-
Thermodynamically-Informed Iterative Neural Operators for Heterogeneous Elastic Localization
Authors:
Conlain Kelly,
Surya R. Kalidindi
Abstract:
Engineering problems frequently require solution of governing equations with spatially-varying discontinuous coefficients. Even for linear elliptic problems, mapping large ensembles of coefficient fields to solutions can become a major computational bottleneck using traditional numerical solvers. Furthermore, machine learning methods such as neural operators struggle to fit these maps due to sharp…
▽ More
Engineering problems frequently require solution of governing equations with spatially-varying discontinuous coefficients. Even for linear elliptic problems, mapping large ensembles of coefficient fields to solutions can become a major computational bottleneck using traditional numerical solvers. Furthermore, machine learning methods such as neural operators struggle to fit these maps due to sharp transitions and high contrast in the coefficient fields and a scarcity of informative training data. In this work, we focus on a canonical problem in computational mechanics: prediction of local elastic deformation fields over heterogeneous material structures subjected to periodic boundary conditions. We construct a hybrid approximation for the coefficient-to-solution map using a Thermodynamically-informed Iterative Neural Operator (TherINO). Rather than using coefficient fields as direct inputs and iterating over a learned latent space, we employ thermodynamic encodings -- drawn from the constitutive equations -- and iterate over the solution space itself. Through an extensive series of case studies, we elucidate the advantages of these design choices in terms of efficiency, accuracy, and flexibility. We also analyze the model's stability and extrapolation properties on out-of-distribution coefficient fields and demonstrate an improved speed-accuracy tradeoff for predicting elastic quantities of interest.
△ Less
Submitted 10 November, 2024;
originally announced November 2024.
-
Recurrent Localization Networks applied to the Lippmann-Schwinger Equation
Authors:
Conlain Kelly,
Surya R. Kalidindi
Abstract:
The bulk of computational approaches for modeling physical systems in materials science derive from either analytical (i.e. physics based) or data-driven (i.e. machine-learning based) origins. In order to combine the strengths of these two approaches, we advance a novel machine learning approach for solving equations of the generalized Lippmann-Schwinger (L-S) type. In this paradigm, a given probl…
▽ More
The bulk of computational approaches for modeling physical systems in materials science derive from either analytical (i.e. physics based) or data-driven (i.e. machine-learning based) origins. In order to combine the strengths of these two approaches, we advance a novel machine learning approach for solving equations of the generalized Lippmann-Schwinger (L-S) type. In this paradigm, a given problem is converted into an equivalent L-S equation and solved as an optimization problem, where the optimization procedure is calibrated to the problem at hand. As part of a learning-based loop unrolling, we use a recurrent convolutional neural network to iteratively solve the governing equations for a field of interest. This architecture leverages the generalizability and computational efficiency of machine learning approaches, but also permits a physics-based interpretation. We demonstrate our learning approach on the two-phase elastic localization problem, where it achieves excellent accuracy on the predictions of the local (i.e., voxel-level) elastic strains. Since numerous governing equations can be converted into an equivalent L-S form, the proposed architecture has potential applications across a range of multiscale materials phenomena.
△ Less
Submitted 21 September, 2021; v1 submitted 29 January, 2021;
originally announced February 2021.
-
Emergence of the Circle in a Statistical Model of Random Cubic Graphs
Authors:
Christy Kelly,
Carlo Trugenberger,
Fabio Biancalana
Abstract:
We consider a formal discretisation of Euclidean quantum gravity defined by a statistical model of random $3$-regular graphs and making using of the Ollivier curvature, a coarse analogue of the Ricci curvature. Numerical analysis shows that the Hausdorff and spectral dimensions of the model approach $1$ in the joint classical-thermodynamic limit and we argue that the scaling limit of the model is…
▽ More
We consider a formal discretisation of Euclidean quantum gravity defined by a statistical model of random $3$-regular graphs and making using of the Ollivier curvature, a coarse analogue of the Ricci curvature. Numerical analysis shows that the Hausdorff and spectral dimensions of the model approach $1$ in the joint classical-thermodynamic limit and we argue that the scaling limit of the model is the circle of radius $r$, $S^1_r$. Given mild kinematic constraints, these claims can be proven with full mathematical rigour: speaking precisely, it may be shown that for $3$-regular graphs of girth at least $4$, any sequence of action minimising configurations converges in the sense of Gromov-Hausdorff to $S^1_r$. We also present strong evidence for the existence of a second-order phase transition through an analysis of finite size effects. This -- essentially solvable -- toy model of emergent one-dimensional geometry is meant as a controllable paradigm for the nonperturbative definition of random flat surfaces.
