Switch on a solid-state linear amplifier cold and let it work into a steady carrier for ten minutes, then watch the idle current. On most fixed-bias designs it will have crept upward, sometimes by a third or more of where it started. The transistor heats, its threshold shifts, more current flows for the same gate voltage, the extra current makes more heat, and the loop tightens on itself. At best the amplifier wanders off its operating class and the distortion changes through the over. At worst the spiral never stops and the device destroys itself in a slow thermal avalanche. Adaptive bias treats the idle current not as a number you set once but as a target the circuit defends continuously.
The principle reverses the failure. Instead of holding the control voltage fixed and letting temperature drag the current wherever it likes, an adaptive circuit measures temperature, predicts how the device wants to drift, and moves the control voltage in the opposite direction by exactly enough to cancel the drift. Done well, the idle current sits within a few percent of its set point whether the heat sink is at room temperature or scorching after an hour of heavy duty cycle.
Why a fixed gate voltage is a slow trap
The trouble begins with the temperature behavior of the active device. In a bipolar transistor the base-emitter voltage needed for a given collector current falls with temperature at a well-known rate:
dVbe/dT = -2 mV/degC (approximately)
Hold Vbe fixed and a warming junction draws steadily more current. The collector current depends on Vbe exponentially:
Ic = Is exp(Vbe / VT), VT = kT/q ~ 25.85 mV at 300 K
so a small forced offset in Vbe moves the current by a large factor. In a field-effect power device such as the LDMOS transistors that fill modern finals, the gate threshold drifts similarly, and the datasheet states a gate-voltage compensation rate directly, typically a few millivolts per degree.
Put numbers to a bipolar to feel the exponential. If the junction warms 40 degrees and Vbe is held fixed, the device behaves as though Vbe were raised by
dVbe = 2 mV/degC * 40 degC = 80 mV
The current multiplies by
Ic_new / Ic_old = exp(0.080 / 0.02585) = exp(3.09) = 22
a more than twentyfold runaway in principle, checked in practice only by series resistance and supply limits. This is why a fixed-bias bipolar final cannot survive without emitter degeneration or active compensation.
The danger sharpens at high idle currents in FET finals too, because the transconductance gm grows with current, so the same gate offset produces a larger current swing. A 1 to 2 percent drift in the gate voltage of a high-current LDMOS translates into a large swing in idle current, which is why builders of legal-limit amplifiers treat thermal bias control as mandatory.
A small transistor pressed against the hot one
The most common adaptive circuit is also one of the simplest. A small signal transistor, costing pennies, is mounted in physical contact with the heat sink or the case of the power transistor. This little transistor becomes a thermometer. Its own base-emitter voltage falls at the same predictable 2 millivolts per degree, and that falling voltage is wired into the bias network so that as the sensor warms it pulls the power device gate voltage down by a proportional amount.
The sensor and the power device share the same thermal environment, so when one heats the other heats and the correction tracks the cause. A working example from the design literature uses a small PNP transistor against the baseplate to generate a compensating voltage of about 8 millivolts per degree, fed into the gate bias to hold the operating point. The required compensation slope is set by the device, and the sensor delivers its natural slope scaled by a resistor network. To turn the sensor's natural -2 mV/degC into a needed -S mV/degC, the network multiplies by
gain = S / 2
So a device wanting 8 mV/degC needs the sensor voltage amplified by a factor of 4 before it reaches the gate. A trimmer sets the room-temperature idle current, and the thermal tracking defends it from there. Different idle currents need different slopes, so a network tuned for 1 ampere will not correctly compensate the same device at 2 amperes, a subtlety that catches builders who change operating point without retuning.
The thermal lag that no clever circuit can wish away
A subtle limitation governs every temperature-sensing scheme, and it is geometric rather than electrical. The sensor measures its own temperature, not the silicon inside the power transistor. Between the junction where heat is born and the case where the sensor sits lies a thermal resistance and a thermal mass, behaving exactly like a resistor-capacitor low-pass filter with time constant
tau = Rth * Cth
where Rth is the thermal resistance in degrees per watt and Cth the thermal capacitance in joules per degree of the path between junction and sensor. Heat takes time tau to propagate, so during a sudden transition from idle to full output the junction heats almost instantly while the sensor still reports the old cool temperature. For roughly one time constant the compensation has not caught up and the idle current overshoots. The closer the sensor sits to the die, the smaller Cth of the intervening path and the shorter tau. Manufacturers who integrate the thermal-tracking transistor onto the same die, right beside the active area, achieve near-instant tracking. A builder with discrete parts cannot match that but gets close by mounting the sensor against the transistor case and accepting a few seconds of lag as the price of a simple circuit.
Active sensing of the current itself
A more sophisticated branch abandons the temperature proxy and measures the thing it cares about, the current. The drain or collector current flows through a small sensing resistor Rs, the voltage across it reveals the current as
Vsense = Iq * Rs
and a control circuit compares Vsense to a reference and adjusts the bias to hold Iq constant. This closes the loop on the real quantity, so it compensates not just thermal drift but supply variation and aging.
The trade is a new pitfall. The sensing resistor sits in the current path and drops voltage. A gate resistor placed for low-frequency stability, for instance, can develop a drop of 200 millivolts or more when the idle current rises at high temperature, pushing the gate the wrong way. The cure is a feedback path that compensates for the drop, with one constraint: it must be slow relative to the modulation. The compensation pole
fp = 1 / (2 pi R * C)
must sit below the lowest modulation frequency, so
1 / (R C) < 2 pi * f_mod_min
because a current loop fast enough to follow the envelope would fight the signal and manufacture distortion and memory effects.
