Physics-InformedDiffusion Modelsfor Climate Emulation in the Caribbean Basin
Can physics-based AI models reproduce the accuracy, stability, and physical credibility of regional climate models, while reducing the time to usable projections by at least one order of magnitude?
Adrian Dunkley · Department of Physics, University of the West Indies, Mona · Kingston, Jamaica
Research Question
Can Physics-Based AI Replace Regional Climate Models?
Can physics-based AI models reproduce the accuracy, stability, and physical credibility of the regional climate models (RCMs) currently used in Caribbean climate studies, while reducing the time to usable projections by at least one order of magnitude?
Proposed Title
Physics-Informed Diffusion Models for Climate Emulation in the Caribbean Basin
Secondary Research Questions
Accuracy
How well do AI models replicate RCM climatology, variability, and extremes (rainfall, temperature, hurricanes)?
Stability
Do AI projections remain consistent and drift-free over decades of climate simulations?
Physical Credibility
Do AI outputs conserve physical laws (energy, water, mass budgets) and capture realistic teleconnections?
Efficiency Gain
What is the computational speed-up relative to RCMs, and does it enable ensemble-scale experiments?
Hypothesis Outcomes
Physics-informed Climate Emulator reproduces RCM climatology within 5% bias in seasonal temperature and precipitation over 15 km resolution
Extreme rainfall distributions captured within 10% error for 50-year return levels
Island-mean bias and RMSE for daily precipitation within 10% of best RCM
Runtime reduced by 75% compared to RCM baselines
Hit Rate for physics-based AI model > RCM, for the same false alarm rates
Physical credibility preserved via constrained learning and conservation losses
Methodology
From Data to Climate Emulator
A systematic pipeline from raw climate data to a physics-informed AI emulator that can credibly replace traditional RCMs for Caribbean climate studies.
7-Stage Research Pipeline
Collect → Analyse → Pre-process → Design → Train → Test → Evaluate
Data Collection
Coarse-scale predictors: wind, humidity, temperature, geopotential, SST, topography, flash flood metrics. Output targets: 10, 25, 50 km resolution daily rainfall, temperature, and extremes across the Caribbean.
Data Quality & Exploratory Analysis
Perform comprehensive data quality checks and exploratory analysis on climate data from CORDEX, PRECIS, ERA5, and observational datasets (GPCC, HadCRUT, TRMM/GPM).
Pre-processing for ML
Pre-process climate data for machine learning development including normalisation, spatial regridding, temporal alignment, and train/validation/test splits.
Physics Rationalisation Layer
Design an architecture with an embedded physics rationalisation layer: loss functions to penalise mass/energy imbalance, negative rainfall, unrealistic land-sea constraints, Clausius-Clapeyron violations, and geostrophic imbalance.
Train Model Ensemble
Train an ensemble of deep learning models: LSTMs for temporal sequences, U-Net backbone + Diffusion for spatial generation, cGANs for super-resolution, and Reinforcement Learning for adaptive constraint weighting.
Multi-Resolution Testing
Test at coarse (25 km), super-resolution (10 km), and sub-10 km resolutions. Stress-test findings for physics rationale across all spatial scales.
Evaluation & Benchmarking
Evaluate against RCM outputs, stress-test under future warming scenarios, and measure computational speed-up (CPU/GPU hours vs RCM runtime).
Data Sources
Evaluation Framework
RMSE, Mean Bias, Correlation Coefficient
EVT/GEV fit, K-S test, Q-Q comparisons
Brier Score, Heidke, CRPS
Trend bias detection, variance consistency
Energy/water budget closure, cross-variable correlation
Wall-clock time, energy consumption, FLOPs
Deliverable
A physics-informed AI-based climate emulator that is a credible, efficient alternative to RCMs for Caribbean climate studies, enabling faster and larger-scale risk assessments and forecasts for Caribbean meteorological services, CSGM, the Caribbean Institute for Meteorology and Hydrology, and national agencies whose computing resources are constrained.
Draft · March 2026
Literature Review
Physics-Informed AI Models as Alternatives to Regional Climate Models: Accuracy, Stability, and Efficiency for Caribbean Climate Studies
Adrian Dunkley · Climate Studies Group Mona (CSGM) · Department of Physics, University of the West Indies, Mona
Prepared for doctoral supervisor review
Surveying four intersecting domains: Caribbean climate dynamics, climate model emulation, deep learning architectures, and physics-informed machine learning.
1. Introduction
Can physics-based AI models reproduce the accuracy, stability, and physical credibility of the regional climate models currently used in Caribbean climate studies, while reducing the time to usable projections by at least one order of magnitude? That is the question this thesis sets out to answer. It is a question that matters urgently for the Caribbean, where climate risk is not an abstraction but an annual reality: hurricanes that strip rooftops, droughts that empty reservoirs, rainfall extremes that trigger landslides across terrain too steep and too small for the global models to see.
