Charting the Weather of Mars with a Single AI Model
30 May 2026, Yanjiang
A unified AI foundation model could forecast Martian dust storms, map low-level jets, and fill satellite data gaps across the Red Planet’s atmosphere.
A single AI model that can forecast Martian dust storms, predict low‑level jets, and fill gaps in satellite observations — it sounds like the stuff of science fiction. But the building blocks are already here, scattered across research groups and data archives. Only in the past few years have Earth scientists begun to train so‑called foundation models for our own planet’s atmosphere. Now a team led by Rahul Ramachandran at NASA’s Marshall Space Flight Center, together with Sujit Roy at the University of Alabama in Huntsville and collaborators across five institutions, is asking a harder question: what would it take to build one for the Red Planet? Their answer, laid out in a preprint (arXiv:2605.28851), is part design study, part honest reckoning with the limits of artificial intelligence when data are sparse and physics stubborn.
The martian atmosphere is a puzzle that has resisted simple solutions for decades. It hosts phenomena that range from planet‑encircling dust storms — opaque veils that can blot out the Sun for months — to delicate orographic clouds that cling to the slopes of Arsia Mons, and nocturnal low‑level jets that race through valleys under a carbon‑dioxide sky. Global circulation models, the workhorses of terrestrial weather prediction, can simulate these features, but they become painfully expensive when run at resolutions fine enough to capture the mesoscale — the intermediate scale where topography meets turbulence, where cloud belts and dust‑laden fronts first organize themselves. At the same time, the observational record is thin. Satellites scatter their measurements across different instruments, different orbits, and different years; the archive is as fragmented as a half‑finished mosaic.
This is where a foundation model enters the picture. The term, borrowed from the world of large language models, refers to a neural network that is pre‑trained on enormous and diverse data — not for one narrow task, but to serve as a common scaffold that can later be fine‑tuned for many specific jobs. On Earth, such models have already learned to forecast weather, downscale climate projections, and fill missing satellite pixels, all from a single trained network. The dream for Mars is analogous: build one model that can be prompted to perform a ten‑day forecast, to reconstruct the atmosphere where observations are missing, or even to detect anomalous dust activity before a full‑blown storm erupts. But the design space is vast, and Ramachandran, Roy, and their colleagues set out to map it.
The team’s roadmap is built on three layers: available data, candidate phenomena, and plausible AI architectures. The key data source is OpenMARS, a reanalysis dataset that fuses satellite retrievals with a global circulation model to produce a best‑estimate history of the martian atmosphere, gridded at a resolution of 5 degrees — about 300 kilometres. That is enough to see the great dust belts and planetary waves, but far too coarse for the mesoscale features that the researchers ultimately want to capture. Their ambition is a future model that could operate at 2‑degree resolution, roughly halving the grid spacing, which would begin to resolve cloud belts, dust fronts, and the sharp wind gradients that buffet a descending lander. It is the difference between seeing the outline of a storm and glimpsing its internal circulation — a difference that matters enormously for both science and safety.
To explore what is possible with today’s tools, the team tested three deep‑learning architectures — in effect, three different blueprints for the same imagined building. They trained each model on twenty years of OpenMARS reanalysis and evaluated them on a year‑long holdout period, measuring how well each could run an autoregressive forecast out to one martian day.
The first candidate, a transformer‑based model that the team calls Mars SpectFormer, treated the atmosphere as a set of spectral coefficients — not unlike the spherical harmonics that traditional circulation models themselves use. It preserved large‑scale structure admirably, but over twenty‑four forecast hours it gradually smoothed out finer details, blurring the jet streams and temperature gradients that are the grist of a working meteorology.
The second candidate, Mars GraphCast, was adapted from Google DeepMind’s celebrated Earth‑weather model. It represents the atmosphere as a graph draped over the planet, with edges carrying information between neighbouring grid cells. The result was sobering. Temperature errors exceeded 16 kelvin just two hours into the forecast, a failure so large and so early that it suggested a systematic bias — the model had simply not learned to track the diurnal cycle at all. On Mars, the day‑night temperature swing is ferocious, often exceeding 50 kelvin. A model that cannot reproduce this rhythm is deaf to the planet’s daily heartbeat.
The third candidate, Mars Prithvi‑WxC, built on a vision‑transformer architecture originally developed for Earth weather, showed the most promise. Its zonal‑wind errors remained around 5 to 7 metres per second over the full 24‑hour forecast, and it captured the broad latitudinal structure of temperature — off by 5 to 11 kelvin, certainly, but preserving the shape of the polar and tropical gradients. This is progress. Yet even Prithvi‑WxC stumbled badly on dust opacity. The model’s estimates of how much dust was suspended in the atmosphere quickly diverged from the reanalysis truth, growing far beyond the persistence baseline. Surface pressure also drifted by tens of pascals — a seemingly small number, but alarming given that pressure is one of the easiest variables to persist on a wind‑blown desert planet.
Why does dust remain so elusive? The paper does not fully answer this, but it points toward a deeper challenge: the martian dust cycle is intimately tied to unresolved, sub‑grid surface processes — saltation, sand‑blasting, electrostatic lifting — that a coarse reanalysis cannot see. A model that only sees the atmosphere, without a direct line to the surface physics, may never learn the right causal chain. This is not a failure of the neural network; it is a diagnosis that the input data themselves carry a form of irreducible ignorance.
That honesty — the willingness to show where AI stumbles against simple physical benchmarks — is what sets this work apart from the breathless headlines that often accompany foundation models. The paper does not announce a finished product. It sketches a design space and then populates it with painstaking, quantitative examples of what works and what doesn’t. In doing so, it offers what is essentially an architectural drawing for a building that does not yet exist — complete with annotations about where the foundations will need to be reinforced and which load‑bearing walls are still uncertain.
The study also gestures toward a larger, more provocative question. On Earth, the explosion of satellite data and the availability of high‑quality reanalysis products have allowed data‑driven models to flourish, sometimes even bypassing explicit physical equations. Mars, with its thinner data and harsher environment, forces a different bargain: a foundation model there will almost certainly need to be infused with stronger physical constraints, either through hard‑wired conservation laws or through hybrid approaches that couple neural networks to traditional dynamical cores. The team discusses data assimilation — the art of blending observations with a physical model — as a crucial bridge, and they point to recent AI‑based approaches that could learn the assimilation process itself from data. In this light, the martian atmosphere becomes a crucible for testing how much physics an AI must know in advance, and how much it can be trusted to learn.
For the reader who simply wants to know whether a martian weather service is on the horizon, the answer is: not yet, but the first serious blueprint is now in the public domain. Ramachandran and Roy’s team has placed on the table a set of concrete benchmarks and a clear‑eyed assessment of where current AI falls short. The next step, which the paper outlines, involves better integration of satellite observations, higher‑resolution reanalysis, and architectures that can explicitly handle the sharp diurnal cycles and sporadic dust events that make Mars so unlike Earth.
Perhaps one day, a single model, pre‑trained on the whole of martian meteorological history, will run on a modest computer inside a habitat on the edge of Jezero Crater, issuing warnings of an approaching dust front with the same routine confidence that today’s terrestrial models predict tomorrow’s rain. The road from a handful of scattered architectures — SpectFormer, GraphCast, Prithvi‑WxC — to that vision is long. But the cartography of that road, drawn with unusual candour, is now in print. And the first steps will be taken not by imagining a finished cathedral, but by understanding precisely where the ground is soft.
— Yanjiang
Yanjiang is an online editor of LoomSci.com.
References
- Sujit Roy et al., Towards a Foundation Model for the Martian Atmosphere, arXiv:2605.28851