Neural Modal Decomposition: Architectural Priors from Observables
summary
The gist
Neural Modal Decomposition: Architectural Priors from Observables proposes a neural framework that learns multi-port pole-residue decomposition end-to-end from system-level observables without…
In short
The framework learns how to decompose multi-port system responses into intrinsic resonant modes and port-dependent couplings using a neural network trained only on system observables. This method establishes a general design principle for RF cavities and photonic devices by predicting complex poles end-to-end, achieving zero-shot generalization to unseen port counts without needing explicit supervision of the underlying physical modes.
Key concepts
- Pole-Residue Form
- This is the mathematical structure describing how a system's response matrix is built. Each entry in this matrix is calculated by summing contributions from a small number of intrinsic resonant modes, defined by their poles and residues. This algebraic form applies universally to RF, photonic, and quantum systems.
- Port-Independent Pole Predictor
- This module of the neural network learns to predict the complex poles ($ ext{p}_k$) that are common across all ports. These poles represent the intrinsic physical resonant modes of the system. They are learned solely from structural information (cavity tokens) and do not depend on how many ports are connected.
- Port-Dependent Coupling Predictors
- These modules learn to predict the coupling coefficients ($ ext{l}_{k,i}$ and $ ext{r}_{k,j}$) that describe how each specific port interacts with the predicted intrinsic modes. They take the predicted poles and port locations as input to output these coupling entries, separating intrinsic physics from geometry.
- Zero-Shot N-Port Generalization
- The model can accurately predict responses for a number of ports (N) that it was never trained on, provided the underlying physical structure is similar. This is achieved because the network learns the fundamental decomposition structure rather than memorizing specific port counts, overcoming scaling barriers.
Terminology used across episodes
This episode discusses
The paper
Neural Modal Decomposition: Architectural Priors from Observables · Read on arXiv
Juho Park, Kaushik Sengupta
Princeton University
Many engineering building blocks behave as multi-port linear time-invariant systems. RF cavities, photonic devices, and superconducting quantum chips, despite their different underlying physics, all share a common mathematical structure for their port-level response. Each entry of the response matrix is a sum of contributions from a small number of intrinsic resonant modes, the pole-residue form. A model capable of predicting such responses for arbitrary geometries and arbitrary port configurations, while simultaneously extracting the underlying eigenmode structure, would therefore establish a foundational design principle spanning all these domains. We propose a neural framework that learns this modal decomposition end-to-end, supervised only by system-level observables and without supervising the modal parameters themselves. The architecture decomposes into a port-independent pole predictor and two port-dependent coupling predictors whose outputs are combined entry-wise, separating intrinsic from port-dependent features. This factorization yields a single trained model that generalizes to port counts unseen during training, dissolving the O(N 2) scaling barrier of direct regression. Despite no modal supervision, the freely-parameterized poles converge to physically meaningful eigenmodes, verified by cross-validation against the AAA rational approximation algorithm. We instantiate the framework in radio-frequency electromagnetic surrogate modeling. A model trained only on 2-port data accurately predicts N-port responses unseen during training.
Transcript
Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.
Kai: I'm Kai.
Mira: And I'm Mira, here with Lev.
Kai: We're looking at the paper: "Neural Modal Decomposition".
