Source-linked AI summary
Machine-learning-assisted multiscale topology optimization of functionally graded superimposed lattice structures
Prashant Kumar Gupta, Jonathan Stollberg, Dominik Schillinger, Mohammad Ashraf Iqbal
TL;DR
Multiscale topology optimization of functionally graded lattices is limited by repeated computational homogenization and the need for physically admissible stiffness predictions. The paper combines a parameterized BCC–FCC–simple-cubic unit cell with Cholesky-constrained stiffness and density surrogates in a two-stage optimization framework. On a three-dimensional MBB beam, it generates spatially varying lattice fields while avoiding repeated online homogenization and satisfying the prescribed material constraint.
Problem
Repeated homogenization limits computationally attractive multiscale topology optimization, while unconstrained stiffness surrogates can produce non-physical tensors.
Method
The framework uses a BCC–FCC–simple-cubic unit cell, Cholesky-constrained stiffness and density neural surrogates, and two-stage macroscale–microscale optimization.
Results
The MBB beam benchmark produces spatially varying lattice parameters and relative-density fields while avoiding repeated online homogenization and satisfying the prescribed material constraint.
Takeaways & Limitations
The three-parameter superimposed lattice provides an interpretable design space for spatially tuning effective stiffness and relative density.
Takeaways & Limitations
The optimized fields are approximate because surrogate error propagates through the assembled stiffness matrix, and direct validation against homogenization remains future work.
Abstract
from arXiv · showhide
Functionally graded lattice structures enable lightweight designs with spatially tunable stiffness and density, but their use in multiscale topology optimization is limited by the cost of repeated computational homogenization. This work presents a machine learning-assisted multiscale optimization framework for regular superimposed lattice structures. The unit cell is formed by combining body-centered cubic, face-centered cubic, and simple cubic lattice components, each controlled by an independent geometric parameter. Offline computational homogenization is used to generate effective stiffness data, which are then used to train a Cholesky-constrained neural network surrogate. This representation reconstructs the homogenized stiffness tensor in a physically admissible form. A separate neural network is trained to predict relative density from Monte Carlo-based density estimates. We incorporate our surrogates into a two-stage topology optimization strategy. First, a macroscale topology is obtained using the solid isotropic material with penalization (SIMP) method. The resulting solid region is then used for microscale lattice optimization, where the local lattice parameters are updated using the method of moving asymptotes (MMA). The trained stiffness and density surrogates replace repeated online homogenization during this stage. The method is demonstrated on a three-dimensional Messerschmitt-Bölkow-Blohm (MBB) beam benchmark, producing spatially varying lattice parameters and relative density fields consistent with compliance minimization under a material constraint.
1. Introduction
The paper motivates a multiscale topology-optimization framework for functionally graded lattices, addressing the cost and physical-admissibility challenges of surrogate-based homogenization while introducing a superimposed lattice design space.
- Research gap: Conventional SIMP determines macroscale material placement but does not directly control microstructural architecture.This motivates coupling structural layout with microscale lattice design.
- Research gap: Functionally graded lattices vary unit-cell parameters spatially, assigning stiffer or denser microstructures to critical regions and lighter configurations elsewhere.Their effective stiffness, density, anisotropy, and manufacturability depend strongly on unit-cell geometry.
- Research gap: Homogenization-based multiscale optimization links representative-volume-element behavior to the macroscale continuum but repeatedly evaluating unit cells is computationally costly.Parametric lattice cells reduce design-variable counts while retaining homogenization-based microscale-to-macroscale coupling.
- Surrogate modeling: Neural-network surrogates can accelerate stiffness and density prediction, but unconstrained stiffness-entry predictions may violate symmetry or positive definiteness.Physically constrained architectures instead reconstruct stiffness from constrained factors.
- Proposed framework: The proposed framework superimposes BCC, FCC, and simple-cubic components with independent geometric parameters and embeds stiffness and density surrogates in two-stage optimization.Its contributions include Monte Carlo-based density estimation, Cholesky-constrained stiffness prediction, and a three-dimensional benchmark demonstration.
