Technical Summary: Automated Dimensional Analysis for PDEs

Problem Statement

Physical units are fundamental to scientific computing, yet many finite element frameworks lack built-in support for dimensional analysis and unit consistency checking. This deficiency is particularly critical in the Finite Element Method (FEM) for solving Partial Differential Equations (PDEs), where complex multiphysics models often involve quantities spanning many orders of magnitude and diverse physical dimensions. The absence of robust unit handling can lead to catastrophic errors, as historically demonstrated by the 1999 Mars Climate Orbiter disaster, where a mismatch between pound-force-seconds and newton-seconds resulted in the loss of a spacecraft. While existing libraries (e.g., Pint, Boost.Units, Unitful.jl) address unit handling for standalone calculations, they do not integrate seamlessly with PDE solvers or variational form languages. Consequently, there is a need for a systematic framework to integrate physical units directly into the formulation of variational problems to ensure dimensional homogeneity, automate nondimensionalization, and improve numerical conditioning.

Methodology

The authors present a systematic framework implemented within the Unified Form Language (UFL), the domain-specific language used by finite element frameworks such as FEniCSx, Firedrake, and DUNE. The implementation is part of the Dolfiny package. The core methodology relies on three main components:

1. Symbolic Representation (Quantity Class):
   The framework introduces a Quantity class that extends ufl.Constant. This class stores a symbolic representation (based on sympy.Symbol) alongside physical units using SymPy's units module. Physical dimensions are treated as vectors in $\mathbb{Q}^n$ (where $n=7$ for SI base dimensions), exploiting the abelian group structure of physical dimensions. This allows multiplication, division, and powers of quantities to correspond to vector addition, subtraction, and scalar multiplication, respectively.

2. Transformation (Visitor Pattern):
   Since UFL expressions are inherently dimensionless, the framework employs a UnitTransformer visitor pattern to traverse the UFL expression tree (Directed Acyclic Graph). This transformer injects reference scales (e.g., $l_{ref}$, $v_{ref}$, $p_{ref}$) into the expressions. For instance, spatial gradients are scaled by $l_{ref}^{-1}$, and integration measures are scaled by $l_{ref}^d$ (where $d$ is the topological dimension). This step converts dimensionless UFL nodes into scaled expressions containing Quantity objects.

3. Factorization and Normalization:
   A QuantityFactorizer traverses the transformed DAG in post-order to extract dimensional factors. It checks for dimensional consistency (ensuring terms in a sum have identical dimension vectors) and factor consistency (ensuring terms have comparable magnitudes).

   • Inconsistent Dimensions: Raised as a severe error indicating a modeling mistake.
   • Inconsistent Factors: Raised to alert the user to potential numerical conditioning issues, suggesting strategies such as choosing a common reference scale or reformulating the problem.
   • Normalization: The framework divides the entire variational form by a chosen reference factor (e.g., the dominant term) to produce a dimensionless form. This process automatically derives dimensionless groups (e.g., Reynolds, Euler numbers) directly from the variational formulation.

Key Contributions

• Integration with UFL: The work provides the first systematic approach to handle physical units within the UFL expression tree, enabling automatic unit tracking and consistency checks for finite element variational forms.
• Automated Nondimensionalization: The framework automates the derivation of dimensionless forms and the identification of dimensionless groups (Buckingham $\Pi$ groups) directly from the weak form of PDEs, without requiring manual derivation.
• Full Operator Preconditioning: The authors demonstrate that this automated nondimensionalization functions as a form of Full Operator Preconditioning. By equilibrating the scales of linear systems prior to assembly, it acts as a physics-aware diagonal preconditioner.
• Multiphysics Support: The framework extends to coupled multiphysics problems, allowing for the selection of reference factors for individual equations to ensure dimensional consistency across the coupled system.

Results

The framework was validated through numerical experiments on several PDE systems:

1. Incompressible Navier–Stokes Equations:

   • The framework successfully identified classical dimensionless numbers (Reynolds, Froude, Euler, Strouhal) from the weak form.
   • Conditioning Analysis: Experiments showed that the condition number of the saddle-point matrix is highly sensitive to the choice of the pressure reference scale ($p_{ref}$). A naive choice (e.g., $p_{ref} = 1$ Pa) led to an ill-conditioned system, whereas selecting $p_{ref}$ based on dynamic pressure ($\rho v_{ref}^2$) significantly improved the condition number. This confirms the role of automated nondimensionalization as a preconditioning strategy.

2. Neo-Hooke Hyperelasticity:

   • Analysis of small deformation regimes highlighted the framework's ability to detect floating-point cancellation errors. By identifying terms with inconsistent factors (large differences in magnitude), the system flags potential numerical instability before assembly.

3. Poisson–Nernst–Planck System:

   • The framework successfully handled a coupled multiphysics problem involving derived scaling parameters, demonstrating its capability to manage complex interactions between different physical fields.

Significance and Claims

The paper claims that integrating physical units into the variational formulation is not merely a bookkeeping exercise but a fundamental step toward robust and efficient scientific computing. The authors assert that:

• Error Prevention: The framework prevents modeling errors by detecting dimensional inconsistencies (e.g., adding quantities with different dimensions) at the formulation stage.
• Numerical Stability: By automatically equilibrating scales, the method acts as a "physics-aware" preconditioner, improving the condition number of linear systems and reducing the risk of floating-point cancellation.
• Generalizability: Although implemented for FEniCSx, the concepts are general and adaptable to other UFL-based libraries (Firedrake, DUNE).
• Theoretical Insight: The automated derivation of dimensionless groups from the variational form provides a rigorous link between the mathematical structure of the PDE and the physical scaling laws governing the system, facilitating better experimental design and numerical simulation.

The work concludes that strong type checking and dimensional analysis should be fundamental aspects of scientific computing frameworks to avoid the types of costly errors seen in historical engineering failures.