Chris Munley, Wenchao Ma, Johannes E. Fröch et al.
Meta-optics have rapidly become a major research field within the optics and photonics community, strongly driven by the seemingly limitless opportunities made possible by controlling optical wavefronts through interaction with arrays of sub-wavelength scatterers. As more and more modalities are explored, the design strategies to achieve desired functionalities become increasingly demanding, necessitating more advanced design techniques. Herein, the inverse-design approach is utilized to create a set of single-layer meta-optics that simultaneously focus light and shape the spectra of focused light without using any filters. Thus, both spatial and spectral properties of the meta-optics are optimized, resulting in spectra that mimic the color matching functions of the CIE 1931 XYZ color space, which links the distributions of wavelengths in light and the color perception of a human eye. Experimental demonstrations of these meta-optics show qualitative agreement with the theoretical predictions and help elucidate the focusing mechanism of these devices.
We present an algorithm to compute the Jordan chain of a nearly defective matrix with a $2\times2$ Jordan block. The algorithm is based on an inverse-iteration procedure and only needs information about the invariant subspace corresponding to the Jordan chain, making it suitable for use with large matrices arising in applications, in contrast with existing algorithms which rely on an SVD. The algorithm produces the eigenvector and Jordan vector with $O(\varepsilon)$ error, with $\varepsilon$ being the distance of the given matrix to an exactly defective matrix. As an example, we demonstrate the use of this algorithm in a problem arising from electromagnetism, in which the matrix has size $212^2\times 212^2$. An extension of this algorithm is also presented which can achieve higher order convergence [$O(\varepsilon^2)$] when the matrix derivative is known.
Lu Lu, Raphael Pestourie, Steven G. Johnson et al.
Deep neural operators can learn operators mapping between infinite-dimensional function spaces via deep neural networks and have become an emerging paradigm of scientific machine learning. However, training neural operators usually requires a large amount of high-fidelity data, which is often difficult to obtain in real engineering problems. Here, we address this challenge by using multifidelity learning, i.e., learning from multifidelity datasets. We develop a multifidelity neural operator based on a deep operator network (DeepONet). A multifidelity DeepONet includes two standard DeepONets coupled by residual learning and input augmentation. Multifidelity DeepONet significantly reduces the required amount of high-fidelity data and achieves one order of magnitude smaller error when using the same amount of high-fidelity data. We apply a multifidelity DeepONet to learn the phonon Boltzmann transport equation (BTE), a framework to compute nanoscale heat transport. By combining a trained multifidelity DeepONet with genetic algorithm or topology optimization, we demonstrate a fast solver for the inverse design of BTE problems.
This paper is the first in a series devoted to the development of efficient and highly accurate algorithms, with multiprecision \texttt{Fortran} implementations, for the computation of Bessel functions. In this first part, we present a \emph{novel, self-contained, efficient,} and \emph{multiprecision} algorithm for evaluating the modified Bessel function of the first kind, $I_ν(z)$. The method integrates several analytic representations of $I_ν(z)$, carefully selected to ensure both high accuracy and suitability for high-precision computation, together with optimally determined transition boundaries between computational regions. This design achieves high efficiency while fully preserving numerical accuracy. Unlike other widely used algorithms and libraries, such as AMOS, Boost, and GSL, which either reject negative orders $ν$ or rely on special-case symmetries valid only for integer orders, the present algorithm provides a stable approach for evaluating $I_ν(z)$ for arbitrary real orders, including $ν< 0$, and complex arguments $z$. The developed robust \texttt{Fortran} implementation provides support for both double and native quadruple-precision arithmetic. The availability of quadruple precision further enhances numerical stability, extends the reliable computational domain in $(ν, |z|)$ by approximately an order of magnitude in each direction, and enables accuracies exceeding 26 significant digits. This advancement substantially broadens the applicability of the method to demanding high-precision problems in science and engineering. Compared to AMOS (Algorithm~644), which is restricted to double precision, the present algorithm exhibits superior accuracy and efficiency, with benchmark tests demonstrating execution times reduced to 38--71\% of those of AMOS in double precision.
This paper, the second in a series, presents an efficient, self-contained algorithm for computing the modified Bessel function of the second kind, \(K_ν(z)\) for complex argument and real orders, building on Part~I (for \(I_ν\)). The method adaptively selects among analytic representations such as power series, large-\(|z|\) asymptotics, uniform asymptotics for large \(|ν|\), and numerically stable forward recurrence with region boundaries tuned for accuracy and efficiency. A robust \texttt{Fortran} implementation supports double precision and quadruple precision. The use of quadruple precision extends the reliable computational domain and improves stability in challenging regimes. Accuracy is validated against high-precision \texttt{Maple} results, and benchmarks show runtimes significantly superior to those of established methods, in the literature, while avoiding their numerical failure modes across several decades of the parameter domain. Together with Part~I, this work provides a comprehensive, multiple-precision toolkit for \(\{I_ν,K_ν\}\) across wide parameter ranges.
