Finite Element-Based Machine Learning

• Finite Element-Based Machine Learning (FEBML) is a computational framework that integrates finite element discretizations and machine learning to predict physical responses and accelerate simulations.
• FEBML leverages FE-generated data, mesh geometry, and variational operators as inductive biases to enhance surrogate modeling and efficiently solve inverse problems.
• FEBML spans diverse applications—from structural mechanics to tactile sensing—by replacing or augmenting traditional FE solvers with neural networks for improved speed and accuracy.

Finite Element-Based Machine Learning (FEBML) designates a family of computational strategies in which finite-element discretizations, FE-generated simulation data, FE meshes, or FE variational operators are combined with machine learning to predict physical response, assess mesh quality, infer missing constitutive or transport relations, or accelerate many-query analyses. In the literature surveyed here, the defining feature is not a single algorithm but the use of finite-element structure as either the source of supervision, the representation space, the optimization constraint, or the deployment environment itself (Kononenko et al., 2018, Meethal et al., 2022, Ouyang et al., 23 Jun 2025, Farsi et al., 21 Jul 2025).

1. Scope and taxonomic structure

FEBML spans multiple positions in the simulation pipeline. Some methods use FEM as an offline “teacher” that generates accurate labels for supervised surrogates; some operate directly on finite-element meshes and adjacency relations; some treat the FE residual or weak form as the learning objective; and some insert trainable operators or reusable learned elements inside the FE variational framework itself. A recurrent simplification is to equate FEBML with replacing an FE solver by a neural network, but the research record is broader: preprocessing, constitutive modeling, multiscale coupling, inverse problems, and real-time deployment all appear as legitimate FEBML modes (Kononenko et al., 2018, Sprave et al., 2021, Meethal et al., 2022, Farsi et al., 21 Jul 2025).

FEBML pattern	Role of FEM	Representative papers
FE-generated surrogate modeling	High-fidelity data generator for supervised learning	(Kononenko et al., 2018, Shaikh et al., 2023, Sáenz et al., 2 Jun 2025)
Mesh-centric learning	Source of geometry, topology, and neighborhood structure	(Sprave et al., 2021)
FE-residual or weak-form learning	Discrete physics operator and loss function	(Meethal et al., 2022, Sunil et al., 2024)
Solver-integrated learned operators	Variational host for neural constitutive or subdomain operators	(Ouyang et al., 23 Jun 2025, Farsi et al., 21 Jul 2025)
FE-basis coefficient learning	Function-space representation for field prediction	(Backmeyer et al., 25 Jul 2025)
FE-trained inverse sensing or multiscale surrogates	Label generator for otherwise unobservable fields or micro-scale responses	(Sferrazza et al., 2019, Helmut et al., 2024, Gupta et al., 2022)

This variety implies that FEBML is best understood as a methodological interface between computational mechanics and statistical learning rather than as a single architecture class. The unifying idea is that FE structure carries physically meaningful inductive bias: basis functions, mesh connectivity, weak forms, stiffness operators, and parameterized geometries can all be exploited rather than discarded.

2. Mathematical foundations and coupling mechanisms

The finite element side is the familiar discretization of a PDE or ODE into basis functions and algebraic systems. Typical formulations in the surveyed work include the standard FE approximation

$$\mathbf{u}_h(\mathbf{x}) = \sum_j N_j(\mathbf{x}) \mathbf{U}_j,$$

stiffness assembly of the form

$$K_{ij} = \int_{\Omega_e} \mathbf{B}_i^T \mathbf{C}\mathbf{B}_j\, d\Omega,$$

and weak formulations written as

$$\text{Find } u\in H \text{ such that } a(u,v)=L(v)\quad \forall v\in H_0,$$

or, equivalently in energy form,

$$u=\arg\min_{w\in H} J[w],\qquad J[w]=\frac12 a(w,w)-L(w)$$

(Kononenko et al., 2018, Ouyang et al., 23 Jun 2025).

FEBML modifies this pipeline in several distinct ways. In surrogate settings, the FE solver produces pairs $(x_i,y_i)$ and machine learning approximates the map

$x\mapsto y,$

for example frequency to amplitude or maximum displacement (Kononenko et al., 2018). In FE-residual learning, the network predicts FE coefficients directly and training minimizes the algebraic residual

$\delta=\|K\,U-F\|_2,$

so the FE system itself becomes the loss rather than a source of labels (Meethal et al., 2022). FE-PINNs generalize this idea by using the finite-element residual of the weak form,

$$R(\tilde d)\equiv G(\tilde d)=F_{\mathrm i}(\tilde d)-F_{\mathrm e},$$

as the physics loss, while stencil convolution uses the inverse isoparametric map and FE shape functions to define CNN-like operations on arbitrary meshes (Sunil et al., 2024).

