AGENT CLIMATE · HANGZHOU

SEQUENTIAL PROMPT EXPERIMENT

A forecast scenario.
Every revision preserved.

From a single request to explicit source data,3-D transport, numerical checks and documented limitations.

Scope matters.

Real forecast data and published source documents do not make every assumption verified. EIA design rates assume operation. Background is excluded. Local sampling has not converged. An identical-input precipitation RMSE of 0 mm is an implementation check, not rainfall-forecast validation.

LATEST COMPLETED MODEL CASE · CPU

360 hours · 353,894,400 primary particles.

Full-count native hourly queries and a separate sampled near-path residence display. Original stages and failed criteria remain preserved. Video-layout review is pending.

Open the latest3D scene → · Native Scientific results · Lossless hourly arrays

Preserved versions

Download version matrix (CSV) · JSON with source hashes

STAGE 04 · FROZEN

3-D transport and plume rise

3-D conditional particle scenario

Horizontally uniform forecast profiles, parameterized turbulent diffusion and buoyant rise; flow basis unresolved; no observed-concentration or convergence validation.

Open geographic scenario → · Original native analysis view

STAGE 09 · FROZEN

Release implementation gates

Implementation agreement only

Exact-zero controls, nonnegativity and budget checks pass. Zero precipitation RMSE compares the same broadcast input; it is not an independent weather-solver validation.

Open geographic scenario → · Original native analysis view

STAGE 11 · FROZEN

Horizontal estimator refinement

Local grid convergence not demonstrated

2 km, 1 km and 500 m density bins use shared particle paths; the meteorological input resolution is unchanged. Fine-bin receptor differences remain above the declared criterion.

Open geographic scenario → · Original native analysis view

STAGE 12 · FROZEN

Particle-count refinement

Sampling noise reduced; criteria not all met

5.53 to 353.89 million particles per realization at fixed physical emission mass. Two-seed daily-field differences fall to 8.42%, but successive-count field and receptor criteria still fail.

Open geographic scenario → · Original native analysis view

STAGE 13 · FROZEN

Vertical estimator refinement

Native profiles not converged

13, 26 and 52 layers with shared trajectories. The lowest layer is 6.25 m; 26-to-52-layer global and outlet profile changes exceed the declared limits. Fixed 0–25 m agreement is a common-volume consistency check.

Open geographic scenario → · Original native analysis view

STAGE 15 · FROZEN

Terrain and surface exchange

CPU terrain scenario; new GPU agreement failed

Actual bilinear DSM boundary, contact uptake and reversible/retained stores. Forecast re-release is zero under the assumed wetness gate. The CPU case has only 345,600 primary particles; previous higher-count convergence does not qualify it. Surface chemistry is uncalibrated; wind is not terrain-resolved.

Open geographic scenario → · Original native analysis view · Native Scientific panels

Equations & approximations

These equations distinguish the original flat-ground particle model from stage15 terrain and surface exchange. Later numerical refinements change discretization and sampling; original failures remain preserved. No later correction was supplied to the initial blind baseline.

Initial2-D scenario

\frac{\partial b}{\partial t}+u_{10}\frac{\partial b}{\partial x}+v_{10}\frac{\partial b}{\partial y}=K_h\nabla_h^2b+q-\left(\frac1{18}+0.08P\right)b

The first versions use a2-D column burden, single-column10m winds, prescribed diffusion and empirical removal. Time is in hours and rainfall P in mm/h. The semi-Lagrangian discretization is not strictly mass-conservative. Relative indices in stages00-02 are not physical concentration.

Particle motion from stage 04

x_{n+1}=x_n+u(z_n,t_n)\Delta t+\sqrt{2K_h\Delta t}\,\xi_x,\qquad \xi_x\sim\mathcal N(0,1)

The analogous equation advances y. Forecast wind changes with height and time but is spatially uniform across this model domain. There is no concentration feedback on wind, resolved ambient vertical velocity or simulated building flow.

Vertical stochastic transport

z_{n+1}=\left|z_n+\frac{\partial K_z}{\partial z}\Delta t+\sqrt{2K_z\Delta t}\,\xi_z+\Delta h_{\mathrm{plume}}\right|

For the flat-ground stages 04-09, absolute value implements ground reflection. The gradient drift approximates the well-mixed condition. Prescribed, stability-dependent diffusivities and a finite timestep introduce model error; this is not LES.

