Multiscale Data-Driven Framework

• Multiscale data-driven framework is a modeling approach that transforms fine-scale simulations and measurements into reduced representations and closure relations.
• It organizes hierarchical scales by decomposing, averaging, and learning unresolved dynamics to integrate micro and macro-level behaviors in applications like fracture mechanics and porous media.
• The framework leverages machine learning tools, including Gaussian processes and deep neural networks, to build surrogate models that enhance efficiency, accuracy, and interpretability.

A multiscale data-driven framework denotes, in current computational science usage, a class of workflows in which fine-scale simulations, measurements, or heterogeneous observations are transformed into reduced representations, closure relations, or surrogates that can be embedded in a coarser-scale solver, optimizer, or digital twin. Across fracture mechanics, porous media, reduced-order modeling, topology optimization, stochastic dynamics, turbulence, and musculoskeletal modeling, the common objective is to retain the influence of unresolved structure—microstructure morphology, subgrid fluctuations, hidden modes, or patient-specific tissue state—without repeatedly solving the full fine-scale problem (Deng et al., 2022, Coltman et al., 2023, Vasilyeva et al., 24 Jun 2026, Paccini et al., 13 Jun 2025).

1. Scale organization and problem classes

The term multiscale is used in several distinct but structurally related senses. In concurrent multiscale fracture simulation, each macroscopic integration point is assigned an RVE, and macroscale deformation gradients are exchanged with microscale effective stress responses in a nested FE$^2$-style simulation; the unresolved variability is manufacturing-induced porosity, which changes strain localization, damage initiation, fracture path, and load capacity (Deng et al., 2022). In porous-media parametrization, the coarse description is the REV-scale model, while pore-scale simulations provide the data needed to infer closure terms and effective parameters (Coltman et al., 2023). In operator-learning and reduced-order settings, the scale split is expressed in terms of fine grids and coarse grids, or resolved and unresolved modal subspaces, rather than physical specimens and RVEs (Vasilyeva et al., 24 Jun 2026, 2208.00059).

Hierarchical formulations make this scale separation explicit. One porous-media framework assumes $\ell \ll l \ll L$, where $\ell$ is the heterogeneity scale, $l$ the REV scale, and $L$ the macroscale body scale (Ashworth et al., 2021). Reduced-order multiscale formulations instead distinguish large resolved modes, small resolved modes, and unresolved modes, or large, medium, and small scales, depending on the closure architecture (Mou et al., 2020, Mou et al., 2022). Digital-twin formulations extend the notion further by combining Macroscale, Mesoscale, Microscale, and Subcellular/Nanoscale data into a patient-specific representation (Paccini et al., 13 Jun 2025).

This breadth shows that “multiscale” does not refer to a single numerical technique. It refers to a modeling commitment: important dynamics or constitutive effects are distributed across levels of description, and the framework must specify how information moves between those levels.

2. Mathematical structure: decomposition, averaging, and closure

A defining feature of these frameworks is the conversion of unresolved content into an explicit closure object. In REV-scale porous-media averaging, a pore-scale field is decomposed as

$m=\langle m\rangle+\tilde m,$

and the averaged model differs from the closed REV model by a closure term $\mathcal{F}(\boldsymbol{\alpha},\dots)$, where the effective parameters $\boldsymbol{\alpha}$ are identified by minimizing the mismatch between averaged pore-scale fluctuations and the REV-scale closure (Coltman et al., 2023). In variational multiscale formulations, the same principle appears as a scale split in function space,

$u=u_h+u',$

or, for POD-ROMs,

$X^d=X^L\oplus X^S,$

so that unresolved or small resolved scales enter the coarse equations through explicit residual or interaction terms (Pradhan et al., 2021, 2208.00059).

Several frameworks make the closure structure itself the main learning target. Data-driven VMS-ROMs identify exact ROM closure terms induced by truncation and then approximate them with learned operators—vectors, matrices, and tensors—rather than with phenomenological eddy-viscosity prescriptions (Mou et al., 2020). Residual-based variants go further by learning a map from the ROM residual to the unresolved subscale effect, thereby tying the closure to the defect of the resolved equations rather than to reduced coefficients alone (2208.00059). In stiff multiscale random dynamics, the unresolved term appears as a state-dependent conditional expectation in an exact reduced-order PDF equation, and sparse-data discrepancies are represented by a kinetic-like defect term added to the RoPDF evolution equation (Maltba et al., 2023).

