Machine learning is revolutionizing weather forecasting – the next step is a change in how we work

Peter Dueben, Peter Bauer, Oliver Fuhrer, Nikolay Koldunov, Jørn Kristiansen · 2026

Read Machine learning is revolutionizing weather forecasting – the next step is a change in how we work by Peter Dueben, Peter Bauer, Oliver Fuhrer, Nikolay Koldunov, Jørn Kristiansen (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-25076.txt.

Introduction

Peter Dueben and four colleagues argue that machine learning's revolution in weather forecasting is about to change how forecasters work, not just what forecasts say. Machine learning systems now rival traditional physics based models in skill. This article, a perspective piece, turns from outputs to practice. The authors contend that learning methods and digital technology will reshape the forecasting value chain. Coding, observations, data management, verification, and services are all in scope.

Six areas of change are surveyed. Agentic software engineering, shared verification workflows, open and compressed data, interactive computing, and generative methods could make modelling, evaluation, and service creation faster and more accessible. The authors are clear on the burden this places on weather and climate centres. Infrastructure, data stewardship, quality assurance, skills, and service delivery all have to adapt. Scientific understanding, operational reliability, human expertise, and the public service role must survive the transition.

Catalog record

CREATORPeter Dueben, Peter Bauer, Oliver Fuhrer, Nikolay Koldunov, Jørn Kristiansen
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE43 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2606.25076

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Machine learning is revolutionizing weather forecasting – the next step is a change in how we work Peter Dueben, Peter Bauer, Oliver Fuhrer, Nikolay Koldunov, Jørn Kristiansen Abstract Following the success of machine learning in producing weather predictions with competitive skill compared to complex traditional systems, this article shifts attention from forecast output to the working practices that make prediction systems possible. We argue that machine learning and recent digital technologies will reshape the forecasting value chain: how models are coded and developed, how observations and Earth-system data are exploited, how data and computing are managed, how systems are verified, and how information is created, evaluated and turned into services. We discuss six non-exhaustive areas in which agentic software engineering, open and compressed data, shared verification workflows, interactive computing and generative methods may make modelling, evaluation and service creation faster, more interactive and more widely accessible. These changes will require weather and climate centres to adapt their infrastructures, data stewardship, trust and quality-assurance frameworks, skills and service delivery while maintaining scientific understanding, operational reliability, human expertise and their public-service role. [type=editor, auid=000,bioid=1, orcid=0000-0002-4610-3326] [type=editor, auid=000,bioid=2, orcid=0000-0002-3205-6055] [type=editor, auid=000,bioid=3, orcid=0000-0002-0682-1374] [type=editor, auid=000,bioid=4, orcid=0000-0002-3365-8146] [type=editor, auid=000,bioid=5, orcid=0000-0002-7210-9166] † † highlights: Agentic coding, generative methods and modern software engineering will change the tools and workflows used to develop weather and climate models. Open, compressed and better curated Earth-system data will change how observations are exploited, systems are verified and forecast information is accessed. Weather and climate centres will need agile infrastructures, skills and trust frameworks while preserving reliability, human expertise and public-service delivery. keywords Numerical weather prediction ,Machine learning ,Ways of working † † credit: Writing - Original draft preparation † † credit: Conceptualization of this study, Writing - Original draft preparation † † credit: Writing - Original draft preparation † † credit: Writing - Original draft preparation † † credit: Writing - Original draft preparation † † corresponding: Corresponding author † † note: This article compiles key areas where machine learning and recent technology developments will change methods by which weather prediction will be developed and implemented in the future. † † note: The views expressed in this article are those of the authors and do not necessarily reflect the views, positions or policies of their affiliated institutions. Introduction In the last three years, machine learning (ML) has revolutionised numerical weather prediction ( 14 ; 50 ; 28 ) . The initial push from technology companies has led to machine-learned forecast models running operationally as public services ( 35 ) . Training efforts aside, ML produces often superior medium-range predictions at negligible time-critical production cost ( 34 ) . However, while the capability of end-to-end ML forecast systems is progressing ( 3 ) , current systems still largely rely on initialisation and training data from traditional models, and ML predictions still produce a much smaller amount of output fields at coarser steps in space and time. The reliance on traditional models is crucial for driving science and understanding ML model performance, but this reliance ties the pace of ML model development to the rather incremental evolution of traditional systems. As the rate of score improvements from ML models for global NWP is starting to slow down after the initial surge, a dramatic, ML assisted change in the way we design cutting-edge systems to forecast weather and do climate projections is yet to come. The revolution discussed here is not a single technical breakthrough, but a change in the operating model and forecasting value chain: who develops models, how data and compute are accessed, how systems are verified, how forecast information is turned into services, and how trust is maintained. A big part of the transition concerns the way we work and how this is changing through ML and recent technological developments. The ’how’ will determine the speed and cost-effectiveness of the adaptation to the ML revolution, as well as the extent to which digital technologies and software development practices from outside the weather domain can be adopted without introducing undesirable dependencies or loss of control over key capabilities and service quality. The traditional approach in Earth science has been that domain scientists with specialisations in mathematics, physics, chemistry and computer science jointly work on large mature codes. These are incrementally improved over time, tested using standard metrics, and then turned into data and products disseminated to users ( 9 ) . The transfer cycle from research to operations is typically around a year for weather and much longer for climate models. Today, this working model appears increasingly strained in a context where ML requires new skills across all science domains, software development is becoming faster-paced and aligned with industry standards, and large-scale computing architectures are driven more by AI than by traditional scientific requirements. As Moore-style cost-performance and Dennard-style power-density/clock-frequency scaling are expiring, and energy and data-movement costs of modern computing are exploding, it is evident that the next generation of supercomputers will not automatically provide more usable compute power for traditional physics-based weather and climate applications per Euro or Dollar spent ( 54 ) . The ‘how’ will therefore determine whether weather prediction and weather centres can remain productive and cost-effective. We discuss six non-exhaustive areas, selected as a horizon scan because they cut across the forecasting value chain and are already visible in parts of current ML, software, data and computing practice. While the individual topics in isolation may not revolutionize weather prediction, their combination will fundamentally change the ways we work and how much impact is created. The discussion therefore combines developments that are already happening, emerging near-term trends, and longer-term scenarios. This is not meant as technological determinism ( 56 ) , but as one plausible scenario for the future of our field: concrete enough to be useful for preparation, but not a forecast of an inevitable path. Most importantly, operational centres should start preparing for mixed human-AI system environments today because ’the how’ will touch almost all areas of their work. 1 Science & codes: AI-assisted and agentic development workflows disrupt the working culture Today, code development is performed manually by a large number of scientists and software developers affecting code quality, performance, portability across different high-performance computing (HPC) hardware configurations, and development pace. LLMs are already accepted as interactive research agents and text generators, but also emerge as primary code developers (’vibe coding’, ’agentic coding’) and integrators, code testers and workflow executors ( 51 ) . This trend has been seen in other domains earlier 1 1 1 https://www.anthropic.com/research/vibe-physics , as their models and workflows are typically smaller when compared to today’s Earth-system models (ESM 2 2 2 The term Earth-system model is used here to denote the numerical model components representing the physical and chemical processes in atmosphere, ocean, land and cryosphere that are relevant for weather prediction from days to seasons and climate projections; it also includes the methodological framework to create the forecasts’ initial conditions through the assimilation of observational data into models. ). But there are already examples showing that LLMs can understand the overall context of ESMs and data, and overcome our slow-paced manual approach ( 48 ; 32 ) . To prepare large ESMs for the change, an important prerequisite

Read the complete text