- Variational Matrix-Learning Fourier Networks (VMLFNs) eliminate the need for high-order differentiation and penalizing tuning by combining fixed log-space sine features, embedded boundary conditions, and variational form solutions to linear matrix in forming parametric multiphysics surrogates.
- The network's efficiency is validated via high reconstruction accuracy and computational speed up for objectives like heat conduction, solid mechanics, and Helmholtz wave propagation, by methodically replacing finite element methods or comparable parameter optimization networks.
- Spatio-torque parametric surrogates with zero-shot predictions at untested material parameters demonstrating significant improvement over traditional techniques.
Motivation and problem setting
Multiphysics simulation is a central component of system-technology co-optimization (STCO) for chiplet-based design, but the governing partial differential equations (PDEs) must typically be solved repeatedly across parametric design sweeps. The paper targets this parametric surrogate-modeling setting, where conventional physics-informed neural networks (PINNs) incur two well-known costs: iterative training with high-order automatic differentiation of PDE residuals, and manual tuning of penalty coefficients that balance residual losses against boundary-condition losses. The authors propose a variational matrix-learning Fourier network (VMLFN) that removes both costs by construction rather than by algorithmic improvement.
Log-space sine representation with fixed hidden layers
The first component is a log-space sine neural-network representation. Frequency points are randomly sampled within a prescribed range in logarithmic space, and a frequency-dependent decay factor regulates the contribution of each spectral component to the trial solution. Dirichlet boundary conditions are embedded directly into the trial solution, so they are satisfied exactly rather than enforced through loss terms. The hidden-layer parameters are fixed after sampling; only the output-layer weights are trainable. This design choice is what later enables a closed-form solution for those weights.
A practical difficulty in Fourier-feature methods is choosing the spectral bandwidth. VMLFN addresses this with a heuristic scanning algorithm that determines a problem-adaptive maximum frequency, intended to ensure the basis covers the dominant eigenfrequency range of the target problem. This replaces hand-tuned frequency hyperparameters with an automated, per-problem selection procedure.
The second component reformulates the governing PDEs in their variational weak forms. Applying the zero-gradient stationarity condition to the resulting functional yields a linear matrix system in the output-layer weights, which is solved directly rather than iteratively. Three consequences follow from this formulation:
- First-order derivatives suffice: because integration by parts transfers derivatives onto test functions, no high-order automatic differentiation is required.
- No penalty tuning: Dirichlet conditions live in the trial solution, while Neumann boundary conditions and interface traction continuity enter naturally through the weak form's boundary terms, eliminating the residual-versus-boundary loss balancing that plagues strong-form PINNs.
- Direct solvability: the output weights are obtained via direct matrix solution, avoiding backpropagation-based optimization entirely.
The combination means the "training" step reduces to assembling and solving one linear system per parameter configuration, which is where the computational speedup originates.
Validation and results
The framework is validated on three representative classes of multiphysics problems — heat conduction, solid mechanics, and Helmholtz wave propagation — across five benchmark cases. The reported outcomes are high field-reconstruction accuracy and substantial computational speedup relative to both conventional PINNs and repeated finite-element simulations. A notable capability is a 4-D spatio-parametric surrogate that supports zero-shot prediction at unseen material parameters, which is the property that enables real-time parametric design exploration in STCO workflows. The paper does not report specific accuracy or timing figures in the material provided here, so the magnitude of the speedup cannot be quantified independently.
Limitations and open questions
Several constraints on generality are evident from the formulation itself. The variational route requires that a weak form exists for the target PDE, so problems lacking a standard variational principle fall outside the framework as presented. The heuristic frequency-scanning algorithm determines the maximum frequency but its robustness on spectra with widely separated dominant modes is not established. The benchmarks cover scalar elliptic-type physics (heat conduction, elasticity, Helmholtz); strongly nonlinear or coupled multiphysics systems, where the stationarity condition may not reduce to a linear matrix solve, remain untested. Finally, the zero-shot parametric prediction relies on the assumption that the sampled log-space frequencies span the relevant solution manifold across the parameter domain; how far this assumption extends is left open.
Conclusion
VMLFN combines a fixed log-space sine feature map, embedded Dirichlet conditions, a variational weak-form reduction to a linear matrix system, and automated maximum-frequency selection into a surrogate framework that trains without backpropagation, high-order differentiation, or penalty tuning. Validation on heat conduction, solid mechanics, and Helmholtz benchmarks shows accurate reconstruction with substantial speedups over PINNs and repeated FEM, and the 4-D spatio-parametric variant demonstrates zero-shot generalization over material parameters. The main open questions concern extension to nonlinearly coupled multiphysics and the reliability of the automatic frequency-selection heuristic on broader spectral content.