A related but more explicit strategy is to augment the coarse model with new modeled small-scale state variables. In the turbulence effective-field-theory formulation, the subgrid stress is decomposed as

$\ell \ll l \ll L$0

and the genuinely unresolved part $\ell \ll l \ll L$1 is promoted to an additional dynamical field with its own evolution law, rather than being algebraically slaved to the resolved velocity field (Choi et al., 13 Nov 2025). This is still a closure strategy, but one in which the closure is a dynamical subsystem rather than a static constitutive map.

3. Recurrent workflow patterns

Taken together, the literature exhibits a recurring workflow. First, fine-scale or heterogeneous data are generated or acquired. Second, those data are reduced, averaged, or compressed into a lower-dimensional representation. Third, a surrogate, closure law, or constitutive map is trained or calibrated. Fourth, the learned object is embedded in an online simulation, optimization loop, or digital twin. The specific implementations differ, but the architecture is recognizably shared.

Some frameworks state this structure explicitly. The DuMu\textsuperscript{x}-based porous-media parametrization is divided into four parts: Pore-scale data development, Averaging and REV-scale model parametrization, Generalization with machine learning, and Analysis using descriptive metrics (Coltman et al., 2023). The nonlinear-flow neural-operator framework uses an offline stage that generates random permeability realizations, solves the fine nonlinear PDE, constructs local multiscale basis functions via local POD and partition of unity, and trains a branch network to predict reduced coefficients; its online stage then evaluates the learned operator and optionally applies local correction iterations (Vasilyeva et al., 24 Jun 2026). The process-guided composite workflow follows the sequence process model → microstructure descriptors → RVE database → DMN training/transfer learning → online constitutive law in FE simulation (Liu et al., 2020).

Sequential updating appears when observations arrive in time. Dynamic data-driven Bayesian GMsFEM partitions the temporal domain into subintervals, computes a fixed solution with permanent basis functions, and then uses residuals plus dynamic observations to select additional basis functions probabilistically from the posterior distribution (Cheung et al., 2018). The musculoskeletal digital twin uses a two-level integration system in which multimodal acquisition is first fused into a common patient assessment and then organized into a graph structure for inference about diagnosis, treatment, and trajectory (Paccini et al., 13 Jun 2025).

A plausible implication is that “data-driven” in this context is rarely synonymous with end-to-end black-box regression. More often it denotes a staged architecture in which data intervene at a specific multiscale bottleneck: basis construction, closure identification, constitutive parameter regression, or state assimilation.

4. Learning mechanisms and surrogate architectures

The learning component varies widely, but several families recur. Gaussian-process models are prominent when the closure depends on mixed continuous and categorical inputs. Latent map Gaussian processes calibrate reduced-order fracture models as functions of microstructure and fidelity level, so that clustered ROMs faithfully surrogate DNS across different clustering resolutions (Deng et al., 2022). Latent-variable Gaussian processes embed discrete microstructure classes into a continuous latent space, yielding a differentiable stiffness map that can be optimized jointly with volume fraction in cellular-composite topology optimization (Wang et al., 2021). Standard Gaussian-process surrogates are also used to emulate effective Lamé constants from pore morphology, or constitutive responses of spinodoid and soft functionally graded materials, thereby avoiding repeated homogenization inside the optimization loop (Deng et al., 2022, Deng et al., 29 Jun 2025, Deng et al., 29 Jun 2025).

Image-based and operator-learning architectures appear when the fine-scale geometry itself is the natural input. The DuMu\textsuperscript{x} framework trains a convolutional neural network on binary pore-geometry images, with custom periodic padding to respect periodic microstructures and with optional descriptive metrics injected into dense layers (Coltman et al., 2023). For nonlinear flow in random heterogeneous porous media, a DeepONet-like construction uses a fixed multiscale basis as the trunk and a branch MLP that maps log-permeability fields to reduced coefficients; the reconstructed pressure is then obtained as $\ell \ll l \ll L$2 (Vasilyeva et al., 24 Jun 2026). Self-similar dynamics are treated with a two-stage multiscale neural framework that first learns microscopic dynamics and then learns a coarse-graining map under a self-similarity prior, using consistency between coarse-grain-then-evolve and evolve-then-coarse-grain as the key diagnostic (Tao et al., 2023).

