U-Cast: Forecasting, Antenna & Satellite INAC

• U-Cast is a multi-context term that encompasses a U-Net based weather forecasting framework, a unicast beamforming technique for circular mm-wave antennas, and a uni-cast satellite integrated navigation and communication design.
• In weather forecasting, U-Cast employs a two-stage training process—deterministic MAE pre-training followed by CRPS fine-tuning—and achieves ensemble performance with over 10× reduced computational cost.
• In wireless and satellite applications, U-Cast enables agile beam steering through sum-of-modes phase control and optimizes power allocation for uni-cast prioritization, enhancing both data throughput and positioning accuracy.

U-Cast is a term with multiple contexts across contemporary scientific research, signifying (1) a state-of-the-art probabilistic weather forecasting method based on a U-Net backbone, (2) a flexible unicast beamforming technique for circular millimeter-wave antenna arrays, and (3) a scenario in satellite integrated navigation and communication leveraging non-orthogonal multiple access for uni-cast prioritization. This article systematically addresses each usage within its technical domain, following the terminology and methodologies as established in the cited literature.

1. U-Cast in Probabilistic AI Weather Forecasting

U-Cast denotes a simple yet frontier-level AI-based probabilistic weather forecasting framework that leverages a standard U-Net backbone architecture, advancing beyond previous approaches that required domain-specific complexity or large computational resources (Cachay et al., 10 Apr 2026).

U-Net Architecture and Design

The forecasting model employs a "Dhariwal" U-Net featuring four encoder-decoder down/up-sampling stages and self-attention at the two coarsest resolution levels. Residual convolutional layers are stacked within each resolution block; feature channels increase by factors [1, 2, 3, 4] from a base width of 320, amounting to ∼895 million parameters. The architecture integrates circular longitudinal padding for atmospheric continuity, bilinear decoder upsampling for 121×240 (1.5° resolution) grids, and omits adaptive LayerNorm in favor of standard Dropout for parameter efficiency and simplicity.

Two-Stage Training Process

U-Cast establishes its probabilistic skill through a two-stage training methodology:

1. Stage 1 (Deterministic Pre-training): The network is first trained as a deterministic 1-member forecaster by minimizing the Mean Absolute Error (MAE) between the predicted and target atmospheric states. The loss function

$$\mathcal{L}_{\text{det}} = \frac{1}{HW} \sum_{h,w} a_h \| f_\theta(x)_{h,w} - u_{t+1,h,w} \|_1$$

employs standard latitude area weights, with training data drawn from ERA5 (1979–2019, 12 h input/output).

2. Stage 2 (Probabilistic CRPS Fine-Tuning): Upon deterministic convergence, fine-tuning is performed using the Continuous Ranked Probability Score (CRPS) as the objective, employing Monte Carlo Dropout for stochastic ensemble generation. Skill and spread are computed over $M$ dropout samples, and the probabilistic loss is:

$$\mathcal{L}_{\text{prob}} = \frac{1}{HW} \sum_{h,w} a_h \left[ \text{Skill}_{h,w} - \frac{1}{2} \text{Spread}_{h,w} \right]$$

This stage converges rapidly (≈8 epochs, ≈1.2 H200-days).

Deep ensembling is achieved by performing multiple (K=4) fine-tuning restarts with distinct seeds, combining the results with multiple dropout draws (N=15), yielding a 60-member ensemble.

Comparative Performance and Efficiency

U-Cast's probabilistic skill, as quantified by CRPS, is reported to match or exceed leading models such as GenCast and IFS ENS at 1.5° resolution. In 60-step (15 day) forecasts, U-Cast DeepEns achieves, for example, z500 10d CRPS of 253—surpassing IFS ENS (262) and slightly ahead of GenCast (254). Efficiency gains are substantial: training time is reduced by over 10× (sub-12 H200 GPU-days versus 100–150 accelerator days for comparators), and inference latency for a 60-member ensemble is 11 seconds, over 10× faster than diffusion-based models.

