Machine Learning Models For Predictive Analytics

So excited to share our new Nature paper on GenCast, an ML-based probabilistic weather forecasting model: https://lnkd.in/enzPUFbn It represents a substantial step forward in how we predict weather and assess the risk of extreme events. 🌪️ GenCast uses diffusion to generate multiple 15-day forecast trajectories for the atmosphere. It assigns more accurate probabilities to possible weather scenarios than the SoTA physics-based ensemble system from ECMWF, across a 2019 evaluation period. It’s vital that we ensure these new ML weather systems are safe and reliable. One thing I'm proud of is our range of evaluation experiments: per-grid-cell skill & calibration, spatial structure, renewable energy, extreme cold/heat/wind, and the paths of tropical cyclones (i.e. hurricanes). For example, we created a dataset of simulated wind power data at wind farm sites across the globe, and found that GenCast outperforms ENS by 10–20% up to 4 days ahead. This is promising, because better weather forecasts can reduce renewable energy uncertainty and accelerate decarbonisation. We also compared cyclone tracks from GenCast and ENS with ~100 cyclones observed in 2019. GenCast's ensemble mean cyclone track has a 12-hour position error advantage over ENS out to 4 days, and more actionable track probability fields out to 7 days. Cyclone maximum wind speeds are still generally underestimated (a common problem for ML weather models), but this performance on tracks is really promising. One recent devastating cyclone was Hurricane Milton, which caused >$85 billion in damages. GenCast predicted ~70% probability of landfall in Florida 8.5 days before the hurricane struck (and ~2 days before it even formed). A GenCast ensemble member takes 8 minutes on a TPU chip, versus hours on a supercomputer for physics-based models. This opens up the possibility of large ensembles (eg 1000s of members) which could better estimate risks of extreme events. We don't yet know how much value this will yield over conventional ensemble sizes (~50 members). Like its predecessor (GraphCast), the weights & code of GenCast have been made publicly available: https://lnkd.in/eg78dd7T. We’re looking forward to seeing how the community builds on this! It's been an honour to work on this study led by Ilan Price with such a talented team ✨: Alvaro Sanchez Gonzalez, Ferran Alet Puig, Andrew El-Kadi, Dominic Masters, Timo Ewalds, Jacklynn Stott, Shakir Mohamed, Peter Battaglia, Rémi Lam, & Matthew Willson

For those of you following developments with AI & Science, particularly around weather forecasting… At Google Research and Google DeepMind we have introduced an experimental model for tropical cyclone prediction, which can predict a cyclone’s formation, track, intensity, size and shape – generating 50 possible scenarios, up to 15 days in advance. And as we head into this year’s cyclone season, we’re partnering with the US National Hurricane center to support their forecasts and warnings. We’re publicly sharing this experimental model in Weather Lab, a new platform to access experimental weather forecast visualizations, and we hope to gather feedback and enable researchers and forecasters to leverage our models and predictions to inform their own work. You can learn more in our blog post (https://lnkd.in/geG62c2v) or this New York Times story (https://lnkd.in/gAFPbUrD).

Google DeepMind created a Gen AI model to predict extreme heat, and cyclones -- and it's faster and more accurate than traditional prediction models. It's going to be a huge deal as the climate crisis keeps getting worse. The model's called GenCast, and it uses a diffusion model, similar to those in image generation, adapted for Earth's spherical geometry. The model was trained on four decades of weather data from ECMWF's ERA5 archive. It generates 50+ possible weather scenarios, giving probabilistic ensemble forecasts. These forecasts predict daily weather and extreme events like cyclones with high accuracy. GenCast operates faster and more efficiently than traditional systems, needing just 8 minutes per forecast using TPUs. GenCast outperformed ECMWF’s ENS on 97.2% of forecasting targets, especially for extreme heat, wind, and cyclones. Its speed and precision help safeguard lives, improve renewable energy reliability, and support climate resilience. #GenAI #AI