Mechanistic neural architectures occupy an intermediate position between pure surrogates and purely analytical reduction. The variational super-resolution neural network factorizes the unresolved state into functions of coarse-scale coefficients and physics-informed parameters such as the cell Peclet number $\ell \ll l \ll L$3, CFL, or $\ell \ll l \ll L$4, so that learning the subscales becomes equivalent to learning projected element Green’s-function coefficients in the linear setting (Pradhan et al., 2021). Deep Material Networks represent an RVE as a binary tree of two-layer mechanistic building blocks with closed-form homogenization and rotation laws; trained on linear-elastic RVE data, they extrapolate to nonlinear constitutive behavior and can be concatenated into three-scale networks for CFRP systems (Liu et al., 2019, Liu et al., 2020).

A separate line of work uses data not to predict effective properties directly, but to identify closure operators inside reduced equations. This includes least-squares identification of VMS-ROM closure matrices and tensors (Mou et al., 2020), residual-based closure maps in parameter-dependent convection-diffusion (2208.00059), and conditional linear closures that make the hidden medium-scale dynamics conditionally Gaussian and therefore analytically tractable for filtering (Mou et al., 2022).

5. Modes of coupling: simulation, optimization, assimilation, and digital twins

Once learned, the multiscale object must be coupled back to a larger computation. In concurrent multiscale mechanics, this usually means constitutive updates at macroscale quadrature points. The fracture framework replaces direct microscale DNS with a mechanistic ROM calibrated by LMGP, so that the macroscale FE$\ell \ll l \ll L$5-style analysis sees a calibrated microscale damage response rather than raw clustered dynamics (Deng et al., 2022). The process-guided composite framework similarly maps process-induced descriptors to a DMN database that returns homogenized stress and tangent stiffness at each structural integration point during concurrent simulation (Liu et al., 2020). In dual-porosity flow, the learned constitutive map is injected explicitly into the Newton residual of the macroscale simulator as a replacement for the classical linear inter-porosity transfer law (Ashworth et al., 2021).

In topology optimization, the learned multiscale model becomes part of the design map. One framework uses a six-bar unit-cell database plus a neural surrogate for geometry-to-property mapping, then reconstructs a high-resolution design from homogenization-based optimization via a sawtooth-function-based conformal mapping that supports oriented and compatible tiling (Wang et al., 2021). Another uses latent vectors as continuous design variables so that the optimizer can move smoothly between microstructure classes while maximizing the first natural frequency (Wang et al., 2021). Spinodoid and soft-FGM frameworks go further by reformulating the design field as neural-network weights, letting automatic differentiation propagate sensitivities through descriptor fields, surrogate constitutive laws, finite-element equilibrium, and compliance or potential-energy objectives (Deng et al., 29 Jun 2025, Deng et al., 29 Jun 2025).

Assimilation-oriented frameworks instead couple the learned model to streaming or sparse observations. Dynamic Bayesian GMsFEM uses residual-driven priors and observation-driven likelihoods to activate additional basis functions beyond permanent GMsFEM modes (Cheung et al., 2018). RoPDF-based uncertainty propagation augments reduced PDF equations with learned defect terms and then assimilates sparse, low-fidelity PDF observations via nudging or deep neural networks (Maltba et al., 2023). The musculoskeletal digital twin uses the coupling layer to fuse motion capture, ultrasound, sEMG, CT, MRI, and EHR-derived features into a graph-based patient representation that supports inference and interactive exploration (Paccini et al., 13 Jun 2025).