Model	z500 1d CRPS	z500 10d CRPS	10 m u 1d CRPS	Training	Inference (60 steps)
U-Cast	20.3	256	0.349	8.2 H200-day	11 s/H200
U-Cast (DE)	19.6	253	0.345	11.8 H200-day	11 s/H200
IFS ENS	22.4	262	0.406	—	—
GenCast	20.2	254	0.332	112+48 TPU-d	16 min/TPU

These findings demonstrate that highly efficient and scalable architectures can achieve "frontier" probabilistic weather forecasting performance without resorting to specialized components or significant computational expense (Cachay et al., 10 Apr 2026).

Key Implications

• Architectural excess (e.g., diffusion loops, spherical graph nets, adaptive LayerNorm) is not required for SOTA probabilistic skill.
• The two-stage (MAE → CRPS) curriculum accelerates convergence and reduces computational burden.
• Monte Carlo Dropout, coupled with a spread-aware loss, enables ensemble generation without additional architectural machinery.
• U-Cast's low requirements democratize access to high-quality weather models, opening capabilities to smaller research centers.

2. U-Cast in Millimeter-Wave Circular Antenna Arrays

U-Cast further denotes a steering and beamforming technique for unicast data delivery in circular millimeter-wave antenna arrays, relevant to advanced wireless applications and high spectral efficiency systems (Assimonis et al., 2020).

Array Factor and Unicast Excitation

A twelve-element microstrip patch array, arranged on a circle in the $xy$-plane, uses an excitation strategy that exploits the orthogonality of azimuthal modes. In the unicast (U-Cast) mode, the excitation weights

$w_n = 1 + 2\sum_{k=1}^{M} \cos(k\phi_n)$

(sum over $m=-M$ to $M$) focus energy into a single directed lobe, as opposed to $m=0$ (broadcast/omnidirectional). Steering the main lobe to an arbitrary azimuth $\phi_0$ is achieved via a common phase shift:

$$w_n(\phi_0) = 1 + 2\sum_{k=1}^{M} \cos[k(\phi_n - \phi_0)]$$

Performance Characteristics

Measured and simulated performance confirm that, for U-Cast mode ($M$0 or $M$1 configurations), the array achieves:

• Peak horizontal-plane field: approximately 42.8 dB V/m
• Directivity: 8.9 dBi (simulation), with measured values within 0.5 dB
• Half-power beamwidth: approximately 60°–70° in azimuth
• Sidelobe level: –12 to –15 dB below the main lobe
• Radiation efficiency: >90%
• S₁₁ < –10 dB over a 9.5% bandwidth
• Mutual coupling: Sₙ,ₙ₊₁ ≈ –24 dB at 28 GHz

The feed network requires only phase control per element, as amplitude is equal-weighted. Steering is dynamically achievable via electronic phase shifters (Assimonis et al., 2020).

U-Cast Mode	Directivity (dBi)	HPBW (deg)	Sidelobe Level (dB)
M = 5 (A)	~8.6	~50	~–10
M = 3 (B)	~8.9	60–70	~–12 to –15

Design Trade-Offs

Increasing the number of modes $M$2 narrows the main beam at the expense of higher sidelobe levels. The absence of amplitude tapering simplifies feed design and avoids efficiency losses, while highly efficient element patterns and tight mutual coupling control are essential for maintaining the desired beam profiles.

3. U-Cast–Oriented INAC in Satellite Integrated Navigation and Communication

Within the field of satellite signal design, "U-Cast" refers to the uni-cast-oriented scenario in Multi-/Uni-Cast Non-Orthogonal Multiple Access–based Integrated Navigation and Communication (INAC) (Zhang et al., 26 May 2026).

Signal Structure and Power Allocation

The unified DSSS-NOMA waveform superposes navigation and high-rate uni-cast communication signals using a shared pseudo-noise (PN) sequence:

$M$3

with $M$4, $M$5. The uni-cast-oriented (UO-INAC) case prioritizes the communication stream, $M$6, allocating more power to uni-cast.

Receiver Design and SIC Decoding

At the receiver:

1. Uni-cast data are first decoded (with correlator output and hard decision);
2. The decoded uni-cast component is reconstructed and subtracted;
3. The navigation (multi-cast) stream is recovered from the residual.

This successive interference cancellation (SIC) approach is required due to the superposed, non-orthogonal nature of the signals.

