Hybrid Fourier Neural Operator-Plasma Fluid Model for Fast and Accurate Multiscale Simulations of High Power Microwave Breakdown

arXiv:2509.05799 · physics.plasm-ph, cs.AI, cs.LG, physics.comp-ph · Submitted 2025-09-06 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "Hybrid Fourier Neural Operator-Plasma Fluid Model for Fast and Accurate Multiscale Simulations of High Power Microwave Breakdown".

Jane: The paper was written by Kalp Pandya, Pratik Ghosh, Ajeya Mandikal, Shivam Gandha and Bhaskar Chaudhury from Group in Computational Science and HPC, DA-IICT, Dhirubhai Ambani University and Smart Energy Learning Center, DAU.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Title: Tom: Welcome back to the show, everyone. Today we're diving into a paper that's got a mouthful of a title: "Hybrid Fourier Neural Operator–Plasma Fluid Model for Fast and Accurate Multiscale Simulations of High Power Microwave Breakdown." Jane, I gotta say, just reading that title out loud makes me feel smarter.

Jane: Ha! It does have that effect, Tom. But let's break it down for our listeners who might not spend their weekends reading plasma physics papers. This is about high power microwaves — think intense electromagnetic waves — and what happens when they hit air and turn it into plasma. That's called breakdown, and it's a big deal for things like communications, propulsion, even aerospace applications.

Tom: Right, and the paper is from a group at DA-IICT in India, led by Bhaskar Chaudhury. They've been working on this problem for years. The challenge is that simulating this breakdown is brutally expensive on a computer. We're talking days for a small 2D simulation, and months for anything realistic.

Jane: Exactly. And that's where the "Fourier Neural Operator" part comes in. That's a type of machine learning model that can learn to predict how electromagnetic fields behave without solving the full physics equations every single time. So they're basically teaching a neural network to be a fast, approximate physicist.

Tom: And the "hybrid" part is clever. They don't replace everything with the neural network. They keep the traditional plasma solver — the part that tracks how the electron density evolves — and they only replace the expensive electromagnetic field solver with the neural network. It's like keeping your reliable old engine but swapping in a turbocharger for the part that was slowing you down.

Jane: That's a great way to put it, Tom. And the results are pretty stunning. They're getting speedups of around sixty times compared to the traditional method. A simulation that took over thirty-five hours can now run in under forty minutes.

Tom: Sixty times. Let me just sit with that number for a second. That's not a small improvement. That's the difference between running one simulation and running sixty simulations in the same amount of time. For researchers who need to explore different parameters — different electric field strengths, different pressures — that's a game changer.

Jane: Absolutely. And we should mention that the neural network isn't just fast, it's accurate. They tested it on electric field values it had never seen during training, and the predictions matched the traditional solver almost perfectly. The streamer shapes, the growth rates, the temporal evolution — all in excellent agreement.

Tom: So we've got speed and accuracy. What's not to love? Let's bring in Lu from Tsinghua to give us the big picture on what this means for the field.

Lu: Thanks, Tom. I think this paper is significant because it shows a practical path forward for multiscale simulation. These problems — where you have fast electromagnetic oscillations happening at picosecond scales and slow plasma evolution happening over nanoseconds — they're notoriously hard to simulate. The traditional approach forces you to resolve the fastest timescale, which makes everything painfully slow. This hybrid approach sidesteps that bottleneck.

Jane: So it's not just about this one application. It's a template for how to combine machine learning with physics-based simulation in general.

Lu: Exactly. And the fact that they used a Fourier Neural Operator is smart, because FNOs are designed to learn mappings between function spaces. They're not just doing image-to-image translation; they're learning the underlying operator that maps plasma density to the resulting electric field. That's why it generalizes so well to unseen inputs.

Tom: I love that. So we've got a paper that's both practically useful and conceptually elegant. Now, Meng, I know you're the engineer in the room. What's your take on actually implementing something like this?

