XiChen: AI Weather Forecast System

• XiChen is a fully AI-driven global weather forecasting system that employs a single pretrained neural foundation model for data assimilation and medium-range forecasting.
• It leverages a Conditional Hybrid Neural Operator architecture that combines Fourier and convolution branches with cross-attention for both global and local feature extraction.
• The system delivers rapid, end-to-end forecasts in approximately 17 seconds on a single GPU while achieving competitive accuracy and scalability with integrated satellite observation modules.

XiChen is a fully AI-driven global weather forecasting system in which both data assimilation and medium-range forecasting are implemented by a single pretrained neural “foundation” model that is subsequently fine-tuned for multiple operational roles (Wang et al., 12 Jul 2025). It is described as the first observation-scalable fully AI-driven global weather forecasting system, replacing the core numerical components of a traditional Numerical Weather Prediction pipeline with neural components while retaining training dependence on ERA5 reanalysis and observations. XiChen operates on a global $128 \times 256$ grid at $1.40625^\circ$ resolution, predicts 49 variables, assimilates GDAS prepbufr conventional observations together with raw radiances from AMSU-A and MHS, and produces 6-hourly forecasts out to 10 days. The reported end-to-end latency from observations to analysis to 10-day forecast is approximately 17 seconds on a single NVIDIA A100, while the reported skillful forecasting lead time exceeds 8.25 days under the criterion $\mathrm{ACC}(Z500) > 0.6$ (Wang et al., 12 Jul 2025).

1. Scope, targets, and defining properties

XiChen is designed as a complete global forecasting pipeline rather than as an isolated forecast model. Its defining property is that no physical NWP model is used either for the forecast or for data assimilation. The forecast model is a neural network, the data assimilation solver is a neural network, and the satellite observation operators are neural networks. The only NWP contribution identified in the system description is the training data, specifically ERA5 reanalysis together with observations (Wang et al., 12 Jul 2025).

The target configuration consists of a global grid at $1.40625^\circ$ resolution, corresponding to $128 \times 256$, with 49 variables. These comprise 5 atmospheric fields—$Z$, $T$, $Q$, $U$, and $V$—over 9 pressure levels, plus 4 surface variables: $1.40625^\circ$0, $1.40625^\circ$1, $1.40625^\circ$2, and $1.40625^\circ$3. Forecast output is generated every 6 hours up to a horizon of 10 days (Wang et al., 12 Jul 2025).

The term “observation-scalable” has a specific architectural meaning in XiChen. New observation types, especially satellites, can be incorporated by fine-tuning a reusable foundation model into new observation operators and new data-assimilation components, without redesigning or retraining the entire system. The data-assimilation pipeline is therefore cascaded and modular: modules can be added for new instruments or skipped when a data stream is temporarily unavailable. This suggests that XiChen is intended to mimic an operational plug-in workflow for evolving observing systems rather than a fixed single-dataset benchmark configuration.

2. Neural architecture and end-to-end workflow

All core components in XiChen share the same neural architecture, termed a Conditional Hybrid Neural Operator built on top of a Hybrid Neural Operator (Wang et al., 12 Jul 2025). The HNO combines an AFNO-type Fourier branch for global interactions with a convolutional branch for local features. The CHNO augments this backbone with cross-attention that injects conditional information such as lead time, satellite scan geometry, or 4D-Var gradients. Through this conditioning, the same backbone is used as a forecast model, a satellite observation operator, and a data-assimilation model.

At analysis time $1.40625^\circ$4, the background field $1.40625^\circ$5 is provided by the previous cycle’s XiChen forecast. Observations within the data-assimilation window $1.40625^\circ$6 consist of conventional observations $1.40625^\circ$7 from GDAS prepbufr and satellite radiances $1.40625^\circ$8 from AMSU-A and MHS. XiChen then forms a 4D-Var-like cost function using the AI forecast model $1.40625^\circ$9 as the time-evolution operator and the learned AI observation operator $\mathrm{ACC}(Z500) > 0.6$0 for satellite radiances. Automatic differentiation through the forecast and observation networks yields the gradient of the cost with respect to the background state, $\mathrm{ACC}(Z500) > 0.6$1, and the data-assimilation model then maps this gradient, conditioned on the background field, to an analysis field $\mathrm{ACC}(Z500) > 0.6$2 (Wang et al., 12 Jul 2025).

The forecast stage uses two fine-tuned instances of the same backbone: XiChen-Short and XiChen-Medium. XiChen-Short is used for the first 3 days, and XiChen-Medium for days 3–10, with lead-time conditioning. This yields 6-hourly forecasts out to 10 days.

