Memory Effect

Caused by dynamic self-heating of the transistor channel in response to signal envelope variations. As instantaneous power dissipation changes, the junction temperature fluctuates, modulating transconductance and threshold voltage.

• Time constant: Microseconds to milliseconds
• Dominant in: GaN HEMT and Si LDMOS amplifiers
• Manifests as: Asymmetric intermodulation distortion sidebands
• Mitigation: Thermal compensation circuits and envelope-dependent biasing

Originate from frequency-dependent impedances in the bias network and matching circuits. The envelope-frequency components of the modulated signal interact with non-ideal decoupling capacitors and bias inductors, creating dynamic supply voltage modulation.

• Time constant: Nanoseconds to microseconds
• Key contributors: Baseband impedance at the drain/gate terminals
• Effect: Intermodulation asymmetry that changes with tone spacing
• Design fix: Low-impedance baseband terminations using quarter-wave bias lines

Specific to compound semiconductor devices like GaN HEMTs. Electrons become captured in deep-level traps within the epitaxial layers or at the surface, creating a virtual gate that modulates the channel current.

• Recovery time: Microseconds to seconds
• Consequence: Gate-lag and drain-lag phenomena
• Impact: History-dependent knee voltage walkout
• Characterization: Pulsed I-V measurements at different quiescent bias points

Memory effects are categorized by their duration relative to the modulation envelope period. Short-term memory spans a few symbol periods and is dominated by electrical reactance. Long-term memory persists over many symbols and is driven by thermal and trapping dynamics.

• Short-term: Correctable with memory polynomial models (3–5 taps)
• Long-term: Requires recurrent neural networks or Volterra series with deep memory
• Crossover region: ~1 MHz envelope bandwidth typically separates regimes
• Practical impact: Long-term memory limits the effectiveness of simple LUT-based DPD

Memory effects fundamentally limit the linearization bandwidth achievable with digital predistortion. A memoryless DPD can only correct static AM-AM and AM-PM distortion. To suppress distortion in adjacent channels, the predistorter must invert the full nonlinear dynamic transfer function.

• Model requirement: Must capture both nonlinearity and dispersion
• Failure mode: Residual memory effects appear as uncorrected ACLR asymmetry
• Solution: Generalized memory polynomial or neural network models with cross-terms
• Validation metric: Adjacent Channel Error Power Ratio (ACEPR) specifically quantifies memory-induced residual distortion

The classic method for exposing memory effects uses a two-tone test with variable tone spacing. As the spacing between f1 and f2 is swept, the relative amplitudes and phases of the third-order intermodulation products (IM3) change if memory is present.

• Memoryless behavior: Symmetric IM3 sidebands, constant with tone spacing
• Memory behavior: Asymmetric IM3 sidebands, magnitude and phase vary with Δf
• Measurement: Vector network analyzer with modulated source
• Interpretation: The asymmetry pattern reveals the dominant memory mechanism (thermal vs. electrical)

An extension of the memory polynomial that incorporates cross-terms between different time delays and nonlinear orders. This structure captures more complex memory interactions where the nonlinearity and memory effects are not separable.

• Addresses lagging envelope effects
• Improves accuracy for strong memory scenarios
• Adds signal and envelope lagging cross-terms

The most comprehensive mathematical framework for modeling nonlinear dynamic systems with memory. It uses multi-dimensional convolution kernels to represent the complete input-output relationship, serving as the theoretical foundation for most behavioral models.

• Full nonlinear dynamic system representation
• Kernel complexity grows exponentially with order
• Pruned versions form the basis of practical models

A recurrent neural network architecture specifically designed to capture long-range temporal dependencies in power amplifier behavior. LSTM cells maintain an internal state that can remember past inputs over extended sequences.

• Effectively models long-term thermal memory
• Learns complex nonlinear dynamics from data
• Requires careful training to avoid overfitting

Techniques for modeling and correcting both short-term and long-term thermal memory effects in power amplifiers. Thermal memory arises from the temperature-dependent behavior of semiconductor junctions, particularly in GaN and GaAs devices.

• Short-term: millisecond-scale self-heating
• Long-term: second-scale substrate temperature changes
• Requires specialized model structures for each timescale

The primary regulatory metric for quantifying spectral regrowth caused by nonlinear distortion. ACPR measures the ratio of power leaked into adjacent frequency channels relative to the power in the main channel.

• Directly impacted by memory effect severity
• Memory effects cause asymmetric spectral regrowth
• Key validation metric for DPD performance