Voltage limitations on AREF for Arduino

I don't see any problem with applying an external voltage, through a 5 kohm resistor, to the Arduino reference input. Or better, with using a resistor divider, so that you turn 5 V into your desired AREF voltage, while at the same time exhibiting a source resistance of approximately 5 kohm. This second requirement does not have to be accurate. That is just to limit the current that will flow from AVCC to ground, through the external circuitry.

If you want to end up with 1.8 V at the AREF input of the MCU, just choose R1 and R2 so that \$ V_{AREF} = 5·\frac{R2||32000}{R1+(R2||32000)} =\$ 1.8 V and \$ R_{source} = R1||R2 \approx\$ 5 kohm.

When you need to work with the [0, 1.8] V range, disable the references internal to the ATMega, and when you need to work with the [0, 5] V, enable the internal AVCC reference (if that is 5 V). If the MOSFET shown in Fig. 24-1 (that connects the internal references to the AREF line) has an on resistance much lower than 5 kohm (which I suppose it has), the internal circuitry will see AVCC. In this second situation, the current drain from the internal AVCC (assumed 5 V) to your external resistor divider will be \$\leqslant\$ 1 mA, but that is not a problem.

In summary: it would be bad advice if something could get damaged, but 1 mA won't damage anything.

The offset with the zener is of little use. The best zeners have a 1 % tolerance, which is 100 mV for a 10 V zener. On a 10-bit ADC this is a 20 count error, on a 12-bit ADC an 82 count error. You could trim the error away if you can measure the voltage accurately enough, but there are other factors. The BZX84-A10 has an 8 mV/°C temperature coefficient, giving a 2 count error per °C change in temperature for the 10-bit ADC, and 7 counts for the 12-bit. Looks like it's better suited as thermometer than as voltage meter. When you use a 10 V zener you'll also need a higher voltage power supply.

The resistance divider will do much better. Resistors also have a tolerance, but at 25 ¢ a 0.1 % resistor is still affordable. (Better that 0.1 % becomes expensive quickly: a 0.05 % costs almost 1 dollar.) At 10-bit resolution that will give you a 1 count error, 4 counts at 12-bit. Temperature coefficient will be less of a problem if both resistors are from the same series and placed close to each other: since the divider is ratiometric resistance changes will cancel each other out.

The numbers indicate that a higher than 10-bit resolution is of little use; component tolerances and variations will cause extra bit to be unreliable. A few extra bits may help to increase noise immunity, though, by averaging a series of measurements, or using a sigma-delta ADC, which averages the input signal anyway.

There's also something more philosophical: we always want better, but why on earth would you want to know a 12 V battery's voltage to a precision of better than 10 mV. You'll have a hard time getting the required resolution, and you'll always be uncertain about that last digit.

The ADS1000 is a low cost 12-bit ADC which will operate from a single 5 V supply.