The term MEMS was first coined in the 1980s and has since brought tons of innovative/insane fabrication techniques and incredible devices. Among my favorite examples of this is the DLP technique invented by TI and a DNA nanoinjector for cells made at BYU (pictured below). Obviously, I won’t be making anything this complicated but had some decent results on my first attempts that are worth sharing and continuing to work on.
All of my MEMS devices are fabricated with chemical etching and maskless photolithogrpahy. The most basic free-standing mechanical device is the cantilever, which can be made using a sacrificial underlayer made of oxide or photoresist (SiO2 in my case) and a harder upper layer (SU-8 epoxy resist in my case) which remains when the sacrificial layer is removed with chemical etching (HF for SiO2). These can be made into oscillators, nano balances, any many other useful instruments.
There are many variables here to optimize, so it is best to run many experiments at once.
The next MEMS technique to demonstrate is bulk Silicon Micromachining which leverages the high level of order in a monocrystalline Silicon wafers to make controlled and often self-terminating etches. On a standard <100> orientation wafer, a solution of KOH (KOH pellets from amazon/ebay dissolved in water) can anisotropically etch Silicon to expose the <111> crystal face at at an angle of 54.7º from the wafer surface.
The setup for this is quite simple, 60% KOH solution with a dash of isopropyl alcohol is brought to 80°C in a hotplate with stirring and etches about 90um in 1hr, or about one-third the thickness of a Silicon wafer. Nature’s fume hood is used for this experiment.
Suspended cantilever-like structures can also be made here. The first KOH etching attempt shows some undercutting, but in the future stirring conditions will be optimized to provide more undercutting. The circled areas in the high-tilt angle SEM image below show 500nm thick SiO2 overhanging 90um of airgap above the Silicon surface.
Memristors are interesting devices first posed by Leon Chua in 1971 as the 4th basic circuit element, after the resistor, capacitor, and inductor. It would theoretically satisfy a relationship between charge and flux, and in simple terms is a 2-terminal device whose instantaneous resistance depends on the device’s electrical past (total charge or current that has flowed through the device). The possibility for existence of a memristor in our universe is highly debated and claims of “discovering” the missing memristor were circulated by HP Labs in 2008. The resistance-memory effect they noted was met with huge amounts of press and recognition, but in reality it was first noted in the late 1960s in similar materials and was not valid under the original mathematical definitions for a memristor. Since then, the definition and requirements for a memristor have been relaxed by some and even Leon Chua himself has recognized HP’s memristor. As with most politics and similar situations in academia, you should read about this on your own and come to your own conclusions, but it is safest to describe such devices with a resistance-memory as “memristive devices” or “memristive systems”, because while they may not adhere to the original definition of a memristor, they are still worth exploring and have proved to be useful for machine learning and artificial neuron applications. Interestingly, an ideal memristor fits as a perfect electrical model for Axons/synapses.
Most of these memristors are made with some metal oxide stack with top and bottom electrodes and rely on the electrical current moving oxygen vacancies within the metal oxide to create a sort of hysteresis in resistivity. The results in a unique pinched-figure-eight hysteresis type IV curve whose loop area decreases with increasing frequency:
Basic memristive devices can also be made easily at home out of Copper Sulfates, as described here and here. Of course, it is advantageous for lots of reasons to be able to manufacture these on a parallelized, industry standard CMOS-type fab process, where thousands of these devices can be made at once using photolithography.
About 2 years ago when I was setting up my sputtering chamber, none of the coatings were conductive. This was frustrating but eventually I figured out the films were metal oxides, not pure conductive metal coatings, due to too much Oxygen being present during the sputtering. Recently, I read that memristive devices can be made out of materials such as Titanium and Copper oxides, so I dug out some old failed sputtering attempts of these materials, put them back in a chamber to deposit a top Copper electrode, and measured their IV characteristics.
Both Titanium and Copper oxide based devices worked as memristive systems and showed similar characteristics. The bottom electrode in both cases was doped Silicon which causes the devices to be slightly rectifying (asymmetry in curve) and highly sensitive to light (built-in PN junction) but they still show a resistance-memory hysteresis effect. In the future, more devices will be built on insulating substrates and should yield much better results. Rectifying memristors made on Si substrates have applications in high density 3D array stacks, even though they deviate highly from an “ideal” memristor.
Here, differences in slope (1 / resistance) for the on (blue) and off (red) regions can clearly be seen, as well as the switching region (yellow).
As previously mentioned, these devices are highly light-sensitive due to the doped Si substrate and a have very strange response:
After these first encouraging results, I made Copper Sulfide based devices that actually exhibit more ideal characteristics and can be made for a few dollars. Excess pure Sulfur powder from Amazon is dissolved in Ethanol or Isopropyl alcohol and a bare Copper PCB was placed in solution for 12 hours. The experiment was in a closed vessel to avoid evaporation and a black CuₓSᵧ film was noted. Then, the device was washed in alcohol for 3 hours to remove any residual Sulfur particles before being probed with a function generator and oscilloscope wired as a curve tracer.
