Scaling atom-by-atom inverse design with nano-topology optimization and diffusion models
Chun-Teh Chen, Denvid Lau · 2026
Read Scaling atom-by-atom inverse design with nano-topology optimization and diffusion models by Chun-Teh Chen, Denvid Lau (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-03276.txt.
Introduction
Chen and Lau design tiny metal structures atom by atom, mixing nano-topology optimization with conditional diffusion models. Treating each atom as a design variable and reading stiffness from energy curvature, their method handles designs above 650,000 atoms. Tests on aluminum cantilevers reveal a surface-driven selection rule about walls versus trusses, and diffusion models trained on that data yield diverse strong candidates.
Catalog record
| CREATOR | Chun-Teh Chen, Denvid Lau |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 96019 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.03276
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Scaling atom-by-atom inverse design with nano-topology optimization and diffusion models Chun-Teh Chen, Denvid Lau Abstract The mechanical properties of metallic nanostructures are governed not only by topology but also by crystal symmetry and face-specific surface physics, which are typically absent from continuum topology optimization. We develop an atom-by-atom inverse design framework that combines Nano-Topology Optimization (Nano-TO) with conditional denoising diffusion probabilistic models. Nano-TO treats each atom as a discrete design variable and evaluates stiffness from the symmetric curvature of the total energy, removing residual surface-stress bias. A crystallography-aligned multi-shell sensitivity filter stabilizes the optimization and enables designs containing more than 6.5 × 10 5 6.5\times 10^{5} atoms. Using aluminum nanocantilevers, we identify a surface-physics-driven topology selection rule: thickness-periodic beams favor brace-dominated trusses, whereas finite-thickness beams favor nearly closed walls that provide efficient shear paths and reduce surface penalties. At sufficiently small scales, these walls become mechanically unstable, and truss-like layouts reappear. In nanopillar studies, atomistic optimization outperforms continuum topology-optimized designs. Finally, conditional diffusion models trained on Nano-TO data generate diverse high-performance candidates near the optimization frontier. These results establish nanoscale inverse design as a coupled problem of topology and surface physics. At nanometer scales, topology not only influences how forces are transmitted through a nanostructure but also dictates which crystallographic facets, edges, and low-coordination atomic sites are exposed. This is especially important in micro- and nano-electromechanical systems (MEMS/NEMS), including resonators, sensors, and scanning probes, whose performance depends on elastic response [ 1 , 2 ] . At macroscopic scales, continuum elasticity is usually adequate as atomic details can be averaged without significantly altering the predicted behavior. At nanoscale dimensions, however, a large fraction of atoms are located at free surfaces, where reduced coordination, relaxation, and facet-dependent bonding alter residual stress and elasticity. Mechanical response, therefore, depends jointly on topology, crystal symmetry, and face-specific surface physics [ 3 , 4 , 5 ] . Experiments and atomistic simulations on nanowires have shown pronounced size effects in the effective axial modulus, with the modulus either decreasing or increasing with radius, depending on material, wire orientation, and exposed facets [ 6 , 7 , 8 , 9 , 10 ] . Nanoscale inverse design must therefore optimize not only the distribution of atoms but also the atomic surfaces this topology creates. Topology optimization (TO) provides a powerful framework for structural layout design [ 11 , 12 , 13 , 14 , 15 ] , and density-based implementations now scale to very large continuum problems [ 16 , 17 , 18 ] . Standard TO, however, treats the solid as a homogeneous medium. It optimizes the coarse geometry of a structure, but it does not specify which crystallographic facets, edges, or local atomic motifs are created by that geometry. Continuum extensions based on surface elasticity, including Gurtin–Murdoch models [ 19 , 20 , 21 ] , and higher-order theories such as strain-gradient and couple-stress formulations [ 22 , 23 ] , can capture partial size effects by introducing effective surface constitutive laws and intrinsic length scales. These approaches have been valuable for predicting how surfaces change the mechanical responses of prescribed nanostructures and, in some cases, for incorporating surface effects into continuum-level shape or topology