XiChen also implements a dual data-assimilation framework. A 12-hour DAW, using observations at 0, 3, 6, and 9 hours, is used to run the data-assimilation cycle and update the background. A 3-hour DAW, using observations only at the analysis time, is used to produce real-time initial conditions for medium-range forecasting, thereby avoiding the latency associated with waiting for the full 12-hour window. The paper explicitly likens this to avoiding the inherent lag of classical 4D-Var (Wang et al., 12 Jul 2025).

3. Data assimilation and the meaning of “4D variational knowledge”

XiChen constructs an explicit 4D-Var-like objective,

$\mathrm{ACC}(Z500) > 0.6$3

Here $\mathrm{ACC}(Z500) > 0.6$4 is the control vector, $\mathrm{ACC}(Z500) > 0.6$5 is the background state, $\mathrm{ACC}(Z500) > 0.6$6 is the neural forecast model, and $\mathrm{ACC}(Z500) > 0.6$7 is the observation operator, decomposed into identity plus masking for conventional gridpoint observations and a learned neural operator $\mathrm{ACC}(Z500) > 0.6$8 for satellite radiances (Wang et al., 12 Jul 2025).

In practice, XiChen does not require an explicit background covariance $\mathrm{ACC}(Z500) > 0.6$9 in the data-assimilation update because only the gradient with respect to $1.40625^\circ$0 is used, and $1.40625^\circ$1 is effectively absorbed into the neural DA model. For conventional observations, the observation covariance is written as $1.40625^\circ$2; for satellites, $1.40625^\circ$3, using observation-operator error estimated from training (Wang et al., 12 Jul 2025).

The phrase “4D variational knowledge” refers to the fact that XiChen emulates classical 4D-Var structure without solving an adjoint PDE. It does so by propagating the initial state through the AI forecast model across the data-assimilation window, mapping model states to observation space through AI observation operators, computing the 4D-Var cost and its gradient through automatic differentiation, and training a learned DA operator to convert that gradient signal into analysis increments. The paper identifies the gradient as encoding temporal structure, spatial correlations, and multivariable couplings. In this formulation, the data-assimilation network is explicitly trained to invert the 4D-Var gradient signal into an analysis increment field (Wang et al., 12 Jul 2025).

The data-assimilation pipeline is cascaded. A prepbufr DA model is applied first, an AMSU-A DA model second using the post-prepbufr analysis as background, and an MHS DA model third using the post-AMSU-A analysis as background. Each stage uses its own 4D-Var gradient based on its observation subset. This modular ordering is the operational basis of observation scalability.

Observation handling is likewise specified in detail. GDAS prepbufr observations are pre-interpolated to the model grid and to 3-hourly times using a $1.40625^\circ$4 hour window, represented as a tensor $1.40625^\circ$5, with unobserved entries set to NaN. AMSU-A channels 5–10 and MHS channels 3–5 are nearest-neighbor interpolated from raw swath data and auxiliary satellite information to the same grid and temporal spacing. Brightness temperatures below 150 K or above 350 K are discarded, and latitudes beyond $1.40625^\circ$6 are ignored to avoid sea-ice issues. In addition, observations whose innovation exceeds $1.40625^\circ$7 of the relevant variable are excluded from the 4D-Var cost to prevent over-corrective gradients (Wang et al., 12 Jul 2025).

The reported satellite impacts are asymmetric. AMSU-A strongly improves temperature and geopotential analyses and also improves winds via dynamical balances. MHS improves humidity fields but, when used alone, tends to degrade temperature and geopotential, with the paper attributing this likely to cloud, precipitation, and surface effects not fully modeled, together with imbalance induced by humidity adjustments. When AMSU-A and MHS are assimilated together, errors in temperature, geopotential, and humidity are reduced, and the MHS-only degradation is mitigated. This suggests that XiChen’s observation-scalable design is not merely a software property but materially affects forecast-relevant balance relationships across variables.

4. Foundation-model pretraining and task-specific fine-tuning

XiChen’s foundation model is pretrained for medium-range forecasting on ERA5 reanalysis from 2010–2021, with 2022 used for validation and 2023 for test data (Wang et al., 12 Jul 2025). The pretraining task is conditional one-step forecasting: given a state $1.40625^\circ$8, the model predicts $1.40625^\circ$9 for $128 \times 256$0 hours using lead time as a conditioning scalar. The single-step loss is

$128 \times 256$1

with

$128 \times 256$2

and latitude weighting

$128 \times 256$3

The pressure weight $128 \times 256$4 is described as being as in GraphCast and FuXi. Pretraining requires approximately 45 hours on $128 \times 256$5A100-40GB (Wang et al., 12 Jul 2025).