Meng: Well, Tom, the thing that caught my eye is how they handled the integration. They took an existing C-based simulation code and wrapped it with a Python controller. The neural network runs in Python, the plasma solver runs in C, and they pass data back and forth using shared libraries. That's a really pragmatic approach because it means you don't have to rewrite your entire simulation from scratch.

Jane: That's a huge point. A lot of research groups have legacy code that's been validated over years. If you tell them they need to rewrite everything in Python, they'll never adopt it. But this approach lets them keep their trusted C code and just bolt on the machine learning part.

Meng: Right. And the performance overhead of that Python-C interface is minimal. The data is passed as contiguous memory buffers, so there's no expensive copying. It's a clean, efficient design.

Tom: So we've got speed, accuracy, and a practical implementation strategy. This is looking like a win-win-win. What do you think, Lalam? Where does this take us?

Lalam: I think the most impactful vision here is democratizing access to high-fidelity plasma simulation. Right now, only well-funded labs with serious computing resources can run these simulations. With a sixty-fold speedup, smaller research groups, universities in developing countries, even undergraduate students could explore this physics. That could accelerate discovery across the entire field.

Tom: That's a beautiful thought to end this segment on. We'll come back in a moment to dig deeper into the methodology and what makes this hybrid approach tick. Stay with us.

Summary: Tom: Welcome back. We're still on "Hybrid Fourier Neural Operator–Plasma Fluid Model for Fast and Accurate Multiscale Simulations of High Power Microwave Breakdown." Jane, we touched on the big picture, but let's get into the actual physics problem they're solving.

Jane: Gladly, Tom. So picture this: you've got a microwave beam hitting a small pocket of air. If the electric field is strong enough, it ionizes the air — strips electrons off molecules — and you get a plasma. But here's the fascinating part: that plasma then interacts with the microwave field. It reflects it, it absorbs it, it modifies it.

Tom: And that feedback loop is what creates these structures called "microwave streamers." The plasma starts as a little blob, but because the electric field gets enhanced at the poles of the blob, ionization happens faster there, and the blob stretches out into a filament along the field direction.

Jane: Exactly. And in this paper, they're simulating that process in 2D. They have a rectangular domain, they send two identical microwaves in from opposite sides to create a standing wave, and they seed a small Gaussian plasma blob in the center. Then they watch it evolve into a streamer.

Tom: Now, the traditional way to simulate this is to solve Maxwell's equations — that's the physics of electromagnetic waves — coupled with a plasma continuity equation that tracks the electron density. The Maxwell solver uses something called FDTD, which stands for Finite-Difference Time-Domain. It's accurate but slow.

Jane: And here's the key insight from the paper: the electromagnetic solver takes up more than ninety-nine percent of the computation time. The plasma solver is almost free by comparison. So if you could speed up the electromagnetic part, you'd speed up the whole simulation dramatically.

Tom: That's exactly what they did. They trained a Fourier Neural Operator to predict the scattered electric field — the part of the field that's modified by the plasma — given the plasma density and the incident wave. And then they plugged that neural network into the simulation loop in place of the FDTD solver.

Meng: Can I jump in here? I want to talk about the data. They generated training data by running the full FDTD simulation for different electric field strengths — from two point five to three point zero megavolts per meter. Each simulation produces thousands of snapshots of plasma density and the corresponding electric field. That's how they built their dataset.

Jane: Right, and they were careful about validation. They held out five electric field values — two point five five, two point six five, two point seven five, two point eight five, and two point nine five megavolts per meter — that the model never saw during training. Then they tested the hybrid model on those unseen values to make sure it generalizes.

Lu: That's the right way to do it. If you only test on data the model has seen, you're just measuring memorization. By testing on entirely new field strengths, they're actually testing whether the model has learned the physics.