△ Less
Submitted 20 January, 2021; v1 submitted 26 August, 2020;
originally announced August 2020.
-
Self-Assembly of Geometric Space from Random Graphs
Authors:
Christy Kelly,
Carlo A Trugenberger,
Fabio Biancalana
Abstract:
We present a Euclidean quantum gravity model in which random graphs dynamically self-assemble into discrete manifold structures. Concretely, we consider a statistical model driven by a discretisation of the Euclidean Einstein-Hilbert action; contrary to previous approaches based on simplicial complexes and Regge calculus our discretisation is based on the Ollivier curvature, a coarse analogue of t…
▽ More
We present a Euclidean quantum gravity model in which random graphs dynamically self-assemble into discrete manifold structures. Concretely, we consider a statistical model driven by a discretisation of the Euclidean Einstein-Hilbert action; contrary to previous approaches based on simplicial complexes and Regge calculus our discretisation is based on the Ollivier curvature, a coarse analogue of the manifold Ricci curvature defined for generic graphs. The Ollivier curvature is generally difficult to evaluate due to its definition in terms of optimal transport theory, but we present a new exact expression for the Ollivier curvature in a wide class of relevant graphs purely in terms of the numbers of short cycles at an edge. This result should be of independent intrinsic interest to network theorists. Action minimising configurations prove to be cubic complexes up to defects; there are indications that such defects are dynamically suppressed in the macroscopic limit. Closer examination of a defect free model shows that certain classical configurations have a geometric interpretation and discretely approximate vacuum solutions to the Euclidean Einstein-Hilbert action. Working in a configuration space where the geometric configurations are stable vacua of the theory, we obtain direct numerical evidence for the existence of a continuous phase transition; this makes the model a UV completion of Euclidean Einstein gravity. Notably, this phase transition implies an area-law for the entropy of emerging geometric space. Certain vacua of the theory can be interpreted as baby universes; we find that these configurations appear as stable vacua in a mean field approximation of our model, but are excluded dynamically whenever the action is exact indicating the dynamical stability of geometric space. The model is intended as a setting for subsequent studies of emergent time mechanisms.
△ Less
Submitted 30 April, 2019; v1 submitted 28 January, 2019;
originally announced January 2019.
-
Direct Imaging of Dynamic Glassy Behavior in a Strained Manganite Film
Authors:
Worasom Kundhikanjana,
Zhigao Sheng,
Yongliang Yang,
Keji Lai,
Yue Ma,
Yong-Tao Cui Michael A. Kelly,
Masao Nakamura,
Masashi Kawasaki,
Yoshinori Tokura,
Qiaochu Tang,
Kun Zhang,
Xinxin Li,
Zhi-Xun Shen
Abstract:
Complex many-body interaction in perovskite manganites gives rise to a strong competition between ferromagnetic metallic and charge ordered phases with nanoscale electronic inhomogeneity and glassy behaviors. Investigating this glassy state requires high resolution imaging techniques with sufficient sensitivity and stability. Here, we present the results of a near-field microwave microscope imagin…
▽ More
Complex many-body interaction in perovskite manganites gives rise to a strong competition between ferromagnetic metallic and charge ordered phases with nanoscale electronic inhomogeneity and glassy behaviors. Investigating this glassy state requires high resolution imaging techniques with sufficient sensitivity and stability. Here, we present the results of a near-field microwave microscope imaging on the strain driven glassy state in a manganite film. The high contrast between the two electrically distinct phases allows direct visualization of the phase separation. The low temperature microscopic configurations differ upon cooling with different thermal histories. At sufficiently high temperatures, we observe switching between the two phases in either direction. The dynamic switching, however, stops below the glass transition temperature. Compared with the magnetization data, the phase separation was microscopically frozen, while spin relaxation was found in a short period of time.
△ Less
Submitted 19 January, 2016; v1 submitted 13 June, 2013;
originally announced June 2013.
