Eero Vainikko

Ivan G. Graham, Euan A. Spence, Eero Vainikko

In this paper we give new results on domain decomposition preconditioners for GMRES when computing piecewise-linear finite-element approximations of the Helmholtz equation $-Δu - (k^2+ {\rm i} \varepsilon)u = f$, with absorption parameter $\varepsilon \in \mathbb{R}$. Multigrid approximations of this equation with $\varepsilon \not= 0$ are commonly used as preconditioners for the pure Helmholtz case ($\varepsilon = 0$). However a rigorous theory for such (so-called "shifted Laplace") preconditioners, either for the pure Helmholtz equation, or even the absorptive equation ($\varepsilon \not=0$), is still missing. We present a new theory for the absorptive equation that provides rates of convergence for (left- or right-) preconditioned GMRES, via estimates of the norm and field of values of the preconditioned matrix. This theory uses a $k$- and $\varepsilon$-explicit coercivity result for the underlying sesquilinear form and shows, for example, that if $|\varepsilon|\sim k^2$, then classical overlapping additive Schwarz will perform optimally for the absorptive problem, provided the subdomain and coarse mesh diameters are carefully chosen. Extensive numerical experiments are given that support the theoretical results. The theory for the absorptive case gives insight into how its domain decomposition approximations perform as preconditioners for the pure Helmholtz case $\varepsilon = 0$. At the end of the paper we propose a (scalable) multilevel preconditioner for the pure Helmholtz problem that has an empirical computation time complexity of about $\mathcal{O}(n^{4/3})$ for solving finite element systems of size $n=\mathcal{O}(k^3)$, where we have chosen the mesh diameter $h \sim k^{-3/2}$ to avoid the pollution effect. Experiments on problems with $h\sim k^{-1}$, i.e. a fixed number of grid points per wavelength, are also given.

Stefan Kuhn, Vandana Dwarka, Przemyslaw Karol Grenda et al.

We introduce a reversible deep learning model for 13C NMR that uses a single conditional invertible neural network for both directions between molecular structures and spectra. The network is built from i-RevNet style bijective blocks, so the forward map and its inverse are available by construction. We train the model to predict a 128-bit binned spectrum code from a graph-based structure encoding, while the remaining latent dimensions capture residual variability. At inference time, we invert the same trained network to generate structure candidates from a spectrum code, which explicitly represents the one-to-many nature of spectrum-to-structure inference. On a filtered subset, the model is numerically invertible on trained examples, achieves spectrum-code prediction above chance, and produces coarse but meaningful structural signals when inverted on validation spectra. These results demonstrate that invertible architectures can unify spectrum prediction and uncertainty-aware candidate generation within one end-to-end model.