In this paper, we propose a generic algorithm to train machine learning-based subgrid parametrizations online, i.e., with a posteriori loss functions, but for non-differentiable numerical solvers. The proposed approach leverages a neural emulator to approximate the reduced state-space solver, which is then used to allow gradient propagation through temporal integration steps. We apply this methodology on a single layer quasi-geostrophic system with topography, known to be highly unstable in around 500 temporal iterations with offline strategies. Using our algorithm, we are able to train a parametrization that recovers most of the benefits of online strategies without having to compute the gradient of the original solver. It is demonstrated that training the neural emulator and parametrization components separately with different loss quantities is necessary in order to minimize the propagation of approximation biases. Experiments on emulator architectures with different complexities also indicates that emulator performance is key in order to learn an accurate parametrization. This work is a step towards learning parametrization with online strategies for realistic climate models.
Large language models rely on attention mechanisms with a softmax activation. Yet the dominance of softmax over alternatives (e.g., component-wise or linear) remains poorly understood, and many theoretical works have focused on the easier-to-analyze linearized attention. In this work, we address this gap through a principled study of the single-location regression task, where the output depends on a linear transformation of a single input token at a random location. Building on ideas from statistical physics, we develop an analysis of attention-based predictors in the high-dimensional limit, where generalization performance is captured by a small set of order parameters. At the population level, we show that softmax achieves the Bayes risk, whereas linear attention fundamentally falls short. We then examine other activation functions to identify which properties are necessary for optimal performance. Finally, we analyze the finite-sample regime: we provide an asymptotic characterization of the test error and show that, while softmax is no longer Bayes-optimal, it consistently outperforms linear attention. We discuss the connection with optimization by gradient-based algorithms.
In this work, we revisit the Generalised Navier Boundary Condition (GNBC) introduced by Qian et al. in the sharp interface volume-of-fluid context. We replace the singular uncompensated Young stress by a smooth function with a characteristic width $\varepsilon \gt 0$ that is understood as a physical parameter of the model. Therefore, we call the model the 'contact region GNBC' (CR-GNBC). We show that the model is consistent with the fundamental kinematics of the contact angle transport described by Fricke, K & ouml;hne and Bothe. We implement the model in the geometrical volume-of-fluid solver Basilisk using a 'free angle' approach. This means that the dynamic contact angle is not prescribed, but reconstructed from the interface geometry and subsequently applied as an input parameter to compute the uncompensated Young stress. We couple this approach to the two-phase Navier-Stokes solver and study the withdrawing tape problem with a receding contact line. It is shown that the model allows for grid-independent solutions and leads to a full regularisation of the singularity at the moving contact line, which is in accordance with the thin film equation subject to this boundary condition. In particular, it is shown that the curvature at the moving contact line is finite and mesh converging. As predicted by the fundamental kinematics, the parallel shear stress component vanishes at the moving contact line for quasi-stationary states (i.e. for $\dot heta _d=0$ ), and the dynamic contact angle is determined by a balance between the uncompensated Young stress and an effective contact line friction. Furthermore, a nonlinear generalisation of the model is proposed, which aims at reproducing the molecular kinetic theory of Blake and Haynes for quasi-stationary states.
We prove that for almost all symmetric spaces X and for any sequence of compact locally symmetric spaces Y_n which is uniformly discrete, has a uniform spectral gap, and converges in the sense of Benjamini–Schramm to X, the joint eigenfunctions of all invariant differential operators on Y_n delocalize on average when their spectral parameters are taken to lie in a fixed spectral window.
Generative AI (GenAI) holds potential for organizations, offering transformative opportunities while simultaneously raising concerns about its associated risks. Like many emerging technologies, GenAI presents organizations with a significant challenge: navigating uncertainty before making large-scale decisions about which systems to adopt and how to implement and leverage them. Managers cannot rely solely on general knowledge of GenAI; they require insights tailored to their specific organizational context. Drawing on an 18-month study of sandbox experiments conducted within a large international service organization, this paper presents CORE-sandbox experiments as a structured framework for systematically learning about the critical dimensions of uncertainty surrounding GenAI. The framework organizes learning into four key domains: Capabilities, Opportunities, Risks, and Ecosystem. The paper also advances the discourse on organizational learning and dynamic capabilities by demonstrating how in-situ and ex-situ learning cycles reinforce one another and how second and third-order organizational learning emerge under conditions of high uncertainty before GenAI rollout decisions are made.