6. Accuracy, efficiency, and interpretability

Reported gains are strongly problem-dependent, but the numerical evidence is substantial. In multiscale fracture, a single DNS micro-damage simulation took about 29.8 hours, whereas ROMs took about 68.8, 27.9, and 15.6 minutes for $\ell \ll l \ll L$6, and the full multiscale analysis with calibrated ROMs was reported to be about 172.6× faster than the DNS-based FE$\ell \ll l \ll L$7 simulation; the same study reported that including spatially varying micro-porosity reduced an L-bracket’s load-carrying capacity from 70.86 N to 63.62 N (Deng et al., 2022). In nonlinear flow, the learned multiscale neural operator was about 3 orders of magnitude faster than the Darcy fine solve, about 4 orders of magnitude faster than nonlinear fine solves, and about 3 orders of magnitude faster than the corresponding Galerkin reduced solve at the finest $\ell \ll l \ll L$8 resolution (Vasilyeva et al., 24 Jun 2026). In generalized de-homogenization, the proposed multiscale design pipeline was reported to be over 5700× faster than direct fine-mesh optimization (Wang et al., 2021). In the SFRC setting, a DNS RVE with about 360,000 tetrahedral elements took about 1100 s for one loading path on 8 CPUs, whereas the DMN surrogate with 33 DOFs took about 3 s on 1 CPU (Liu et al., 2020).

Interpretability is not uniform across the literature, but it is a recurrent design goal. The DuMu\textsuperscript{x} framework correlates learned REV-scale parameters with descriptive pore metrics such as porosity, specific surface area, directionality, tortuosity, and maximum flow (Coltman et al., 2023). LVGP-based topology optimization interprets distance in latent space as a measure of correlation between microstructure classes (Wang et al., 2021). The turbulence effective-field-theory framework preserves explicit equations for both resolved scales and the modeled small-scale tensor field $\ell \ll l \ll L$9, rather than hiding the closure in an opaque network (Choi et al., 13 Nov 2025). Spinodoid frameworks are explicit that anisotropic and isotropic microstructures are favored in different regions for mechanistic reasons tied to local stress trajectories, and that the descriptors—density, orientation, anisotropy indices, and frequency-related parameters—are physically interpretable (Deng et al., 29 Jun 2025, Deng et al., 29 Jun 2025).

This body of evidence suggests that efficiency and interpretability need not be mutually exclusive. Many successful multiscale data-driven frameworks achieve speed by learning only the multiscale bottleneck, not by replacing the entire computational pipeline.

7. Misconceptions, limitations, and open issues

A common misconception is that coarsening is merely a computational expedient. In fracture ROMs, clustering is physically intrusive: clustered models smear local strain concentrations, reduce plastic strain peaks, delay damage initiation, and make UTS and toughness artificially higher unless the ROM is calibrated back to DNS (Deng et al., 2022). A related misconception is that adding more reduced modes or larger learned representations necessarily improves prediction. In nonlinear-flow neural operators, increasing the trunk dimension does not always improve accuracy because the higher coefficient dimension can make learning harder (Vasilyeva et al., 24 Jun 2026).

Another misconception is that “data-driven” implies “black-box.” Several frameworks explicitly reject that reading. VMS-based super-resolution defines unresolved scales through mathematically specified finite-dimensional subspaces and $\ell$0-projected data (Pradhan et al., 2021). DMN uses mechanistic building blocks rather than unrestricted function approximation (Liu et al., 2019). The turbulence effective theory is constrained by translations, rotations, and Galilean invariance through symmetry-aware sparse discovery (Choi et al., 13 Nov 2025). Even when deep learning is used, it is often embedded in a structured closure or homogenization pipeline.

The limitations are equally explicit. The musculoskeletal digital twin notes that accuracy depends on input data quality and completeness, that standardization and interoperability remain difficult, and that continuous updating is required to maintain twin fidelity (Paccini et al., 13 Jun 2025). VMS super-resolution emphasizes that predictive success is not guaranteed for severely underresolved flows, and that offline reconstruction does not automatically yield a robust online closure unless the training and online states are consistent (Pradhan et al., 2021). VMS-stabilized DDCM approximations show that noisy data can violate local thermodynamic consistency, with the sign condition $\ell$1 failing in some regions (Codina et al., 23 Dec 2025). In spinodoid design, very strong anisotropy can become truss-like and cease to be spinodoid-like, so admissible parameter ranges are deliberately restricted (Deng et al., 29 Jun 2025).

Open directions follow directly from these limitations. The literature points toward tighter integration of surrogate learning with solver-consistent training data, stronger uncertainty quantification and data assimilation, better management of noisy or sparse measurements, and broader extensions from linear or weakly nonlinear homogenization to fracture, damage, and strongly path-dependent multiscale dynamics. A plausible implication is that the field is converging on a hybrid principle: preserve mechanistic structure where it is reliable, and learn only the unresolved operators, parameters, or state variables that block practical multiscale computation.