Neural Network Finite Elements

Overview

The Neural Network Finite Element (NNFE) method is a variational SciML framework that replaces the constant parameters defined in traditional finite elements with neural networks parameterized by boundary conditions. This allows a single trained model to represent an entire family of PDE solutions over a range of boundary conditions simultaneously, rather than solving one instance at a time.

Technical Formulation

In standard FE, forming and solving the stiffness matrix for multiple parameter instances is computationally intractable at scale. NNFE sidesteps this by training network parameters directly through gradient-based optimization, where the FE residual serves as the loss function. This variational formulation requires no training data — the physics itself provides the error metric. The method leverages JAX-native automatic differentiation to compute the spatial derivatives required for residual evaluation, and depends on CARDIAX for the underlying finite element machinery.

Accuracy & Validation

Because the loss is the FE residual, the method has an innate error metric built in. Validation compares NNFE solution manifolds against high-fidelity FE baselines across the trained range of boundary conditions, verifying that the learned solution family remains physically consistent throughout.

Clinical Application

The primary clinical motivation is speed — cardiac inverse problems require many-query evaluations where traditional FE solvers are prohibitively expensive. By learning solution families rather than individual instances, NNFE enables the kind of rapid parameter sweeps needed for patient-specific clinical workflows.

References & Resources