A more intrusive coupling learns operators that live inside the PDE model. In the neural-operator element method, a subdomain operator

$$\mathcal{G}:\text{boundary data}\mapsto u|_{\Omega_j^{\text{NOE}}}$$

is approximated by a neural operator and inserted into a mixed FE–NOE variational space; the global solution remains the minimizer of the FE energy over this mixed space (Ouyang et al., 23 Jun 2025). In fully differentiable missing-physics discovery, an unknown operator is represented as an encode–process–decode map

$\mathcal{G}_\theta(f)=D\circ P_\theta\circ E(f),$

acting over FE degrees of freedom and embedded directly in the weak form, so constitutive laws or conductivity functions are learned inside the solver rather than outside it (Farsi et al., 21 Jul 2025). Closely related is the isogeometric-field setting, where the field is learned through spline-basis coefficients,

$$K_{ij} = \int_{\Omega_e} \mathbf{B}_i^T \mathbf{C}\mathbf{B}_j\, d\Omega,$$0

followed by POD compression and coefficient regression (Backmeyer et al., 25 Jul 2025).

These formulations differ in where the learning occurs, but they share a common premise: FE spaces, operators, and assemblies are used as machine-learning priors rather than treated as external simulators only.

3. FEM as teacher: supervised surrogates and response emulators

The earliest and most explicit FEBML pattern is the supervised surrogate built from FE or analytic data. "Machine Learning and Finite Element Method for Physical Systems Modeling" formalized a workflow in which FEM generates accurate simulation outputs across sampled parameters, these are used to train fully connected neural networks, and the trained model then returns fast predictions without rerunning FEM (Kononenko et al., 2018). In that note, the examples are deliberately simple: a driven harmonic oscillator and a 3D stainless steel beam under harmonic load. The FE role is one-way and offline; the neural network learns a global parameter-to-response map rather than participating in assembly or constitutive evaluation.

Subsequent work broadened this pattern to higher-value engineering workflows. For carbon-fiber battery enclosures, high-throughput thermoforming and crash FEA were chained into a process–structure–property dataset, and Gradient Boosting, XGBoost, and Random Forest surrogates were trained to predict crush load efficiency, absorbed energy, intrusion, and maximum deceleration. The reported predictive accuracy was strong, with $$K_{ij} = \int_{\Omega_e} \mathbf{B}_i^T \mathbf{C}\mathbf{B}_j\, d\Omega,$$1 for the crashworthiness metrics cited in the study (Shaikh et al., 2023). In craniosynostosis planning, patient-specific synthetic skulls were generated from 3D photographs, population-average anatomical quantities were used to build FE models, and a multi-output support vector regressor predicted postoperative head-shape modes from age, osteotomy parameters, spring parameters, and pre-operative shape coefficients. The reported model achieved $$K_{ij} = \int_{\Omega_e} \mathbf{B}_i^T \mathbf{C}\mathbf{B}_j\, d\Omega,$$2 with MSE and MAE below $$K_{ij} = \int_{\Omega_e} \mathbf{B}_i^T \mathbf{C}\mathbf{B}_j\, d\Omega,$$3 (Sáenz et al., 2 Jun 2025).

A closely related line uses FE to generate labels for inverse sensing tasks that are experimentally inaccessible. For optical tactile sensing, FE models of soft elastomeric media were calibrated and then used to produce dense normal and shear force distributions over the sensor surface; image-based networks subsequently learned to infer these distributions in real time. In one case, a U-net predicted force maps for the GelSight Mini from raw images, with MAE below $$K_{ij} = \int_{\Omega_e} \mathbf{B}_i^T \mathbf{C}\mathbf{B}_j\, d\Omega,$$4 N on total normal force and below $$K_{ij} = \int_{\Omega_e} \mathbf{B}_i^T \mathbf{C}\mathbf{B}_j\, d\Omega,$$5 N on shear, while mean inference time was $$K_{ij} = \int_{\Omega_e} \mathbf{B}_i^T \mathbf{C}\mathbf{B}_j\, d\Omega,$$6 ms compared with 10–120 minutes per FEA case (Helmut et al., 2024). An earlier vision-based tactile study used hyperelastic FE labels derived from an indentation setup, validated the model against a commercial force–torque sensor, and showed that the learned force-distribution predictor could run at 40 Hz on a standard laptop CPU (Sferrazza et al., 2019).

Within this paradigm, FEBML does not alter the FE solve itself. Its contribution lies in converting costly many-query analysis into cheap inference, provided the simulation campaign spans the operational parameter space well enough.

