TimePrism: Multi-Domain Temporal Architectures

• TimePrism is a multifaceted concept that defines scalable, programmable architectures for mapping temporal information in both electromagnetic metasurfaces and probabilistic forecasting models.
• In the electromagnetic domain, TimePrism employs cascaded supercells with digital control to achieve nonreciprocal gain (>20 dB) and precise frequency-to-angle mapping across a 5.73–5.98 GHz band.
• In time-series forecasting, TimePrism integrates trend and seasonal streams to generate explicit output scenarios with probabilities, delivering state-of-the-art performance with low computational cost.

TimePrism, as a term in the contemporary scientific literature, refers to fundamentally distinct concepts in both wave manipulation (metaphotonics and metamaterials) and time-series forecasting. In each context, TimePrism denotes a scalable, programmable, or modular architecture for mapping temporal structure—either frequencies or scenarios—into a spatial, angular, or probabilistic representation. The following sections detail these distinct manifestations, focusing first on @@@@1@@@@ for electromagnetic wave engineering and space–time modulated metamaterials, and then on discrete probabilistic scenario modeling for time-series forecasting.

1. TimePrism in Programmable Nonreciprocal Metasurfaces

TimePrism, as introduced by Taravati and Eleftheriades, characterizes a programmable nonreciprocal metasurface system that realizes frequency-to-angle mapping with tunable nonreciprocal gain and digital reconfigurability. The design replaces the traditional, bulky, passive prism with an ultra-thin, actively controlled surface composed of cascaded supercells, each implementing transistor-based amplification and phase shifting (Taravati et al., 2020).

1.1 Physics of Frequency-to-Angle Mapping

Central to the operation is the generalized Snell’s law for metasurfaces with a spatial phase gradient:

$$k_0 \sin\theta_t(\omega) - k_0 \sin\theta_i = \frac{\partial\phi(x,\omega)}{\partial x},$$

where $k_0 = \omega/c_0$, $\theta_i$ is the angle of incidence, and $\theta_t(\omega)$ is the transmitted angle for frequency component $\omega$. In the zero-incidence case, the transmission angle is directly determined by the spatial derivative of the imparted phase.

A frequency-dependent, linear phase profile can be discretized into $N$ supercells of size $d$, yielding stepwise control over the beam steering angle as a function of temporal frequency. Extension to two spatial dimensions enables three-dimensional angular control.

1.2 Metasurface Architecture

The metasurface consists of a periodic arrangement of five-stage supercells, each integrating:

• Receiving patch antenna (insertion loss $G_1(\omega)$, phase $\phi_1(\omega)$)
• Two unilateral amplifiers (each a Darlington transistor pair, gains $G_2(\omega)$ and $G_4(\omega)$, phases $\phi_2(\omega)$ and $\phi_4(\omega)$)
• Gradient phase shifter (varactor-loaded, provides $\phi_3(\omega)$, loss $G_3(\omega)$)
• Transmitting patch antenna (loss $G_5(\omega)$, phase $\phi_5(\omega)$)

The net phase and amplitude for any supercell at frequency $\omega$ are:

$$\phi_k(\omega) = \sum_{m=1}^5 \phi_{m,k}(\omega),\qquad T_k(\omega) = \prod_{m=1}^5 G_{m,k}(\omega).$$

The system's circuit model admits analytical S-parameter (S21/S12) characterization, capturing both amplification and multiple internal reflections.

1.3 Nonreciprocity and Gain

Nonreciprocal operation is engineered by two cascaded unilateral amplifiers per supercell, resulting in forward power transmission exceeding +20 dB and back-propagation attenuation below −20 dB, yielding isolation $>40$ dB over a $5.73$–$5.98$ GHz band. The operation is further interpreted through a generalized susceptibility tensor (biaxial GSTC) framework with nonreciprocal electric–magnetic coupling terms, formally violating Lorentz reciprocity.

