Met2Net: Forecasting, Photonics & Electromagnetics

• Met2Net is a term used to label distinct research frameworks in meteorology, photonics, and reconfigurable metasurfaces.
• Its meteorological variant, Met²Net, employs variable-specific encoders, latent translators, and implicit two-stage training to improve spatio-temporal forecasting.
• In photonics and electromagnetics, the term describes a benchmarking repository and a mapping from physical metasurfaces to communication channels.

Met2Net is a term used in multiple, non-equivalent senses in recent arXiv literature. Most specifically, it denotes “Met$^2$Net: A Decoupled Two-Stage Spatio-Temporal Forecasting Model for Complex Meteorological Systems,” a multivariable forecasting architecture that combines variable-specific encoders and decoders, a latent-space Translator, and an implicit two-stage training procedure for complex weather prediction (Li et al., 23 Jul 2025). The same string also appears as a misspelling of “MetaNet,” an openly accessible repository for photonic inverse design and benchmarking (Jiang et al., 2020), and as “Met2Net mapping,” the multiport-network-theoretic route from reconfigurable metasurfaces to end-to-end communication channels in wireless systems (Renzo et al., 2024). In contemporary usage, therefore, “Met2Net” is not a single unified framework but a label attached to distinct lines of research in meteorology, photonics, and programmable electromagnetics.

1. Terminological scope and disambiguation

The principal source of ambiguity is orthographic rather than conceptual. In meteorology, “Met$^2$Net” is the formal name of a forecasting model introduced for multivariable spatio-temporal prediction (Li et al., 23 Jul 2025). In photonics, “MetaNet” is the intended name of an online database of photonic devices and inverse-design codes, but the repository is explicitly noted as being “sometimes misspelled as ‘Met2Net’” (Jiang et al., 2020). In electromagnetics and wireless communications, “Met2Net mapping” refers to the transformation from a physical metasurface model, expressed with multiport network theory, into a communication-theoretic end-to-end channel description (Renzo et al., 2024).

Usage of “Met2Net”	Domain	Meaning
Met$^2$Net	Meteorology	Decoupled two-stage spatio-temporal forecasting model
MetaNet (misspelled as “Met2Net”)	Photonics	Repository of photonic device layouts and inverse-design codes
Met2Net mapping	Reconfigurable metasurfaces	MNT-based mapping from physical RIS models to end-to-end channels

A common misconception is to treat these as variants of one framework. They are instead separate research programs that share only a near-identical name. The meteorological model concerns multivariable latent representations and forecasting losses; the photonics platform concerns data sharing, benchmarking, and inverse design; the metasurface usage concerns electromagnetically consistent circuit abstractions for RIS optimization. This suggests that precise citation by arXiv identifier is especially important when the term appears without context.

2. Met$^2$Net in meteorological forecasting

In its most direct sense, Met$^2$Net addresses multivariable weather forecasting under the observation that meteorological variables such as near-surface air temperature (T2M), relative humidity (R), wind (U/V or U10/V10), pressure (MSL), and cloud cover (TCC) interact through nonlinear, scale-dependent processes across space and time (Li et al., 23 Jul 2025). The central diagnosis is “representation inconsistency”: heterogeneous variables have different distributions, temporal rates of change, and spatial patterns, so forcing them into a single end-to-end feature space causes the learned representation to drift away from the characteristics of each source variable. The paper further argues that conventional two-stage multimodal pipelines are suboptimal for forecasting because their two stages optimize incongruent tasks, creating “task inconformity.”

The model therefore treats each meteorological variable as an independent modality. Each variable has its own encoder $E_i(\cdot)$ and decoder $D_i(\cdot)$, while a dedicated Translator $T(\cdot)$ operates in a shared latent space to capture inter-variable interactions and spatio-temporal dynamics. Inputs and targets are multivariable spatio-temporal tensors

$$X \in \mathbb{R}^{T\times N\times C\times H\times W}, \qquad Y \in \mathbb{R}^{T'\times N\times C\times H\times W},$$

where $T/T'$ are input and prediction lengths, $^2$0 is the number of variables, $^2$1 is the number of channels per variable, and $^2$2 is spatial resolution. Variable-wise encodings are formed as

$^2$3

The stated purpose of this decomposition is twofold. First, it preserves variable-specific structure by avoiding destructive homogenization in the first representation layer. Second, it postpones multivariable interaction modeling to a dedicated latent-space module, rather than entangling representation learning and cross-variable coupling inside one monolithic encoder-decoder. In the paper’s framing, this decoupling is the mechanism by which Met$^2$4Net addresses both representation inconsistency and task inconformity.

