Computational Multiscale Methods

Björn Engquist, Daniel Peterseim, Barbara Verfürth, Yunan YangOberwolfach 2025 · DOI

Abstract

Many scientific and engineering problems exhibit complex interactions over a wide range of inseparable scales in space and time. Direct numerical simulations to solve such multiscale problems are often beyond current computational capabilities. The difficulties are exacerbated by the presence of uncertainty, randomness, and disorder and are hardly manageable for multiscale inverse problems. Therefore, the simulation of novel phenomena using multiscale models requires a new generation of multiscale computational methods. These must account for under-resolved scales, cross-scale couplings, and stochasticity in a hierarchical and adaptive manner and be able to integrate probabilistic, data-driven, and machine learning approaches. The workshop enhanced the development of a new generation of efficient multiscale computational methods and their rigorous mathematical and numerical analysis so that reliable and fast simulations of challenging multiscale problems from applications eventually become a reality.

Introduction

Scientific and engineering disciplines frequently encounter systems characterized by complex interactions across a vast range of inseparable spatial and temporal scales. From the microscopic pores in geological formations to the macroscopic flow of groundwater, or the intricate lattice structures in composite materials, the behavior of these systems is dictated by fine-scale details that are computationally prohibitive to resolve directly. Direct numerical simulations (DNS) often require trillions of variables, exceeding the capacity of even the most advanced supercomputing clusters.

Historically, researchers relied on homogenization—a process of averaging microscopic properties to derive effective macroscopic equations. However, classical homogenization often assumes scale separation and periodicity, conditions rarely met in disordered, random, or high-contrast real-world materials. The 2025 Oberwolfach Workshop on Computational Multiscale Methods highlighted a paradigm shift toward "super-localized" numerical homogenization and the integration of machine learning and quantum computing to address these challenges.

Methodological Innovations

The workshop showcased a "new generation" of algorithms designed to bridge the gap between micro-scale physics and macro-scale simulation:

Super-Localized Numerical Homogenization

A significant development is the refinement of Localized Orthogonal Decomposition (LOD) methods. By constructing basis functions that exhibit "super-exponential" decay, researchers can now compute local solutions on extremely small patches without sacrificing global accuracy. This localization is critical for handling materials where material properties vary significantly over short distances without a clear repeating pattern.

Operator Learning and Kernel Methods

While neural networks are popular, the workshop emphasized the efficacy of data-efficient kernel methods for learning solution operators. These methods have demonstrated the ability to "learn" the underlying differential equations with higher accuracy and less training data than standard deep learning architectures, providing a more robust framework for multiscale inverse problems.

Quantum Multilevel Preconditioning

Quantum algorithms are emerging as a potential solution to the "curse of dimensionality." New quantum gradient descent algorithms for Gaussian Processes and multilevel preconditioning for elliptic PDEs suggest a path toward polylogarithmic complexity relative to dataset size—a massive theoretical leap over the cubic complexity inherent in classical methods.

Probabilistic Identification of Dynamical Systems

For chaotic systems, matching individual trajectories is often impossible due to extreme sensitivity to initial conditions. Instead, researchers are utilizing "invariant measures"—the statistical fingerprints of a system—to identify governing dynamics from sparse and noisy data, focusing on where the system spends its time rather than its exact path.

Benchmarks and Applications

The validity of these new methods was tested against several high-fidelity physical benchmarks and synthetic models:

  • Subsurface Flow: Simulations of Darcy’s law using log-normal random fields to model complex soil conductivity in aquifers.
  • Structural Mechanics: Modeling the deformation of large-scale fiber networks (e.g., paper and cardboard) involving over 600,000 edges and 400,000 nodes.
  • Hamiltonian Systems: Benchmarks like the Fermi-Pasta-Ulam-Tsingou problem and the gravitational three-body problem to test parallel-in-time stability.
  • Condensed Matter Physics: Studying vortex structures in type-II superconductors using Ginzburg-Landau energy minimizers.

Key Insights and Results

The integration of randomization and Gaussian concentration of measure has turned high dimensionality from a "curse" into a "blessing" for certain local ansatz spaces, allowing for the breakdown of complexity bottlenecks.

Empirical results presented at the workshop include:

  • Efficiency Gains: New super-localized spectral methods (MS-GFEM) were found to be up to 8x cheaper to construct than previous state-of-the-art techniques while maintaining identical accuracy.
  • Accuracy Boosts: Kernel-based operator learning achieved one to two orders of magnitude improvement in accuracy over equivalent neural network approaches.
  • Quantum Scaling: Proposed quantum algorithms achieved polylogarithmic complexity in relation to dataset size, offering a theoretical path to bypass classical computational limits.
  • Stability: The importance of "stable" neural network architectures was emphasized, ensuring that small weight fluctuations do not lead to divergent outputs in physical simulations.

Limitations and Future Outlook

Despite these advances, several hurdles remain. The theoretical speedups offered by quantum algorithms currently outpace available hardware, with most tests conducted on simulators or small-scale systems like ibm_fez. Furthermore, "channeled" coefficients—where high-conductivity paths snake through a material—still pose challenges for standard localization strategies, requiring further refinement of partition-of-unity techniques.

Future research will focus on extending these methods to complex 3D geometries, where the computational overhead of higher-order multiscale basis functions remains significant, and refining iterative schemes for non-monotone quasilinear problems. The ultimate goal remains the realization of reliable, fast simulations for the most challenging multiscale problems in science and engineering.

Summary updated

Abstract

Many scientific and engineering problems exhibit complex interactions over a wide range of inseparable scales in space and time. Direct numerical simulations to solve such multiscale problems are often beyond current computational capabilities. The difficulties are exacerbated by the presence of uncertainty, randomness, and disorder and are hardly manageable for multiscale inverse problems. Therefore, the simulation of novel phenomena using multiscale models requires a new generation of multiscale computational methods. These must account for under-resolved scales, cross-scale couplings, and stochasticity in a hierarchical and adaptive manner and be able to integrate probabilistic, data-driven, and machine learning approaches. The workshop enhanced the development of a new generation of efficient multiscale computational methods and their rigorous mathematical and numerical analysis so that reliable and fast simulations of challenging multiscale problems from applications eventually become a reality.
BibTeX
@inproceedings{engquist2025computational,
  author = {Björn Engquist and Daniel Peterseim and Barbara Verfürth and Yunan Yang},
  title = {{Computational Multiscale Methods}},
  booktitle = {Oberwolfach},
  year = {2025},
  doi = {10.4171/owr/2025/22},
}
Figure