Minimally Intrusive Machine Learning Accelerates Non-Linear Structural Mechanics Simulations to Seconds

A novel computational strategy developed by computational mechanics researchers combines proper orthogonal decomposition with machine learning to accelerate structural simulations while maintaining high numerical fidelity. Published in Advanced Modeling and Simulation in Engineering Sciences, the study demonstrates a minimally intrusive reduced order modeling technique capable of handling both linear and geometrically non-linear structural mechanics problems. By extracting core system matrices directly from commercial finite element solvers without invasive internal code alterations, the method addresses a longstanding trade-off between computational speed and implementation complexity. The investigation was initiated to advance predictive digital twins for structural health monitoring, where engineers require real-time feedback, continuous virtual asset tracking, and parametric optimization. Traditional high-fidelity finite element modeling remains computationally prohibitive for repeated parametric evaluations, while purely data-driven alternatives often lack physical consistency or demand an intractable volume of precomputed training data.

In modern engineering design, model order reduction serves as an alternative to solving large-scale systems generated by finite element, finite volume, or finite difference discretizations. Conventional intrusive reduction techniques exhibit high mathematical accuracy and physical stability, yet they are rarely deployed as standalone industrial tools because their implementation requires deep mathematical expertise and extensive low-level solver customization. Furthermore, the computational advantages of classical intrusive approaches diminish considerably when tackling severe non-linear behaviors, where repeated operator evaluations incur substantial overhead. Conversely, fully non-intrusive surrogate methods avoid direct interaction with proprietary commercial software by training regression or deep learning models strictly on external input and output snapshots. However, non-intrusive models frequently struggle to achieve adequate accuracy in non-linear regimes unless engineers execute an exceptionally large design of experiments, generating hundreds or thousands of high-dimensional snapshots that render offline preparation computationally impractical.

To overcome these limitations, the research team formulated a minimally intrusive framework that combines the rigor of physics-based projection methods with the speed and flexibility of machine learning algorithms. The approach preserves physical characteristics by interfacing with commercial solvers, such as Abaqus, through standard outputs: the global system coefficient stiffness matrix, the right-hand load vector, and the applied Dirichlet boundary conditions. The reduction workflow begins by collecting solution snapshot vectors obtained from solving full-order problems across selected points in a defined parameter space. Applying singular value decomposition to these collected displacement fields extracts a compact set of dominant orthogonal basis vectors, known as proper orthogonal decomposition modes. This reduced basis allows engineers to project the high-dimensional equilibrium equations onto an equivalent low-dimensional coordinate subspace, dramatically decreasing the matrix dimensions needed to represent structural deformation while retaining essential physical governing properties.

The central theoretical innovation of the proposed framework centers on how the reduced stiffness operators are treated and learned. In traditional surrogate workflows, machine learning models attempt to predict full displacement fields or the forward stiffness matrix directly. However, predicting the forward stiffness matrix introduces approximation discrepancies that can be severely magnified during matrix inversion, especially when equations exhibit poor numerical conditioning. Direct displacement regression, on the other hand, frequently demands hundreds of precomputed simulation runs to generalize across complex parameter spaces. The authors avoided both pitfalls by explicitly computing the inverse of the reduced stiffness matrix, flattening each inverted matrix into a high-dimensional vector, and performing a second principal component analysis reduction across all parametric snapshots. Machine learning models are subsequently trained to predict only the compact set of modal weighting coefficients that reconstruct the flattened inverted matrix for any arbitrary parameter combination.

Selecting an appropriate learning algorithm was crucial for maintaining the goal of minimal user intervention and seamless operational adoption across diverse engineering structures. The researchers evaluated several candidate regression algorithms, including artificial neural networks, support vector machines, and extreme gradient boosting, before selecting random forest regression. While artificial neural networks achieved high precision, they required extensive customization of network depth and architecture whenever dataset dimensions or problem types changed. Support vector machines and gradient boosting also demanded meticulous hyperparameter tuning to yield competitive results. In contrast, the random forest regressor provided exceptional prediction accuracy and robustness using standard default parameters from the scikit-learn library, including one hundred decision trees and standard splitting criteria. The algorithm required no manual hyperparameter calibration, trained within seconds on compact snapshot databases, and delivered instantaneous coefficient predictions during online deployment.

To demonstrate the step-by-step implementation, the team first applied the method to a linear static structural benchmark featuring a thin metal plate measuring one square meter. Discretized with an eleven-by-eleven mesh of linear shell elements yielding 726 degrees of freedom, the plate had all four outer edges constrained against vertical displacement while bearing a concentrated vertical static point load at its center. The problem incorporated three variable material and geometrical parameters: Young’s modulus varied between 100 and 300 gigapascals, Poisson’s ratio ranged from 0.30 to 0.49, and plate thickness spanned from 1 to 10 millimeters. The parametric domain was sampled using 65 snapshot experiments determined by the roots of fourth-order Chebyshev polynomials across all three parameter dimensions plus the central domain coordinate. Fifty-four randomly selected simulation snapshots formed the training dataset, while the remaining eleven served as unseen validation instances.