🌍 Climate scientists often face a trade-off: Global Climate Models (GCMs) are essential for long-term climate projections — but they operate at coarse spatial resolution, making them too crude for regional or local decision-making. To get fine-scale data, researchers use Regional Climate Models (RCMs). These add crucial spatial detail, but come at a very high computational cost, often requiring supercomputers to run for months. ➡️ A new paper introduces EnScale — a machine learning framework that offers an efficient and accurate alternative to running full RCM simulations. Instead of solving the complex physics from scratch, EnScale "learns" the relationship between GCMs and RCMs by training on existing paired datasets. It then generates high-resolution, realistic, and diverse regional climate fields directly from GCM inputs. What makes EnScale stand out? ✅ It uses a generative ML model trained with a statistically principled loss (energy score), enabling probabilistic outputs that reflect natural variability and uncertainty ✅ It is multivariate – it learns to generate temperature, precipitation, radiation, and wind jointly, preserving spatial and cross-variable coherence ✅ It is computationally lightweight – training and inference are up to 10–20× faster than state-of-the-art generative approaches ✅ It includes an extension (EnScale-t) for generating temporally consistent time series – a must for studying events like heatwaves or prolonged droughts This approach opens the door to faster, more flexible generation of regional climate scenarios, essential for risk assessment, infrastructure planning, and climate adaptation — especially where computational resources are limited. 📄 Read the full paper: EnScale: Temporally-consistent multivariate generative downscaling via proper scoring rules ---> https://lnkd.in/dQr5rmWU (code: https://lnkd.in/dQk_Jv8g) 👏 Congrats to the authors — a strong step forward for ML-based climate modeling! #climateAI #downscaling #generativeAI #machinelearning #climatescience #EnScale #RCM #GCM #ETHZurich #climatescenarios

Further progress in AI+climate modeling "Applying the ACE2 Emulator to SST Green's Functions for the E3SMv3 Global Atmosphere Model". Building on ACE2 model which uses our spherical Fourier neural operator (SFNO) architecture, this work shows that ACE2 can replicate climate model responses to sea surface temperature perturbations with high fidelity at a fraction of the cost. This accelerates climate sensitivity research and helps us better understand radiative feedbacks in the Earth system. Background: The SFNO architecture was first used in training FourCastNet weather model, whose latest version (v3) has state-of-art probabilistic calibration. AI+Science is not just about blindly applying the standard transformer/CNN "hammer". It is about carefully designing neural architectures that incorporate domain constraints like geometry and multiple scales, while being expressive and easy to train. SFNO accomplishes both: it incorporates multiple scales, and it respects the spherical geometry and this is critical for success in climate modeling. Unlike short-term weather, which requires only a few autoregressive steps for rollout, climate modeling requires long rollouts with thousands or even greater number of time steps. All other AI-based models fail for long-term climate modeling including Pangu and GraphCast which ignore the spherical geometry. Distortions start building up at the poles since the models assume domain is a rectangle, and they lead to catastrophic failures. Structure matters in AI+Science!

Another new milestone is broken by AI. Just two years after producing a better short-term forecast model, a recent model has now comprehensively beaten the existing weather models much, much less compute power. 💥 But how its beaten it, and when, is where the interesting part begins. .. A new model from Google's Deepming ('GenCast') has beaten our best models in forecasting weather events. This is a big deal for two reasons: 1. Climate is changing, and bigger storms, and bigger economic events are happening. Accurate and fast weather predictions are crucial to mitigating the economic impact of extreme conditions. Faster, more precise forecasts can help governments, businesses, and communities act sooner—reducing the damage from floods, heatwaves, and other disruptive events. The forecast of weather, specifically winds, is also critical for our use of wind energy, and turbine management. To avoid damaging infrastructure, and optimally switch these on and off for example (Which I'm told is extremely expensive to do regularly). 2. Traditionally, weather prediction relies on numerical weather prediction (NWP) models, which simulate complex atmospheric processes using physical equations. While powerful, NWPs require enormous computational resources and often struggle to deliver high-resolution forecasts quickly enough, particularly for extreme conditions. So importantly, we are not talking about the 'middle' of the distribution, but rather around the edges, and unusual events. GenCast, by contrast, is a diffusion model that generates highly accurate forecasts at lower computational costs and with faster turnaround times. Fourty years of training, it generates many scenarios in parallel, effectively creating distributions of uncertainty. This approach enhances the prediction of severe weather events, allowing for more timely warnings and better preparation. Importantly, and again, the advent of machine learning means we are allowing data and processes to learn patterns (potentially not linked or driven by our understanding of the underlying causal processes). This should make us nervous, because again, 'it just works'. The paper: https://lnkd.in/gavxZvjX (they open sourced the model!) What are your thoughts on the role of AI in tackling climate risks? Bureau of Meteorology Andrew Huang, CFA Well done the team at: Google DeepMind #climate #esg #ai #artificialintelligence #genai Marcos Lopez de Prado