4. FE-native learning on meshes, residuals, and variational operators

A second cluster of methods uses FE-specific structures more directly. Mesh-quality evaluation for structural mechanics is formulated as element-wise binary classification on the element neighborhood graph

$$K_{ij} = \int_{\Omega_e} \mathbf{B}_i^T \mathbf{C}\mathbf{B}_j\, d\Omega,$$7

with features built from skewness, aspect ratio, warping, area, curvature proxies, triangle and border flags, aggregated over $$K_{ij} = \int_{\Omega_e} \mathbf{B}_i^T \mathbf{C}\mathbf{B}_j\, d\Omega,$$8-ring frontiers. This converts irregular mesh neighborhoods into fixed-size feature vectors usable by Extremely Randomized Trees and feedforward neural networks (Sprave et al., 2021). Here, FEM enters not through the PDE solve but through mesh geometry and topology, and the learning target is expert judgment about rework regions.

Transient FE surrogates can also be FE-aware without explicitly using FE residuals. DeepFEA targets transient 2D and 3D structural mechanics, predicts both nodal and element quantities, and uses a multilayer ConvLSTM backbone branching into two CNN heads. Its Node-Element Loss Optimization couples node and element errors over all timesteps while scheduled sampling stabilizes long autoregressive rollout. The reported results show less than 3% normalized mean and root mean squared error for 2D and 3D scenarios, with inference times two orders of magnitude faster than FEA (Triantafyllou et al., 2024). Although this framework is mesh-specific, it treats nodal coordinates, loads, boundary flags, and element topology as first-class design objects.

Residual-based and weak-form methods bring the FE operator into training itself. The FEM-enhanced neural network of (Meethal et al., 2022) uses the finite-element residual as the primary loss, so a network is trained by minimizing $$K_{ij} = \int_{\Omega_e} \mathbf{B}_i^T \mathbf{C}\mathbf{B}_j\, d\Omega,$$9 rather than by matching precomputed FE solutions. FE-PINNs go further by defining physics loss as the FE residual of the weak form and replacing grid-based convolution with stencil convolution on arbitrary FE meshes via the inverse isoparametric map (Sunil et al., 2024). These approaches retain the FE assembly, boundary-condition handling, and discrete operator structure, but move optimization into a neural parameter space.

The most integrated schemes embed learned operators inside the solver. Finite Element Network Analysis replaces element models by pre-trained bidirectional recurrent networks that can be concatenated without retraining, preserving an assembly-like workflow for structural systems (Jokar et al., 2020). The multiscale U-Net framework of (Gupta et al., 2022) replaces micro-scale FE solves in an $$\text{Find } u\in H \text{ such that } a(u,v)=L(v)\quad \forall v\in H_0,$$0-style setting by predicting microstructural stress tensor fields, supplying both effective properties for upscaling and local stress fields for downscaling. The neural-operator element method constructs reusable neural-operator elements that replace clusters of fine elements inside a global FE variational problem, and reports approximately $$\text{Find } u\in H \text{ such that } a(u,v)=L(v)\quad \forall v\in H_0,$$1 speed-up in 2D heat problems with 100 holes (Ouyang et al., 23 Jun 2025). Fully differentiable FEBML then generalizes this logic to unknown physics discovery: constitutive laws and temperature-dependent conductivity are learned as operator blocks inside a Firedrake-based FE solver, with end-to-end differentiation through assembly and solution (Farsi et al., 21 Jul 2025).

5. Domains of application and reported performance

FEBML has been applied across structural dynamics, multiscale mechanics, electromagnetics, biomedical planning, tactile sensing, metal forming, and inverse PDEs. The diversity of application domains is itself consequential: it shows that the common denominator is not a single governing equation but the exploitation of FE structure wherever the numerical model is expensive, high-dimensional, or only partially observable.

In structural and solid mechanics, applications range from harmonic beam response (Kononenko et al., 2018) to transient FEA surrogates (Triantafyllou et al., 2024), rolling-induced deflection in austenitic stainless steel 316L where an ANN-based flow-stress model feeds a finite-difference rolling-pressure model and then an FE beam model of roll deflection (Lotfinia et al., 2021), and automotive shell-mesh quality control (Sprave et al., 2021). In composite crashworthiness, FEBML couples thermoforming FEA, draping, crash simulation, and tree-based surrogates for rapid exploration of laminate and process parameters (Shaikh et al., 2023). In craniofacial surgery, the FE–ML chain is used to move from patient-specific synthetic skulls to real-time postoperative head-shape prediction without CT-based FEM planning (Sáenz et al., 2 Jun 2025).