1.4 Digital Control and Reconfigurability

Each supercell is digitally controlled via an FPGA, which sets bias voltages to the amplifiers and varactor phase shifters. The procedure is as follows:

1. Specify the desired angle map $\theta_t(\omega_j)$ for each frequency.
2. Calculate required local phase gradients $\xi_k(\omega_j)$.
3. Integrate to recover spatial phase profiles $\phi_k(\omega_j)$.
4. Use calibration to map control voltages to component phase/gain settings.
5. Upload DAC values to the FPGA; real-time reconfiguration is possible in less than microseconds.

1.5 Performance Characteristics

• Operational frequency range: $5.73$–$5.98$ GHz
• Angular resolution: $\Delta\theta \sim \lambda/(N d)$; with $N=16$ (x-direction), step size $\sim2$–$3^\circ$
• Maximum forward gain: $>20$ dB (system-level, supercell chain $\sim+10$ dB)
• Isolation: $>40$ dB
• Power consumption: $\sim5$ W (32 amplifiers at 30 mA each, 5 V)
• Thickness: $\delta \sim \lambda_0/28$ (ultrathin)
• Capable of independent 3D steering via vectorial phase-collinearity

1.6 Applications

• 5G/6G beam squint correction, aligning subcarrier frequency beams to user direction
• Full-duplex MIMO and spectrum-isolated antenna arrays
• Frequency-multiplexed radar (independent, simultaneous chirp scanning)
• Dynamic (reconfigurable) holography for real-time 3D imaging
• High-efficiency, directionally selective wireless power transfer

2. TimePrism as a Space–Time Fresnel Prism

A complementary TimePrism concept arises in the context of space–time modulated metamaterials, wherein “Fresnel-reduced” space–time prisms are engineered for frequency conversion via periodic spatial–temporal modulation (Li et al., 2023).

2.1 Theoretical Underpinnings

The foundation is the interaction of an incident wave with a moving step-interface in a dielectric, leading to frequency shifts expressed as:

$$\omega_t = \omega_i \frac{1 - \beta n_1}{1 - \beta n_2},\qquad \beta = v_m / c,$$

with $n_1, n_2$ refractive indices, $v_m$ the interface velocity, and $c$ the speed of light.

Implementing an ideal, semi-infinite moving interface is impractical. The “Fresnel-reduced” approach periodically re-injects copies of the target modulation within a finite-depth device, preserving anharmonic and nonreciprocal conversion effects over realistic device lengths.

2.2 Construction and Modulation Schemes

Two main modulation schemes are used:

• Prism I: Identical velocity $v_m$, zones indexed by $m$, depth $d$, with repeated “on” intervals:

$$n(z,t) = n_1 + (n_2 - n_1) \sum_{m=0}^{N-1} \Theta(z - v_m t - m d) \Theta((m+1)d - (z - v_m t))$$

• Prism II: Oblique segments with a secondary velocity $v_m'$ eliminate undesired pure-time modulation.

Prism depth and the number of sections $N$ are chosen to balance phase-matching, interaction time, and broadband performance. Total transmission through $N$ matched facets approximately factorizes, maintaining the target frequency shift.

2.3 Conversion Efficiency and Practical Realization

Conversion efficiency is set by geometry:

• For Prism I: $\eta_I = 1 - n_2 \beta$
• For Prism II: $\eta_{II} = (1-n_2\beta)/(1-n_1\beta)$

Microwave implementation uses switched transmission lines and low-loss dielectrics; at optical frequencies, free-carrier injection and EO modulation supply the required index contrast.

2.4 Functional Applications

• Nonreciprocal isolation (strong directionality without magnets)
• Frequency mixers and arbitrarily tuneable shifters
• Space–time “chirp” or time-lens for ultrafast signal processing
• Systems attaining or surpassing fundamental bounds (e.g., beyond Chu limits)
• Temporal cloaking and time-domain wave refocusing

3. TimePrism as a Probabilistic Scenario Model for Time-Series Forecasting

In a distinct domain, TimePrism refers to a probabilistic forecasting model built on the “Probabilistic Scenarios” paradigm, which eschews sampling-based uncertainty quantification in favor of explicit scenario–probability tuples (Dai et al., 24 Sep 2025).

3.1 Motivation and Paradigm Shift

Traditional probabilistic sequence forecasting relies on sampling to approximate the predictive distribution, which leads to three main limitations: absence of explicit scenario probabilities, incomplete coverage of possible futures, and inference costs that grow linearly with sample size.

