Editing Tropical Cyclone EVOlution Model



'''TC Evolution''' is an experimental tropical cyclone current-intensity estimation model developed in 2025 and currently described as '''Beta 1.6'''. It is designed to estimate a storm's present maximum sustained wind (V<sub>max</sub>) from a short sequence of storm-centred environmental and satellite inputs rather than produce a long-range forecast.

The system combines gridded environmental data, storm-centred infrared and water-vapour satellite imagery, sea-surface temperature, land masking, and recent storm-history scalars into a single neural architecture. In its Beta 1.6 form, TC Evolution uses a multimodal sequence model with separate encoders for full-storm satellite structure, zoomed inner-core satellite structure, and environmental fields, followed by a temporal [[Transformer (deep learning architecture)|Transformer]] encoder.

== History ==
TC Evolution began as a '''2025 prototype''' focused on tropical cyclone current-intensity analysis from storm-centred gridded data. The earliest versions were built to answer a practical question: whether a neural model could infer present intensity from recent storm evolution, rather than rely on a single image or on long-range forecast logic.

The project later evolved into a two-stage training workflow:

=== Stage A ===
The first major training stage used a satellite-only pretraining approach based on long-term storm-centred IR and WV imagery from '''1998 to 2024'''. This stage was intended to teach the model broad tropical-cyclone structural recognition before environmental data were added.

=== Stage B ===
The second major stage fine-tuned the model on a multimodal dataset spanning '''2015 to 2024'''. This stage added environmental wind and height fields, sea-surface temperature, land masking, and scalar storm-history features. The resulting architecture became the basis of the operational Beta line.

=== Beta 1.6 ===
'''Beta 1.6''' refers to the operational inference implementation using the Stage B multimodal architecture. In this version, the model is driven by:
* storm-centred GFS environmental fields,
* storm-centred GOES infrared and water-vapour imagery,
* OISST sea-surface temperature,
* land masking,
* and ATCF storm-history data for recent motion and prior intensity.

== Operation ==
TC Evolution is designed for '''current intensity analysis'''. It does not operate as a traditional long-range forecasting model. Instead, it takes a short sequence of recent storm-centred frames, processes their structure and environment, and predicts the storm's present intensity.

In operational use, the system:
* reads storm history from ATCF best-track or operational b-deck style data,
* gathers recent environmental fields from GFS,
* gathers recent satellite imagery from GOES,
* gathers sea-surface temperature data,
* builds storm-centred grids for several recent synoptic times,
* normalises the data using checkpoint statistics,
* and estimates the current V<sub>max</sub> in knots.

== Principles ==
TC Evolution was built around several design principles.

=== Evolution over snapshot analysis ===
The model is intended to represent the idea that tropical cyclone intensity is not purely a visual snapshot problem. Instead, intensity depends on how the storm has been evolving over recent hours. For that reason, the model uses:
* previous intensity,
* recent intensity tendency,
* storm motion,
* and a sequence of recent frames rather than a single time step.

=== Inner-core structure matters ===
The system uses both a full storm-centred satellite crop and a zoomed inner-core crop. This was intended to let the model separate broad-scale storm organisation from features such as eye definition, eyewall structure, and core symmetry.

=== Environment matters ===
The model does not treat the cyclone as an isolated image. Environmental wind fields, low- and upper-level flow, shear-related diagnostics, SST, and land interaction are included because storm intensity is strongly linked to environmental context.

=== Operational practicality ===
The inference system was designed around data sources that can be retrieved operationally, especially for storms not yet present in finalized archival best-track files.

== Architecture ==
The Beta 1.6 implementation uses a neural network named '''TropicalCurrentIntensityNet'''.

=== Input format ===
The model ingests a storm-centred sequence with shape:

<pre>
[channels, time, height, width]
</pre>

In Beta 1.6, the input uses:
* '''16 channels'''
* '''41 × 41''' spatial grids
* a short sequence of recent 6-hourly frames

=== Channel structure ===
The 16 channels are arranged as follows:

{| class="wikitable"
! Channel group !! Channels !! Description
|-
| Environmental fields || 10 || u10, v10, u850, v850, u200, v200, gh850, rh850, shear200_850, shear850_sfc
|-
| Satellite full field || 2 || infrared (IR), water vapour (WV)
|-
| Satellite zoom field || 2 || zoomed infrared, zoomed water vapour
|-
| Surface context || 2 || sea-surface temperature (SST), land mask
|}

=== Scalar features ===
In addition to the gridded input, TC Evolution uses a separate scalar feature vector. In Beta 1.6, this includes:
* latitude,
* longitude encoded as sine and cosine,
* month encoded as sine and cosine,
* basin one-hot encoding,
* zonal and meridional storm motion,
* V<sub>max</sub> at t−6 h,
* V<sub>max</sub> at t−12 h,
* 6-hour intensity change,
* and availability flags for recent intensities.

=== Branch encoders ===
The model uses three separate 2D residual backbones:

==== Full satellite encoder ====
The '''full_sat_enc''' branch processes the two-channel full storm satellite input.

==== Zoom satellite encoder ====
The '''zoom_sat_enc''' branch processes the two-channel inner-core zoom satellite input.

==== Environmental encoder ====
The '''env_enc''' branch processes the remaining twelve channels:
* ten environmental fields,
* SST,
* and land mask.

Each branch is built from:
* convolution layers,
* [[Group normalization|GroupNorm]],
* GELU activation,
* residual 2D blocks,
* progressive downsampling,
* and adaptive average pooling.

