Abstract
Accurate modeling of nonlinear microwave devices is critical for the reliable design of modern microwave circuits and systems. This paper proposes a general neuro-space mapping (Neuro-SM) method designed to address the increased modeling complexity of contemporary transistors. By incorporating both dynamic voltage and dynamic current mappings, the proposed technique generalizes previous Neuro-SM methods, producing more accurate models with enhanced dynamics and nonlinearity. Validated through a 2 × 50 μm GaAs pHEMT example, the model demonstrates high accuracy and computational efficiency suitable for standard circuit simulators.
Introduction
As wireless communication technologies advance toward 5G and beyond, microwave transistors are increasingly operated in complex, nonlinear regimes. Traditional "coarse" models—typically physics-based equivalent circuits—often fail to capture high-order dynamic effects such as non-quasi-static behavior or capacitive nonlinearities.
While previous iterations of Neuro-Space Mapping (Neuro-SM) attempted to bridge this gap using Artificial Neural Networks (ANNs), they were often limited to static mappings or focused exclusively on voltage dynamics. There has been a clear need for a generalized framework capable of handling the full range of dynamics—both voltage and current—to produce "fine" models that match real-world measurements across DC, small-signal, and large-signal conditions.
The General Neuro-SM Framework
The core of the proposed "General Neuro-SM" is the use of Time Delay Neural Networks (TDNNs) to augment a traditional coarse model. Unlike previous versions, this method employs a dual-mapping strategy. One TDNN maps input voltages (including their temporal history) to the coarse model, while a second TDNN maps the resulting coarse currents back to the fine model's output.
As shown in Figure 1, this architecture allows the model to "mop up" errors that the physics-based model cannot explain, combining physical intuition with the flexible learning power of neural networks. This unified framework is truly general; by setting specific parameters like delay buffers to zero, the formulation mathematically collapses back into any previously published Neuro-SM method.
Methodology and Training
To ensure the General Neuro-SM is practical for engineering workflows, the authors derived analytical formulations ensuring compatibility across three critical domains: DC, small-signal (S-parameters), and large-signal (Harmonic Balance).
A significant challenge in training such models is the computational overhead. To address this, the authors derived exact gradients (sensitivity analysis) for the neural network weights. This allows for fast, gradient-based optimization rather than relying on slower, black-box training methods. This approach ensures that the model can be trained efficiently using a combination of DC, small-, and large-signal data.
Experimental Validation
The technique was validated using a real-world 2 × 50 μm GaAs pseudomorphic high-electron mobility transistor (pHEMT). The training dataset included DC measurements, bias-dependent S-parameters, and Harmonic Balance data at 7.5 GHz with varying input power levels.
Key Results
- Accuracy: The General Neuro-SM significantly outperformed traditional methods. For the pHEMT example, the training/test error dropped from 38.12% (coarse model) to just 1.50% (General Neuro-SM with 30 hidden neurons).
- Large-Signal Performance: The proposed model tracks Power Out vs. Power In (Pout/Pin) curves and harmonic distortions much more closely than previous static or dynamic Neuro-SM iterations.
- Efficiency: Despite its increased complexity, the model remains computationally efficient. A Monte Carlo analysis of 200 simulations showed the general model took 57.17 seconds, compared to 48.32 seconds for the coarse model, confirming its readiness for integration into standard CAD tools like Keysight ADS.
Conclusion
The General Neuro-SM technique provides a robust framework for modeling the complex nonlinearities of modern microwave devices. By integrating dynamic current and voltage mappings into a unified sensitivity-based training approach, it achieves high accuracy without sacrificing the simulation speed required for practical circuit design. Future work may explore scaling this approach to n-port networks and addressing specific phenomena like trapping effects.
Summary updated