The probabilistic scenarios paradigm addresses these by directly learning $K$ possible outcomes (“scenarios”) $\{s_i\}$, each with an explicit probability $p_i$ such that $\sum_i p_i = 1$:

$f(x) = (\{s_i\}_{i=1}^K, \{p_i\}_{i=1}^K)$

3.2 Architecture and Scenario Construction

TimePrism implements this by a minimal neural model comprised of three parallel linear layers:

1. Trend Extraction: $x_{\rm trend}$ from average pooling, $x_{\rm season} = x - x_{\rm trend}$
2. Trend Stream: $M$ trend forecasts via $y_{t,m} = W_t x_{\rm trend} + b_t$
3. Seasonal Stream: $K$ seasonal forecasts via $y_{s,k} = W_s x_{\rm season} + b_s$
4. Scenario Assembly: All $N = M \times K$ scenario combinations $s_{(m, k)} = y_{t,m} + y_{s,k}$
5. Probability Assignment: $p = \textrm{softmax}(W_p x_{\rm flat} + b_p)$

All outputs are produced in a single forward pass.

3.3 Training Objective

TimePrism is trained with:

• Winner-Takes-All (WTA) Reconstruction Loss: Finds the scenario $s_{n^*}$ closest to ground truth and optimizes its fit.
• Cross-Entropy Probability Loss: Assigns higher probability to the winner scenario.
• Total Loss: $$\mathcal{L}_{\text{Prism}} = \mathcal{L}_{\text{recon}} + \lambda\,\mathcal{L}_{\text{prob}}$$, with typically $\lambda = 1$.

A relaxed-WTA smoothing term stabilizes gradients for non-winners.

3.4 Empirical Evaluation

Across five GluonTS benchmark datasets (Electricity, Exchange, Solar, Traffic, Wikipedia), TimePrism delivers state-of-the-art or near-state-of-the-art performance on weighted CRPS and distortion metrics compared to methods including DeepAR, TimeGrad, Transformer-TempFlow, and TACTiS-2. Its inference cost remains constant regardless of $N$, with $5.1 \times 10^5$ FLOPs per forward pass, several orders of magnitude lower than sampling-based baselines at equivalent $S$.

4. Comparative Table: TimePrism in Three Contexts

Context	Functionality	Reference
Programmable nonreciprocal metasurface (microwave/EM)	Frequency-to-angle mapping, active gain, reconfigurability, nonreciprocal isolation	(Taravati et al., 2020)
Space–time Fresnel Prism	Space–time frequency conversion, compact multi-facet geometry, nonreciprocity	(Li et al., 2023)
Probabilistic time-series forecasting	Discrete scenario–probability forecasting, no sampling, explicit uncertainty quantification	(Dai et al., 24 Sep 2025)

Each instantiation maintains programmability or modular architecture as a core feature, whether in physical metasurfaces, modulated transmission lines, or neural scenario ensembles.

5. Key Challenges and Research Directions

5.1 Metamaterial and Metasurface Implementations

Further scaling of active metasurface TimePrisms to higher frequencies (mmWave to optical) requires developments in high-speed, low-loss amplification and phase shifting. Integration with system-on-chip digital controllers and extension to larger metasurface apertures are under active investigation.

5.2 Space–Time Modulated Systems

Design optimizations for conversion efficiency, minimization of spurious interval effects, and integration with advanced material platforms are ongoing. Applications in time-lens, temporal cloaking, or nonreciprocal electronics depend on robust, broadband, and low-loss realization.

5.3 Probabilistic Forecasting

Limitations of linear scenario backbone in TimePrism motivate integrating the probabilistic scenarios paradigm with richer models such as transformers or flow-based architectures. Adaptive scenario count and clustering represent next steps for robust and scalable uncertainty quantification. Broader application domains—including finance, energy systems, supply-chain risk, and RL—are anticipated (Dai et al., 24 Sep 2025).

A plausible implication is that the convergence of TimePrism architectures across domains reflects a wider shift toward explicit, programmable representations of temporal information—whether in electromagnetic processing or machine learning—enabling novel functionalities and analytic tractability.