3. Architecture, latent fusion, and implicit two-stage training

The Translator is the architectural center of Met$^2$5Net. It performs self-attention over the variable axis rather than over flattened spatial tokens, with queries, keys, and values obtained through lightweight 2D CNN projections:

$^2$6

The variable-attention operator is

$^2$7

where the softmax acts over the variable axis. The resulting fused latent tensor is then processed by a spatio-temporal block $^2$8, instantiated in the paper with TAU:

$^2$9

Variable-specific components are sliced from $^2$0 and decoded as

$^2$1

The training strategy is described as an implicit two-stage paradigm implemented within a single continuous training cycle (Li et al., 23 Jul 2025). In Stage I, the Translator is frozen while the encoders and decoders are trained to minimize forecasting error in data space:

$^2$2

The paper specifies standard regression metrics,

$^2$3

In Stage II, the encoders and decoders are frozen while the Translator is trained to predict the latent encodings of future frames:

$^2$4

with $^2$5 computed by the frozen encoders. The combined loss is

$^2$6

A distinctive element is the use of momentum updates for frozen modules:

$^2$7

The paper presents this as a way to avoid naive stop-gradient pitfalls while coordinating the two stages. The claim of “implicit” two-stage training therefore rests on two linked properties: both stages remain aligned to forecasting rather than to heterogeneous objectives such as reconstruction and denoising-generation, and freeze/unfreeze dynamics are embedded within a single loop rather than separated into disconnected training phases.

4. Datasets, baselines, and reported empirical performance

Met$^2$8Net is evaluated on a range of meteorological and non-meteorological datasets (Li et al., 23 Jul 2025). The weather benchmarks include Weather_L, Weather_H, Weather_HA, and an ERA5 regional crop over $^2$9–$^2$0 and $^2$1–$^2$2. The paper also reports experiments on TaxiBJ and on MvMmfnist, a constructed multivariable benchmark.

Dataset	Core specification	Split details
Weather_L	32×64, variables UV10, T2M, TCC, R; $^2$3, $^2$4	52,559 train, 17,495 test
Weather_H	64×128, same variables and sequence lengths as Weather_L	52,559 train, 17,495 test
Weather_HA	150, 500, 850 hPa; variables U, V, T, R; $^2$5, $^2$6	54,019 train, 2,883 test
ERA5	128×128; variables MSL, U10, V10, T2M; $^2$7, $^2$8	train 2017–2021, val 2022, test 2023
TaxiBJ	32×32; $^2$9, $^2$0	stated as single-variable urban flow
MvMmfnist	64×64; $^2$1; $^2$2, $^2$3	10,000 train, 10,000 test

The baselines include ConvLSTM, PredRNN/PredRNN++, MAU, SimVP, ConvNeXt, HorNet, TAU, Wast, and MogaNet, depending on the task. On Weather_L, the paper reports state-of-the-art performance across UV10, T2M, TCC, and R. Relative to TAU, MSE reductions are given as UV10: $^2$4 ($^2$5), T2M: $^2$6 ($^2$7), TCC: $^2$8 ($^2$9), and R: $E_i(\cdot)$0 ($E_i(\cdot)$1). On Weather_HA, improvements over TAU are reported for all listed variables: R: $E_i(\cdot)$2, T: $E_i(\cdot)$3, U: $E_i(\cdot)$4, and V: $E_i(\cdot)$5. On Weather_H, the paper gives UV10: $E_i(\cdot)$6, T2M: $E_i(\cdot)$7, TCC: $E_i(\cdot)$8, and R: $E_i(\cdot)$9.

