AI-Based Modulation (HEPS)

• AI-based modulation is a technique that applies machine learning to design and optimize signal schemes, surpassing traditional analytical methods.
• HEPS, a hybrid extended phase shift method, leverages ensemble learning, PSO, and fuzzy inference to achieve full zero-voltage switching and efficiency above 97% in dual active bridge converters.
• Deep learning techniques, including autoencoders and joint optimization frameworks, enable robust performance enhancements in complex wireless and power electronics systems.

AI-based modulation, and in particular the class of "Hybrid Extended Phase Shift" (HEPS) techniques, refers to the application of modern machine learning and neural computation methods to the design and optimization of modulation schemes for signal processing and wireless or power electronics systems. AI-based modulation leverages data-driven modeling, automatic feature extraction, and optimization heuristics to exceed the limitations of classical analytical approaches, delivering gains in performance, robustness, or efficiency under diverse operational constraints.

1. Foundations of AI-Based Modulation

AI-based modulation represents a paradigm shift in the design of physical-layer signal generation, mapping, and adaptation. Unlike deterministic or piecewise-analytical modulation developed from first principles and circuit models, AI-based approaches replace hand-crafted derivations with learned surrogates, experimental feedback, and global optimization.

In wireless communications, such as high-order quadrature amplitude modulation (QAM), AI-based autoencoders (“symbol-wise” or convolutional) replace analytically designed constellations and demappers with fully-learned encoders and decoders. At the system level, the entire transmitter-to-receiver chain—including channel impairments—is modeled as a deep neural network autoencoder, with training objectives directly grounded in bit or symbol error rate (BER/SER) (Pham et al., 31 May 2025, Cheng et al., 29 Oct 2025, Abdel-Qader et al., 30 Jun 2025, Zheng et al., 6 Jan 2026).

In power electronics, specifically isolated converters such as dual active bridge (DAB) topologies, the "Hybrid Extended Phase Shift" modulation—termed HEPS—deploys ensemble learning (e.g., XGBoost), surrogate neural networks, and metaheuristics (particle swarm, fuzzy inference) to map operating points and switching degrees of freedom to optimal modulation waveforms (Li et al., 2023, Li et al., 2023). The goals are to maximize efficiency, ensure full-range Zero-Voltage Switching (ZVS), and minimize device stress without the laborious tuning of analytical models.

2. HEPS for Dual Active Bridge Topologies

The HEPS approach to DAB modulation exemplifies the integration of AI in physical-layer power conversion (Li et al., 2023). The architecture comprises:

• Dual full-bridge structures interfacing through a transformer, with phase-shift waveform control on both bridges.
• Extended phase-shift (EPS) modulation including both outer ($D_o$) and inner ($D_{in}$) phase shifts, enabling three-level voltage waveforms that suppress circulating current.
• HEPS as a hybrid automaton: automatically selecting between two EPS strategies (EPS1, EPS2), each optimal in distinct voltage and load regions, and reverting to simple single-phase-shift (SPS) at unity gain.

The full workflow is as follows:

1. Surrogate Modeling: Large-scale simulation data (e.g., from PLECS) is used to train ensemble regressors (XGBoost) that predict device losses ($P_{loss}$) and ZVS status ($n_{zvs}$) as functions of control parameters ($P$, $V_2$, strategy flag $S$, and $D_{in}$).
2. Global Optimization: Particle Swarm Optimization with State-based Adaptive Velocity Limit (PSO-SAVL) searches for $(S^*, D_{in}^*)$ yielding full ZVS (all switches soft switching) and minimal $P_{loss}$ across the operational map.
3. Real-Time Deployment: Optimized results are mapped into continuous Fuzzy Inference Systems (FIS), allowing fast, quantization-free generation of inner-phase shifts as a function of $P$, $V_2$.