Analytical Performance Metrics

Closed-form expressions are derived for bit error rate (BER) and ranging accuracy:

• Uni-cast BER:

$M$7

• Navigation BER (after SIC):

$M$8

• Positioning accuracy:

$M$9

Increasing $$\mathcal{L}_{\text{prob}} = \frac{1}{HW} \sum_{h,w} a_h \left[ \text{Skill}_{h,w} - \frac{1}{2} \text{Spread}_{h,w} \right]$$0 reduces uni-cast BER exponentially, yields lower navigation BER in many regimes (due to improved SIC), and enhances ranging accuracy by increasing code-phase SNR. However, higher communication rates ($$\mathcal{L}_{\text{prob}} = \frac{1}{HW} \sum_{h,w} a_h \left[ \text{Skill}_{h,w} - \frac{1}{2} \text{Spread}_{h,w} \right]$$1) degrade both navigation performance and BER due to diminished integration times (Zhang et al., 26 May 2026).

Practical and System-Level Implications

UO-INAC facilitates high-rate uni-cast services alongside continuous navigation, useful in scenarios such as satellite backhaul or broadband feeder links. The power allocation parameter $$\mathcal{L}_{\text{prob}} = \frac{1}{HW} \sum_{h,w} a_h \left[ \text{Skill}_{h,w} - \frac{1}{2} \text{Spread}_{h,w} \right]$$2 enables dynamic trading of navigation robustness for communication throughput. For $$\mathcal{L}_{\text{prob}} = \frac{1}{HW} \sum_{h,w} a_h \left[ \text{Skill}_{h,w} - \frac{1}{2} \text{Spread}_{h,w} \right]$$3, sub-meter ranging and $$\mathcal{L}_{\text{prob}} = \frac{1}{HW} \sum_{h,w} a_h \left[ \text{Skill}_{h,w} - \frac{1}{2} \text{Spread}_{h,w} \right]$$4 are attainable at typical MEO link budgets.

4. Methodological Themes and Cross-Domain Insights

Across all U-Cast contexts, the following methodological themes emerge:

• Simplicity and Modularity: U-Cast weather forecasting and unicast array beamforming both illustrate that general, simple architectures—U-Net or sum-of-modes phase control—can match or exceed domain-specific complexity. This supports a "complexity-by-necessity" approach: employ only justifiable specialized components (Cachay et al., 10 Apr 2026, Assimonis et al., 2020).
• Resource Efficiency: Each methodology is designed to achieve frontier-level performance with minimal computational/architectural overhead, democratizing access and deployment. U-Cast weather forecasting, in particular, offers >10× savings in both training and inference time (Cachay et al., 10 Apr 2026).
• Performance/Resource Trade-Offs: Every instance—forecasting ensemble size, array beamwidth/sidelobe levels, or $$\mathcal{L}_{\text{prob}} = \frac{1}{HW} \sum_{h,w} a_h \left[ \text{Skill}_{h,w} - \frac{1}{2} \text{Spread}_{h,w} \right]$$5 in satellite integration—manages explicit trade-offs to meet application-specific requirements.

5. Implications, Limitations, and Future Prospects

U-Cast methodologies have distinct domain implications:

• AI Weather Prediction: The U-Cast approach positions SOTA ensemble weather forecasting within reach of academic and smaller industry groups by minimizing hardware requirements, while challenging the necessity of specialized deep learning designs for probabilistic skill (Cachay et al., 10 Apr 2026). Further research may examine scaling to finer resolutions or hybridizing with physics-based post-processing.
• Antenna Arrays: U-Cast phase-sum beamforming supports agile, high-efficiency spatial selectivity for mm-wave wireless, with straightforward integration in 5G/6G networks (Assimonis et al., 2020). Extensions may examine massive MIMO, dynamic environments, or hybrid analog-digital beamforming.
• Satellite Communication: The UO-INAC scenario enables tightly integrated PNT+comms signaling with tunable quality-of-service and spectrum reuse for code-limited constellations. Research directions include robust adaptive SIC, code design for multiple overlapping streams, and hardware realization for very high-rate services (Zhang et al., 26 May 2026).

No evidence was found indicating controversies surrounding the U-Cast approaches. A plausible implication is that, as these approaches lower the barrier to entry and operational cost, they will accelerate broader adoption of probabilistic ensemble forecasting, flexible mm-wave deployments, and integrated satellite-communication architectures, potentially shifting research focus from architectural complexity to data curation, interpretability, and system integration.