In tactile and robotic sensing, FEBML provides one of the clearest “FEM as oracle” use cases. Dense force distributions are inaccessible experimentally, so FE models of soft contact generate the supervisory signal. The resulting learned inverse maps estimate normal and shear force fields directly from images and are sufficiently fast for contact-rich manipulation (Sferrazza et al., 2019, Helmut et al., 2024). In multiscale materials, U-Net surrogates trained on microstructural FE stress fields accelerate FE-squared-style coupling and were reported to be more than $$\text{Find } u\in H \text{ such that } a(u,v)=L(v)\quad \forall v\in H_0,$$2 faster than an FE analysis of the same microstructure, with macro-scale examples showing seconds versus hours or days for traditional multiscale alternatives (Gupta et al., 2022). In electromagnetics, learning spline-basis coefficients for a permanent magnet synchronous machine yields field surrogates with mean relative errors around $$\text{Find } u\in H \text{ such that } a(u,v)=L(v)\quad \forall v\in H_0,$$3 in the air-gap setting and $$\text{Find } u\in H \text{ such that } a(u,v)=L(v)\quad \forall v\in H_0,$$4 for full-field prediction on the test set, while online speed-ups were reported as about $$\text{Find } u\in H \text{ such that } a(u,v)=L(v)\quad \forall v\in H_0,$$5 and $$\text{Find } u\in H \text{ such that } a(u,v)=L(v)\quad \forall v\in H_0,$$6, respectively (Backmeyer et al., 25 Jul 2025).

Inverse and data-assimilation settings are equally prominent. The Dirichlet-to-Neumann framework for inverse elliptic PDEs combines an unsupervised neural parameterization of the coefficient field with an FEM forward solve in the inner loop, allowing reconstructions from full or partial boundary observations and accommodating discontinuities through regularized losses (Park et al., 4 Apr 2025). The distinguishing trait in such work is that the learning signal is not a direct label on the unknown coefficient, but boundary data filtered through a differentiable or adjointable FE forward model.

6. Limitations, misconceptions, and emerging directions

The literature is explicit that FEBML does not eliminate the classical difficulties of computational mechanics; it rearranges them. In surrogate settings, large FE campaigns remain necessary to cover parameter space, and the examples in (Kononenko et al., 2018) already note that realistic applications will require very large datasets, multiple geometries, and different boundary or initial conditions. High-dimensional outputs remain a central obstacle: predicting scalar amplitudes or a few summary quantities is substantially easier than learning “full field of deformations or electromagnetic fields on all the nodes of the computational mesh,” which is described as “a real challenge” (Kononenko et al., 2018). DeepFEA likewise acknowledges that mesh size and density are hard-wired in the ConvLSTM weights, so a different network must be trained for each mesh configuration (Triantafyllou et al., 2024).

A second misconception is that FE-trained models are automatically physics-consistent. That is only true for methods that keep FE residuals or variational constraints in the loop. Purely supervised surrogates can inherit FE structure through their labels or representations, but they generally do not enforce conservation laws, PDE residuals, or admissibility outside the sampled regime. Several papers identify this explicitly: the beam surrogate note imposes no PDE structure in the network and warns that extrapolation outside the training range may be unreliable (Kononenko et al., 2018); tactile-force surrogates inherit biases from simplified contact and material models, so label quality bounds surrogate quality (Sferrazza et al., 2019, Helmut et al., 2024).

There are also domain-specific issues. Mesh-quality labels derived from expert rework markings are noisy because marked regions do not always correspond one-to-one with individually bad elements, so element-wise precision and recall can misrepresent practical utility (Sprave et al., 2021). Biomedical FE–ML planning tools may be limited by population-average thicknesses, simplified growth assumptions, or missing subject-specific material properties (Sáenz et al., 2 Jun 2025). Solver-integrated schemes face their own burden: end-to-end differentiable FEBML and neural-operator elements preserve FE structure, but they introduce adjoint, memory, and implementation complexity (Ouyang et al., 23 Jun 2025, Farsi et al., 21 Jul 2025).

The forward-looking agenda is correspondingly rich. Several works point toward prediction of full fields, mesh-size-independent surrogates, graph neural networks for mesh-structured learning, broader multi-domain datasets, and tighter incorporation of PDE structure into network design or loss functions (Kononenko et al., 2018, Sprave et al., 2021, Triantafyllou et al., 2024). FE-PINNs explicitly identify variable boundary conditions, material properties, and body forces as the next step toward a general-purpose surrogate modeling framework (Sunil et al., 2024). Differentiable missing-physics discovery suggests a route from FE-trained black-box closures to reusable operator models that transfer across geometries and loading scenarios without retraining (Farsi et al., 21 Jul 2025). Basis-coefficient learning in isogeometric settings suggests another trajectory: rather than predicting fields pointwise, future FEBML systems may increasingly learn directly in FE or spline coefficient spaces, where continuity, interface conditions, and function-space structure are already encoded (Backmeyer et al., 25 Jul 2025).

Taken together, these developments suggest that FEBML is evolving from simple FE-trained emulators toward a layered discipline in which learning may occur on FE meshes, in FE bases, through FE residuals, or within FE solvers. The common ambition is not merely speed, but controlled approximation of PDE-governed systems in representations that remain compatible with the numerical and variational structure of finite elements.