Running a single physics simulation can take hours or days, depending on the complexity of the geometry and the equations involved. For engineers iterating through hundreds of design variations, that wait stacks up fast. But a growing body of peer-reviewed research shows that a new class of AI models, trained not on images or text but on the mathematics of physical systems, can approximate those simulation results in seconds. The technology is not yet a plug-and-play replacement for conventional solvers, but the results published so far suggest it could reshape how engineers approach design optimization across fields from materials science to fluid dynamics.
The core idea: learning the physics, not just the answer
Traditional simulation software solves partial differential equations (PDEs) numerically, grinding through millions of calculations for each specific set of conditions. Change the shape of a wing or the composition of an alloy, and the solver starts over. Two research efforts, published between 2019 and 2021, proposed a fundamentally different approach: train a neural network to learn the underlying mathematical operator so it can produce approximate solutions for entire families of PDEs at once.
The first, the Fourier Neural Operator (FNO), introduced an architecture that maps between function spaces rather than memorizing individual solutions. On benchmark problems, the FNO reported speedups of up to three orders of magnitude compared to conventional numerical solvers, though those gains were measured on specific, well-controlled test cases rather than full industrial workflows. Because the model learns the operator itself, it can generalize across different input conditions without retraining from scratch, a property that matters enormously when engineers need to evaluate thousands of parameter variations in a design loop.
The second approach, Deep Operator Networks (DeepONet), was detailed in a paper later published in Nature Machine Intelligence in 2021. DeepONet established that neural networks can learn mappings from functions to functions, grounded in a universal approximation theorem for operators. “What makes operator learning fundamentally different from standard supervised learning is that you are learning a mapping between infinite-dimensional spaces,” said Lu Lu, an assistant professor of chemical and biomolecular engineering at Yale University who was a lead author on the DeepONet paper. “That is what gives these models the ability to generalize across different input functions without retraining.” Together, these two lines of work laid the theoretical and practical foundation for what researchers now call operator learning.
Where it has already been tested
The clearest demonstration of real engineering value appeared in a 2024 peer-reviewed study published in npj Computational Materials, a Nature Portfolio journal. That paper applied a hybrid version of the Fourier Neural Operator to phase-field modeling, a simulation technique used in materials science to track how microstructures evolve during processes like solidification and grain growth. These simulations are notoriously expensive with conventional solvers. The study benchmarked the hybrid model against those solvers and found that modest sacrifices in fidelity could yield large reductions in wall-clock time, a tradeoff many materials engineers would accept during early-stage design exploration.
Anima Anandkumar, the Bren Professor of Computing and Mathematical Sciences at Caltech and a senior director of AI research at NVIDIA, has been a driving force behind the Fourier Neural Operator line of work. In a May 2025 interview with the Caltech Science Exchange, Anandkumar noted that the real bottleneck for adoption is not inference speed but trust: “Engineers need to know when the surrogate is wrong, not just that it is fast. Building reliable uncertainty quantification into these models is the next critical step.”
A separate effort called PROSE-FD pushed the concept toward fluid dynamics, targeting multiple PDE operators including Navier-Stokes variants. That work, still a preprint as of early 2025, signals a shift from narrow, single-task surrogates toward broader foundation-style models that could transfer across different geometries and physical conditions. If the approach holds up under wider testing, it would reduce the need for a separately trained model for every new engineering problem. But the work has not yet undergone peer review, and its claims about cross-geometry transfer have not been independently replicated.
What engineers should know before adopting these tools
The gap between benchmark performance and production deployment remains wide. No published case study, as of May 2026, quantifies how these models perform when embedded in commercial computer-aided design software or integrated into real manufacturing pipelines. Error behavior when the models encounter PDE families outside their training distribution is not systematically reported, leaving open questions about robustness and failure modes that matter most in practice.
Training cost is another consideration that the primary papers tend to understate. Building a Fourier Neural Operator or DeepONet requires significant GPU time and large datasets of pre-computed solutions generated by the very conventional solvers the AI is meant to replace. The published research focuses on inference speed, the time to produce a new prediction once the model is trained, but does not provide detailed accounting of the total computational budget needed to prepare these models for a new engineering domain. For smaller firms without access to large compute clusters, the upfront investment could offset downstream time savings, especially if each new class of geometry or material behavior demands a fresh training campaign.
Safety-critical applications raise the bar further. The npj Computational Materials study discussed accuracy tradeoffs in practical terms, but the specific thresholds at which surrogate error becomes unacceptable for aerospace structural analysis, nuclear reactor modeling, or biomedical device certification are not defined in the current literature. Engineers in those fields will need domain-specific validation before replacing conventional solvers, including sensitivity studies that link surrogate error to design margins, fatigue life, and regulatory safety factors.
George Karniadakis, the Charles Pitts Robinson and John Palmer Barstow Professor of Applied Mathematics at Brown University and a co-author of the DeepONet paper, has cautioned that the field needs more rigorous benchmarking against industrial-grade solvers. “We have shown these models work beautifully on canonical problems,” Karniadakis said in an April 2026 presentation at the SIAM Conference on Computational Science and Engineering. “But the engineering community rightly demands validation on problems with messy geometries, noisy data, and strict certification requirements. That work is still in its early stages.”
It is also worth noting what these models are competing against. Engineers already use reduced-order methods, kriging surrogates, and other approximation techniques to speed up design loops. Operator-learning models offer a more general and potentially more powerful alternative, but they are entering a space where “good enough” surrogates already exist for many applications. The value proposition depends on whether the new models can handle more complex, higher-dimensional problems than existing methods, and early results suggest they can, though head-to-head comparisons remain sparse.
Separating headline speedups from real workflow gains
For anyone evaluating these tools, the most important distinction is between raw inference time and end-to-end workflow improvement. A surrogate that evaluates in milliseconds may still deliver modest overall benefit if data preprocessing, mesh generation, or post-processing dominate the project schedule. Conversely, in workflows where the PDE solve is the clear bottleneck, such as parametric sweeps in thermal management or topology optimization, operator-learning models could meaningfully compress design cycles even if some accuracy is traded away.
Commercial interest is growing. NVIDIA’s Modulus platform already provides tools for building physics-informed neural network surrogates, and simulation software companies including Ansys have begun exploring AI-accelerated workflows. But these commercial efforts are largely in early or pilot stages, and publicly available performance data from industrial deployments remains scarce.
The current evidence supports a clear but bounded conclusion: operator-learning models like the Fourier Neural Operator and DeepONet have convincingly demonstrated that they can approximate families of PDE solutions far faster than traditional solvers on specific benchmark problems. Early domain applications, particularly in phase-field simulations, show those gains can translate into realistic engineering scenarios. But questions about training cost, out-of-distribution robustness, and safety-critical validation remain unresolved. Engineers and decision-makers evaluating these tools today are best served by treating them as powerful experimental accelerators, not yet as drop-in replacements for the numerical solvers they trust. The technology is real, the math is sound, and the trajectory is promising. The production-grade proof is still being written.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
