AMS 2026 session highlights how Tomorrow.io’s ICGen project is turning massive volumes of satellite observations into faster, smarter forecasts
At AMS 2026, Randy Chase of Tomorrow.io pulled back the curtain on one of the company’s most ambitious research efforts yet: using artificial intelligence to transform how satellite observations are converted into better weather forecasts.
The session focused on ICGen, short for “Initial Condition Generator,” an internal Tomorrow.io project exploring how AI-driven data assimilation methods can help numerical weather prediction (NWP) systems keep pace with a rapidly expanding volume of satellite data. As the weather industry moves toward higher-frequency forecasting and larger observational datasets, ICGen represents a glimpse into the next era of operational forecasting.
The challenge is no longer simply collecting weather observations. It’s determining how to efficiently use them.
A New Era of Microwave Observations
Tomorrow.io’s growing constellation of microwave sounder satellites sits at the center of this effort. The company recently achieved hourly global refresh coverage using 11 satellites currently in orbit — a milestone that dramatically increases the availability of atmospheric observations worldwide.
Unlike optical imagery, microwave observations can penetrate many cloud layers, enabling retrievals of temperature and humidity profiles even in cloudy conditions. This provides critical information for forecasting systems that rely on accurate representations of the atmosphere.
“These satellites are filling observational gaps that have traditionally existed in government-operated microwave sensor networks,” Chase explained during the session.
The impact is significant. Existing public-sector microwave constellations often leave large temporal and spatial gaps between observations. By adding more frequent microwave measurements, Tomorrow.io aims to create a denser and more continuous stream of atmospheric data that can improve forecast initialization around the globe.
The satellites themselves are intentionally compact — roughly the size of two stacked cereal boxes — but the data volume they generate is anything but small.
The Growing Data Assimilation Bottleneck
Modern forecasting systems already struggle with computational constraints during data assimilation (DA), the process of incorporating observations into weather models. In fact, Chase noted that the community often cites a staggering statistic: nearly 90% of available observations never make it into operational DA systems due to computational limitations.
That problem becomes even more pronounced with hourly global microwave updates.
Traditional NWP workflows were not designed to rapidly ingest this volume of data at such high cadence. Operational global models like the GFS and ECMWF IFS require enormous computational resources, making frequent updates difficult.
This is where AI enters the picture.
“The methods that lend themselves to rapid updates are AI and machine learning,” Chase said.
Rather than replacing traditional physics-based forecasting systems, Tomorrow.io’s approach seeks to augment them — using machine learning to generate improved atmospheric initial conditions that can then feed into existing forecast models.
Inside ICGen: Using AI to Improve Initial Conditions
The core idea behind ICGen is relatively straightforward in concept, though highly sophisticated in execution.
The system starts with an existing weather state — such as a six-hour-old forecast or analysis — and then incorporates new Tomorrow.io microwave sounder (TMS) observations to produce an updated atmospheric state. That improved analysis can then be passed into downstream AI weather models like Aurora or ForecastNet.
Initially, the team explored a cutting-edge approach called score-based data simulation, a type of diffusion-model methodology inspired by advances in generative AI.
Diffusion models learn how to generate realistic atmospheric states after training on massive historical datasets like ERA5. Once trained, the models can then be “guided” toward real-world observations using Bayesian techniques.
In early toy-model experiments, the approach showed promise. Microwave observations successfully nudged humidity fields toward more realistic atmospheric conditions, demonstrating that the method could physically alter weather states in meaningful ways.
But scaling quickly became the primary obstacle.
The Reality of Scaling AI Weather Systems
Operational weather prediction involves enormous complexity. AI weather models often require dozens of atmospheric variables across multiple levels and resolutions. Attempting to run diffusion methods on full-scale operational datasets proved computationally expensive, especially for a lean research team.
To address this, Tomorrow.io experimented with latent-space compression techniques using a pretrained variational autoencoder (VAE). Similar to methods used in image-generation AI systems, the VAE compressed atmospheric data into a smaller representation before applying diffusion methods.
At first glance, the results looked encouraging.
Temperature and humidity fields evolved realistically, and downstream forecast models remained stable. But deeper evaluation revealed a critical issue: subtle reconstruction artifacts introduced during compression degraded forecast skill significantly.
In fact, the diffusion-based ICGen implementation performed roughly 50% worse than simply using the original forecast state.
For many teams, that might have marked the end of the experiment. Instead, Tomorrow.io pivoted.
A Simpler, More Operational Path Forward
Recognizing the readiness gap between advanced AI research and operational forecasting, the team shifted toward a more pragmatic machine learning strategy.
Rather than relying entirely on diffusion methods, the newer ICGen approach directly injects microwave sounder channels into machine learning workflows as additional inputs.
The goal became less about generating entirely new atmospheric states and more about nudging existing forecast states closer to ERA5 reanalysis — effectively aligning operational model outputs with a more observation-rich atmospheric truth.
This simpler method produced encouraging results.
Training experiments showed consistent reductions in atmospheric error across multiple variables, including humidity, temperature, wind, and geopotential heights. The strongest improvements appeared in middle- and upper-tropospheric humidity fields, precisely where the current five-channel TMS configuration is most sensitive.
One particularly interesting finding emerged in the stratosphere, where the machine learning system appeared to correct notable representation differences between GFS and ERA5 atmospheric structures.
Does Better Analysis Actually Improve Forecasts?
For meteorologists, improving the analysis state is often assumed to improve the forecast. But Chase emphasized that this relationship is not always guaranteed — especially within autoregressive AI weather models that may smooth over unrealistic features automatically.
Fortunately, the ICGen experiments demonstrated that the improved atmospheric analyses did propagate into forecast improvements.
Forecast verification showed measurable reductions in weighted root mean squared error for both two-meter temperature and 700-mb specific humidity when TMS observations were included. Experiments masking out the microwave observations confirmed that the satellite data itself contributed directly to the gains.
The results suggest that even relatively lightweight AI-based assimilation approaches can meaningfully improve downstream forecast performance when paired with high-frequency microwave observations.
Building the Future of Forecasting
Tomorrow.io’s ICGen work reflects a broader transformation underway across the weather industry.
For decades, forecasting improvements were driven largely by advances in physics, resolution, and supercomputing. Today, AI is opening an entirely new frontier: one where machine learning helps bridge the gap between exploding observational datasets and operational forecasting constraints.
The company’s microwave constellation provides the observational backbone for that future. Meanwhile, projects like ICGen explore how to turn those observations into actionable forecasting improvements at operational speed.
Chase closed the session by emphasizing that the journey is still in its early stages.
The score-based diffusion approach remains compelling long term, particularly as GPU resources and AI infrastructure continue to improve. But for now, Tomorrow.io is focused on building systems that are operationally practical while still pushing the boundaries of what AI-enhanced forecasting can achieve.
With hourly global microwave coverage now a reality, the next challenge is no longer whether the data exists — it’s how intelligently the industry can use it.