Mira: Here's the gist of the paper. Neural Modal Decomposition: Architectural Priors from Observables proposes a neural framework that learns multi-port pole-residue decomposition end-to-end from system-level observables without supervising the modal parameters themselves, which is significant because it establishes a foundational design principle spanning engineering domains like RF cavities and photonic devices The model achieves zero-shot N-port generalization by decomposing the response into a port-independent pole predictor and two port-dependent coupling predictors, which dissolves the O(N 2) scaling barrier of direct regression The Core Mathematical Prior The fundamental structure exploited is the pole-residue form, where each entry of the response matrix is a sum over intrinsic modes with port-independent poles and port-dependent coupling coefficients This algebraic prior applies to multi-port LTI systems including RF, photonic, and quantum domains The decomposition is written as Hij (ω) = Xk lk,i rk,j jω − pk, which separates intrinsic from port-dependent features The Network Architecture The neural framework decomposes the computation into three modules: a CavityNet for structure encoding, a PoleNet to map tokens to complex poles, and an AmpNet for predicting per-port coupling entries The PoleNet maps cavity tokens to K complex poles, parameterized as pk = −γk + j ωk, where γk > 0 ensures stability The AmpNet uses a hybrid scheme encoding port locations to query the predicted poles and output the left (lk,i) and right (rk,j) coupling entries Training Objectives and Stabilization The total training loss combines observable-level reconstruction terms with a gauge-fix term L = LS + λZ LZ,log + λg Lgauge, where LS is a per-entry MAE on the scattering matrix and LZ,log is a log-magnitude error on the impedance matrix The factorization lk,irk,j makes the reconstruction loss invariant to a discrete sign symmetry of the residues A gauge-fixing term Lgauge is introduced to fix a canonical gauge by penalizing predicted modal amplitudes when queried ports lie on an anchor edge Generalization and Validation The framework demonstrates zero-shot N-port generalization, where a model trained only on 2-port data accurately predicts 3-, 4-, and 5-port responses unseen during training The freely-parameterized poles converge to physically meaningful eigenmodes, which are verified against the AAA rational approximation algorithm
15: This emergent agreement suggests that embedding the algebraic structure architecturally can substitute for explicit supervision of latent parameters The in-band physical poles selected by ranking based on domk and sZω∈band are precisely those extracted by AAA, indicating the architectural prior drives the network toward the correct solution manifold Broader Implications The framework extends beyond port-level observables to spatially resolved ones, suggesting that whenever an observable is a Green’s function, the same factorization applies verbatim This points to a more general reading of the same construction where coupling lk,i is the modal field evaluated at a location and the port index is merely a discrete sampling of it The work contributes a surrogate modeling framework for multi-port linear time-invariant systems with primary applications in radio-frequency, photonic, and superconducting quantum hardware design Positive impacts include faster device design iteration and reduced computational cost of full-wave electromagnetic simulation Limitations and Future Directions The framework learns K independent poles, which can be larger than the number of physical poles extracted by AAA, suggesting that additional poles absorb the frequency-independent contribution of the pole-residue form The model targets LTI systems with discrete modal spectra, meaning continuous spectrum or strongly nonlinear regimes lie outside its scope The pole count K is a fixed hyperparameter rather than learned, leaving adaptive selection open for future work A Dataset Details The cavity dataset is generated by uniformly sampling binary metal patterns over the interior 16 × 16 region of the 18 × 18 pixel array, together with port pixel locations along the top and bottom edges The full simulation and dataset construction details follow
1: The training set comprises approximately 291,000 cavity samples Test sets for zero-shot port-count extension consist of 3-, 4-, and 5-port configurations on cavity patterns never seen at any port count during training, each 1000 samples respectively The dataset generation took around 1.5 days under parallel simulation instances ran in cluster equipped with Intel Ice Lake CPUs A Dataset Details The cavity dataset is generated by uniformly sampling binary metal patterns over the interior 16 × 16 region of the 18 × 18 pixel array, together with port pixel locations along the top and bottom edges The full simulation and dataset construction details follow
1: The training set comprises approximately 291,000 cavity samples Test sets for zero-shot port-count extension consist of 3-, 4-, and 5-port configurations on cavity patterns never seen at any port count during training, each 1000 samples respectively The dataset generation took around 1.
Paper summary: Kai: So we're looking at this paper called Neural Modal Decomposition: Architectural Priors from Observables, and basically the core idea is that many different engineering systems—RF cavities, photonic stuff, quantum chips—they all share the same math structure for how they respond at each port.
Mira: Right. Each entry in that response matrix is actually a sum of contributions from some small number of intrinsic resonant modes, and it's described by what we call the pole-residue form.
Lev: So, this paper claims you can build a neural framework that learns this modal decomposition end-to-end, using only system level observables as supervision and not supervising the modal parameters themselves. That seems pretty ambitious for something like that.
