CLIMATE STUDIES GROUP MONA · THE UNIVERSITY OF THE WEST INDIES, MONA
Physical constraints, learned models and climate evidence
Adrian Dunkley · Working manuscripts · November 2026
These eight papers examine physical constraints in rainfall allocation, temperature prediction and local flux models. They compare learned methods with simpler alternatives and define tests for a proposed atmospheric reconstruction.
The exact-constraint network records 14.23 mm/month RMSE. Seasonal allocation reaches 9.73 using the same supplied coarse total. This is conditional CHIRPS downscaling.
Both Timestral checkpoints have lower next-day mean absolute error than the four local controls. Persistence and ridge regression perform better on root-mean-square error; the heavy-day comparison also favours persistence.
P6: Compare the safeguard with exact finite differences
The learned-flux model records 0.591 °C mean block RMSE against 0.590 °C for exact finite differences. The oracle-assisted gate has slightly higher mean reconstruction error.
This study divides a known monthly rainfall total among smaller cells in Jamaica. Forcing the estimates to add back to the supplied total improves the neural network. A seasonal rainfall pattern has lower error against CHIRPS. When the supplied total is wrong, exact conservation passes that error into the detailed map.
The study compares seven-day temperature forecasts at three Jamaican grid cells. It tests satellite predictors, nearby cells and a reduced physical rule. The rule improves its matched graph network in 2026, but the full-period interval includes zero. Neither satellite augmentation nor the larger graph gives a reliable gain.
Three experiments test old inputs, recalibration on an earlier period and scarce local training labels. At the Jamaica mean, the neural advantage over climatology is absent with seven days of added delay. Isotonic recalibration worsens later neural wet-day scores at all seven cells. Pooling helps local neural fits, but local ridge still wins at Kingston.
Two pretrained Timestral checkpoints receive the same Jamaican rainfall history as four local controls. Both have lower next-day mean absolute error than the four local controls. Persistence has lower next-day RMSE and detects more heavy days. Ridge has lower RMSE at three and seven days. With a 45-day observation delay, seasonal climatology has the lowest RMSE everywhere tested.
This study separates moving rain across a map from allowing it to grow or fade. Its numerical update keeps a correct account of every transfer, but that does not prove the rain forecast is right. In a small constructed test, learning growth helped when growth existed. The proposed adaptive tolerance did not beat a simpler method given the same information.
This study hides the soil temperature at 20 cm depth and tries to recover it from other measurements. Replacing a known flux calculation with a small neural network did not improve the result. A safeguard kept the replacement's local error inside its limit but had slightly higher mean reconstruction error. A conventional numerical reconstruction matched the exact-flux PINN at the reported precision.
D-LEGOS divides an atmospheric reconstruction into local blocks and gives each shared boundary one recorded transfer. That makes neighbouring blocks agree on what crosses the boundary. It does not guarantee that the transfer is correct. The paper proves the accounting rules, tests them on a small constructed heat system and sets out the observations and comparisons needed to test an atmospheric reconstruction.
This pilot asks whether converting temperature and dew point into moisture variables helps a rainfall model at a Caribbean grid cell excluded from training and calibration. The change was small and uncertain. A simple model using the season and yesterday's rainfall performed best. The results support comparisons with rainfall-history baselines and tests at locations excluded from model development.
Daily satellite pictures generate plausible next-day fields, but the model does not establish an advantage over probability forecasts drawn from the rainfall record.
Completed retrospective experiment · Draft for supervisor review
Landsat maps compare surface temperatures across three seasonal windows. The original display rasters are shown on this page and as a seasonal animation.
Completed mapping workflow · Methods and visualisation note
The full review connects Caribbean moisture physics, observations, downscaling, learned models, conservation and data assimilation with the thesis research questions.
Full critical literature review · Author and supervisory review draft
D-LEGOS: reconstructing states from incomplete observations
D-LEGOS proposes local neural fields connected through shared physical transfers. Paper P7 collects the mathematics, the constructed graph test and a falsifiable evaluation protocol for retrospective atmospheric reconstruction.
The atmospheric reconstruction has not been tested. The remaining test is whether coupling recovers more withheld physical information under matched observations, model capacity and compute. The component studies have their own questions and evidence; P7 explains their place in the programme.
Two different tests for a coupled model
Paper P7 proves the conditions for shared-face accounting and defines the atmospheric comparison still to be run. A balanced ledger does not establish an accurate flux, forecast or uncertainty estimate. Read P7.