Tom: And the results? Jane, you mentioned the speedup earlier. Let's put some numbers on it. For the test case at two point five five megavolts per meter, the traditional FDTD simulation took about two thousand one hundred thirty-one minutes — that's over thirty-five hours. The hybrid model did it in about thirty-seven minutes. That's a fifty-seven-fold speedup.

Jane: And at two point seven five megavolts per meter, it was even better — almost sixty-three times faster. The hybrid model finished in about twenty-two minutes versus over twenty-three hours for the traditional method.

Meng: What about accuracy? Because a fast model that's wrong isn't useful.

Jane: That's the beautiful part. The model achieved an average SSIM of zero point nine nine nine nine — that's a structural similarity metric where one point zero means perfect match. And the average percent error was just zero point one two. So we're talking about predictions that are visually and quantitatively indistinguishable from the full physics simulation.

Tom: And they didn't just compare the electric field predictions. They ran the entire hybrid simulation — the full coupled loop — and compared the streamer evolution. The streamer length over time, the growth rate, the electric field at the streamer tip — all of it matched the traditional simulation almost perfectly.

Lu: That's the real test. A single prediction being accurate is one thing, but running the whole simulation forward and having it stay on track over hundreds of nanoseconds — that's where errors usually accumulate and blow up. The fact that they didn't see that drift is very encouraging.

Tom: So the summary is: they identified the bottleneck, replaced it with a neural network, kept the physics-based solver for the rest, and got a sixty-fold speedup with essentially no loss in accuracy. That's a clean result.

Jane: And it opens the door to simulations that were previously impossible. We'll talk about that in a moment. But first, let's hear what Lalam thinks about the broader implications of this approach.

Lalam: I think the most exciting implication is what happens when you combine this speedup with larger domains. The paper notes that scaling to a ten by ten wavelength domain would take about five hundred days with the traditional method. With this hybrid approach, that drops to about a week. Suddenly, simulations that were completely out of reach become routine.

Tom: A week versus a year and a half. That's the kind of change that shifts what questions researchers even think to ask. We'll explore that next.

Improvements: Tom: We're back with "Hybrid Fourier Neural Operator–Plasma Fluid Model for Fast and Accurate Multiscale Simulations of High Power Microwave Breakdown." Jane, we've covered what they did and how well it worked. Now let's talk about what this enables — the improvements and the new possibilities.

Jane: Right, Tom. And I think the biggest improvement is simply the scale of what you can simulate. The paper makes this concrete: a one by one wavelength domain with five hundred twelve grid points per wavelength takes about five days with the traditional solver. But a ten by ten wavelength domain — which is what you'd need for realistic applications — would take roughly five hundred days.

Meng: And that's just not practical. Nobody's going to wait a year and a half for one simulation. But with the hybrid approach, that same simulation could be done in about a week. That changes the entire research workflow.

Lu: It does. And it's not just about bigger domains. It's about parameter sweeps. If you want to study how breakdown changes with pressure, or frequency, or gas composition, you need to run dozens or hundreds of simulations. With the traditional method, that's months of compute. With this hybrid method, you could do it in days.

Tom: So researchers could actually explore the parameter space instead of just picking a few points and hoping for the best.

Jane: Exactly. And there's another improvement the paper hints at: the modular architecture. They built this as a Python controller that calls C functions through shared libraries. That means you could swap out the FNO for a different neural network, or swap out the plasma solver for a more sophisticated one, without rewriting everything.

Meng: That's a big deal for maintainability. In my experience, research code is often a tangled mess. But this design — clean separation between the ML component and the physics component — makes it much easier to iterate and improve.

Lu: I'd add that the FNO architecture itself is worth discussing. They used four Fourier layers with sixteen frequency modes each. The input is a two-channel image: plasma density and incident electric field. The output is the scattered electric field. It's an image-to-image mapping, but because it operates in the frequency domain, it captures global spatial dependencies that convolutional networks might miss.

Jane: And that's important for this physics, because the electric field at one point depends on the plasma distribution everywhere, not just locally. The Fourier transform naturally captures those long-range interactions.