A numerical walk through one correction cycle
Numbers turn the principle into a design. Take an LDMOS final set to idle at 1.2 amperes, with a datasheet compensation rate of 7 mV/degC and a transconductance near the operating point of
gm = 4 A/V
Start cold at 25 degrees with the idle trimmed to 1.2 A. Drive it hard until the junction reaches 65 degrees, a 40 degree rise. Without compensation the gate would need to fall by
dVg = 7 mV/degC * 40 degC = 280 mV
to hold current steady. Since the gate stays fixed in an uncompensated design, that missing 280 mV acts as excess gate drive, and the current excess is
dIq = gm dVg = 4 0.280 = 1.12 A
The idle climbs from 1.2 A toward 2.3 A, dumping over a hundred extra watts into a device already struggling with heat.
Now add the loop. The sensor, scaled to 7 mV/degC, pulls the gate down by 280 mV over the same rise, cancelling the drift. The residual is whatever the slope misses. Trim the network to within 5 percent of ideal slope and the leftover mismatch is
dVg_residual = 0.05 * 280 mV = 14 mV
producing
dIq_residual = 4 * 0.014 = 0.056 A = 56 mA
The idle ends the over at 1.26 A instead of 2.3 A, a swing under 5 percent rather than nearly 100 percent. That single comparison, 56 mA against 1.12 A, is the whole case for adaptive bias in one calculation. Halving the slope error to 2.5 percent halves the residual; getting the slope wrong by 20 percent leaves a quarter of the original problem in place.
Setting and verifying the compensation in the real world
A bias scheme is only as good as its calibration. First, find or measure the compensation rate. Many devices state a gate-voltage compensation rate on the datasheet; if not, measure it by setting the idle cold, warming the device in a controlled way, recording the current drift dIq over a temperature rise dT, and computing the gate correction that would have cancelled it:
S = dIq / (gm * dT) (mV/degC)
With S known, scale the sensor network to deliver it and set the trimmer for the target room-temperature idle. Then verify, which separates a design that works from one that merely looks right. Run the amplifier through a realistic duty cycle, watch the idle as the heat sink climbs, and confirm it stays in band. A well-built circuit holds to within a few percent across a 40 or 50 degree rise; a poorly matched one overcompensates and drags the current down until the final goes soft, or undercompensates and creeps toward runaway. Accuracy demands patience, because uncontrolled ambient temperature makes the bias voltage wander until steady state, so measure a warm circuit in a stable room and recheck after a full thermal soak.
Predicting the junction temperature the loop must chase
A bias loop defends against junction temperature, so it helps to predict that temperature from first principles before the amplifier is ever switched on. The junction temperature rises above ambient by the power dissipated multiplied by the total thermal resistance from junction to air:
Tj = Ta + Pdiss * (Rth_jc + Rth_cs + Rth_sa)
where Rth_jc is the junction-to-case resistance from the datasheet, Rth_cs the case-to-sink resistance of the mounting interface, and Rth_sa the sink-to-air resistance of the heat sink. The dissipated power is the difference between DC input and RF output:
Pdiss = Vdd * Id - Pout
Walk an example. An LDMOS final on a 50 volt rail drawing 8 amperes at full output, delivering 250 watts of RF, dissipates
Pdiss = 50 * 8 - 250 = 150 W
Suppose Rth_jc is 0.25 degC/W, the insulator and thermal compound add Rth_cs of 0.15 degC/W, and a generously sized sink gives Rth_sa of 0.20 degC/W. The total is
Rth_total = 0.25 + 0.15 + 0.20 = 0.60 degC/W
so over a 25 degree ambient the junction sits at
Tj = 25 + 150 * 0.60 = 25 + 90 = 115 degC
That 90 degree rise is what the bias loop must track. Notice how the mounting interface, the Rth_cs term, sits directly in the path. A sloppy insulator pad that doubles Rth_cs from 0.15 to 0.30 degC/W adds
dTj = 150 * 0.15 = 22.5 degC
to the junction for nothing, which both stresses the device and changes how hard the bias loop has to work. The same calculation tells the builder where to place the sensor. A sensor on the sink reads near Ta + PdissRth_sa, here 25 + 1500.20 = 55 degC, fully 60 degrees cooler than the junction it is meant to track, which is precisely why the scaling network and the sensor placement must be chosen together rather than in isolation.
The quiet discipline behind a stable final
Adaptive bias is not a glamorous part of an amplifier. No one admires it on a spectrum analyzer the way they admire low intermodulation or high efficiency. Yet those headline numbers rest on it. Gain depends on the operating point, linearity depends on the operating point, efficiency and thermal survival depend on the operating point. A final with beautiful specifications and a drifting idle current delivers those specifications only until the heat sink warms, and then it becomes a different amplifier.
The deeper truth is that thermal stability is a control problem disguised as a component choice. The drift has a direction and a rate, both knowable and measurable through dVbe/dT, gm, and the thermal time constant. A circuit that senses the cause, scales the correction to match, and places its sensor close enough to act before the drift runs away turns an unstable device into a stable system. Whether the sensor is a penny transistor on the heat sink or a sensing resistor feeding a current loop, the discipline is the same: measure the drift, predict it, and cancel it faster than it can compound. That is what keeps a hungry final sitting calmly on its bias point while everything around it gets hot.
Our Telegram Channel
Our YouTube Channel