Memory Effect

Caused by dynamic impedance variations at the transistor's bias and matching networks across the modulation bandwidth. Key mechanisms include:

• Bias circuit inductance: Inductors in the drain/collector bias path create a low-frequency pole that modulates the instantaneous supply voltage as the envelope changes
• Baseband impedance: The impedance presented to the envelope-frequency components generated by the detection process, typically in the MHz range for wideband signals
• Video bandwidth (VBW): The frequency range over which the bias network can supply current without significant impedance variation, directly limiting the bandwidth of memory-free operation

Electrical memory effects dominate at envelope frequencies and scale with signal bandwidth, making them the primary concern for wideband 5G signals.

Originate from dynamic junction temperature fluctuations that alter transistor parameters on microsecond to millisecond timescales. Critical aspects:

• Self-heating: Instantaneous power dissipation raises the channel temperature, reducing carrier mobility and transconductance
• Thermal time constants: GaN HEMTs exhibit multiple thermal time constants corresponding to different layers (junction, substrate, package), ranging from microseconds to seconds
• Gain and phase variation: Temperature changes shift both the AM-AM and AM-PM characteristics, creating a signal-history-dependent distortion

Thermal memory is particularly significant in GaN power amplifiers operating at high power densities, where junction temperature swings can exceed 50°C within a single OFDM symbol.

Semiconductor charge trapping and detrapping in surface states, buffer layers, or barrier regions creates slow-memory dynamics unique to compound semiconductors. Key characteristics:

• Gate lag: Slow recovery of drain current after a gate voltage step, caused by surface traps between gate and drain
• Drain lag: Slow current recovery following drain voltage changes, attributed to buffer traps in the epitaxial layers
• Kink effect: An abrupt change in output conductance at specific drain voltages, common in GaAs pHEMT devices

Trapping time constants range from nanoseconds to milliseconds, creating a complex, multi-timescale memory that interacts with both electrical and thermal effects. This is a dominant memory mechanism in GaN HEMT and LDMOS technologies.

Memory effects are categorized by their time constants relative to the modulation envelope period:

Short-term memory (fast dynamics):

• Time constants shorter than or comparable to the symbol period
• Caused by reactive matching networks and intrinsic device capacitances
• Manifests as frequency-dependent AM-AM/AM-PM within the signal bandwidth
• Correctable with memory polynomial models using a few taps

Long-term memory (slow dynamics):

• Time constants spanning multiple symbols to entire packets
• Caused by thermal impedance, bias circuit charging, and deep trapping states
• Creates hysteresis in the amplifier transfer characteristic
• Requires augmented models such as the generalized memory polynomial or Volterra series with long-delay taps

Memory effects fundamentally limit static DPD performance and dictate the required model complexity:

• Memoryless DPD: A simple AM-AM/AM-PM look-up table cannot compensate for memory effects, typically achieving only 5-10 dB ACLR improvement in wideband scenarios
• Memory polynomial DPD: Introduces delay taps to capture short-term memory, achieving 15-20 dB ACLR improvement with 3-5 taps at moderate bandwidths
• Generalized memory polynomial: Adds cross-terms between different delays to capture the interaction of memory mechanisms, essential for signals exceeding 100 MHz bandwidth
• Model complexity trade-off: Each additional memory tap increases FPGA resource utilization linearly, requiring careful optimization for real-time implementation

Accurate behavioral modeling of memory effects is the prerequisite for effective DPD coefficient extraction.

Memory effects are quantified through dynamic envelope-domain measurements:

• Two-tone envelope test: Varying the tone spacing reveals the frequency-dependent nonlinearity, with asymmetry in IM3 sidebands indicating memory
• Pulsed I-V characterization: Applying short-duration pulses with varying quiescent bias points isolates trapping and thermal dynamics from DC self-heating
• Wideband modulated stimulus: Using OFDM or multi-carrier signals with high PAPR excites the full range of memory dynamics
• Envelope-domain modeling: Extracting the relationship between instantaneous envelope magnitude and the resulting complex gain deviation as a function of envelope frequency

Vector network analyzers with envelope-domain capability and large-signal network analyzers (LSNAs) are the primary instruments for memory effect characterization.