Kai: Exactly. The thesis is that a model doing this would be a foundational design principle across all these different engineering domains because it lets you predict responses for any geometry and any number of ports while figuring out the underlying eigenmode structure at the same time.
Mira: It really boils down to decomposing the computation into three parts: a port independent pole predictor, and two port dependent coupling predictors whose outputs are combined entry by entry, which separates what's intrinsic from what depends on the ports.
Lev: That factorization is key because it means you get a single trained model that can generalize to port counts you never saw during training, which really cuts down on the scaling barrier of direct regression, which is O of n squared.
Kai: And they achieved this by training it only on two port data and then showing that this model accurately predicts responses for three, four, and five ports that it had never seen before.
Mira: But what's interesting is that even though they didn't supervise the modal parameters directly, the poles themselves end up converging to physically meaningful eigenmodes when you let them be free parameters.
Lev: That convergence is validated by cross validation against the AAA rational approximation algorithm, and they found that these physical poles line up with what AAA extracts when we rank them based on certain criteria like domk and sZ omega in the band.
Kai: So it suggests that embedding this algebraic structure into the architecture can act as a substitute for explicitly supervising those latent parameters.
Mira: And they even extend this beyond just port level observables, suggesting that whenever you have an observable that's a Green's function, the same factorization applies to it verbatim.
Paper summary: Lev: That points towards a more general way of thinking about this construction where the coupling coefficient is like the modal field evaluated at a location and the port index is just discrete sampling of that location.
Kai: For me, I find this really interesting because it moves us from just looking at port responses to spatially resolved observables, which opens up new modeling possibilities for things like electromagnetic fields.
Mira: It definitely gives us a way to build surrogate models for multi-port linear time invariant systems with primary applications in radio frequency and photonic hardware design.
Lev: The limitation they mention is that the framework learns K independent poles, and K can be larger than the number of physical poles extracted by AAA, which suggests extra poles are just absorbing the frequency independent contribution of that pole residue form.
Kai: And they flag that this model targets LTI systems with discrete modal spectra, meaning continuous spectrum or strongly nonlinear regimes are outside its scope.
Mira: So it’s a very specific tool for linear systems where the number of modes is finite and well-defined, which is important to keep in mind when applying it elsewhere.
Lev: The dataset they used was quite large—about two hundred ninety-one thousand cavity samples generated by sampling binary metal patterns over an eighteen by eighteen pixel array.
Kai: It took about fifteen days to generate that data using parallel simulation instances on a cluster with Intel Ice Lake CPUs, which is a significant amount of work for creating the training set.
Mira: The structure of the training set includes port pixel locations along the top and bottom edges alongside sampling in the interior region.
Lev: And for testing that zero shot generalization, they used three, four, and five port configurations on patterns never seen before during training, with one thousand samples for each test case.
Kai: That's a solid demonstration of its zero shot N port generalization capability. It shows the model is learning the underlying physics structure rather than just memorizing specific input/output pairs.
Mira: The convergence towards physical poles when they are free parameters is a strong indicator that this architectural prior guides the network toward the correct solution manifold, even without direct supervision of those modal parameters.
Lev: I wonder how much overhead there is in running this framework on actual hardware compared to something like a full wave electromagnetic simulation, since that's what we actually use for design iterations.
Paper summary: Kai: That's a fair question. The paper focuses on the modeling aspect, so it doesn't really give us concrete figures on computational cost for the full inference run, but the goal is to speed up those simulations significantly.
Mira: Ultimately, this framework contributes a surrogate modeling approach for multi-port linear time invariant systems that aims to make device design iteration faster and reduce the computational cost of those full wave simulations.
Lev: So if you're looking at designing a new RF cavity or photonic chip, this paper suggests that instead of running every single simulation for every possible port configuration, you might train this decomposition model once and then use it to quickly predict the response across many configurations.
Kai: Exactly. It’s about using system observables to infer the fundamental modal structure so you don't have to brute force the full matrix calculation repeatedly for different port counts.