Experimental prototypes (1 kW, 200 V bus, 20 kHz) demonstrate that HEPS achieves:

• Peak converter efficiencies up to 97.1%
• Universal full-ZVS from 100 W to 1 kW and 160–240 V
• 1.5–4% higher efficiency than SPS or baseline EPS at light and medium loads
• Fast, stable transitions between modulation regimes (Li et al., 2023)

3. Deep Learning Techniques in AI-Modulation

AI-modulation for communication channels employs variants of the neural network autoencoder (AE) paradigm:

• Symbol-wise AE: Input $k$-bit blocks are encoded into $M$-ary complex symbols via convolutional or fully connected networks, transmitted over an AWGN or fading channel, and decoded via neural receivers with trainable demapping (Pham et al., 31 May 2025, Abdel-Qader et al., 30 Jun 2025, Cheng et al., 29 Oct 2025).
• Residual and Convolutional Backbones: For resource-spread modulation (OFDM or MIMO systems), deep convolutional networks act as neural modulators, learning to distribute energy across time–frequency or spatial channels to maximize diversity under fading, Doppler, or no-pilot constraints (Ankireddy et al., 21 Jun 2025, Cheng et al., 29 Oct 2025).
• Loss Surrogate Optimization: AI-based modulators directly minimize communication-centric losses (e.g., categorical cross-entropy on symbols, BCE on bits), and can incorporate regularization terms such as Peak-to-Average-Power Ratio (PAPR) constraints in OFDM (Cheng et al., 29 Oct 2025).

AI-modulation systems routinely utilize Adam or similar optimizers, supervised or unsupervised training (including variational autoencoders for modulation signature embedding (Abbasloo et al., 2019)), and model selection via accuracy, BER/SER, and complexity/parameter metrics (Jafarigol et al., 7 Feb 2025).

4. Optimization Strategies and Performance

The hybrid AI-based modulation paradigm enables several emergent properties and optimization strategies:

• Full-Layer Joint Optimization: Cross-module frameworks—incorporating modulation, channel coding, precoding, and CSI feedback as trainable blocks—maximize end-to-end throughput by removing inter-module boundaries and enabling gradient flow across all levels (Zheng et al., 6 Jan 2026).
• Curriculum and Adaptive Training: Curriculum SNR training, where low-SNR examples are introduced gradually, yields improved generalization and robustness to SNR regime transitions (Jafarigol et al., 7 Feb 2025). Training at multiple anchor SNRs and combining via ensembling further enhances average BER.
• Lightweight Adaptation: The inclusion of channel-adaptation modules (e.g., parameter-efficient bottleneck adapters in the neural receiver) allows rapid adaptation to new or mismatched channel statistics with minimal parameter updates (<5% of total), reducing the need for retraining large models (Cheng et al., 29 Oct 2025).
• Operational Flexibility: Multi-modulation-order support within a single neural model (by appending zero-padding or auxiliary input selection) achieves near-identical BER to per-order trained models while drastically reducing parameter storage (Cheng et al., 29 Oct 2025).
• Complexity Control: Unlike classical demappers scaling as $O(M\log_2 M)$ with constellation size, neural network demappers achieve $O(k^2)$ per symbol, offering efficiency for large $M$ (Pham et al., 31 May 2025).

Empirical benchmarking demonstrates (for communications):

• BER/SER equivalent to, or lower than, MAP-LLR demapping in classical systems up to 16/64-QAM, with improved computational efficiency (Pham et al., 31 May 2025).
• BLER and goodput gains of 1–3 dB and 14% at high Doppler, pilot-sparse, or pilotless OFDM (Ankireddy et al., 21 Jun 2025).
• Throughput improvements of +26.4% over pilot+CP baselines, with BER improvement of 2–5 dB over LS+LMMSE in NextG scenarios (Cheng et al., 29 Oct 2025).

For DAB and HEPS:

• >97% efficiency, full ZVS over the entire domain, and controller automation (Li et al., 2023).