Stability diagnostic

G=\frac{T_{80}-T_2+0.0098(80-2)}{80-2},\qquad Ri_b=\frac{(g/T_2)G\,78^2}{\max\!\left[(u_{80}-u_{10})^2+(v_{80}-v_{10})^2,1\right]}

Stable: Ri>0.05; unstable: Ri<−0.05; otherwise neutral. This regularized bulk diagnostic uses two temperature levels and is not an observed stability class. Elevated inversions and valley circulations are unresolved.

Buoyant plume rise

F=\frac{gQ_v}{\pi}\frac{\max(T_s-T_a,0)}{T_s},\qquad \Delta h(a)=h_\infty\min\!\left(\frac a{t_f},1\right)^{2/3}

A simplified Briggs-style rise law uses stack temperature and a conditional volume-flow interpretation; stable hours limit the final rise. Flow standard/actual basis is unverified, so two scenarios are retained. Explicit jet momentum, entrainment integration and building downwash are absent.

Competing mass-loss processes

\Delta m=m_n\left(1-e^{-(\lambda_d+\lambda_w+\lambda_c)\Delta t}\right),\qquad \Delta m_j=\Delta m\frac{\lambda_j}{\lambda_d+\lambda_w+\lambda_c}

From stage05, dry deposition, wet deposition and chemical loss are recorded separately for particles remaining inside the domain. Dry deposition is a lowest25m slab parameterization, wet removal uses forecast rain, and the18h chemical lifetime is assumed. Chemical products are not transported or mislabeled as deposited SO2.

Concentration reconstruction

\overline C_{d,kji}=\frac{10^9}{24V_{kji}}\sum_{h\in d}\sum_p m_p(t_h)\,\mathbf1[p(t_h)\in(k,j,i)]

Particle masses are kg and cell volumes are m³, giving micrograms/m³. These saved daily means average24 hourly endpoint snapshots. They are not2m point measurements, plume photographs or independently observed concentrations. No missing hourly fields are invented for playback.

Mass accounting

M_0+M_{\mathrm{emitted}}=M_{\mathrm{air}}+M_{\mathrm{dry}}+M_{\mathrm{wet}}+M_{\mathrm{chemical}}+M_{\mathrm{export}}+\varepsilon

Lateral and model-top export are counted. Source-only initialization sets M0=0; it does not mean the real atmosphere was clean. A small residual tests accounting consistency, not local atmospheric accuracy.

Display clock

t_{\mathrm{model}}=360\,\mathrm h\;\frac{t_{\mathrm{play}}}{15\,\mathrm s}

Default playback maps15days to15seconds and supports speed adjustment. Daily scalar values are interpolated before each independently rendered frame. Adjacent rendered images, RGB frames and video frames are never blended.

Stage15 terrain coordinates

h(x,y)=(1-a)(1-b)h_{00}+a(1-b)h_{10}+(1-a)bh_{01}+ab h_{11},\quad z_{\rm AGL}=z-h(x,y)

The actual GLO90 DSM is bilinearly sampled on500m nodes; a and b are cell fractions. Heights use EGM2008 orthometric datum. Valid negative values are retained. This changes particle collision geometry, not the central forecast wind into a terrain-resolved meteorological solution.

Stage15 terrain-normal reflection

\mathbf n=\frac{(-\partial_xh,-\partial_yh,1)}{\sqrt{1+|\nabla h|^2}},\qquad\Delta\mathbf r_{\rm rem}^{\prime}=\Delta\mathbf r_{\rm rem}-2(\Delta\mathbf r_{\rm rem}\cdot\mathbf n)\mathbf n

DDA traverses crossed bilinear cells and checks quadratic path clearance, including ridges between two above-ground endpoints. At most eight reflections are processed per30s step. The original case recorded nine collision-limit stops; it is not claimed geometrically exact for arbitrarily many bounces.

Stage15 contact uptake

D_n=K_h(n_x^2+n_y^2)+K_z n_z^2,\quad p_c=\min\!\left[1,v_d\sqrt{\frac{\pi\Delta t(1-\tau_c)}{\max(D_n,10^{-12})}}\right],\quad\Delta m_c=p_c m

Here tau_c is the collision fraction of the current remaining segment. This Robin-inspired assumption replaces the old25m dry slab. Ten percent of contact uptake enters a reversible pool R, ninety percent enters an irreversible SO2-equivalent store I. Assumed contact velocity is0.003m/s,0.006 at RH>90%, or0.01 when rain>0.1mm/h; no Hangzhou surface calibration is established.

