Data-Driven Whitney Forms for Structure-Preserving Control Volume Analysis.

Jonas Actor, Nathaniel Trask, Ravi G. Patel, Xiaozhe Hu, Xujiao Gao, Andy Huang2022 · DOI

This research poster from Sandia National Laboratories presents a novel framework for creating "physics-guaranteed" surrogate models. By combining algebraic topology with machine learning, the authors demonstrate how to learn from data while strictly adhering to the conservation laws that govern the physical world.

TL;DR

The authors introduce a data-driven approach to building surrogate models using Whitney forms and chain complexes, ensuring that physical laws like conservation and Stokes' Theorem are satisfied by construction. This method allows for massive model reduction—shrinking a battery simulation from 5.9 million degrees of freedom to just 136—without sacrificing the underlying physics or stability.

What Problem Does This Paper Address?

High-fidelity physics simulations (like those used in battery design or semiconductor modeling) are notoriously expensive and compute-intensive. While standard machine learning (ML) can create fast "surrogate" models, these "black-box" approaches often fail in critical ways: they might "leak" mass, violate energy conservation, or become unstable when faced with data outside their training set.

The challenge is building a model that is data-driven (learned from observations) but structure-preserving (guaranteed to obey the rules of calculus and physics). Specifically, the authors tackle three hurdles:

  1. Enforcing exact conservation from discrete, sampled data.
  2. Extracting generalizable models from small datasets.
  3. Guaranteeing the same stability and convergence properties found in rigorous numerical methods like the Finite Element Method (FEM).

Methods

The core of the approach is "baking" the physics into the mathematical architecture of the model using Discrete Exterior Calculus (DEC) concepts.

  • Building a Chain Complex: The model is structured as a sequence of spaces connected by operators (like gradient, curl, and divergence). By ensuring this sequence is "exact," the model automatically satisfies Stokes' Theorem.
  • Whitney k-Forms: These are used to interpolate data across different geometric dimensions (points, edges, faces, volumes). The authors use Partitions of Unity (POU) to define these forms, allowing the model to represent physical fields (like pressure or flux) in a way that is mathematically compatible.
  • Free-Knot B-Splines: Instead of using a fixed grid, the model uses splines where the "knots" (the points where spline segments meet) are trainable parameters. This allows the model to adaptively "move" its resolution to where the physics is most complex, such as at material interfaces or shocks.
  • Learnable Metrics: The boundary and coboundary operators are parameterized, allowing the model to "learn" the specific metric properties of the system from the data while keeping the topological structure fixed.

Datasets

The framework was evaluated on several classic and complex physics benchmarks:

  • Poisson Problem: A standard test for stability and convergence in numerical analysis.
  • Five Strip Problem: A multi-material interface problem used to test how well the free-knot splines adapt to sharp changes in properties.
  • Battery Problem: A high-fidelity electrochemical model used to demonstrate extreme model reduction.
  • PN Diode: A semiconductor problem used to show the method's extensibility to nonlinear and advection-dominated physics.

Key Insights and Ideas

  • Topology as a Constraint: The most powerful idea here is that conservation shouldn't be a "loss function" in a neural network (which can be ignored); it should be a mathematical necessity of the space the model lives in. By using a chain complex, the model cannot violate conservation.
  • Adaptive Resolution: The use of "free knots" means the model discovers the geometry of the problem. If there is a sharp gradient in the data, the splines will naturally cluster their knots there, providing high resolution only where needed.
  • Bridging ML and FEM: The results show "$hp$-convergence," a hallmark of high-order Finite Element Methods. This means the model isn't just "mimicking" the data; it is behaving with the mathematical rigor of a professional engineering simulation tool.

Key Results

  • Massive Compression: In the Battery Problem, the authors reduced the system from 5.9 million degrees of freedom (DoF) to a mere 136 DoF. Despite this 40,000x reduction, the model preserved the essential "quantities of interest."
  • Stability: Unlike many data-driven surrogates that "blow up" or produce unphysical oscillations, this method maintains the stability of traditional FEM.
  • Accuracy: The Poisson problem results demonstrate that the method achieves the same convergence rates as high-end numerical solvers, proving that "data-driven" doesn't have to mean "less accurate."

Limitations and Future Work

While the poster demonstrates success across several domains, some areas for further exploration include:

  • Scaling to 3D: Most examples shown are 1D or 2D; scaling the free-knot spline optimization to complex 3D geometries may introduce significant computational overhead during the training phase.
  • Nonlinear Complexity: While the PN Diode example shows promise for nonlinear problems, highly turbulent or chaotic systems (like high-Reynolds-number fluid dynamics) would be a rigorous next test for this framework.
  • Training Cost: The "variational problem" solved to discover control volumes is more complex than standard backpropagation, potentially requiring more specialized optimization routines.

Summary updated

Abstract

Abstract not provided.
BibTeX
@misc{actor2022data,
  author = {Jonas Actor and Nathaniel Trask and Ravi G. Patel and Xiaozhe Hu and Xujiao Gao and Andy Huang},
  title = {{Data-Driven Whitney Forms for Structure-Preserving Control Volume Analysis.}},
  year = {2022},
  doi = {10.2172/2004812},
}
Figure