-
New unexpected limit operators for homogenizing optimal control parabolic problems with dynamic reaction flow on the boundary of critically scaled particles
Authors:
Jesus Ildefonso Diaz,
Alexander V. Podolskiy,
Tatiana A. Shaposhnikova
Abstract:
We pass to the limit in the homogenization of an optimal control problem associated with a parabolic equation with a dynamic boundary condition. New unexpected terms appear due to the critical scale.
We pass to the limit in the homogenization of an optimal control problem associated with a parabolic equation with a dynamic boundary condition. New unexpected terms appear due to the critical scale.
△ Less
Submitted 26 June, 2024;
originally announced June 2024.
-
Physics-Guided Problem Decomposition for Scaling Deep Learning of High-dimensional Eigen-Solvers: The Case of Schrödinger's Equation
Authors:
Sangeeta Srivastava,
Samuel Olin,
Viktor Podolskiy,
Anuj Karpatne,
Wei-Cheng Lee,
Anish Arora
Abstract:
Given their ability to effectively learn non-linear map**s and perform fast inference, deep neural networks (NNs) have been proposed as a viable alternative to traditional simulation-driven approaches for solving high-dimensional eigenvalue equations (HDEs), which are the foundation for many scientific applications. Unfortunately, for the learned models in these scientific applications to achiev…
▽ More
Given their ability to effectively learn non-linear map**s and perform fast inference, deep neural networks (NNs) have been proposed as a viable alternative to traditional simulation-driven approaches for solving high-dimensional eigenvalue equations (HDEs), which are the foundation for many scientific applications. Unfortunately, for the learned models in these scientific applications to achieve generalization, a large, diverse, and preferably annotated dataset is typically needed and is computationally expensive to obtain. Furthermore, the learned models tend to be memory- and compute-intensive primarily due to the size of the output layer. While generalization, especially extrapolation, with scarce data has been attempted by imposing physical constraints in the form of physics loss, the problem of model scalability has remained.
In this paper, we alleviate the compute bottleneck in the output layer by using physics knowledge to decompose the complex regression task of predicting the high-dimensional eigenvectors into multiple simpler sub-tasks, each of which are learned by a simple "expert" network. We call the resulting architecture of specialized experts Physics-Guided Mixture-of-Experts (PG-MoE). We demonstrate the efficacy of such physics-guided problem decomposition for the case of the Schrödinger's Equation in Quantum Mechanics. Our proposed PG-MoE model predicts the ground-state solution, i.e., the eigenvector that corresponds to the smallest possible eigenvalue. The model is 150x smaller than the network trained to learn the complex task while being competitive in generalization. To improve the generalization of the PG-MoE, we also employ a physics-guided loss function based on variational energy, which by quantum mechanics principles is minimized iff the output is the ground-state solution.
△ Less
Submitted 15 February, 2022; v1 submitted 12 February, 2022;
originally announced February 2022.
-
Boundary control and homogenization: optimal climatization through smart double skin boundaries
Authors:
J. I. Díaz,
A. V. Podolskiy,
T. A. Shaposhnikova
Abstract:
We consider the homogenization of an optimal control problem in which the control is placed on a part of the boundary and the spatial domain contains a thin layer of "small particles", very close to the controlling boundary, and a Robin boundary condition is assumed on the boundary of those "small particles". This problem can be associated with the climatization modeling of Bioclimatic Double Skin…
▽ More
We consider the homogenization of an optimal control problem in which the control is placed on a part of the boundary and the spatial domain contains a thin layer of "small particles", very close to the controlling boundary, and a Robin boundary condition is assumed on the boundary of those "small particles". This problem can be associated with the climatization modeling of Bioclimatic Double Skin Façades which was developed in modern architecture as a tool for energy optimization. We assume that the size of the particles and the parameters involved in the Robin boundary condition are critical (and so they justify the occurrence of some "strange terms" in the homogenized problem). The cost functional is given by a weighted balance of the distance (in a H^1-type metric) to a prescribed target internal temperature u_{T} and the proper cost of the control (given by its L^2 norm). We prove the (weak) convergence of states u_{ε} and of the controls v_{ε} to some functions which are completely identified: u_0 satisfies an artificial boundary condition on the control boundary and v_0 is the optimal control associated to a limit cost functional J_0 in which the "boundary strange term" on the control boundary arises. This information on the limit problem makes much more manageable the study of the optimal climatization of such double skin structures.
△ Less
Submitted 30 December, 2021; v1 submitted 23 December, 2021;
originally announced December 2021.
-
Homogenization of a net of periodic critically scaled boundary obstacles related to reverse osmosis "nano-composite" membranes
Authors:
Jesús Ildefonso Díaz,
David Gómez-Castro,
Alexander V. Podolskiy,
Tatiana A. Shaposhnikova
Abstract:
One of the main goals of this paper is to extend some of the mathematical techniques of some previous papers by the authors showing that some very useful phenomenological properties which can be observed to the nano-scale can be simulated and justified mathematically by means of some homogenization processes when a certain critical scale is used in the corresponding framework. Here the motivating…
▽ More
One of the main goals of this paper is to extend some of the mathematical techniques of some previous papers by the authors showing that some very useful phenomenological properties which can be observed to the nano-scale can be simulated and justified mathematically by means of some homogenization processes when a certain critical scale is used in the corresponding framework. Here the motivating problem in consideration is formulated in the context of the reverse osmosis. We consider, on a part of the boundary of a domain $Ω\subset \mathbb R^n$, a set of very small periodically distributed semipermeable membranes having an ideal infinite permeability coefficient (which leads to Signorini type boundary conditions) on a part $Γ_1$ of the boundary. We also assume that a possible chemical reaction may take place on the membranes. We obtain the rigorous convergence of the problems to a homogenized problem in which there is a change in the constitutive nonlinearities. Changes of this type are the reason for the big success of the nanocomposite materials. Our proof is carried out for membranes not necessarily of radially symmetric shape. The definition of the associated critical scale depends on the dimension of the space (and it is quite peculiar for the special case of $n=2$). Roughly speaking, our result proves that the consideration of the critical case of the scale leads to an homogenized formulation which is equivalent to have a global semipermeable membrane, at the whole part of the boundary $Γ_1$, with a "finite permeability coefficient of this virtual membrane" which is the best we can get, even if the original problem involves a set of membranes of any arbitrary finite permeability coefficients.
△ Less
Submitted 19 July, 2018;
originally announced July 2018.
