Source-linked AI summary
Introducing SINFONIA: Symplectic, slimplectic and Magnusian (Neural) Flows for Orbital Numerical Integration and Acceleration
Lidia J. Gomes Da Silva
TL;DR
Long-duration inspiral modelling needs numerical maps that resolve fast orbital motion while preventing small errors from becoming secular phase drift. The paper constructs structure-preserving neural flow maps, tests them through chained inspirals, and finds that a constrained secular channel governs long-time accuracy while also enabling recovery of hidden dissipative forces. Within the tested regime, this provides a proof of concept for fast repeated integration and physics inference.
Problem
Long-duration gravitational-wave modelling must control secular phase drift while resolving fast orbital motion and slow dissipative evolution across many compositions.
Method
The paper constructs symplectic, Taylor-anchored, and Magnusian neural flow maps with analytic structure, learned corrections, secular constraints, and network-off controls.
Results
Encoding the secular channel brings analytic and learned generators to the 10^-6-rad tier, while leaving it free recovers a hidden dynamical-friction-like force from chained phase.
Takeaways & Limitations
The results establish a proof of concept for structure-preserving learned evolution maps in fast long-duration integration and gravitational-wave source physics inference.
Takeaways & Limitations
The conclusions are limited to quasi-circular, planar, non-spinning, weak-field motion with a single secular channel and known balance structure.
Abstract
from arXiv · showhide
Long-duration gravitational-wave modelling must resolve fast orbital motion together with slow dissipative evolution while preventing small numerical errors from accumulating into secular phase drift. Here we ask whether the finite-time evolution map itself can be learned as an explicit, differentiable, structure-preserving object and then repeatedly composed through a complete inspiral. We construct three neural-flow architectures: a symplectic and slimplectic flow on Galley's doubled phase space, [SINFONIA-J0]; a Taylor-anchored flow, [SINFONIA-J1]; and a Magnusian flow that learns the finite-time dissipative correction in the interaction picture, [SINFONIA-J2]. Applied to a 2.5PN neutron-star inspiral, all three expose the same controlling mechanism: long-time accuracy is governed not by pointwise map error alone, but by its signed projection onto a single secular channel fixed by energy--angular-momentum balance. Encoding this structure allows the learned maps to remain accurate through $10^{2}$--$10^{5}$ window compositions to coalescence at timesteps of a full orbital period and beyond, reaching chained phase errors orders of magnitude below a benchmark slimplectic integrator at lower cost. The same secular structure can also be exploited for physics inference: when the channel is left unconstrained, the accumulated phase retains enough information to recover an un-modelled dynamical-friction-like force, both parametrically and as a learned function of separation. Network-off controls isolate the contribution of learning from the analytic structure already built into each map. These results establish a proof of concept for structure-preserving learned evolution maps as tools for fast long-duration integration and physics inference in gravitational-wave source modelling.
1. INTRODUCTION
The paper asks whether explicit, differentiable neural flow maps can preserve long-time accuracy when repeatedly composed through dissipative inspirals. It combines geometric structure, analytic flow components, and learned corrections to control secular phase error and support repeated calculations and force inference.
- Long-term binary modelling must resolve fast orbital motion and slow radiation-reaction evolution without confusing physical secular drift with numerical drift.
- Radiation reaction motivates non-Hamiltonian structure-preserving methods, including slimplectic doubled-variable formulations and generator-based finite-time descriptions.
- Repeated composition can amplify small structural errors, so a learned flow map must retain the geometry responsible for good long-time behaviour.
- The programme tests whether structure-preserving maps control secular error, amortize repeated-integration cost, isolate learned contributions, and recover an un-modelled dissipative force from chained phase.
- The study compares symplectic, slimplectic, and Magnusian neural-flow architectures by chaining their maps through an inspiral and scoring accumulated phase and secular dissipation against independent numerical solutions.
2. NON-CONSERVATIVE MECHANICS: DISCRETIZATION AND RE-ORGANIZATION
The paper reorganizes dissipative orbital dynamics through Galley’s doubled variational mechanics and finite-time flow maps. It then separates the dominant conservative motion from a small perturbation so analytic structure can carry most of the dynamics while learning targets the remaining correction.
- Galley’s doubled action: Galley’s doubled-variable formalism converts dissipative dynamics into an initial-value variational construction, with the physical sector selected only after variation.
- Galley’s doubled action: The doubled histories are auxiliary, and the physical limit imposes the corresponding physical-sector conditions on the doubled phase-space variables.
- Slimplectic discretization: Discretizing the doubled action yields slimplectic variational integrators whose physical maps are non-symplectic but retain long-time variational advantages for dissipative systems.
- Doubled-space structure: Canonical evolution on the doubled space does not imply symplectic or volume-preserving evolution after projection to the dissipative physical phase space.
- Interaction picture: The interaction picture factors the full map into an exact conservative flow and a near-identity perturbative map, leaving the learned object to represent the transported correction.
- Interaction picture: For the benchmark, the dissipative force is 10^4–10^5 times smaller than the conservative force, making this factorization important for exposing the correction that remains to be learned.
3. NEURAL FLOW MAPS
The paper distinguishes geometric, approximation-order, and analytic-anchor constraints, then adds a force-law secular constraint directly to the trained map. Its architectures target long-time phase accuracy by controlling the secular error channel rather than relying on pointwise or one-step accuracy alone.
