Moscow Institute of Thermal Technology

Abstract Differentiable Earth system models (ESMs) enable powerful applications such as sensitivity analysis, gradient‐based calibration, state estimation, boundary flux inversions, uncertainty quantification, and online machine learning. Reverse‐mode automatic differentiation (AD) efficiently provides gradients for such tasks, yet models have rarely included this capability because of complex, bespoke numerical algorithms. As part of the Differentiable programming in Julia for Earth system modeling (DJ4Earth) initiative, we present improved capabilities of the AD tool Enzyme.jl and the new compiler transpilation tool Reactant.jl, augmented by sophisticated checkpointing algorithms, which, together make general‐purpose AD tractable and efficient for full‐fledged ESM components written in Julia. Operating at the low‐level virtual machine intermediate representation or multi‐level intermediate representation compiler levels, these frameworks support mutable memory, custom kernels, and compiler optimizations before and after differentiation. Julia‐specific challenges related to just‐in‐time compilation and garbage collection are handled efficiently. Reactant further enables automatic performance portability across central processing units, graphics processing units, and tensor processing units, facilitating use of emerging AI‐customized high‐performance computing architectures. We demonstrate these frameworks on four Julia‐based ESM components featuring diverse spatial discretizations and numerical algorithms: the rotating‐sphere shallow water model ShallowWaters.jl, the finite‐volume ocean model Oceananigans.jl, the finite‐element ice sheet model DJUICE.jl, and the spectral atmospheric model SpeedyWeather.jl. Across these ESM components, our tools compute efficient and correct gradients. These results establish a foundation for differentiable, high‐performance and performance‐portable ESMs that can integrate neural networks for unresolved processes, trained online, enabling next‐generation hybrid physics–machine learning ESMs constrained by physical dynamics and observations.

Abstract Islands in the Caribbean are vulnerable to anthropogenic warming due to sea level rise and their reliance on rainfall for agriculture. These islands are particularly prone to rainfall extremes, such as the 1,029 mm of daily maximum rain in Puerto Rico due to Hurricane Maria in 2017. Rainfall extremes mostly occur in the early rainy season (ERS) from April–June and late rainy season (LRS) from August–November. While global climate models project reduced rainfall in the Caribbean by the end of the century, they are too coarse to properly resolve the complex coastline and terrain of Puerto Rico and associated convection that is often induced by sea‐breeze convergence and orographic uplift. Here, we resolve this issue by running the Model for Prediction Across Scales‐Atmosphere (MPAS‐A) using a 60–3 km global variable mesh centered over the Caribbean to downscale extreme rainfall days from coarser transient simulations during 2001–2021 and 2041–2061. This model configuration allows for the evaluation of dynamical and thermodynamic future changes at convection‐permitting scales over Puerto Rico using MESACLIP as forcing data, although these simulations underestimate extreme rainfall amounts. Results show that by mid‐century, rainfall extremes increase in the ERS but decrease in the LRS, mainly associated with changes in isolated convection. Stronger upward motion and sea breeze convergence support future increased rainfall in the ERS, while stronger subsidence likely reduces LRS rainfall extremes. These results suggest that more attention needs to be given to the increasing risk of ERS rainfall extremes over Puerto Rico.

Molecular solar thermal (MOST) materials store and release solar energy through light-induced reversible reactions involving molecular photoswitches. Solid-state crystalline MOST materials can offer higher energy densities and easier device integration than their liquid counterparts. However, their photochemical mechanisms remain poorly understood. Norbornadiene (NBD), which undergoes a [2 + 2]-photocycloaddition to form its photoisomer quadricyclane (QC), has been proposed as a candidate for MOST applications. We used multiconfigurational quantum mechanical calculations and non-adiabatic molecular dynamics to investigate the mechanism of a push-pull NBD-derivative, 1,5,6-trimethyl-2,3-dicyanonorbornadiene (TMDCNBD). This study demonstrates a cutting-edge multiscale ONIOM(QM/QM ') nonadiabatic molecular dynamics framework in TMDCNBD crystals. The crystal packing of TMDCNBD preserves molecular flexibility, enabling ultrafast [2 + 2]-photocycloaddition via energetically accessible S1/S0 conical intersections, with negligible exciton transport. Simulations predict product quantum yields of 57% for TMDCNBD and 37% for its metastable quadricyclane (QC) form, TMDCQC, which stores 0.36 MJ kg-1. This work demonstrates push-pull norbornadiene photoswitches are promising crystalline MOST candidates and establishes a transferable computational protocol for modelling ultrafast photochemistry in the solid state.

The capability of Deep Neural Networks (DNNs) to recognize objects in orientations outside the distribution of the training data is not well understood. We present evidence that DNNs are capable of generalizing to objects in novel orientations by disseminating orientation-invariance obtained from familiar objects seen from many viewpoints. This capability strengthens when training the DNN with an increasing number of familiar objects, but only in orientations that involve 2D rotations of familiar orientations. We show that this dissemination is achieved via neurons tuned to common features between familiar and unfamiliar objects. These results implicate brain-like neural mechanisms for generalization.