Tom: Let me ask something practical. The paper mentions that the plasma solver takes less than one percent of the computation time. So even if the FNO were perfect, the maximum speedup you could get is about a hundred times. They're getting sixty. Is there room to push further?

Meng: There might be. The gap between sixty and a hundred could come from the FNO inference time itself, plus the overhead of the Python-C communication. But honestly, sixty times is already transformative. Pushing to ninety would be nice, but it's not necessary for the approach to be useful.

Lu: I think there's also room to improve the FNO itself. They used a relatively simple architecture. More sophisticated neural operators — or even training on larger datasets with more diverse physics — could improve both speed and accuracy.

Jane: And there's another direction: extending to three dee. This paper is 2D, but real-world applications are three dee. The computational cost of three dee is dramatically higher, so the speedup from the hybrid approach would be even more valuable there.

Tom: That's a great point. The paper is a proof of concept in 2D, but the methodology should transfer. The FNO doesn't care whether the input is a 2D or three dee image — it just needs enough training data.

Meng: And that's the catch. Generating training data in three dee would require running the expensive three dee simulations first. But you'd only need to do that once to train the model, and then you could use it for many simulations.

Lu: There's also the question of whether the model could be trained to handle different gas mixtures or pressures. The current model is trained for air at atmospheric pressure. But if you could train a model that takes pressure as an additional input, you'd have a much more general tool.

Tom: So the improvements are both incremental — better architectures, more training data — and fundamental — extending to three dee, adding more physics. Lalam, what do you see as the most impactful direction?

Lalam: I think the most impactful vision is using this hybrid approach to build a real-time simulation capability. Imagine a system that can predict microwave breakdown as it's happening, in real time, for control or safety applications. With a sixty-fold speedup, you're getting close to that. And with further optimization, it could become truly real-time.

Jane: That's a compelling vision. Real-time prediction of plasma behavior could be used in anything from protecting aircraft electronics from electromagnetic interference to optimizing plasma-assisted combustion in engines.

Tom: And that's the kind of impact that goes beyond academic research. We'll wrap up with our final thoughts in just a moment.

Conclusion: Tom: And we're back for our final segment on "Hybrid Fourier Neural Operator–Plasma Fluid Model for Fast and Accurate Multiscale Simulations of High Power Microwave Breakdown." Jane, let's pull it all together.

Jane: Let's do it, Tom. This paper tackles a fundamental problem in plasma physics: simulating high power microwave breakdown is computationally brutal. The traditional approach — solving Maxwell's equations with FDTD — eats up ninety-nine percent of the compute time. The authors replaced that solver with a Fourier Neural Operator trained on simulation data, and kept the plasma continuity solver as the physics backbone.

Tom: And the results speak for themselves. Sixty-fold speedup on average, with accuracy that's essentially indistinguishable from the full physics simulation. Streamer shapes, growth rates, temporal evolution — all matching perfectly across unseen electric field values.

Meng: I think the engineering contribution is just as important as the physics. The Python-C integration strategy means this can be adopted by groups with existing legacy code. You don't need to rewrite everything from scratch.

Lu: And the scientific contribution is showing that hybrid data-driven and physics-based modeling works for multiscale problems. This isn't just about microwaves and plasma. It's a template for tackling any problem where you have fast dynamics coupled with slow dynamics.

Jane: That's a great way to frame it, Lu. The multiscale challenge is everywhere — from plasma physics to climate modeling to materials science. Anywhere you have processes operating on wildly different timescales, this hybrid approach could help.

Tom: And Lalam, you had a vision about democratizing access. Want to close us out?

Lalam: I do, Tom. The sixty-fold speedup means that simulations which required supercomputing resources are now feasible on a good desktop workstation. That opens the door for smaller research groups, for universities in developing countries, for students who want to explore this physics. It lowers the barrier to entry for an entire field.

