How We Built MLOps for Climate Downscaling on AWS
In the face of accelerating climate risks, the demand for localized, high-resolution climate information has never been greater. As decision-makers across sectors seek to assess physical climate risks and design adaptation strategies, global climate models remain too coarse to support actionable insights at the asset or city level. That’s where Deep Learning for climate downscaling comes into play.
After many months of focus development we’ve taken a significant step forward in our technological capabilities by building a scalable, automated MLOps platform to support climate model downscaling and emulation by combining deep learning, cloud-native tools, and operational rigor.
Super Resolution downscaling methods
At the core of our approach is a class of deep learning techniques known as Super-Resolution, which originated in the field of computer vision. In short, super-resolution models use deep learning to learn how to reconstruct fine-scale images from low-resolution input images.
We’ve built our super-resolution pipeline using a progressive mix of SR models. Starting with established methods and advancing to cutting-edge techniques:
- U-Net: A reliable deep learning model that enhances blurry climate data by learning both large patterns and fine details. It’s fast, proven, and forms the foundation of many super-resolution systems.
- Vision Transformers: A newer approach that views climate data like images — not just locally, but globally. This allows the model to understand large-scale climate dynamics, such as how sea surface temperatures can affect rainfall patterns far inland.
- Diffusion Models: One of the most advanced AI techniques today, diffusion models generate realistic high-resolution projections step-by-step from random noise. Because they can produce multiple plausible outcomes (ensembles), they’re especially useful for quantifying uncertainty — a key need in climate risk analysis and planning for extreme events.
These deep learning downscaling techniques complements traditional statistical and dynamical methods by offering a fast, reproducible, and data-driven alternative.
Industrializing super resolution downscaling with MLOps
To scale these models and ensure reproducibility across hundreds of climate simulations with millions of data points, we developed a robust MLOps pipeline built on AWS cloud infrastructure. This includes:
- Automated model training and validation workflows
- Versioned data pipelines and model registries
- Continuous integration & deployment (CI/CD) for climate workflows
- Scalable inference across regions and time horizons
This framework enables us to serve large-scale climate data processing, with high-frequency updates, auditable data lineage, quality control, and full traceability — critical for reproducibility and auditing as we serve organisations under important compliance constraints.
Built on AWS, Co-Designed with ANEO
This project came to life with the strategic and technical support of AWS, whose cloud-native services provide the backbone of our AI infrastructure. With AWS’s sponsorship, we teamed up with ANEO, a digital consulting firm with deep expertise in MLOps and data engineering, to co-design and implement the core components of our platform.
Together, we built a robust, scalable system that enables TCDF to deliver an operational-grade climate data processing system as we’re not MLOps experts — and that’s exactly where ANEO’s strong expertise made the difference. Their team supported us in navigating every stage of the architecture: from cloud infrastructure design and data pipelines to CI/CD workflows, scalable training and inference environments, model versioning, and monitoring tools. Their ability to bridge our scientific requirements with best practices in software engineering was key to making this platform production-ready.
Our MLOps capabilities now allow us to be proactive and innovative in integrating the latest advances in machine learning into our products and services — delivering even more value to our clients through faster iteration, higher reliability, and continuous improvement.
Supporting Resilience and Risk Disclosure
Organizations increasingly need to understand how climate change may affect their operations, assets, and long-term strategies. Whether for strategic risk assessment, resilience planning, or regulatory disclosure, there is a growing demand for granular, forward-looking climate data that can inform real decisions.
Our deep leaning climate data downscaling work supports this shift, and aligns with evolving frameworks like the TCFD, ISSB, and CSRD, which emphasize the importance of scenario-based physical risk analysis.
By connecting climate science with operational decision-making, our platform helps clients and partners:
- Integrate future hazard projections into enterprise risk and strategy
- Run localized scenario analyses across different warming pathways
- Support both compliance and climate-resilient planning across sectors
We see this as a practical step toward making climate data more useful, relevant, and impactful.
What’s Next?
With this foundation in place, we’re already working on the next phase of our platform — expanding its capabilities and deepening its impact. In the months ahead, we’re focusing on:
- Enhancing our ability to simulate compound and extreme events
- Emulation of regional climate model with uncertainty quantification
- Implementing real-time downscaling for operational forecast use cases
As part of our commitment to transparency, we’re also preparing a series of technical papers that will detail what’s under the hood — including our modeling approaches, training datasets, and evaluation methods.
Our goal is to continuously bring the latest developments in machine learning and climate science into production — so our clients can make smarter, faster, and more resilient decisions.
