Atmospheric world models for infrastructure scenarios in Jamaica

Inspect a miniature scenario table. Filled isometric scientific teaching instrument. The table is a constructed scenario with model-space dimensions. Local flood decisions need validated hydrology, terrain, exposure and vulnerability evidence. — Constructed explanatory instrument; it is not an observed atmospheric field or a trained-model result. Geometry uses Anatomy (MIT). The paper states its evidence and limits.

Research papers · Reformatted 31 Oct 2026Executed experiments and negative results ↗Theories and hypotheses ↗

Abstract

We define a staged framework for using an atmospheric reconstruction in paired infrastructure and imposed-storm scenarios. A reduced surface-storage model and prescribed vortex provide executable conceptual examples. Their equations, units and limitations are explicit. Procedural buildings do not constitute a surveyed urban model, and the visual companion supplies no flood, damage or building-resolving forecast. The paper specifies the additional geometry, closures, observations, comparison methods and computational measurements required to establish planning value. This extension depends on the physical inference and verification developed in Papers 0–6.

1. Decision question and evidence boundary

The proposed atmospheric world model is intended to support controlled questions about place, forcing and infrastructure. A planner might ask how a coastal development changes heat storage, or how a prescribed storm circulation exposes Kingston differently from an offshore domain. Those questions require different physical models. A temperature forecast benchmark does not validate building aerodynamics, flood depth, damage or storm-track counterfactuals.

This paper defines a hierarchy: historical replay; uncertain atmospheric reconstruction; imposed-forcing sensitivity; coupled intervention experiments; and externally validated impact analysis. Only the first and a small analytical sensitivity are implemented in this visual companion. The three-dimensional buildings are procedural geometries, not a surveyed Kingston inventory. Their value here is to expose the data and closures needed to turn an attractive scenario into a falsifiable scientific experiment.

2. State, intervention and causal interpretation

Let x be the atmospheric state, b the building and surface descriptors, f the large-scale forcing, and η the unresolved-process parameters. A transition or continuous evolution model maps these quantities to a trajectory. An intervention modifies b while retaining a paired forcing and uncertainty draw. The resulting model difference is conditional on the equations and assumptions; it is not automatically the causal effect in the real city.

\frac{d\mathbf x}{dt}=\mathcal F(\mathbf x,\mathbf b,\mathbf f;\boldsymbol\eta),\quad\Delta Y=Y(\mathbf x[\mathbf b_1,\mathbf f,\boldsymbol\eta])-Y(\mathbf x[\mathbf b_0,\mathbf f,\boldsymbol\eta]).

(1)

Y must be defined in advance: for example, a degree-hour exposure or an area-mean temperature. Paired experiments reduce forcing differences; they do not remove structural error.

Use the same background trajectory, random-number stream, observation treatment and integration window for the paired runs. Report both runs and their difference. If a changed surface alters circulation beyond the domain, lateral conditions held fixed become a limitation. A large urban intervention may require an expanded coupled domain rather than a local correction. Surface heat, humidity, momentum and roughness exchanges must enter the appropriate budgets.

3. A deliberately reduced surface-temperature sensitivity

The implemented thermal demonstration is an anomaly budget per unit surface area. It uses the archived hourly shortwave input as a prescribed forcing. Surface heat capacity, albedo, anthropogenic heat and a linear exchange coefficient are chosen scenario parameters. The anomaly is initialised at zero and is not calibrated to urban observations. It illustrates the direction and timescale of a simple storage process, rather than estimating Kingston’s urban heat island.

C_s\frac{d\vartheta}{dt}=(1-\alpha)S_\downarrow+Q_A-h\vartheta,\quad\vartheta(0)=0.

(2)

Cs: J m⁻² K⁻¹; S↓ and QA: W m⁻²; h: W m⁻² K⁻¹. ϑ is a constructed surface anomaly, not observed 2 m air temperature.

\vartheta_{n+1}=e^{-h\Delta t/C_s}\vartheta_n+\frac{(1-\alpha)S_n+Q_A}{h}(1-e^{-h\Delta t/C_s}).

(3)

Exact update for constant hourly forcing, positive h and Cs. This avoids an unstable explicit step when parameters change.

The illustrative land uses Cs=2×10⁶ J m⁻² K⁻¹ and α=0.20; the illustrative ocean uses Cs=2×10⁸ and α=0.06. Both use h=15 W m⁻² K⁻¹. Development fraction β adds QA=30β W m⁻². These values define a transparent sensitivity experiment. Evaporation, longwave exchange, canopy air, wall orientation, shading and atmospheric feedback are omitted. The plotted quantity must therefore remain a scenario anomaly.

4. Prescribed storm exposure versus a storm simulation

The vortex mode prescribes an axisymmetric tangential speed profile with maximum Vmax at radius R. A translating centre crosses the display domain. Moving the track offshore changes geometric distance, not hurricane formation or intensity physics. The pressure readout is a dynamic-pressure proxy from a prescribed wind, and excludes drag coefficients, gust factors, pressure coefficients, debris and structural response.

V(r)=V_{\max}\frac rR\exp\left(1-\frac rR\right),\quad V(0)=0,\qquad q_d=\frac12\rho V^2.