Tom: That's a beautiful note to end on. So let's say goodbye to "Hybrid Fourier Neural Operator–Plasma Fluid Model for Fast and Accurate Multiscale Simulations of High Power Microwave Breakdown." A paper that's fast, accurate, and opens doors.

Jane: And we'll be back next time with another paper to dissect. Until then, keep asking questions and keep exploring. Thanks for listening, everyone.

Tom: Take care, folks. We'll see you on the next episode.

Kalp Pandya, Pratik Ghosh, Ajeya Mandikal, Shivam Gandha, Bhaskar Chaudhury

Group in Computational Science and HPC, DA-IICT, Dhirubhai Ambani University · Smart Energy Learning Center, DAU

physics.plasm-ph, cs.AI, cs.LG, physics.comp-ph

Submitted: 2025-09-06

Updated: 2026-08-18

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 57/100

Key concepts

Fourier Neural Operator: A type of machine learning model designed to learn how electromagnetic fields behave without solving the full physics equations every time. It learns the mapping between plasma density and resulting electric field.
Hybrid Approach: Combining a traditional plasma solver with a neural network. The traditional solver handles the electron density evolution, while the neural network replaces the slow, expensive electromagnetic field solver to speed up computation.
Multiscale Simulation Bottleneck: The difficulty in simulating high power microwave breakdown because it involves phenomena happening at very fast picosecond scales and slow plasma evolution over nanoseconds. The traditional method is slowed by having to resolve the fastest timescale.
SSIM and Percent Error: Metrics used to measure accuracy. The paper achieved an average Structural Similarity Index of 0.9999, indicating a near-perfect visual match, with an average percent error of just 0.12.

Terminology

Accuracy — 27× in this paper · explained in 1 episode
Fourier Neural Operator — 19× in this paper
High Power Microwave — 15× in this paper
Hybrid Approach — 12× in this paper · explained in 4 episodes
Fidelity — 5× in this paper · explained in 2 episodes
Finite-Difference Time-Domain — 3× in this paper
Generalization — 3× in this paper · explained in 13 episodes
Loss Function — 3× in this paper · explained in 2 episodes
Adam — 2× in this paper · explained in 2 episodes
Latent Space — 2× in this paper · explained in 7 episodes
Modular Architecture — 2× in this paper · explained in 4 episodes
Alignment — 1× in this paper · explained in 1 episode
Benchmark — 1× in this paper · explained in 2 episodes
Benchmarking — 1× in this paper · explained in 3 episodes
Computational overhead — 1× in this paper · explained in 2 episodes
Decoupling — 1× in this paper · explained in 2 episodes
Fast Fourier Transform — 1× in this paper
Latent Representation — 1× in this paper · explained in 2 episodes
Physics-Informed Neural Networks (PINNs) — 1× in this paper · explained in 6 episodes
Robustness — 1× in this paper · explained in 5 episodes

Summary

Summary

This paper presents a hybrid computational framework that integrates a Fourier Neural Operator (FNO) with a plasma fluid model to accelerate simulations of High Power Microwave (HPM) breakdown. The authors state: In this work, we present a hybrid modeling approach that combines the accuracy of a differential equation-based plasma fluid solver with the computational efficiency of FNO (Fourier Neural Operator) based EM solver.

The physical model of HPM breakdown is governed by two coupled phenomena: the interaction of EM wave with plasma described by Maxwell’s equations (Eqs. 1,2 in Fig.1; Block A- EM Solver) and plasma dynamics governed by the plasma continuity equation (Eq. 5 in Fig. 1; Block B - Plasma Solver). The coupling occurs through the electron current density (J). The plasma is assumed quasi-neutral, and ion contribution to current density is neglected. The electron continuity equation is averaged over one EM wave cycle, retaining only the diffusive term in the flux divergence, justified by the significant separation of time scales, as the plasma density evolves much more slowly compared to the period of the EM wave. The model uses an effective field approximation, where electron transport properties depend on the local effective field E eff, given by E eff = sqrt E rms squared / (1 + omega squared / nu m 2). The electron temperature is assumed constant at 2 eV.