optimization. However, they still describe both bulk and surface in homogenized form. Consequently, these approaches cannot directly resolve atomistic realizations of a coarse-grained surface orientation, such as surface terminations, atomic steps and terraces, local coordination changes, or discrete defects. This limitation becomes especially important in atomistic inverse design, as changing the topology at the nanoscale simultaneously alters the populations of exposed facets, edges, and low-coordination sites that jointly determine target mechanical responses. Nano-Topology Optimization (Nano-TO) addresses this atomistic inverse-design gap by treating each atom as a discrete design variable [ 24 ] . Rather than predicting surface effects for a fixed geometry, Nano-TO allows topology and the surfaces created by that topology to be determined together. This formulation can, in principle, resolve the discrete surface and lattice physics that continuum models omit. In practice, however, atomistic inverse design is much harder to scale. Each design update must be evaluated through nonlinear relaxations under interatomic potentials, and the resulting per-atom sensitivities become noisy and unstable if used too locally. This instability limits accessible system size and can disrupt the formation of coherent load-bearing paths. A second challenge is non-uniqueness. For a given set of target properties, there is generally not a single admissible nanostructure, but rather a family of distinct atomistic topologies with comparable performance. Deterministic optimization can find one feasible design, but it does not, by itself, map the broader near-optimal design manifold or expose useful trade-offs among the properties of interest. We address both limitations by combining atomistic topology optimization with generative modeling. Building on our earlier Nano-TO framework [ 24 ] , we formulate stiffness through the symmetric energy-curvature measure that removes residual surface-stress bias, and we introduce a crystallography-aligned multi-shell sensitivity filter that regularizes per-atom sensitivities sufficiently to enable stable large-batch Nano-TO in systems exceeding 6.5\times 10^{5} atoms. More broadly, filtering and minimum-length-scale control have long served as regularization tools in TO [ 25 , 26 , 27 ] , although here the neighborhood is chosen to reflect crystallography connectivity and the interaction range of the interatomic potential. We then couple the resulting optimization data to conditional denoising diffusion probabilistic models (c-DDPMs) [ 28 , 29 , 30 ] . Recent studies have applied generative models, including diffusion-based models, to inverse design problems [ 31 , 32 , 33 , 34 , 35 , 36 , 37 ] . We use c-DDPMs to sample a diverse set of target-consistent, near-optimal designs. Using aluminum nanocantilevers and nanopillars as testbeds, we show that explicit surface physics can qualitatively change the optimal topology: thickness-periodic cantilevers favor truss-like motifs, exposed side surfaces drive nearly closed-wall designs, and at a smaller scale, the optimum shifts back toward truss-like layouts as ultrathin walls lose their ability to carry transverse shear at the nanoscale. These results establish a route to inverse design in which topology and surface physics are optimized simultaneously, while generative models broaden access to high-performing alternatives and multi-objective design trade-offs. Results Nano-TO and c-DDPM frameworks To make surface physics part of the inverse-design problem, we represent each nanostructure atom by atom and evaluate its mechanical properties with an embedded-atom method (EAM) potential [ 38 , 39 ] . This description captures the facet-dependent surface elasticity of FCC metals, whose low-coordination surfaces exhibit distinct symmetries and in-plane elastic responses (Supplementary Notes A.1, Supplementary Figures S1 and S2). Changes in topology modify not only the load path, but also the populations of surfaces that contribute to stiffness. Figure 1a summarizes the resulting Nano-TO workflow, which builds on our previous atom-by-atom inverse materials design formulation [ 24 ] . For a prescribed loading mode, stiffness is evaluated from the symmetric curvature of the total energy, which removes the linear contribution from residual surface stress (Methods). Atomistic relaxation makes the per-atom sensitivities noisy and highly local, especially as system size and geometric complexity increase. We address this with a crystallography-aligned multi-shell sensitivity filter spanning the first 12 FCC shells (Supplementary Notes A.2, Supplementary Table S1, Supplementary Figure S3). This multi-shell filter suppresses atom-scale fluctuations