William S. Moses, Gong Cheng, Valentin Churavy, Maximilian Gelbrecht, Milan Klower, Joseph Kump, Mathieu Morlighem, Sarah Williamson, Dhruv Apte, Paul Berg, Mosè Giordano, Chris Hill, Nora Loose, Alexis Montoison, Sri Hari Krishna Narayanan, Avik Pal, Michel Schanen, Simone Silvestri, Greg Wagner, Patrick Heimbach

Position. A new generation planetary compute engine for climate science needs to overcome two fundamental limitations of current-generation climate models: (1) it should be end-to-end auto-differentiable to be able to harness advanced gradient-based algorithms (adjoint and backpropagation operators) to comprehensively learn from diverse, heterogenous data (observations and/or high-fidelity simulations); (2) it should be portable across a rapidly diversifying landscape of compute hardware, from traditional HPC to AI-oriented cloud accelerators. We argue that both capabilities should be delivered at the compiler level, built around language-agnostic intermediate representations such as LLVM and MLIR, rather than by rewriting science codes. This position is grounded in recent experience differentiating and cross-compiling four production-scale Earth system model components written in Julia and exercised on different architectures (CPUs, GPUs, and TPUs).

Differentiability as infrastructure. A “live computational commons” (as envisioned in the call for submission) that ingests billions of observations is, at its core, a massive inverse-problem. The data integration is provided by a physical model which encapsulates known laws, within which may be embedded a neural operator, encapsulating uncertain closures. Extracting actionable information from the data is achieved via gradient-based parameter calibration, state and boundary-flux estimation, sensitivity analysis, and uncertainty quantification. Each becomes computationally tractable if the adjoint of the physical model – along with the backpropagation operator of the embedded neural architecture is available. Today’s Earth system models (ESMs) are rarely differentiable because their complexity pushes the limits of available, reverse mode-enabled AD tools. The DJ4Earth initiative is building a framework for overcoming this hurdle: the AD compiler Enzyme.jl, operating at LLVM IR, together with MLIR-level transpilation via Reactant.jl and sophisticated checkpointing, produces efficient, correct gradients for a range of Earth system model components: the finite-volume ocean GCM Oceananigans.jl, a Julia implementation of the finite-element Ice Sheet and Sea level Model (ISSM), and the spectral atmospheric model SpeedyWeather.jl.

Portability on the same stack. Earth system simulation is limited by compute availability, yet the exaFLOPs being deployed for AI — TPUs, low-precision GPUs, complex interconnect topologies — are largely inaccessible to Fortran, C++, and Julia codes targeting MPI-over-InfiniBand. Even new “HPC” machines increasingly expose AI accelerators that are hard to utilize with conventional climate model codes. A manual rewrite strategy is not sustainable in view of rapid AI compute hardware evolution. Compiler-level cross-compilation — demonstrated by transparently running the same Julia ocean and atmosphere forward and adjoint models on AMD, NVIDIA, and Google TPU hardware, with automated optimization of the inter-node communication that dominates scaling — offers a more promising path. It enables rapid retargeting of the “commons” on whatever silicon is cheapest, greenest, or most available, without asking scientists to maintain refactored versions of their code.

Implications for the PROPL agenda. Taken together, these threads suggest a concrete architectural roadmap: treat the planetary compute engine as a compiler stack, anchored on a small number of shared IRs (LLVM, MLIR/StableHLO). Above them, scientists write climate models in languages of their choosing (Julia, Python, Fortran, C++); below them, a portability layer targets whatever silicon is available. Differentiation, checkpointing, communication optimization, mixed-precision lowering, and hybrid physics–ML coupling become IR-level passes that any model inherits for free. We suggest the working meeting address four concrete questions: (1) a minimal shared IR contract supporting mutable state, custom kernels, and adjoint-friendly control flow; (2) verification and numerical-stability standards for compiler-transformed simulation code on low-precision AI hardware; (3) checkpointing and storage primitives for adjoint runs over century-scale trajectories constrained by streaming observations; and (4) open governance of the compiler infrastructure a planetary commons will depend on. A commons that is differentiable and portable by construction is, we believe, the shortest path from the observations we already collect to simulations and decisions that actually use them.