The conventional numerical implementation uses a 2D FDTD method based on the scattered field formulation. The computational domain is L x = 1 lambda and L y = 0.5 lambda, discretized with 512 × 256 grid points, satisfying a spatial resolution of N lambda = 500 grid points per wavelength. Two identical linearly polarized EM waves at 110 GHz with amplitude E 0 = 2.5 times 10 6 V/m are injected from top and bottom boundaries. A 2D Gaussian plasma density profile is initialized at the center with peak density n 0 = 10 15 m-3. The simulation uses a multirate approach: "Maxwell’s equations are solved over one EM wave cycle (represented by T M) using the plasma density evaluated at the start of the cycle. The plasma density is then updated for the subsequent cycle using transport coefficients that depend on the RMS field (E rms) obtained from the previous cycle." The simulation terminates when the plasma streamer spans 80% of the domain width.

The authors note the computational bottleneck: the EM solver consumes approximately more than 99% of the overall simulation time, whereas the plasma solver accounts for less than 1%. A single 2D simulation requires approximately five days on a modern desktop, and scaling to a 10λ × 10λ domain would require close to 500 days.

The hybrid framework replaces the EM solver with an FNO surrogate. The authors state: "the FNO is employed as a DL surrogate model that replaces the conventional EM solver (Block A in Figure 1). To account for the EM-plasma interaction, the FNO surrogate is trained to predict the scattered electric field (E rms) using the plasma density and incident field as inputs." The FNO is applied as an image-to-image reconstruction model, taking a two-channel input (plasma density and incident electric field) and predicting the scattered electric field.

The implementation uses a Python controller with C-based subroutines compiled into shared object files, loaded via the ctypes library. This approach "preserves the performance benefits of low-level numerical routines already implemented in C, enables seamless integration with Python-based ML components such as the FNO model, and ensures minimal overhead when sharing data between the two languages."

The FNO architecture consists of a lifting layer (1×1 convolution expanding input to 32 channels), four Fourier layers (each with 16 frequency modes and zero-padding of size five), and a projection layer reducing to a single-channel output. Each Fourier layer performs a 2D FFT, retains the lowest 16 modes in each direction, applies a learnable complex-valued weight matrix, then applies an inverse FFT. A pointwise convolution is added as a bias term to reintroduce spatial nonlinearity.

Training data was generated using the FDTD-based solver for incident electric fields of 2.5, 2.6, 2.7, 2.8, 2.9, and 3.0 MV/m. The training dataset contains 36,864 input-output pairs, validation contains 8,922 pairs, and the test dataset contains 37,530 pairs. Test cases use unseen electric field values of 2.55, 2.65, 2.75, 2.85, and 2.95 MV/m. Data was normalized using min-max normalization and rescaled to 0-255 range. The model was trained using the Adam optimizer with learning rate 10-3, batch size 32, for 300 epochs with early stopping, using MSE loss.

The FNO model achieved an average MSE of 0.0394, an average APE of 0.12, and an average SSIM of 0.9999 on the test dataset, showing excellent agreement with FDTD results.

For the hybrid model validation, the authors compared key parameters: normalized streamer length L s/lambda, streamer growth rate, and scattered E rms at the streamer tip. The results show the hybrid solver predictions exhibit an almost perfect overlap with those from the FDTD-based plasma fluid model across all tested cases. The hybrid model reproduced streamer shape, growth rate, and temporal evolution due to diffusion-ionization mechanism with near-perfect overlap compared to the conventional full-physics DE based EM-plasma fluid solver.

