Monthly Diffusion v0.9: A Latent Diffusion Model for the First AI-MIP

2026-04-16 · Source: cs.AI updates on arXiv.org · Field: Science & Research — Artificial Intelligence & Machine Learning, Environmental Science & Earth Systems · Depth: Expert, extended

Summary

Monthly Diffusion at 1.5-degree grid spacing (MD-1.5 version 0.9) is a climate emulator developed for the first AI Model Intercomparison Project (AI-MIP). This model uses a Conditional Variational Auto-Encoder (CVAE) architecture inspired by spherical Fourier neural operators (SFNOs) and latent diffusion to simulate low-frequency internal atmospheric variability at monthly mean timesteps. Trained on ECMWF Reanalysis version 5 (ERA5) data with oceanic forcings, MDv0.9 was designed for data-sparse regimes and modest computational requirements. It successfully performs stable, multi-decadal autoregressive ensemble rollouts, such as 46.25-year historical, +2K, and +4K SST-forced experiments, in approximately 20 minutes for a 50-member ensemble. While MDv0.9 reproduces key climate features and responds appropriately to SST forcings, it exhibits biases in atmospheric circulation, temperature, and precipitation, particularly over mountainous regions and in the equatorial stratosphere, and shows dampened ENSO teleconnections and NAO-like variability compared to ERA5.

Key takeaway

For AI Scientists developing climate emulators for long-timescale simulations, MDv0.9 demonstrates that stable, multi-decadal autoregressive rollouts are achievable with latent diffusion models even in data-limited settings. You should consider joint optimization of CVAE and predictor components to improve latent space dynamics and explore dual-stream conditioning to better isolate internal variability. Be prepared to address biases in stratospheric circulation and land surface representation, and consider longer spin-up periods for your models.

Key insights

A latent diffusion model emulates monthly atmospheric variability with low computational cost for long-timescale climate simulations.

Principles

• Joint optimization improves latent geometry for diffusion dynamics.
• Stochastic prior-state conditioning enhances robustness in autoregressive rollouts.

Method

MDv0.9 uses a CVAE with SFNO-inspired spectral S2-convolution layers and a conditional latent diffusion model for monthly atmospheric state prediction, incorporating dual-stream conditioning for forcings and seasonality.

In practice

• Use 1.5-degree grid spacing for climate emulation.
• Employ square root or logit transformations for non-Gaussian variables.
• Train for 100 epochs with AdamW optimizer and EMA weights.

Topics

• Latent Diffusion Models
• Climate Emulation
• Spherical Fourier Neural Operators
• Conditional Variational Autoencoders
• AI Model Intercomparison Project

Best for: AI Scientist, Research Scientist

Related on AIssential

• Monthly Diffusion v0.9: A Latent Diffusion Model for the First AI-MIP — MDv0.9 is a climate emulator using latent diffusion and a SFNO-inspired CVAE for atmospheric variability.
• Bias-Constrained Diffusion Schedules for PDE Emulations: Reconstruction Error Minimization and Efficient Unrolled Training — Adaptive noise schedules and proxy unrolled training enhance diffusion model accuracy and stability for PDE emulation.
• Latent Generative Solvers for Generalizable Long-Term Physics Simulation — LGS uses latent space, generative modeling, and an uncertainty knob for stable, generalizable, and efficient long-term PDE simulation.
• The Race to Production-Grade Diffusion LLMs with Stefano Ermon - #764 — Diffusion language models offer significantly faster and more efficient inference for text and code generation than autoregressive LLMs.
• Variance Reduction for Expectations with Diffusion Teachers — CARV reduces Monte Carlo variance in frozen diffusion teacher gradients by hierarchically amortizing expensive computations.
• Quotient-Space Diffusion Models — A new diffusion framework simplifies learning by modeling on quotient spaces, outperforming prior symmetry treatments.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.