For ERA5, Met$D_i(\cdot)$0Net is reported to achieve the lowest MSE/MAE and highest $D_i(\cdot)$1 among baselines. The listed values are MSL MSE: $D_i(\cdot)$2 with $D_i(\cdot)$3; U10 MSE: $D_i(\cdot)$4 with $D_i(\cdot)$5; V10 MSE: $D_i(\cdot)$6 with $D_i(\cdot)$7; and T2M MSE: $D_i(\cdot)$8 with $D_i(\cdot)$9. The paper also states that Typhoon Mawar trajectory forecasting at a 3-hour lead time tracks closer to ground truth than TAU, with fewer parameters. Beyond meteorology, MvMmfnist results include MSE $T(\cdot)$0 versus TAU $T(\cdot)$1, while on TaxiBJ the single-variable variant Met$T(\cdot)$2Net_S is said to match TAU performance.

The reported analysis extends beyond aggregate error. Centered Kernel Alignment across layers is used to show higher representation similarity across encoders and Translator than TAU, and ablations from baseline to “+ MED,” then “+ variable-attention fusion,” then “+ implicit two-stage training” are reported to produce consistent MSE reductions across variables. This suggests that the method’s gains are not attributed to a single architectural change alone, but to the combination of per-variable representation learning, latent fusion, and decoupled training.

5. Other established uses of the name

A second major usage is the photonics platform “MetaNet,” which the source material explicitly notes is “sometimes misspelled as ‘Met2Net’” (Jiang et al., 2020). MetaNet is an online, openly accessible repository designed to coordinate sharing of freeform photonic device layouts and inverse-design codes. Its stated motivation is the lack of common, well-defined benchmarking tasks, the difficulty of reproducing and comparing optimization results across groups, and the limited access to detailed layout files needed to simulate and fabricate reported designs. The initial focus is on periodic metasurfaces, specifically metagratings that deflect normally incident light to the $T(\cdot)$3 diffraction order. The repository contains 135,000 distinct metagrating unit-cell layouts, both 2D and 3D, indexed by operating wavelength, polarization, deflection angle, material, device thickness, dimensionality, design method, and efficiency.

MetaNet provides open-source code for local gradient-based topology optimization and for a global generative neural-network optimizer, GLOnets, both interfaced with the RCWA solver Reticolo. Devices are designed under Maxwell’s equations, with the inverse-design objective expressed as maximizing

$T(\cdot)$4

subject to fabrication and symmetry constraints. The reported benchmarking regime uses polycrystalline silicon, 500–1300 nm wavelengths, and typically 35–85° or 40–80° deflection angles, with TE polarization emphasized. Average efficiencies are reported as 84.7% for 2D and 86.7% for 3D devices across angle–wavelength pairs, while a GLOnet case at 850 nm and 65° reaches 97% efficiency. In this literature, therefore, “Met2Net” does not denote a forecasting model at all; it denotes an erroneous rendering of a photonics data-sharing platform.

A third usage appears in the review “Multiport Network Theory for Modeling and Optimizing Reconfigurable Metasurfaces,” where “Met2Net mapping” names the passage from a physical RIS description to a communication-level network model (Renzo et al., 2024). The starting point is multiport network theory, in which a network with $T(\cdot)$5 ports is described by stacked voltages and currents, $T(\cdot)$6 and $T(\cdot)$7, or by incident and reflected waves, $T(\cdot)$8 and $T(\cdot)$9. Circuit and scattering descriptions are connected through

$$X \in \mathbb{R}^{T\times N\times C\times H\times W}, \qquad Y \in \mathbb{R}^{T'\times N\times C\times H\times W},$$0

with

$$X \in \mathbb{R}^{T\times N\times C\times H\times W}, \qquad Y \in \mathbb{R}^{T'\times N\times C\times H\times W},$$1

or equivalently

$$X \in \mathbb{R}^{T\times N\times C\times H\times W}, \qquad Y \in \mathbb{R}^{T'\times N\times C\times H\times W},$$2

For a composite system containing transmitter, receiver, and RIS ports, terminating the RIS ports with control-dependent loads and eliminating them yields

$$X \in \mathbb{R}^{T\times N\times C\times H\times W}, \qquad Y \in \mathbb{R}^{T'\times N\times C\times H\times W},$$3

This effective scattering matrix is then projected to a baseband channel $$X \in \mathbb{R}^{T\times N\times C\times H\times W}, \qquad Y \in \mathbb{R}^{T'\times N\times C\times H\times W},$$4.

