Estimating surface pressure loads on a hypersonic vehicle from structural strain measurements, where material properties vary with temperature ( $m$ ).
Setup: Quasi-static elastic response with 461,664 DOF. Reduced basis dimensions: $r_m=3$ (temperature), $r_q=20$ (pressure), $d=108$ (sensors).
Performance: NEMO achieved a speedup of >900x (0.006s vs. 5.6s). The method met the 100 Hz real-time requirement for onboard guidance and control, maintaining low relative errors even with observational noise.

Comparison with State-of-the-Art

vs. Forward Surrogates (e.g., MIONet): While MIONet (a multiple-input neural operator) could learn the forward map, solving the inverse problem required online nonlinear optimization (L-BFGS). NEMO provided >10x to >50x speedup over MIONet-based inversion because it avoids iterative optimization, offering fixed, low computational cost.
vs. Direct Inverse Maps: Directly learning the map from observations to parameters (inverse MIONet) resulted in poor adaptability; the model could not adjust to different noise levels or regularization parameters without retraining. NEMO allows principled tuning of the regularization parameter $\gamma$ post-training.
Uncertainty Quantification: Unlike generic neural operators that require iterative sampling for UQ, NEMO provides an explicit, closed-form posterior covariance matrix.

Significance and Claims

The paper claims that NEMO represents a significant advancement for digital twin deployment by enabling real-time data assimilation for complex PDE systems. The primary significance lies in:

Scalability: The method decouples online computational cost from the high-dimensional PDE discretization, relying instead on the low-dimensional reduced basis.
Physical Consistency: By outputting the operator matrix rather than a direct inverse map, NEMO preserves the physical linear structure of the problem, enabling closed-form solutions and rigorous uncertainty quantification.
Practical Viability: The demonstrated speedups (orders of magnitude) and ability to handle real-time constraints (e.g., 100 Hz for hypersonics) make the approach suitable for edge computing and critical control applications where traditional solvers are infeasible.

The authors note that the approach relies on the validity of low-rank approximations for the operator and assumes model parameters $m$ are known or estimated externally. Future work is suggested to explore nonlinear encodings for $m$ and analyze errors introduced by imperfect knowledge of $m$ .