5. Applications and System Integrations

AI-based modulation and HEPS have been validated in:

• Wireless Communications: High-Order QAM, OFDM, MIMO, and code/fading-robust communication, including:
  • End-to-end neural transceivers for AWGN/fading;
  • OFDM frameworks without pilots or cyclic prefix, maximizing spectral efficiency;
  • Adaptive, multi-order, PAPR-constrained systems for NextG (Ankireddy et al., 21 Jun 2025, Cheng et al., 29 Oct 2025, Zheng et al., 6 Jan 2026).
• Power Electronics: High-frequency isolated converters (e.g., EV charging, power conversion), utilizing hybrid phase-shift modulation with AI-based selection and parameterization (Li et al., 2023, Li et al., 2023).
• Cognitive Radio: Modulation recognition via unsupervised deep latent embeddings, enabling online adaptation or anomaly detection for spectrum monitoring (Abbasloo et al., 2019, Jafarigol et al., 7 Feb 2025).
• Meta-Optimization for Facility Design: Nonlinear lattice design in accelerator physics, employing AI (genetic algorithms) to optimize working point and nonlinearities (Jiao, 2015).

6. Limitations and Future Research Directions

While AI-based modulation, including HEPS, offers substantial gains, several limitations and open areas remain:

• Generalization and Data Dependence: Surrogate models’ accuracy (e.g., XGBoost/NN in HEPS) depends on the quality and domain coverage of simulated data; model mismatch against hardware may introduce sub-percent errors (Li et al., 2023).
• Complexity and Parameterization: Training complexity and required dataset size grow with constellation order ($M$), channel diversity, and layers ($L$) in multi-level designs (Abdel-Qader et al., 30 Jun 2025).
• Robustness to Model Drift: Real-time adaptation to hardware aging, temperature variations, or channel drift remains an active area, with online continual learning or digital-twin concepts proposed (Li et al., 2023).
• Scalability to MIMO/Fading Channels: Extension beyond AWGN or SISO systems requires further architectural advances, e.g., explicit channel-state inputs or spatial mixing (Zheng et al., 6 Jan 2026, Cheng et al., 29 Oct 2025).
• Theoretical Limits: Sphere-packing and shaping gains for high-dimensional joint modulation remain underexplored topics; bounds indicate potential for capacity increases via cross-layer designs (Zheng et al., 6 Jan 2026).

A plausible implication is that future work will integrate reinforcement learning for real-time controller adaptation, apply physics-informed neural networks to improve data/physical model fit, develop type-2 fuzzy or neural-fuzzy hybrid logic for uncertain environments, and extend to federated/ensemble governance for extremely heterogeneous deployments (Li et al., 2023, Jafarigol et al., 7 Feb 2025).

7. Summary Table: Representative AI-Based Modulation Use Cases

Application Domain	AI/ML Methodology	System/Topology	Measured Gain	Reference
Wireless (E2E QAM/AE)	Symbol-wise CNN autoencoder	SISO AWGN/Fading	BER ≈ MAP-LLR or lower	(Pham et al., 31 May 2025)
OFDM (Pilotless, PAPR)	ResNet CNN, constraint opt	Pilot/CP-free OFDM	+26.4% throughput, –5dB BER	(Cheng et al., 29 Oct 2025)
Power Conversion (HEPS)	XGBoost + PSO/FIS	DAB – EPS hybrid	97% eff., full ZVS, auto-strat	(Li et al., 2023)
Modulation Recognition	CNN/LSTM/Transformer/AE	SDR, cognitive	+10–20% SNR-robust accuracy	(Jafarigol et al., 7 Feb 2025)
Nonlinear Lattice (HEPS)	Genetic algorithm (NSGA-II)	Synchrotron optics	30–80% ↑ DA, 25% ↑ acceptance	(Jiao, 2015)

In conclusion, AI-based modulation—including hybrid and cross-layer HEPS variants—defines a class of computationally driven, automatically optimized signal generation and recognition strategies that supersede classical deterministic methods in efficiency, adaptability, and system integration. Empirical and theoretical results across both communications and power electronics confirm that these approaches can realize robust, real-time, high-performance systems under demanding operational requirements (Li et al., 2023, Cheng et al., 29 Oct 2025, Pham et al., 31 May 2025, Zheng et al., 6 Jan 2026).