The computational speedup achieved was significant: the hybrid approach achieved speedup factors between 55× to 60×, reducing multi-day simulations to under few hours without loss of accuracy. Specifically, speedups ranged from 57.10× to 62.77× across different electric field values, with FDTD simulation times ranging from 991.28 to 2131.67 minutes reduced to 16.98 to 37.34 minutes using the hybrid approach. The authors note this enables simulations of larger computational domains (e.g., 10λ×10λ) and longer physical timescales that would otherwise be computationally prohibitive with conventional methods.

The authors conclude: "This work establishes a generalizable and scalable hybrid modeling framework that combines the accuracy of physics-based solvers with the computational efficiency of machine learning surrogates for advancing plasma science and engineering. They also note this is the first reported application of such hybridization in the context of plasma fluid modeling and simulation of HPM breakdown."

Improvements for AI systems

Based on the scientific paper, here are the specific improvements I can make to AI systems and what the improved system can do:

Improvement: Replace the pure FNO-based EM solver with a hybrid architecture that couples the FNO surrogate with the physics-based plasma continuity solver, rather than using a single end-to-end neural network.
What it can do: Achieve 57–63× speedup over full FDTD simulations while maintaining near-perfect accuracy (SSIM ≈ 0.9999, APE ≈ 0.12) for unseen electric field amplitudes, enabling multi-day simulations to complete in under 40 minutes.
Improvement: Encode both plasma density and incident electric field as separate input channels to the FNO, rather than using a single scalar field.
What it can do: Generalize across a continuous range of incident field amplitudes (2.5–3.0 MV/m) without retraining, correctly predicting streamer formation, growth rate, and tip field enhancement for entirely unseen field values (e.g., 2.55, 2.75, 2.95 MV/m).
Improvement: Use Fourier layers with 16 frequency modes plus a pointwise convolution bias term to capture both global EM-plasma coupling and localized streamer tip features.
What it can do: Resolve sharp plasma density gradients (on the order of micrometers) and steep electric field enhancements at streamer tips, which traditional convolutional networks fail to capture due to their local receptive fields.
Improvement: Train using MSE loss while simultaneously tracking SSIM and APE metrics, and implement early stopping based on validation loss stagnation.
What it can do: Ensure both pixel-wise accuracy and structural fidelity of predicted scattered fields, preventing small errors from accumulating into large deviations in long-duration simulations (100+ ns physical time).
Improvement: Implement a modular pipeline where the C-based plasma solver is compiled into a shared library (.so) and invoked from Python via ctypes, while the FNO runs in PyTorch.
What it can do: Seamlessly integrate legacy high-performance C/Fortran simulation codes with modern ML frameworks, reducing restructuring overhead and enabling adoption in existing plasma simulation workflows.

Predict microwave streamer formation, elongation velocity (tens of km/s), and tip field enhancement in real-time for arbitrary incident field amplitudes within the trained range, enabling interactive design of HPM protection devices and experiments.
Simulate 10λ × 10λ domains (which would take 500 days with FDTD) in under a week, capturing both picosecond EM oscillations and microsecond plasma evolution without sacrificing resolution (λ/500 grid spacing).
Rapidly sweep incident field amplitudes, pressures, and frequencies to map breakdown thresholds and streamer dynamics, enabling optimization of plasma-assisted combustion, propulsion, and electromagnetic shielding systems.
Provide confidence bounds on predicted streamer length and growth rate by leveraging the APE and SSIM metrics, flagging cases where the surrogate may deviate from physics (e.g., near critical field thresholds).
Adapt the hybrid framework to other coupled EM-plasma phenomena (e.g., microwave breakdown in waveguides, plasma-based power limiters, and laser-plasma interactions) by retraining the FNO on new datasets while reusing the same Python-C integration and spectral-domain architecture.
Run on a single GPU (NVIDIA RTX 6000 Ada) with 300 epochs of training, requiring only 37,000 training samples, making it feasible for research groups with limited HPC resources to conduct high-fidelity HPM breakdown simulations.

Sources

Fourier Neural Operator for Parametric Partial Differential Equations