Neural Operator Acceleration

Neural Operators

80-200% CPU savings through AI-enhanced Newton iterations and domain-decomposed PINNs

Scientific machine learning meets traditional numerics: neural operators dramatically accelerate Newton iterations while preserving accuracy, achieving 80-200% CPU time savings.

1. Overview

By integrating neural operators (FNO, DeepONet) with classical Newton-based nonlinear solvers, we’ve achieved substantial computational savings without sacrificing solution quality.

2. Key Achievement

80-200% CPU Savings

FNO-accelerated Newton loops
Domain-decomposed PINNs
Hybrid physics-ML workflows

3. Technical Innovations

3.1. Neural Operator Integration

  • Fourier Neural Operators (FNO): Learn solution operators in spectral space

  • Physics-Informed Neural Networks (PINNs): Enforce PDE constraints during training

  • DeepONet: Operator learning for parametric families of PDEs

  • Hybrid Coupling: Seamless integration with mesh-based discretizations

3.2. Performance Metrics

  • CPU Time Reduction: 80-200% savings depending on problem class

  • Accuracy Preservation: Maintains scientific computing standards

  • Memory Efficiency: Reduced solver iterations = lower memory footprint

  • Scalability: Domain-decomposed training for large problems

4. Hybrid Physics-ML Workflow

The integration strategy:

  1. Offline Training: Neural operators pre-trained on solution families

  2. Online Acceleration: Replace expensive Jacobian operations with learned approximations

  3. Error Control: Traditional solver verifies and corrects predictions

  4. Adaptive Strategy: Switch between neural and classical based on convergence metrics

5. Application Areas

Successfully demonstrated in:

  • Nonlinear PDEs: Navier-Stokes, elasticity, reaction-diffusion

  • Parametric Studies: Design space exploration with varying coefficients

  • Multi-Physics: Coupled thermal-fluid-structural problems

  • Inverse Problems: Parameter identification accelerated by surrogate models

6. Domain Decomposition + Neural Methods

Novel contribution: combining DD-based parallelism with neural operators:

  • Local Neural Models: Each subdomain trains specialized operators

  • Communication Reduction: Fewer interface exchanges per Newton step

  • Load Balancing: Adaptive work distribution based on neural convergence

  • Mesh Coupling: Neural predictions consistent with WP1 discretizations

7. Impact on Computational Science

This breakthrough enables:

  • Real-Time Simulation: Interactive exploration of complex phenomena

  • Design Optimization: Faster optimization loops with reduced simulation cost

  • Uncertainty Quantification: Affordable ensemble methods for UQ campaigns

  • Climate & Weather: Accelerated multi-scale atmospheric models

8. Related Work Packages

9. Software & Training

  • Open-source neural operator libraries (PyTorch, JAX)

  • Integration tutorials with Feel++, FreeFEM

  • Training workshops: March 2026 (planned)

  • Reproducible Jupyter notebooks

Upcoming: March 2026 training workshop on neural operators for HPC, co-organized with CoE Hidalgo2.

← Previous: Exascale Scale | Back to Results | Next: Urban Meshing →