Regional climate models (RCMs) such as PRECIS and the CORDEX suite have served the Caribbean well for two decades. They resolve island-scale features that global models cannot. But they are slow. A single century-long simulation at 25 km resolution ties up a computing cluster for weeks. Running the large ensembles needed to quantify uncertainty in precipitation projections, or to explore the full range of Shared Socioeconomic Pathways, is beyond the computational reach of most Caribbean institutions. The result is a paradox: the region most vulnerable to climate change has the least capacity to generate the projections it needs for adaptation planning (Nurse et al., 2014; ECLAC, 2011).
This review surveys the published work across four domains that intersect in the proposed thesis. Section 2 examines Caribbean climate dynamics and the regional modelling efforts that have shaped current understanding. Section 3 reviews climate model emulation, from pattern scaling to neural surrogates. Section 4 covers the deep learning architectures relevant to the proposed model ensemble. Section 5 reviews physics-informed machine learning. Section 6 identifies the gaps.
The review prioritises literature from 2019 to 2025. Caribbean-specific work is drawn primarily from the Climate Studies Group Mona (CSGM) at the University of the West Indies, the Instituto de Meteorología in Cuba, and international groups that have contributed to Caribbean regional modelling through PRECIS and CORDEX.
Key References
Watt-Meyer et al. (2024)
ACE2: Stable 50-year neural climate simulation
Price et al. (2024)
GenCast: Probabilistic weather forecasting (Nature)
Beucler et al. (2021)
Enforcing conservation in neural emulators
Bassetti et al. (2024)
DiffESM: Diffusion models for climate emulation
Taylor et al. (2018)
Caribbean climates: 1.5 vs 2.0°C dilemma
Campbell et al. (2011)
Future Caribbean climate from PRECIS
Kochkov et al. (2024)
NeuralGCM for weather and climate (Nature)
Raissi et al. (2019)
Physics-informed neural networks
Martinez-Castro et al. (2024)
Drivers of Caribbean precipitation change
Progress & Timeline
Research Milestones
MPhil → PhD Upgrade Timeline
MPhil Enrollment at CSGM
Commenced research in Physics-Informed AI for Climate Emulation at the Climate Studies Group Mona, Department of Physics, UWI
Literature Review & Methodology Design
Comprehensive review of Caribbean RCMs, climate emulation, deep learning architectures, and physics-informed ML. Design of physics rationalisation layer.
Data Collection & Pre-processing
Acquire CORDEX CAM, PRECIS, ERA5, GPCC, HadCRUT, and TRMM/GPM datasets. Quality analysis and ML pre-processing pipeline.
Model Development & Training
Build and train ensemble: LSTM temporal models, U-Net + Diffusion spatial generator, cGAN super-resolution, RL adaptive constraint weighting.
Multi-Resolution Testing
Test at 25 km (coarse), 10 km (super-resolution), and sub-10 km. Stress-test physics rationale across scales and warming scenarios.
Evaluation & Benchmarking vs RCMs
Full evaluation framework: RMSE, bias, skill scores, EVT, stability tests, conservation budget closure, wall-clock benchmarks against PRECIS/CORDEX.
PhD Upgrade Submission
Submit upgrade report demonstrating emulator credibility, efficiency gains, and expanded research scope for doctoral candidacy.
About Me
Building AI Tools for Caribbean Climate
I'm Adrian Dunkley, an MPhil researcher at the Climate Studies Group Mona (CSGM), Department of Physics, University of the West Indies, Mona Campus, Kingston, Jamaica. My work is focused on answering whether physics-informed AI can credibly replace the regional climate models that Caribbean nations depend on for adaptation planning.
Currently pursuing my MPhil with the objective of upgrading to a PhD, I'm developing physics-informed diffusion models that embed conservation laws (energy, water, mass budgets) directly into the architecture. The goal is a climate emulator that produces 10-25 km resolution daily projections across the Caribbean Basin - at a fraction of the computational cost of traditional RCMs like PRECIS and CORDEX.
The Caribbean is the most climate-vulnerable region with the least computing capacity for projections. My deliverable is not just a paper - it's a tool for Caribbean meteorological services, CIMH, and national agencies.
Key Acronyms & Terms
Regional Climate Model - dynamically downscales GCM outputs to capture local features
Physics-Informed Neural Network - embeds governing equations into the loss function
Coordinated Regional Downscaling Experiment - framework for regional climate projections
Providing Regional Climates for Impacts Studies - RCM framework used at CSGM
Small Island Developing States - nations most vulnerable to climate change
Caribbean Low-Level Jet - key driver of regional rainfall variability
Extreme Value Theory - statistical framework for modelling rare climate events
Continuous Ranked Probability Score - measures probabilistic forecast accuracy
Get In Touch
Let's Collaborate
Interested in collaborating on climate research, AI applications in physics, or discussing potential PhD opportunities? I'd love to hear from you. Reach out through the form or connect via the links below.