(4)

r and R use the same distance unit. qd is in Pa when ρ is kg m⁻³ and V is m s⁻¹. The profile is an idealised scenario, not a fitted Hurricane Beryl model.

The maximum follows by differentiating a e^(1−a): d/da=e^(1−a)(1−a), so the peak is at a=r/R=1. Squaring the speed explains why doubling V quadruples the dynamic-pressure proxy. The demonstration does not predict building damage. Adding blocks to the scene shows exposure geometry; they do not alter this prescribed flow. A building-resolving extension would need an appropriate CFD or urban-canopy model and measured validation data.

5. Data and model requirements for an urban extension

Required input	Physical role	Validation needed
Surveyed footprints, heights and terrain	Geometry, flow obstruction, shading and surface area	Coordinate, height and coverage checks; independent survey where possible
Roof/wall properties, land cover and vegetation	Radiation, storage, evaporation and roughness	Property uncertainty; local flux and temperature observations
As-issued wind, pressure, rain and ocean/boundary fields	Initial and lateral forcing	Time provenance, common support and held-out events
Drainage, soil, bathymetry and water levels	Runoff, inundation and coastal interaction	Independent flood depths and discharge; separate solver
Stations, profiles and flux observations	Atmospheric and surface verification	Quality control, siting, uncertainty and assimilation lineage
Infrastructure inventory and response models	Exposure, vulnerability and consequence	Engineering evidence; no inferred damage from animation alone

A defensible sequence begins with an ocean-inclusive mesoscale reconstruction and independent withheld observations. Surface interventions follow only after budget closure and baseline skill are assessed. Building-scale effects require either calibrated parameterisation at regional scale or a nested high-resolution model with compatible boundary transfers. A hydrostatic mesoscale neural field cannot claim building-wake resolution by drawing smaller cubes.

6. Computational design and comparison

Local fields permit caching repeated background features and evaluating only the windows needed for a scenario. Shared parameters can reduce the number of independently trained networks, while a coarse communication level can transmit boundary information across the domain. These are implementation options, not measured speed claims. Record training and inference separately, including source retrieval, observation operators and posterior sampling. [1]

N_{\rm vox}=N_xN_yN_zN_t,\qquad C_{\rm local}\sim N_{\rm active}\,C_{\rm field}+N_{\rm face}\,C_{\rm exchange}.

(5)

These are bookkeeping estimates, not a hardware benchmark. Neural-field cost and interface iterations must be measured in the actual implementation.

Compare against an uncoupled local model, a global field, weak-constraint assimilation and an appropriate regional numerical baseline. Use the same observation window and forcing information. A useful result is an error–cost frontier, including memory and uncertainty quality. Reject a complexity increase when it yields no meaningful improvement in the target quantity. The current browser renderer is deliberately lightweight; its frame rate is unrelated to the cost of validated atmospheric inference.

7. Expected outcomes, regional benefit and limitations

The near-term goal is a source-auditable reconstruction experiment, not an operational urban digital twin. The next deliverables are an observation-support map, a known-truth sparsity benchmark, an interface ablation, a matched-compute comparison and a small real-event study. The expected benefit is better evidence about when the proposed coupling is informative. A null result may establish that a background bias, a missing process or insufficient observations dominates the architecture.

A later Jamaican planning application could compare heat-storage assumptions or stress-test infrastructure under explicitly prescribed events. Extension to other islands would test transfer rather than assume it. The global contribution would be an inspectable method and evidence on error propagation across coupled local representations. Decision value must be measured against a specified planning outcome and uncertainty threshold. Scientific-fiction inspiration motivates the questions; observations, equations and falsification determine the answers.

References and source lineage

Citations resolve to the shared numbered bibliography. The source-access register retains inherited and unappraised records separately. [3] [4] [1] [5]

[1] Dolean, V.; Heinlein, A.; Mishra, S.; Moseley, B. (2022). Finite basis physics-informed neural networks as a Schwarz domain decomposition method. arXiv:2211.05560v2 [preprint]. Included in the revised doctoral literature review; full-text and identity checks vary by source
[2] Beucler, T. et al. (2026). Artificial intelligence pathways from weather to climate. arXiv:2610.09770v1. Preprint submitted 7 October 2026. Primary abstract and relevant full-text sections inspected 16 October 2026; preprint, not journal acceptance
[3] Raissi, Perdikaris & Karniadakis (2019). Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378, 686–707.
[4] Trémolet, Y. (2006). Accounting for an imperfect model in 4D‐Var. Quarterly Journal of the Royal Meteorological Society, 132(621), 2483-2504. Included in the revised doctoral literature review; full-text and identity checks vary by source
[5] Harder, P.; Hernandez-Garcia, A.; Ramesh, V.; Yang, Q.; Sattegeri, P.; Szwarcman, D.; Watson, C.; Rolnick, D. (2023). Hard-Constrained Deep Learning for Climate Downscaling. Journal of Machine Learning Research, 24(365), 1–40. Included in the revised doctoral literature review; full-text and identity checks vary by source

Complete reference library · Executable numerical development

AI-assisted drafting, coding and source reconciliation. Author and supervisory review remain necessary before scholarly submission. No journal acceptance or completed PhD is asserted.