Thrilled to unveil our latest work: multi-modal machine learning to forecast localized weather! We construct a graph neural network to learn dynamics at point locations, where typical gridded forecasts miss significant variation. Paper: https://lnkd.in/eBmfsJin Weather dataset: https://lnkd.in/ejCG8bKs Code: https://lnkd.in/eQg-JzQJ AI weather models have made huge strides, but most still emulate products like ERA5, which struggle to capture near-surface wind dynamics. The correlation between ERA5 and ground weather station data is low due to topography, buildings, vegetation, and other local factors. In this work, we forecast near-surface wind at localized off-grid locations using a message-passing graph neural network ("MPNN"). The graph is heterogeneous, integrating both global forecasts (ERA5) and historical local weather station data as different nodes. What do we find? First off, ERA5 interpolation performs poorly, failing to capture local wind variations, especially in coastal and inland regions with complex conditions. An MLP trained on historical data at a location performs better than ERA5 interpolation, as it learns from the station's past observations. However, it struggles with longer lead times and lacks the spatial context necessary to capture weather patterns. Meanwhile, our MPNN dramatically improves performance, reducing the error by over 50% compared to the MLP. This is because the MPNN incorporates spatial information through message passing, allowing it to learn local weather dynamics from both station data and global forecasts. Interestingly, adding ERA5 data to the MLP does not improve its performance significantly. The MLP struggles to integrate spatial information from global forecasts, while the MPNN excels, highlighting the importance of combining global and local data. Large improvements in forecast accuracy occur at both coastal and inland locations. Our model shows a 92% reduction in MSE relative to ERA5 interpolation overall. This work showcases the strength of machine learning in combining multi-modal data. By using a graph to integrate global and local weather data, we were able to generate much more accurate localized weather forecasts! Congrats to Qidong Yang and Jonathan Giezendanner for the great work, and thanks to Campbell Watson, Daniel Salles Chevitarese, Johannes Jakubik, Eric Schmitt, Anirban C., Jeremy Vila, Detlef Hohl, and Chris Hill for a wonderful collaboration. Thanks also to our partners at Amazon Web Services (AWS) for providing cloud computing and technical support!

AI is cutting weather forecast times from hours to under a minute, while matching or beating traditional models on accuracy. That kind of leap matters. Google’s GraphCast and Huawei’s Pangu-Weather are only a few examples of how machine learning can process massive datasets faster and more efficiently than physics-based systems. But why does it matter? During a major hurricane or flood, even a few hours of early warning can save lives and protect infrastructure. Faster forecasts mean earlier evacuations, better emergency planning, and smarter decisions on the ground. For industries that depend on weather—agriculture, energy, logistics, emergency response—this shift isn’t just interesting, it’s operationally critical. As climate volatility grows, the ability to forecast extreme events faster and more precisely could make all the difference. You can read more about it here: https://lnkd.in/eQi-YY-F #AI #WeatherForecasting #ClimateTech

Coupled chemistry-climate models (CCMs) are essential tools for understanding chemical variability in the climate system, but they are extraordinarily expensive to run. Eric Mei's recent paper shows that linear inverse models (LIMs) can be used to emulate CCMs at a fraction of the computational cost (laptop vs HPC). This opens up new opportunities for strongly-coupled chemistry-climate data assimilation, large ensembles, hypothesis testing, and cost/benefit analysis for nonlinear machine learning emulators of CCMs. In constrast to ML emulators, LIMs have transparent explainability, illustrated by the figure below showing the coupled time-evolving relationship between sea-surface temperature, ozone, and hydroxyl radical for the El Nino mode in the model. Link to the paper: https://lnkd.in/d4DcJHVZ Supported by Schmidt Futures

GenCast is out in Nature Magazine! It's the first high res ML ensemble weather forecast which outperforms the operational state of the art. And a few more things have happened since the preprint was first released ⬇️ ⬇️ ⬇️ The paper shows GenCast provided better probabilistic weather forecasts, including better forecast of extreme weather, than the operational gold standard over our year-long evaluation period. This could mean earlier preparation for extreme events, more reliable wind power, and much more. Alongside publication in Nature Magazine, we are making the code and model weights available to the community (incl. a mini version of the model which gives passable results and can run in a free colab). And soon we'll share an archive of model historical and current forecasts. We also recently fine tuned the model to work on operational inputs so that it can be run live. We conducted a retrospective analysis of this model's forecasts of the track Hurricane Milton 🌪️ . GenCast predicted 60-80% probability of landfall in Florida already from 8.5 days before landfall eventually happened - a couple of days before Milton even formed - and more than 90% from 5.75 days before. *caveats on this example* a) individual examples should always be taken with a pinch of salt - we need rigorous evaluation over an extended period, see the cyclone track section of the paper. b) these are track predictions, not intensity predictions. Overall, GenCast marks something of an inflection point in the advance of AI for weather prediction, with SOTA raw forecasts now coming from AI. I think we can expect them to be increasingly incorporated operationally alongside traditional models (and to continue to improve!) Check out the paper: https://lnkd.in/dsasGbNb The code: https://lnkd.in/dHSjfW-3 And the blog: https://shorturl.at/NPvwL Work with a remarkable team: Alvaro Sanchez Gonzalez, Ferran Alet Puig, Tom Andersson, Andrew El-Kadi, Dominic Masters, Timo Ewalds, Jacklynn Stott, Shakir Mohamed, Peter Battaglia, Rémi Lam, & Matthew Willson