Source-linked AI summary
OptiXDE: A fast optical-inspired solver for differential equations
Yang Yang, Mingjiao Yan, Zongliang Zhang
TL;DR
Differential-equation solvers must handle costly algebraic systems, stability constraints and complex geometries while retaining reliable computation. OptiXDE addresses this with a deterministic matrix-free composition of transform-diagonal propagation, physical-space interactions and constraint operators. Across transform-compatible, nonlinear and embedded-domain problems, the framework establishes distinct accuracy and performance regimes governed by geometry, nonlinear coupling and hardware scale.
Problem
Conventional differential-equation methods face costly algebraic infrastructure, stability-limited time increments and geometry-related meshing or refinement requirements.
Method
OptiXDE composes transform-domain differentiation and analytical modal propagation or inversion with physical-space corrections and constraint operators in a common matrix-free architecture.
Results
Across linear, nonlinear and incompressible-flow benchmarks, accuracy is near floating-point limits for transform-compatible problems, while geometry, nonlinear coupling and hardware scale determine remaining error and efficiency.
Takeaways & Limitations
Transform-diagonal spectral propagation serves as a reusable deterministic backbone for structured and embedded-domain differential-equation solving.
Takeaways & Limitations
OptiXDE is most effective when a dominant operator component is constant coefficient and diagonalizable; strongly heterogeneous, discontinuous, shock-dominated or highly localized problems may require hybrid corrections.
Abstract
from arXiv · showhide
OptiXDE is a matrix-free spectral operator framework for differential equations on uniform grids and embedded domains. Inspired by angular-spectrum propagation in Fourier optics, it maps transform-diagonal spatial operators to analytical modal multipliers and composes them with physical-space operators for nonlinearities, geometry and boundary enforcement. A common transform--operator--inverse-transform backbone is demonstrated across transient diffusion, periodic and embedded-domain Poisson problems, the cubic nonlinear Schr"odinger equation, viscous Burgers dynamics, the two-dimensional Allen--Cahn equation and incompressible flows from the Taylor--Green vortex to embedded-cylinder vortex shedding. Transform-compatible linear problems are recovered near the floating-point limit, whereas errors on the singular L-shaped domain remain localized near the re-entrant corner and regularized interface. Nonlinear benchmarks recover second-order temporal convergence and the expected conservative or dissipative behavior, while incompressibility remains near round-off level during long-time vortex shedding. The matrix-free updates require \(\mathcal{O}(N\log N)\) work and \(\mathcal{O}(N)\) memory. Device-resident transform workloads reach \(94.9\times\) GPU acceleration, and the complete embedded-cylinder solver achieves a \(42.1\times\) CPU--GPU speedup under matched numerical settings. These results establish OptiXDE as a deterministic and extensible operator-centric framework for structured and embedded-domain differential equations.
Introduction
OptiXDE organizes established spectral ingredients into a common matrix-free framework for differential equations, addressing computational costs associated with large systems, stability-limited stepping and complex geometries. Its transform–operator–inverse-transform architecture combines analytical modal treatment with physical-space corrections across diverse linear, nonlinear and embedded-domain problems.
- Motivation: Conventional FEM, FVM and FDM simulations can be dominated by algebraic-system infrastructure, stability-limited time increments or geometry-related meshing and refinement requirements.These costs become especially relevant for high-resolution and complex-geometry simulations.
- Related spectral methods: O(N log N) FFT-based spectral methods provide fast alternatives for global differential operators and support applications including Poisson, wave, quantum-dynamical and accelerator-based simulations.Analytical or semi-analytical modal propagation is also established in exponential time-differencing and integrating-factor schemes.
- Related spectral methods: Existing spectral methods already address non-periodic boundaries and complex geometries through embedded-boundary, Fourier-continuation and immersed-boundary smooth-extension formulations.These approaches retain transform grids while extending spectral calculations to irregular or general domains.
- Related approaches: Neural operator approaches offer rapid inference after training, but their predictive reliability depends on training-data coverage, optimization and architecture choices.This motivates distinguishing OptiXDE’s deterministic operator organization from learned mappings between function spaces.
- Contribution: OptiXDE unifies transform-domain differentiation, analytical modal propagation or inversion, physical-space correction and computational backends within a common matrix-free architecture.The framework treats nonlinearities, heterogeneous coefficients, embedded geometries and boundary corrections as compositions with a spectral core.
- Evaluation: OptiXDE is evaluated across parabolic, elliptic, dispersive, nonlinear and incompressible-flow problems, including transform-compatible benchmarks, singular embedded geometries and long-time vortex shedding.The study separates spectral, temporal-splitting, nonlinear-aliasing, finite-precision and embedded-boundary errors while examining complexity and CPU/GPU performance.
Results
Across linear, embedded-domain, nonlinear and incompressible-flow benchmarks, OptiXDE combines transform-based propagation with physical-space enforcement while preserving accuracy, physical structure and scalable acceleration.
- Linear spectral operators: Transform-compatible diffusion and Poisson problems recover round-off-limited accuracy through analytical modal propagation and direct inverse operators.Diffusion errors are primarily limited by floating-point accumulation under repeated transforms, while Poisson solution errors remain near the floating-point floor.
- Embedded-domain accuracy: Embedded L-shaped-domain errors remain localized near the regularized interface and re-entrant corner while decreasing systematically with spatial refinement.At 768^2, the maximum pointwise discrepancy is approximately 2.4 × 10^-4 near the inner vertical boundary, and fixed interface regularization prevents interpreting the trend as sharp-interface asymptotic convergence.