Mira: The implication is that we can start treating these complex multi-port systems with a shared mathematical language based on those pole-residue forms, which is a common thread across RF, photonics, and even quantum hardware.
Lev: It’s about establishing that foundational design principle you mentioned earlier—a way to model the response of these diverse building blocks consistently regardless of their specific physics.
Kai: That's what I find really compelling about the title: Architectural Priors from Observables. It suggests we don't need to know the exact physics upfront, just have a common mathematical structure to start learning from observables.
Mira: So, to wrap up, this paper introduces Neural Modal Decomposition: Architectural Priors from Observables which proposes a neural framework that learns the pole residue decomposition end-to-end solely from system level observables without supervising the modal parameters themselves.
Kai: The main claim is that by decomposing the computation into a port independent pole predictor and two port dependent coupling predictors, we get zero shot N port generalization, meaning it predicts responses for any number of ports unseen during training.
Lev: They show that even though they don't supervise the modal parameters directly, those free parameters converge to physically meaningful eigenmodes when validated against established methods like the AAA rational approximation algorithm.
Mira: It establishes a foundational design principle because it works across different engineering domains like RF cavities and photonic devices, showing that the underlying mathematical structure is shared.
Kai: It’s about using observable-based modeling to create compact surrogate models for multi port systems, which should lead to faster device design iteration and reduced computational cost in full wave simulations.
Conclusion: Kai: So we're wrapping up our look at Neural Modal Decomposition from arXiv, and what the title tells us is that this paper isn't just building a new model; it’s proposing a way to design these complex systems by focusing on their underlying structure first.
Mira: Right. It suggests that because all these different physical systems—RF cavities, photonic devices—share the same mathematical response structure, we can use those common rules to build something more general for any geometry or port number.
Lev: From an engineering standpoint, I see it as moving away from building a specific simulator for every single design and towards training one flexible model that can handle many different layouts efficiently.
Kai: And the authors are doing this by learning how to take the system level observables, like the measured scattering matrix, and break them down into these intrinsic modes and port dependent couplings without having to tell the AI what those modes actually look like beforehand.
Mira: That’s a huge part of it. They aren't supervising the modal parameters themselves; they’re letting the architecture guide them toward physically meaningful eigenmodes just by forcing the math to work.
Lev: I think that's why it’s interesting for hardware development. If we can train a model that learns this decomposition end-to-end from data, it could cut down the time needed to predict how a new chip will behave across many different connection setups dramatically.
Kai: It really boils down to using observable data as the starting point for understanding the physics of these devices, which is something we need when trying to push the limits of superconducting qubits or high-Q cavities.
Mira: So, it’s about establishing a foundational design principle based on the mathematics of pole residue decomposition that applies broadly across these different engineering domains.
Lev: What this means for us is that instead of running heavy full wave simulations for every single port configuration we want to test, we can use this framework to predict responses much faster and with less computational overhead.
Kai: It’s about using AI to find the underlying structure without needing a perfect theoretical map of every single mode in advance.
Mira: And that opens up new ways to model these systems where we don't have all the exact physics right away, just some measurements we can feed into the system.
More episodes
- 2611.14082-Two atoms with strong light-induced dipolar interactions
- 2611.14040-Characterization of BNLs 20 and 30 micrometer thick LGADs for the Electron Ion Collider
- 2611.13399-Lift-element-based dynamic network of vortex gust-airfoil interactions
- 2611.13476-Hopfion Superlattices in 3D Twisted Moir'e Light Fields
- 2611.13642-Investigation of the timing performance of double-sided p-type 3D silicon sensors produced by CNM
- 2611.13830-SiPM optical readout of a PICOSEC Micromegas precise timing detector
- 2611.14152-Where are the shortest-path nodes in spatial networks?
- 2611.14219-Sub-Ice: semi-automated mapping of basal channels from ice shelf surface elevation data
- 2611.16341-SUMNO: Spatially Unrolled Multigrid Neural Operator for Fluid Flow Surrogates
- 2611.16090-Multi-UTE Myelin Fraction Mapping: Repeatability and Cross-Field-Strength Reproducibility at 1.5T and 3T