Stage15 hourly surface pools

L_R=R\left[1-e^{-(k_d+k_f)\Delta T}\right],\quad\Delta M_{\rm return}=\frac{k_d L_R}{k_d+k_f},\quad\Delta I_{\rm fix}=\frac{k_f L_R}{k_d+k_f},\quad\Delta T=3600\,\mathrm s

Fixation uses k_f=1/(24h). Desorption is allowed only if surface T>=15C, RH<=70%, rain<=0.1mm/h and proxy wetness<=0.05mm. Then k_d=min[1/(3h),exp(0.04(T-15))/(12h)]. Return packets below1e-8kg are retained in R rather than discarded. These are uncalibrated scenario rules; the actual forecast gate was closed throughout360h.

Stage15 mass accounting

M_0+M_{\rm external}=M_{\rm air}+R_{\rm current}+I_{\rm current}+M_{\rm air\,chemical}+M_{\rm lateral}+M_{\rm top}+\varepsilon

Surface return is internal recycling. Gross contact uptake, wet deposition and fixation already enter the current surface stores and must not be added again. The15day surface arrays are day-end stores, not daily averages or independent new deposition. The original CPU maximum closure residual was9.70e-9kg; this does not establish local accuracy or count convergence.

Engineering hourly native query

C_{h,kji}=\frac{10^9}{V_{kji}}\sum_p m_p(t_h)\,\mathbf1[p(t_h)\in(k,j,i)],\qquad C_{0:25}=\frac1{25}\sum_k C_k\,\left|[z_k,z_{k+1}]\cap[0,25]\right|

Post-sequence engineering saves all-particle F64 cell masses at every actual hourly endpoint after surface exchange. The native grid remains500m with52 layers. A nearest-saved-hour query is explicitly labeled; no interpolation or graphics quantization changes its number. The near-ground volume mean is0–25m AGL, not2m.

Accepted-path residence diagnostic

R^{\rm sel}_{h,kji}=\sum_{\ell\in\mathcal S}m_\ell\int_{t_{h-1}}^{t_h}\mathbf1[\mathbf x_\ell(t)\in(k,j,i)]\,dt,\qquad\widehat C^{\rm display}_{h,kji}=\frac{10^9w_{\rm display}R^{\rm sel}_{h,kji}}{3600V_{kji}}

Accepted numerical pieces retain actual masses, duration, source IDs and reflected paths. The diagnostic selects stable IDs modulo particles-per-event; it is neither full-count trajectories nor an unbiased random sample. Its actual kg*s sum and display expansion weight remain separate. Hourly residence and instantaneous hourly endpoints have different temporal support; these overlapping views are never added to the physical mass ledger.

Scalar temporal reconstruction

C(s)=(2s^3-3s^2+1)C_a+(s^3-2s^2+s)d_a+(-2s^3+3s^2)C_b+(s^3-s^2)d_b,\quad 0\le s\le1

Vancouver-style bounded C1 Hermite reconstruction acts on decoded concentration scalars before each frame. Slopes use the harmonic mean of adjacent differences when their signs agree, otherwise zero; the result stays between the two endpoint values. This is display interpolation, not extra physical time integration. No rendered image, RGB frame, smoke texture or video frame is blended.

Geographic display mapping

(x_{\rm view},y_{\rm view},z_{\rm view})=\left(x/1000,\;1.8[h_{\rm context}(x,y)+z_{\rm AGL}]/1000,\;-y/1000\right)

The original early models remain2D columns or flat-ground3D fields. Independent actual geographic terrain and historical cover are context only for those versions, explicitly labeled; they do not retroactively create terrain-driven wind or particle collisions. Only stage15 and its same-physics engineering case use terrain in transport. Unknown city building heights are schematic and do not enter any wind solution. Scientific figures retain physical height scale1×.

Measured activity

Agent activity is a non-tool wall-clock proxy, not a server-inference measurement. Data includes acquisition and numerical work. Rendering includes dedicated figures/browser/export. Other includes queue, environment and inseparable mixed commands. Categories and actors may overlap; elapsed time is their interval union.