To reduce autoregressive error accumulation, XiChen is then fine-tuned with multi-step rollout loss,

$128 \times 256$6

The procedure is staged. Starting from the pretrained model, fine-tuning with $128 \times 256$7 yields XiChen-Short, optimized for 1–5 day forecasts. XiChen-Medium is initialized from XiChen-Short and fine-tuned with $128 \times 256$8 up to 10 to optimize 5–10 day forecasts. Total forecast fine-tuning is reported as approximately 94 hours on $128 \times 256$9A100 (Wang et al., 12 Jul 2025).

The same backbone is then fine-tuned as a satellite observation operator. For each satellite type $Z$0, XiChen learns

$Z$1

where $Z$2 denotes auxiliary satellite information such as scan geometry. The observation-operator loss is

$Z$3

with $Z$4 for AMSU-A and $Z$5 for MHS. Each observation-operator fine-tune costs approximately 13.3 hours on $Z$6A100 (Wang et al., 12 Jul 2025).

Finally, the backbone is fine-tuned as a data-assimilation model. The DA operator $Z$7 is trained to map

$Z$8

with the loss

$Z$9

Training proceeds in stages: a prepbufr DA model is fine-tuned from the forecast foundation model in approximately 45 hours; an AMSU-A DA model is initialized from the prepbufr DA model and trained in approximately 94 hours; and an MHS DA model is initialized from the AMSU-A DA model and trained in approximately 222 hours, with AMSU-A versus MHS order randomized during training to encourage robustness (Wang et al., 12 Jul 2025).

5. Reported forecasting and analysis performance

The primary experimental protocol is a one-year continuous DA cycle over 2023, with 12-hourly analyses at 00 and 12 UTC, a 12-hour data-assimilation window using observations at 0, 3, 6, and 9 hours after analysis time, and backgrounds supplied by the XiChen forecast model (Wang et al., 12 Jul 2025). Medium-range forecasts are run for 10 days at 6-hour steps. Initial conditions are XiChen analysis fields and use timing consistent with WeatherBench and DABench, with 50 initial dates per year spaced every 336 hours. Baselines are operational GFS and ECMWF IFS HRES. Evaluation metrics are latitude-weighted RMSE, ACC, and Bias on $T$0, $T$1, $T$2, $T$3, $T$4, $T$5, $T$6, and $T$7.

For the 2023 DA cycle, XiChen analysis fields using prepbufr, AMSU-A, and MHS together are reported to have RMSE consistently lower than GFS analysis for $T$8, $T$9, $Q$0, $Q$1, $Q$2, $Q$3, and $Q$4, and to approach IFS HRES. Bias remains closer to zero than GFS, with stable behavior after initial spin-up. The paper notes occasional RMSE spikes corresponding to missing observations, followed by recovery in subsequent DA cycles, which is presented as evidence of robustness (Wang et al., 12 Jul 2025).

For medium-range forecasting, XiChen is reported to have lower RMSE than GFS for $Q$5, $Q$6, $Q$7, $Q$8, and $Q$9 for nearly all lead times out to 10 days. At longer lead times, those RMSEs are comparable to or better than IFS HRES. For $U$0, XiChen surpasses GFS after approximately 2 days and surpasses IFS HRES after approximately 4 days. For $U$1, XiChen has higher RMSE than both GFS and IFS HRES during roughly the first 6 days, then surpasses GFS after approximately 7 days and surpasses IFS HRES after approximately 8 days (Wang et al., 12 Jul 2025).

The authors interpret this time-dependent $U$2 behavior as an effect of smoothing: XiChen is described as more smoothed, reducing high-frequency error growth at longer range. Spectral analysis is said to show loss of small-scale variance, similar to that observed in GraphCast, FengWu, and FuXi. A plausible implication is that XiChen’s medium-range competitiveness is currently achieved partly through a scale-selective error structure rather than through uniformly superior fidelity across all resolved scales.

Skillful forecast lead time is defined as $U$3 ACC $U$4. Under that criterion, XiChen exceeds 8.25 days, matching GFS at 8.25 days and exceeding Aardvark, reported as below 8 days, and GraphDOP, reported as approximately 5 days. When initialized directly from ERA5 rather than from XiChen data assimilation, XiChen can maintain ACC $U$5 beyond 10 days, which the paper interprets as evidence that current horizon limitations are dominated by data-assimilation error rather than forecast-model capacity (Wang et al., 12 Jul 2025).