2. Machine-learning surrogate modeling of the superimposed lattice material
The framework represents superimposed lattice materials with stiffness and density surrogates trained offline, enabling multiscale optimization without repeated online homogenization. The lattice combines independently controlled BCC, FCC, and simple cubic components, while Cholesky reconstruction preserves physical admissibility of predicted stiffness.
- Superimposed lattice unit cell: The microscale material is a regular unit cell combining BCC, FCC, and simple cubic lattice components with three independently controlled geometric parameters.The parameters are strut-diameter-to-unit-cell-edge aspect ratios for the three components.
- Superimposed lattice unit cell: Changing the three lattice contributions creates a continuous, interpretable design space for locally tuning stiffness and density across functionally graded structures.Different structural regions can therefore use different effective material responses within the same lattice family.
- Computational homogenization: Computational homogenization uses periodic unit cells under linear small-strain elasticity, with scale separation allowing the lattice to be replaced by an equivalent macroscale continuum.The unit cell is meshed with beam elements, connected through rigid joints, and analyzed using periodic boundary conditions.
- Cholesky-constrained stiffness surrogate: The stiffness surrogate predicts a lower-triangular Cholesky factor with positive diagonal entries and reconstructs the homogenized tensor as C* = G G^T.This architecture automatically preserves symmetry and positive definiteness rather than predicting stiffness entries independently.
- Cholesky-constrained stiffness surrogate: After 50,000 epochs, stiffness-surrogate mean squared error reached 8.9×10^-4 on training data and 8.7 × 10^-4 on test data, with losses decreasing by more than three orders of magnitude.Across 386 held-out samples, the surrogate achieved R2 = 0.974, aggregated relative L2 error of 9.2 %, and RMSE of 0.32.
- Relative density surrogate: The density surrogate predicts relative density from the three aspect ratios, using Monte Carlo density estimates and a sigmoid output constrained to (0, 1).On 300 test samples, it achieved R2 = 0.965, aggregated relative L2 error of 8.0 %, and RMSE of 1.1×10^-2.
3. Two-stage multiscale topology optimization
The framework first determines a macroscale material layout with SIMP, then optimizes lattice parameters within the resulting solid region using differentiable stiffness and density surrogates with MMA.
- Two-stage formulation: The two-stage formulation uses SIMP for macroscale layout and optimizes local lattice parameters only within the converged solid region.The solid region is identified by thresholding the macroscale density field before microscale optimization.
- Macroscale topology optimization: Macroscale compliance minimization uses pseudo-density variables, penalized element stiffness matrices, equilibrium displacements, and a prescribed total-volume upper bound.The pseudo-densities are topological design variables rather than physical relative densities and are updated with the OC method.
- Macroscale topology optimization: The macroscale sensitivities are filtered and used in an optimality-criteria update with move limits, damping, and bisection to satisfy the volume constraint.The optimization terminates when constraints are satisfied and compliance remains stable over successive iterations.
- Microscale lattice optimization: The microscale element stiffness matrix is computed from the neural-network-predicted homogenized stiffness tensor, with the strain-displacement matrix entering its finite-element construction.The local aspect-ratio vector supplies the lattice input while the base-material Young’s modulus and Poisson’s ratio remain fixed.
- Microscale lattice optimization: At the microscale, MMA updates three local lattice parameters while differentiable surrogates provide stiffness and density predictions for sensitivity-based optimization.Forward-mode automatic differentiation supplies sensitivities; predictions assemble the global stiffness matrix, solve equilibrium, evaluate compliance, and enforce the material constraint.
4. Computational results
The MBB beam demonstration combines macroscale SIMP optimization with surrogate-based microscale lattice optimization, producing graded lattice parameters and density fields under a material constraint.
- Benchmark and macroscale optimization: The three-dimensional MBB beam uses a 20 × 5 × 5 domain discretized into 13,500 hexahedral elements, with a downward mid-span load and four pinned lower corners.The isotropic base material has E = 45 and ν = 0.3.
- Benchmark and macroscale optimization: 15.6 % of the initial relative compliance remains after macroscale SIMP optimization, while the volume constraint is satisfied throughout.The relative compliance decreases monotonically, approaching its final value after 12 iterations.