- Nonlinear dynamics: The focusing nonlinear Schrödinger benchmark preserves conservative dynamics, with relative mass drift between 5.33×10^-14 and 8.49 × 10^-13 and diminishing Hamiltonian excursion under refinement.Removing the upper spectral band changes the resolved trajectory by only 2.46 × 10^-10, below the discretization error.
- Nonlinear dynamics: Viscous Burgers and Allen–Cahn benchmarks retain their expected dissipative structures while achieving approximately second-order temporal convergence.Allen–Cahn free energy decreases monotonically for ∆t ≤ 0.5, whereas an energy increase appears at ∆t = 1 despite bounded fields.
- Incompressible flow and embedded vortex shedding: Incompressibility remains near round-off level during long-time embedded-cylinder vortex shedding, with maxt ∥∇·u∥∞ = 9.79×10^-14.Parameter variations show that mask diffuseness affects hydrodynamic quantities: broadening the transition to 1.5h increases mean drag by 4.30%.
- Workload-dependent hardware acceleration: GPU acceleration increases with workload, reaching 94.9× for device-resident diffusion and 42.1× for the complete matched embedded-cylinder solver.A one-dimensional Burgers workload at N = 256 achieves only 0.30× CPU performance, whereas the 1600 × 800 cylinder calculation reaches 42.1×.
Discussion
OptiXDE functions as a reusable deterministic backbone by combining transform-diagonal propagation with physical-space interactions and constraint operators. Its accuracy and efficiency depend on geometry, nonlinear coupling, workload size, hardware scale, and operator structure.
- Linear transform-compatible problems are round-off limited, whereas irregular geometries shift dominant errors toward geometric representation and boundary treatment.The L-shaped results retain interior accuracy but localize discrepancies near the re-entrant singularity and regularized embedded interface.
- Nonlinear and incompressible-flow calculations require composition with physical-space evolution, de-aliasing, geometry enforcement, and incompressibility projection.The framework therefore serves as a computational backbone rather than a complete solver in isolation.
- OptiXDE’s reusable element is the decomposition of differential equations into transform-diagonal propagation, physical-space interactions, and constraint operators.This organization supports nonlinear and incompressible-flow solvers without relying on one propagation formula.
- GPU acceleration is workload dependent: small problems may favor CPUs, while sufficiently large two-dimensional workloads benefit from device-resident transform computation.Transform size, dispatch overhead, transform count, and host–device communication influence realized acceleration beyond the O(N log N) spectral-core complexity.
- OptiXDE is most effective when a dominant operator component has constant coefficients and is diagonalizable under Fourier, sine, or cosine representations.Strong heterogeneity, discontinuities, shocks, and localized geometric features may require iterative correction, splitting, regularization, or hybrid global–local discretizations.
- A proposed next step is a programmable PDE system that translates mathematical descriptions into executable operator compositions while retaining deterministic, interpretable, matrix-free structure.Automatic transform-basis selection, null-mode treatment, and de-aliasing could reduce equation-specific implementation.
Methods
OptiXDE separates transform-diagonal bulk propagation from physical-space operations for nonlinearities, geometry, boundaries and constraints within a common matrix-free architecture. Transform-compatible transient propagation and steady inversion use analytical modal operations, while embedded and nonlinear problems add composed physical-space updates.
- Operator architecture: OptiXDE separates transform-diagonal bulk operators from physical-space operators for nonlinearities, geometry, boundary conditions and constraint projection.The framework organizes propagation, steady inversion, enforcement and projection without assembling global differential, stiffness or mass matrices.
- Spectral propagation: A basis that diagonalizes a constant-coefficient operator converts the equation into independent modal equations governed by the spectral symbol λ(k).Spatial derivatives become analytical multipliers, enabling modewise evolution in transform space.
- Spectral propagation: Exact variation-of-constants propagation handles constant-coefficient linear dynamics, while source-integral approximation controls additional temporal accuracy for time-dependent sources.The same framework covers transient propagation and steady inversion, including solvability and normalization treatment for null modes.
- Boundary treatment: Fourier, sine and cosine transforms are selected according to separable boundary conditions, with lifting fields reducing nonhomogeneous data to homogeneous form.Periodic, homogeneous Dirichlet and homogeneous Neumann conditions are transform-compatible; irregular or embedded boundaries require physical-space enforcement.
- Boundary treatment: Embedded geometries use signed-distance masks, regularized interfaces and localized residual corrections while retaining the transform-diagonal bulk operator.Boundary maps are composed across condition types, and steady embedded problems repeat propagation and enforcement as pseudo-time or fixed-point iterations.
- Nonlinear and flow solvers: Nonlinear terms are evaluated pseudo-spectrally in physical space, with de-aliasing applied when multiplication generates unresolved frequencies.Strang composition combines closed-form linear flows with nonlinear subflows; incompressible embedded flows additionally project corrected velocities onto the divergence-free Fourier subspace.
- Computational cost: O(N log N) work and O(N) memory characterize fixed-stage matrix-free updates, with direct steady spectral inversion requiring O(N log N) complexity.Transient and embedded costs scale as O(NtqN log N) and O(nitqN log N), respectively, while memory stores arrays, workspaces, multipliers and masks.