Dimensional reduction compressed the full 726-by-726 stiffness operator into a compact two-by-two reduced matrix by retaining modes based on an eigenvalue ratio cutoff of one thousand. The random forest regressor successfully learned the two underlying reduced coordinates, producing coefficient estimates whose deviation from reference mathematical projections was virtually imperceptible without extreme visual magnification. Across all 65 experimental cases and 47,190 evaluated degree-of-freedom values, the machine learning predictions aligned closely with full finite element results. Evaluating the reconstructed inverted stiffness matrices demonstrated a relative Frobenius norm error below two percent, with peak discrepancies concentrated primarily in smaller numerical terms. Vertical nodal deflections exhibited maximum relative errors of no more than 0.008 percent, while total elastic strain energy calculations tracked full-order solutions with relative errors remaining strictly under two percent across all evaluated cases.

Building upon this linear foundation, the investigators extended the methodology to geometrically non-linear static scenarios, where structural stiffness continually changes as a function of evolving deformation. In an initial non-linear demonstration with fixed material properties, a four-point vertical load was ramped from zero to 400 kilonewtons over a normalized one-second interval, inducing a maximum plate deflection of 7.32 centimeters. Solving the governing equations required 118 incremental time steps utilizing a generalized trapezoidal integration scheme to construct average tangent stiffness operators. A second proper orthogonal decomposition reduced the inverted matrix snapshots to five governing modal coordinates. The machine learning model learned to predict these reduced coefficients by taking the previous time increment’s reduced displacements, the current time step, and the incremental step size as input features. The reconstructed displacements closely mirrored reference finite element solutions, exhibiting a maximum deviation of less than 7.5 percent.

Fifty-nine simulations were utilized for training the algorithm, leaving six completely independent use cases for rigorous validation testing. In this fully parametric non-linear regime, seven reduced coordinates were extracted and normalized within the unit interval to facilitate model training. Although higher-order modal coefficients displayed increased visual fluctuations across parameter transitions, their overall contribution to physical deformation remained small. Consequently, the predicted displacement fields accurately matched both classical proper orthogonal decomposition reductions and full-order solver trajectories. Across all unseen test cases, the maximum displacement and rotation error never exceeded 8.7 percent, while the average relative error across all test instances remained at approximately three percent.

To verify whether the technique could scale beyond academic plate models to complex aerospace structures, the authors examined an airfoil reinforced by three welded internal truss beams subjected to a uniform lower surface pressure of 2.7 megapascals. Meshed with shell elements, the cantilevered structure comprised 1,050 nodes and 6,300 degrees of freedom, presenting pronounced geometric non-linearity alongside linear elastic material behavior. The analysis considered three structural parameters: shared Young’s modulus from 120 to 300 gigapascals, outer skin thickness from 1 to 5 millimeters, and internal truss thickness from 5 to 7.5 millimeters. One hundred simulations were generated in Abaqus, including eight boundary extremes and 92 Latin hypercube samples. Using grid search optimization to configure random forests predicting nine Z-score normalized modal coefficients, the model achieved training errors below 0.25 percent. Across fifteen unseen validation simulations, average displacement errors remained at 0.7 percent, with only a single extreme case reaching seven percent error.

The computational savings achieved across these benchmarks underscore the practical potential of the minimally intrusive strategy for engineering workflows. Running the one hundred full-order airfoil simulations sequentially in Abaqus required approximately one hour of processing, whereas the trained reduced order model delivered comprehensive multi-step structural predictions in just thirty seconds. Offline data preparation, primarily involving the extraction and processing of stiffness matrices, required roughly fifteen minutes, while actual random forest training concluded in seconds. Similar speedups were documented in the plate tests, where full 65-simulation sequences dropped from 25 minutes to negligible fractions of a minute. Crucially, the method bypasses the expensive recalculation of tangent stiffness matrices during traditional Newton-Raphson non-linear iterations. By combining commercial solver accessibility with rapid machine learning matrix reconstruction, the framework provides a viable pathway toward real-time digital twins for structural health monitoring and rapid industrial design optimization.

Subject of Research: Machine learning reduced order modeling for structural mechanics

Article Title: Machine learning (ML) based reduced order modelling (ROM) for linear and non-linear solid and structural mechanics

Article References: Tannous, M., Ghnatios, C., Fonn, E., Kvamsdal, T., & Chinesta, F. (2025). Machine learning (ML) based reduced order modelling (ROM) for linear and non-linear solid and structural mechanics. Advanced Modeling and Simulation in Engineering Sciences, 12(1), Article 15. https://doi.org/10.1186/s40323-025-00299-1

Image Credits: AI Generated

DOI: 10.1186/s40323-025-00299-1

Keywords: reduced order modelling, proper orthogonal decomposition, structural mechanics, random forest regressor, finite element analysis, geometric non-linearity, digital twins, stiffness matrix inversion, singular value decomposition, principal component analysis, elastostatics, structural health monitoring, surrogate modeling, Abaqus integration, Chebyshev polynomials, Latin hypercube sampling, airfoil deformation