The significance of this “Met2Net mapping” is that it replaces the heuristic diagonal phase-shift model $$X \in \mathbb{R}^{T\times N\times C\times H\times W}, \qquad Y \in \mathbb{R}^{T'\times N\times C\times H\times W},$$5 with a physics-compliant operator that accounts for mutual coupling, losses, structural scattering, passivity, and reciprocity. The paper emphasizes explicit constraints such as $$X \in \mathbb{R}^{T\times N\times C\times H\times W}, \qquad Y \in \mathbb{R}^{T'\times N\times C\times H\times W},$$6 and, for reciprocal media, $$X \in \mathbb{R}^{T\times N\times C\times H\times W}, \qquad Y \in \mathbb{R}^{T'\times N\times C\times H\times W},$$7, $$X \in \mathbb{R}^{T\times N\times C\times H\times W}, \qquad Y \in \mathbb{R}^{T'\times N\times C\times H\times W},$$8, and $$X \in \mathbb{R}^{T\times N\times C\times H\times W}, \qquad Y \in \mathbb{R}^{T'\times N\times C\times H\times W},$$9. Here again, the term “Met2Net” has no relation to meteorological forecasting or to the photonics repository; it is a shorthand for a mapping between metasurface physics and network-level communication models.

6. Limitations, misconceptions, and likely trajectories

For Met$T/T'$0Net in forecasting, the paper identifies several limitations (Li et al., 23 Jul 2025). The implicit two-stage training paradigm adds parameters and computation during training because of momentum copies and frozen paths, even though inference cost is unchanged relative to end-to-end models. Forecast skill is reported to depend on latent dimensionality, the Translator block choice, and attention configuration. The authors also note that performance can vary across regions, seasons, and extreme events as lead time increases, and they identify full physical interpretability and probabilistic uncertainty quantification as future work. Practical guidance includes the use of separate encoders and decoders per variable, variable-axis attention, Adam optimization, batch size 16, about 50 epochs for WeatherBench settings, and learning rates chosen from $T/T'$1.

For MetaNet in photonics, the limitations are of a different kind (Jiang et al., 2020). The current emphasis is on metagratings; the imaginary part of the silicon refractive index is neglected in the initial datasets; material diversity is limited; strict minimum-feature enforcement is easier in 2D than in 3D; and HDF5, while chosen for cross-language support, does not directly solve compatibility with fabrication-oriented formats such as GDSII. The authors also note that machine-learning utility would increase if optimization trajectories, not only final designs, were shared.

For the metasurface “Met2Net mapping,” the limitations stem from modeling assumptions (Renzo et al., 2024). The formulation assumes linear time-invariant operation and a small-signal regime at each frequency, while nonlinear semiconductor effects are absorbed into static control-to-admittance maps. Wideband operation requires frequency-dependent modeling, time modulation and nonreciprocal metasurfaces require generalized MNT, and discrete tuning states introduce mixed-integer optimization. Manufacturing tolerances, temperature drift, and uncertainty in unit-cell calibration or coupling matrices motivate robust formulations rather than purely nominal optimization.

Across all three usages, the most persistent misconception is nominal rather than technical: the name “Met2Net” does not identify one canonical research object. In current arXiv usage it may denote a decoupled multivariable forecasting model, a misspelling of a photonics repository, or a metasurface-to-network mapping. A plausible implication is that the term will remain context-dependent unless one of these usages becomes overwhelmingly dominant. At present, the meteorological Met$T/T'$2Net (Li et al., 23 Jul 2025) is the clearest exact title match, but the competing usages in photonics (Jiang et al., 2020) and programmable electromagnetics (Renzo et al., 2024) remain established enough that unqualified references are intrinsically ambiguous.