Tropical Geometry as a Restricted Architecture for Physics-Informed Neural Networks: Applications in Nonlinear Fluid-Structure Examples

Technical Summary: Tropical Geometry as a Restricted Architecture for Physics-Informed Neural Networks

Problem Statement
Nonlinear algebraic (polynomial) differential equations governing fluid-structure interactions (FSI), such as those modeling vortex-induced vibrations and shock waves, frequently lack analytical solutions. While Physics-Informed Neural Networks (PINNs) offer a mesh-free numerical alternative, they often encounter convergence stagnation when optimizing over high-dimensional landscapes characterized by stiff singularities or topological bifurcations. Standard PINNs suffer from spectral bias and optimization failures in these regimes because their hypothesis spaces are unconstrained, allowing the optimizer to explore non-physical solutions or fail to resolve singular behaviors like shock fronts.

Methodology: The Tropical-PINN Framework
The paper introduces a hybrid methodology, Tropical-PINN, which integrates tropical differential algebraic geometry directly into the neural network architecture. The core approach involves three stages:

1. Theoretical Equivalence (Valuation-Support Equivalence): The authors establish a formal proof that classical Briot-Bouquet indicial analysis (used to find formal power series solutions near singularities) is algebraically equivalent to tropical differential algebraic geometry. Specifically, they prove that the "dominant balance" condition in classical analysis—where the lowest-degree terms of a differential equation must cancel—is identical to the condition in tropical algebra where the minimum tropical evaluation is attained by at least two distinct monomials. This equivalence allows the algorithmic determination of the exact "support" (the set of valid powers in a formal power series solution) without solving the full differential equation.
2. Architectural Restriction: Instead of incorporating physical constraints via the loss function (which leaves the hypothesis space unconstrained), the Tropical-PINN hard-codes these tropical supports into the network's structure. The network's hypothesis space is restricted a priori to the span of basis functions $\{t^k\}$ where $k$ belongs to the algorithmically derived support set $S_{trop}$.
3. Hybrid Training: The network is trained by minimizing the physics residual (PDE loss) over a defined domain. Because the architecture only contains weights for physically permissible terms, the optimization landscape is drastically reduced, bypassing chaotic regions and stiff singularities.

Key Contributions

• Theoretical Proof: The paper proves the exact algebraic equivalence between classical indicial analysis and tropical differential algebraic geometry, demonstrating that tropical methods accurately identify singularity structures and resonance gaps.
• Hard-Constraint Architecture: The authors propose a novel architecture where the "tropical skeleton" (the valid support of the formal power series) is embedded directly into the network. This restricts the hypothesis space to the exact support of valid solutions, eliminating the search for non-physical coefficients.
• Resolution of Optimization Stagnation: By reducing the search space to the exact support, the method overcomes the convergence stagnation typical of standard PINNs in non-homogeneous physical regimes.
• Precision in Coefficient Extraction: The hybrid model extracts highly precise physical Taylor series coefficients directly from the physics loss, matching theoretical values to high decimal precision.

Results
The methodology was validated on several nonlinear fluid-structure examples:

• Van der Pol Oscillator: The tropical analysis correctly predicted that the homogeneous limit cycle is not representable by a simple power series at $t=0$ (returning an empty set), while successfully identifying valid supports for the forced case to model structural entrainment.
• Viscous Burgers' Equation: The framework identified a negative valuation corresponding to a simple pole ($i = -1$), correctly predicting the formation of shock waves. This topological prior prevented the neural network from artificially smoothing the shock front.
• Blasius and Falkner-Skan Equations: The model successfully captured structural phase shifts driven by the pressure gradient parameter $\beta$:
  • Blasius Flow ($\beta = 0$): The algorithm enforced a sparse topological prior ($A = \{2, 5, 8, \dots\}$), restricting the network to specific powers. The model recovered the theoretical coefficient $w_5 = -\alpha^2/240$ with high precision.
  • Falkner-Skan/Hiemenz Flow ($\beta = 1$): The algorithm detected a shift to a dense topological prior ($A = \{2, 3, 4, \dots\}$). The network natively recovered the exact theoretical response $w_3 = -1/6$.
  • In both cases, the Tropical-PINN avoided "phantom physical terms" and spectral bias, converting a continuous optimization problem into a physics-constrained algebraic projection.

Significance and Claims
The paper claims that Tropical Differential Algebra is a practical computational tool for characterizing the singularity structure of nonlinear dynamical systems, rather than just an abstract formalism. By algorithmically pruning non-physical dimensions before compilation, the methodology provides a reliable pathway from "black-box" numerical approximations to "white-box," interpretable analytical series. The authors assert that this approach offers absolute structural fidelity across non-homogeneous physical regimes. They suggest that future applications could extend these topological priors to massive, interconnected physical graphs, potentially enhancing real-time anomaly detection in large-scale fluid distribution networks (Digital Twins) where computational efficiency and physical fidelity are critical.