Finite Element-Based Material Learning via Automatic Differentiation: Learning constitutive neural network models from full-field deformation data

Researchers have developed FE-MAD, a differentiable machine learning framework that integrates neural networks into finite element solvers to identify material properties from experimental deformation data. The method combines the flexibility of neural networks with the physical rigor of finite element analysis, demonstrated on hyperelastic material characterization across multiple experimental datasets without requiring manual surrogate models or analytic adjoints.

FE-MAD represents a meaningful advance in computational materials science by bridging two historically separate domains: physics-based simulation and data-driven learning. The framework tackles a genuine limitation in existing approaches—traditional finite element updating is computationally expensive, weak-form methods are noise-sensitive, and neural operators require massive datasets. By embedding constitutive neural networks directly within a JAX-based FEM solver with automatic differentiation throughout, the researchers eliminate bottlenecks that have constrained material characterization workflows.

The approach addresses a practical engineering challenge: inferring material constitutive models from heterogeneous experimental data rather than simplified lab-controlled stress-strain tests. Full-field deformation measurements from digital image correlation provide richer information than traditional methods, but exploiting this data requires solving inverse problems with high-dimensional parameter spaces. The dual-architecture demonstration—grey-box models for maximum flexibility and white-box interpretable networks—signals maturity in the methodology, balancing predictive accuracy with scientific explainability.

Industrial applications extend across materials science, biomechanics, and mechanical design. Engineers can now rapidly characterize new materials or composites using existing test data without building expensive surrogate models. The validation on multiple experimental scenarios, including matrix-inclusion systems with generalization to unseen samples, demonstrates genuine robustness beyond proof-of-concept.

Longer term, this framework enables real-time material identification during manufacturing or structural health monitoring by seamlessly integrating observations into constitutive model training. The automatic differentiation pipeline could inspire similar architectures across other physics-informed machine learning domains, from fluid dynamics to solid mechanics.

• →FE-MAD eliminates manual surrogate models by integrating neural networks directly into differentiable finite element solvers
• →The framework successfully identifies material properties from heterogeneous experimental data including full-field DIC measurements and reduced-data scenarios
• →Dual architectures enable both maximal flexibility and interpretable, physically-grounded constitutive models
• →Validated generalization to unseen samples demonstrates practical robustness beyond controlled experimental conditions
• →Automatic differentiation throughout the pipeline removes computational bottlenecks inherent in traditional inverse problem solving methods