Tropical cyclone evaluation uses TempestExtremes with a NeuralGCM configuration and IBTrACS as reference. ERA5 identifies 52 out of 81 IBTrACS tropical cyclones, corresponding to 64.2%, with mean track error of approximately $U$6. XiChen data-assimilation analyses with different observation combinations yield numbers of detected tropical cyclones and spatial distributions resembling ERA5, with similar average errors. AMSU-A assimilation increases the number of detected tropical cyclones relative to prepbufr-only assimilation, while MHS-only assimilation reduces detection because of degraded dynamical fields. In 5-day tropical cyclone forecasts initialized at IBTrACS start times, MHS-only assimilation increases track error beyond 48 hours, whereas MHS substantially reduces intensity errors and AMSU-A also helps intensity. The largest intensity improvements occur when AMSU-A and MHS are used together. Minimum-sea-level-pressure errors are described as comparable to FuXi-Extreme, despite FuXi-Extreme operating at much higher $U$7 resolution (Wang et al., 12 Jul 2025).

6. Efficiency, comparison class, limitations, and implications

XiChen’s computational profile is central to its positioning. At $U$8 resolution on a single A100 GPU, the complete prepbufr + AMSU-A + MHS data-assimilation cascade requires approximately 5 seconds, the 10-day forecast using XiChen-Short and XiChen-Medium requires approximately 12 seconds, and the total observation-to-forecast pipeline requires approximately 17 seconds (Wang et al., 12 Jul 2025). The system is also described as more than 400 times faster than an operational system like ECMWF IFS HRES for comparable global medium-range forecasts. In contrast, traditional 4D-Var plus NWP pipelines are described as requiring hours on a supercomputer and as assimilating only 5–10% of available satellite data because of computational constraints.

Relative to traditional NWP, XiChen differs in three ways. First, the forecast component is purely neural rather than PDE-based. Second, the data-assimilation update is purely neural, although it uses a 4D-Var cost and its gradient. Third, the observation operators are neural surrogates rather than hand-crafted radiative transfer models. The reported advantages are reduced latency and resource requirements, elimination of tangent-linear and adjoint maintenance, and easier adaptation to new instrumentation. The reported limitations are dependence on ERA5 for training, smoothing and loss of small scales, limited explicit physics constraints, and the current absence of ensemble or probabilistic forecasting (Wang et al., 12 Jul 2025).

Relative to other AI weather systems, XiChen is distinguished in the paper from GraphCast, Pangu-Weather, FourCastNet, FengWu, and FuXi by the fact that those systems use reanalysis as initial conditions and do not provide an AI data-assimilation front end. XiChen’s novelty is therefore that it creates its own analyses operationally rather than relying on externally produced NWP-based initial conditions. It is also distinguished from Aardvark and GraphDOP as an end-to-end AI system from observations with a longer reported skillful lead time, and from FuXi-DA, FengWu-Adas, DiffDA, and 4DVarFormer by its assimilation of raw satellite radiances, use of the 4D-Var gradient as a DA input, and cascaded modular DA framework (Wang et al., 12 Jul 2025).

Several limitations are explicitly emphasized. XiChen currently operates only at $U$9 resolution, whereas finer-scale weather, convection, and local extremes would require scaling to $V$0 or beyond. Only two satellite instruments, AMSU-A and MHS, are used, while operational NWP systems assimilate much richer satellite constellations. MHS-only assimilation can induce physical imbalance and degrade temperature and geopotential, motivating future inclusion of additional state variables such as cloud water, cloud ice, and rainfall together with explicit physical consistency or balance constraints. XiChen is deterministic only, so probabilistic or ensemble extensions are identified as necessary for decision-making use cases. Finally, because ERA5 reanalysis is required as training truth, XiChen is explicitly framed not as a full replacement for NWP but as a complement to it (Wang et al., 12 Jul 2025).

Operationally, the reported design implies several concrete use cases. A single modern GPU can produce global 10-day forecasts in seconds; the dual-DA approach can produce medium-range products earlier than a full 12-hour 4D-Var cycle; new satellites can be integrated by fine-tuning an observation operator and a DA module; and the system can be run alongside NWP as a fast backup, for scenario generation, or for Observation System Simulation Experiments. The broader significance claimed by the paper is that a single pretrained foundation model, when suitably conditioned and fine-tuned, can span observation operators, data assimilation, and medium-range forecasting within one coherent AI pipeline (Wang et al., 12 Jul 2025).