- Microscale lattice optimization: The thresholded SIMP solid region defines where microscale lattice optimization is performed, excluding void elements from the second-stage design space.Elements with γe ≥ 0.5 are retained as solid, while lower-density elements are treated as void.
- Microscale lattice optimization: The MMA microscale optimization restores feasibility within two iterations and terminates after 32 iterations with final compliance 1.46 times the initial over-material design.The constraint value is 8 × 10^-4 at k = 2 and becomes negative from k = 3 onward; compliance reaches 1.79 times the initial value during feasibility restoration.
- Optimized lattice fields: BCC and simple cubic parameters concentrate near the top and bottom flanges, whereas the FCC parameter remains comparatively large through the web.FCC face-diagonal struts contribute shear stiffness in the web, while BCC and simple cubic struts reinforce flanges against normal stresses.
- Optimized lattice fields: The optimized relative density ranges from 0.21 in the flanges to 0.03 in the central web, concentrating material along high-bending-stress regions.The resulting density trend is consistent with compliance minimization under the prescribed material constraint.
- Discussion: The framework avoids repeated online homogenization by using stiffness and density surrogates, while outputting spatially varying lattice parameters and relative density fields.Compared with SIMP alone, the result includes interpretable microstructural information associated with the BCC, FCC, and simple cubic families.
- Discussion: The stiffness surrogate has 9.2 % aggregate relative L2 error, increasing to 13.0 % for the normal-stiffness factor G33, so optimized fields remain approximate.Error propagation into the converged design was not quantitatively assessed against direct homogenization.
5. Conclusion and outlook
The paper concludes that interpretable superimposed lattice parameters can support spatial tuning of stiffness and density through a surrogate-assisted two-stage optimization framework. The MBB beam results show graded fields consistent with load-carrying behavior, while several validation and manufacturing extensions remain for future work.
- Conclusion: The regular superimposed unit cell combines BCC, FCC, and simple cubic components, each controlled by an independent geometric parameter.This creates an interpretable design space for spatially tuning effective stiffness and relative density.
- Conclusion: Offline homogenization and two neural surrogates replace repeated online unit-cell homogenization within the two-stage SIMP–MMA optimization.The stiffness surrogate predicts physically admissible homogenized tensors, while the density surrogate predicts relative density from Monte Carlo estimates.
- Results: The MBB beam demonstration generates spatially varying a1, a2, and a3 fields together with a corresponding relative density field consistent with expected load-carrying behavior.The microscale convergence history indicates stable optimization while satisfying the prescribed material constraint.
- Outlook: Future work includes stronger macro–micro coupling, process- and geometry-specific manufacturing constraints, and validation through dehomogenized simulations and experiments.Extensions to additional unit-cell families and nonlinear, dynamic, or multiphysics responses are also proposed.
CRediT authorship contribution statement
The authors’ contributions span conception, methodology, software, analysis, visualization, supervision, and manuscript preparation.
- Authorship contributions: Prashant Kumar Gupta led conceptualization, methodology, software, data curation, formal analysis, investigation, visualization, and the original draft.Jonathan Stollberg contributed supervision, conceptualization, methodology, software, visualization, and review and editing.
- Authorship contributions: Dominik Schillinger and Mohammad Ashraf Iqbal contributed supervision, conceptualization, and writing review and editing.
Appendix A. Symmetry relations for the Cholesky factor
Appendix A specifies symmetry-based recovery of dependent Cholesky-factor entries and positivity constraints needed for a physically admissible stiffness representation.
- Symmetry relations: The dependent entries G66, G32, and G31 are recovered from six independently predicted lower-triangular Cholesky entries under tetragonal material symmetry.These relations support reconstruction of the stiffness tensor for the superimposed lattice family.
- Positivity constraints: The Cholesky diagonal entries G11, G22, G33, G44, and G55 are constrained to be positive.These constraints help ensure physically admissible tensor reconstruction.
- Positivity constraints: Additional constraints ensure that the expression inside the square root in Eq. (A.2) is non-negative.This prevents invalid values when recovering dependent factor entries.