=== Temporal encoder ===
For each time step, the outputs of the three branches are concatenated and projected into a shared hidden representation. These per-frame embeddings are then fed into a temporal Transformer encoder with:
* a learned class token,
* learned temporal positional embeddings,
* hidden dimension 384,
* 8 attention heads,
* and 4 Transformer layers.

The temporal encoder is intended to summarize short-term storm evolution rather than a single static frame.

=== Scalar fusion and output ===
The scalar feature vector is processed by a small scalar network and fused with the Transformer summary. The fused representation is then used to produce the final intensity estimate.

The Beta 1.6 operational implementation predicts intensity through a delta formulation:

<math>\hat{V}_{t} = V_{t-6} + \Delta V</math>

This means the model predicts the present intensity as the previous 6-hour intensity plus a learned adjustment.

== Data ==
=== Stage A training data ===
The satellite-only pretraining stage used storm-centred data from '''1998–2024''', consisting of:
* infrared imagery,
* water-vapour imagery,
* zoomed infrared imagery,
* and zoomed water-vapour imagery.

This stage was intended to provide long-horizon structural pretraining for tropical cyclone appearance.

=== Stage B training data ===
The multimodal fine-tuning stage used storm-centred data from '''2015–2024'''. These data included:
* GFS-derived environmental fields,
* satellite IR and WV imagery,
* zoomed satellite imagery,
* SST,
* land mask,
* and best-track intensity labels.

=== Label source ===
Training targets were based on synoptic tropical cyclone intensity labels. For operational 2025-style testing and inference, ATCF best-track or b-deck style records were used when final archival products were not the intended operational source.

=== Spatial setup ===
The storm-centred grid configuration used in Beta 1.6 is:

{| class="wikitable"
! Parameter !! Value
|-
| Output resolution || 0.25°
|-
| Outer box size || 10.0°
|-
| Inner zoom size || 4.0°
|-
| Final grid size || 41 × 41
|}

== Training ==
=== Stage A results ===
The satellite-only stage was treated mainly as a representation-learning stage rather than the final operational model. Development logs reported the following Stage A test metrics:

{| class="wikitable"
! Metric !! Value
|-
| Score || 8.620
|-
| RMSE || 5.73 kt
|-
| RMSE 96+ || 9.91 kt
|-
| RMSE 137+ || 13.91 kt
|}

=== Stage B results ===
The multimodal fine-tuning stage improved performance and became the basis of the operational model. Development logs reported the following final Stage B test metrics:

{| class="wikitable"
! Metric !! Value
|-
| Score || 7.413
|-
| RMSE || 5.47 kt
|-
| RMSE 96+ || 7.24 kt
|-
| RMSE 137+ || 12.54 kt
|-
| RMSE 96–112 || 6.86 kt
|-
| RMSE 113–136 || 7.39 kt
|-
| RMSE 137+ bin || 12.54 kt
|}

These results were interpreted as an improvement over the satellite-only stage, especially in the stronger-storm regime.

== Performance ==
=== Informal 2025 operational spot checks ===
During 2025-style operational testing, several storm cases were manually compared against best-track values and the [[Advanced Dvorak technique|Advanced Dvorak Technique]] (ADT).

{| class="wikitable"
! Storm ID !! Time (UTC) !! Best-track V<sub>max</sub> !! TC Evolution !! Absolute error
|-
| AL132025 || 2025-10-28 12:00 || 155 kt || 151.5 kt || 3.5 kt
|-
| AL082025 || 2025-09-27 00:00 || 120 kt || 100.0 kt || 20.0 kt
|-
| AL022025 || 2025-06-29 18:00 || 40 kt || 37.1 kt || 2.9 kt
|}

In this small informal sample:
* the model performed very well on one extreme-intensity case,
* underestimated one major hurricane case,
* and performed very well on one weak-storm case.

=== ADT comparison ===
The same spot-check sample was also compared against ADT values provided during testing. In that limited comparison:
* TC Evolution outperformed ADT on one extreme-intensity case,
* ADT outperformed TC Evolution on one mid-major hurricane case,
* and both were effectively correct on one weak-storm case.

This comparison was informal and not presented as a formal full-season skill study.

== Operational implementation ==
The Beta 1.6 inference code includes:
* safe PyTorch checkpoint loading,
* ATCF storm-history retrieval,
* GFS file selection and download,
* GOES file selection and processing,
* OISST retrieval and interpolation,
* storm-centred frame construction,
* and final model inference from a saved multimodal checkpoint.

The operational implementation normalises gridded and scalar features using statistics stored in the model checkpoint.

== Strengths ==
Observed and intended strengths of TC Evolution include:
* explicit modelling of storm evolution rather than single-frame regression,
* separate treatment of full-storm and inner-core satellite structure,
* integration of environmental context,
* direct use of previous intensity history,
* and demonstrated ability to represent very intense hurricanes in at least some operational spot checks.

== Limitations ==
Known limitations of Beta 1.6 include:
* case-dependent errors in some major-hurricane situations,
* limited formal operational verification relative to established methods,
* dependence on upstream file availability and storm-centred data extraction quality,
* potential data-distribution differences between training imagery and operational imagery,
* and the fact that the system remains a beta research model rather than an official operational standard.



== Naming ==
The name '''TC Evolution''' reflects the central idea of the model: tropical cyclone intensity is treated as an evolving state shaped by recent storm history, internal structure, and environmental conditions.

== Intended use ==
TC Evolution is intended for:
* experimental tropical cyclone current-intensity analysis,
* case-study review,
* research prototyping,
* and internal benchmarking against satellite-based techniques.

It is not intended to replace official agency products without broader validation and full operational testing.