This paper is the final part in a series devoted to the development of numerically stable and efficient algorithms, together with multi-precision Fortran implementations, for computing Bessel functions with real orders and complex arguments. Parts~I and~II presented stable and efficient algorithms for the modified Bessel functions I_nu(z) and K_nu(z); here we treat Bessel functions J_nu(z) and Y_nu(z). The proposed algorithms support complex arguments for both positive and negative real orders and are implemented in native double and quadruple precision. Quadruple precision substantially increases dynamic range and accuracy (by roughly an order of magnitude in reliably computable |nu| and |z|), thereby extending applicability to problems requiring 20-30 digits. Comprehensive accuracy and performance comparisons are carried out against both the widely used Algorithm~644, restricted to double-precision arithmetic, and the more recent Algorithm~912, which supports both double- and quad-precision arithmetic as well as complex orders. In double precision, the present implementation consistently outperforms Algorithm~644, achieving execution times of approximately 35-67% for J_nu(z) and 44-72% for Y_nu(z), while also producing reliable results where Algorithm~644 fails. In comparison with Algorithm~912, the present algorithm achieves comparable accuracy in double-precision computations and significantly higher accuracy in quad-precision calculations. At the same time, it requires only a small fraction of the computational cost (from a few thousandths to a few hundredths) of the time taken by Algorithm~912, depending on the precision and parameter regime. Furthermore, unlike Algorithm~912, whose applicability is restricted to a limited region of the (Re(nu),z) plane, the present algorithm remains stable and accurate over the full tested domain for which reliable reference values are available.
This course, intended for undergraduates familiar with elementary calculus and linear algebra, introduces the extension of differential calculus to functions on more general vector spaces, such as functions that take as input a matrix and return a matrix inverse or factorization, derivatives of ODE solutions, and even stochastic derivatives of random functions. It emphasizes practical computational applications, such as large-scale optimization and machine learning, where derivatives must be re-imagined in order to be propagated through complicated calculations. The class also discusses efficiency concerns leading to "adjoint" or "reverse-mode" differentiation (a.k.a. "backpropagation"), and gives a gentle introduction to modern automatic differentiation (AD) techniques.
Raphaël Pestourie, Youssef Mroueh, Chris Rackauckas et al.
Many physics and engineering applications demand Partial Differential Equations (PDE) property evaluations that are traditionally computed with resource-intensive high-fidelity numerical solvers. Data-driven surrogate models provide an efficient alternative but come with a significant cost of training. Emerging applications would benefit from surrogates with an improved accuracy-cost tradeoff, while studied at scale. Here we present a "physics-enhanced deep-surrogate" ("PEDS") approach towards developing fast surrogate models for complex physical systems, which is described by PDEs. Specifically, a combination of a low-fidelity, explainable physics simulator and a neural network generator is proposed, which is trained end-to-end to globally match the output of an expensive high-fidelity numerical solver. Experiments on three exemplar testcases, diffusion, reaction-diffusion, and electromagnetic scattering models, show that a PEDS surrogate can be up to 3$\times$ more accurate than an ensemble of feedforward neural networks with limited data ($\approx 10^3$ training points), and reduces the training data need by at least a factor of 100 to achieve a target error of 5%. Experiments reveal that PEDS provides a general, data-driven strategy to bridge the gap between a vast array of simplified physical models with corresponding brute-force numerical solvers modeling complex systems, offering accuracy, speed, data efficiency, as well as physical insights into the process.
Inverse design arises in a variety of areas in engineering such as acoustic, mechanics, thermal/electronic transport, electromagnetism, and optics. Topology optimization is a major form of inverse design, where we optimize a designed geometry to achieve targeted properties and the geometry is parameterized by a density function. This optimization is challenging, because it has a very high dimensionality and is usually constrained by partial differential equations (PDEs) and additional inequalities. Here, we propose a new deep learning method -- physics-informed neural networks with hard constraints (hPINNs) -- for solving topology optimization. hPINN leverages the recent development of PINNs for solving PDEs, and thus does not rely on any numerical PDE solver. However, all the constraints in PINNs are soft constraints, and hence we impose hard constraints by using the penalty method and the augmented Lagrangian method. We demonstrate the effectiveness of hPINN for a holography problem in optics and a fluid problem of Stokes flow. We achieve the same objective as conventional PDE-constrained optimization methods based on adjoint methods and numerical PDE solvers, but find that the design obtained from hPINN is often simpler and smoother for problems whose solution is not unique. Moreover, the implementation of inverse design with hPINN can be easier than that of conventional methods.
Raphaël Pestourie, Youssef Mroueh, Thanh V. Nguyen et al.
Surrogate models for partial-differential equations are widely used in the design of meta-materials to rapidly evaluate the behavior of composable components. However, the training cost of accurate surrogates by machine learning can rapidly increase with the number of variables. For photonic-device models, we find that this training becomes especially challenging as design regions grow larger than the optical wavelength. We present an active learning algorithm that reduces the number of training points by more than an order of magnitude for a neural-network surrogate model of optical-surface components compared to random samples. Results show that the surrogate evaluation is over two orders of magnitude faster than a direct solve, and we demonstrate how this can be exploited to accelerate large-scale engineering optimization.