- 3.1. What “structure-preserving” means for a learned map: Structure-preserving learned maps can impose geometric properties, approximation order, or an analytic baseline before training.These constraints differ in what they guarantee under repeated composition.
- 3.2. The secular channel: The principal constraint is applied after training: a force-law-derived scalar is imposed directly on the realized map output.The analytic dynamics remains important, because transferring it from the learned component to the analytic part improves the error.
- 3.2. The secular channel: Long-time phase error is mediated almost entirely through secular energy–angular-momentum drift, while direct orbital-angle defect is negligible at the tested accuracies.This identifies the signed secular projection as the quantity that must be controlled during repeated composition.
- 3.3. A symplectic and slimplectic neural flow: The plain J0 map has zeroth-order consistency, with one-window error growing linearly and an effective vector-field error of approximately 5 × 10^-3.Its dominant Kepler-motion error exceeds the dissipative scale by about 300 times initially and remains about 20 times larger near the scored interval’s end.
- 3.4. A Taylor-anchored neural flow: J1 confines learning to O(h^6) while shell projection removes its secular projection, reducing chained phase error from 10^-3–10^-1 rad without projection to 10^-6 with projection.The projection-only control reaches 1.35 × 10^-4, and the learned contribution is bounded by analytic maps of the same family.
- 3.5. The Magnusian neural flow: The Magnusian J2 construction supplies the conservative flow and first-order generator analytically, while learning only the higher-order dissipative correction.Its generator also captures a second-order periapsis direction that vanishes at first order; the apparent small-e singularity belongs to the orbital chart, not the dynamics.
4. EXAMPLE: GRAVITATIONAL RADIATION REACTION AND A HIDDEN ENVIRONMENT
The 2.5PN inspiral tests show that long-time phase accuracy is controlled primarily by the signed secular projection of map error, while analytic structure and learned residuals determine cost and transfer. Enforcing this channel structurally improves chained accuracy across architectures, but transfer and inference remain scope-dependent.
- Physical system: Two 1.4 M⊙ neutron stars with leading-order 2.5PN radiation reaction provide the controlled inspiral testbed.The initial separation is r0 = 100 M with symmetric mass-ratio ν = 1/4.
- Secular control: 0.81 rad falls to 8.8 × 10−5 rad when shell projection is applied to the learned Magnusian without retraining.The intervention reduces the three-checkpoint band by 366× to 7×, showing the importance of the signed secular channel.
- Secular control: Below 10−7 rad, the projected order-7 Taylor tower improves from 6.6 × 10−3 rad, with accuracy limited by the reference solution.The same projection reduces the analytic first-order Magnusian from 4.1 × 10−2 rad to 1.7 × 10−7 rad.
- Controlling mechanism: The chained phase error is ordered by the radially integrated signed energy-rate defect rather than pointwise generator error.Across nine J2 generators, the best chained checkpoint has the largest pointwise error, while shell projection aligns the secular defect with long-time accuracy.
- One-orbit deployment: 1.0 × 10−2 rad is achieved by the secular-aware one-orbit network, below Slim4’s 2.6 × 10−2 rad while advancing fifty times farther per map application.The best seed reaches 1.6 × 10−3 rad, and three of five seeds improve over the analytic first-order map.
- Transfer: Transfer at r0 = 200 M is partial: plain networks improve over the analytic map, whereas secular-tuned networks are 3–6× worse without retraining.Short-window transfer is stronger: at 200 M, projected analytic and learned arms reach a common 4.7 × 10−5-rad floor, while the tested range is not a general guarantee.
- Window dependence: At four orbits per window, the learned map reaches 0.126 rad in 379 windows, 56× below its network-off control but about five times above Slim4.At two orbits, the median gain is 13×; at one orbit and below, the same learned correction is less consistently beneficial.
- Analytic controls: A network-free second-order Magnusian gives 1.5 × 10−2, 7.7 × 10−2, and 0.17 rad at one, two, and four orbits, respectively.These errors are factors of 1.9, 5.7, and 41 below the corresponding first-order exponentiated map.
5. DISCUSSION
SINFONIA combines structure-preserving finite-flow maps with learned residual generators and a secular channel tailored to dissipative orbital dynamics. Its results show when learning helps, how the channel supports inference, and why the conclusions remain restricted to controlled orbital regimes.
- Three learned flow maps were chained through a complete radiation-reaction inspiral, with network-off controls isolating learned contributions from analytic structure.
- 10^-6-rad chained error is reached when the single signed secular channel is imposed structurally, while pointwise training objectives remain blind to that channel.
- 2.8× median improvement is obtained by a learned Magnusian generator over the closed form where it was trained, while transfer across separation remains bounded.
- Leaving the secular channel free enables recovery of a dynamical-friction-like force through a two-parameter rate or a learned function of separation, with the vacuum fit returning zero.
- Learning does not reliably improve the sufficiently accurate first-order Magnusian at the benchmark step; it becomes useful as longer windows resolve higher-order generator content.
- The conclusions are limited to quasi-circular, planar, non-spinning, weak-field motion with one secular channel and no transient multi-frequency resonance.