Homemade CuₓSᵧ memristors:
A nice figure-eight is seen to pass through the origin and have a loop area which is decreasing with increasing frequency, as expected by the equations that govern a memristor (time integral of charge or current).
I am very excited to announce the details of my first integrated circuit and share the journey that this project has taken me on over the past year. I hope that my success will inspire others and help start a revolution in home chip fabrication. When I set out on this project I had no idea of what I had gotten myself into, but in the end I learned more than I ever thought I would about physics, chemistry, optics, electronics, and so many other fields. Furthermore, my efforts have only been matched with the most positive feedback and support from the world; I owe a sincere thanks to everyone who has helped me, given me advice, and inspired me on this project. Especially my amazing parents, who not only support and encourage me in any way they can but also give me a space to work in and put up with the electricity costs… Thank you!
Without further ado, I present the first home(garage)made lithographically-fabricated integrated circuit – the “Z1” PMOS dual differential amplifier chip. I say “lithographically-fabricated” because Jeri Ellsworth made the first transistors and logic gates (meticulously hand wired with conductive epoxy) and showed the world that this is possible. Inspired by her work, I have demonstrated ICs made by a scalable, industry-standard, photolithographic process. Needless to say, this is the logical step-up from my previous replication of Jeri’s FET fabrication work.
I designed the Z1 amplifier looking for a simple chip to test and tweak my process. Layout was done in Magic VLSI for a 4 mask PMOS process (active/doped area, gate oxide, contact window, and top metal.) PMOS has advantages over NMOS as far as mobile ionic contamination that lends it to being fabricated in a garage. The masks are designed in 16:9 aspect ratio for easy projection.
The feature (gate) size is approximately 175μm although there are test features as small as 2μm on the chip. Each amplifier section (center and right) contain 3 transistors (2 for long-tailed differential pair and one as current source/load resistor) which means a total of 6 FETs on the IC. The left portion of the IC contains resistors, capacitors, diodes, and other test features used to characterize the fabrication process. Each node of the differential pairs is broken out to a separate pin on the lead frame so it can be analyzed and external biasing can be added as necessary.
EDIT: see update at the bottom, the transistor gate length has been reduced to <5µm (1975 tech. level) which brings an increase in device performance.
There are 66 individual fabrication steps to make this chip and it takes approximately 12 hours for a full run. The process yield can be as high as 80% for these large features, but is largely dependent on my coffee intake that day. I have also made Youtube videos covering semiconductor fabrication theory and discrete MOSFET fabrication.
50mm <100> orientation Silicon wafers with bulk resistivity 1 to 10 Ω-cm (30.8 to 308 Ω/sq for thickness of 325µm) are scored into 5.08 x 3.175mm dies (~16mm^2 area) with an Epilog fiber laser. Polyvinyl Alcohol in water or photoresist can be spun on the wafer prior to laser scribing to “catch” laser ablation debris and the film is later removed in solvent before processing. This die size is chosen to fit into a Kyocera 24pin DIP carrier.
Native oxide is stripped off the wafer with a quick dilute HF dip and then they are extensively cleaned in Piranha solution (H2SO4:H2O2), RCA 1 (H2O:NH3:H2O2), RCA 2 (H2O:HCL:H2O2), followed by another dilute HF dip. Most of these cleaning dips are for 10 minutes and can be facilitated by raising to ~40ºC.
The field oxide is thermally grown in a water vapor ambient (wet oxidation) to a thickness of 5000-8000Å. One may consider mixing the DI water for this step with a few percent HCl. The Chrloine atoms help getter and immobilize ionic contaminants and are also said to increase the growth rate by 5-7%. Together with the fact I am making PMOS devices rather than NMOS, these give a huge edge over contamination control and allow decently preforming devices to be fabricated in a garage.
The oxidized wafer is ready for patterning of the active/doped (P-type) area. Positive photoresist (AZ MiR 701 for SiO2 patterning and AZ 4210 for Al layer) is spun on at around 3000rpm yielding a film of about 1.5μm for the AZ MiR 701 or 3.5μm for the AZ 4210 which is soft baked at 90C on a hotplate.
The resist pattern is then hard baked and a number of other tricks are used to ensure good resist adhesion and chemical stability during the following HF etch step which transfers that pattern to the oxide layer and opens windows to the bare silicon surface for doping. These regions later become the source/drain of the FETs.
Doping is then carried out by either solid or liquid source. The solid source is a Boron Nitride disk that is placed in proximity (<2mm) from the wafer in the tube furnace. Alternatively, spin-on liquid sources can be prepared from Phosphoric or Boric acid in water or solvents and doping is carried out in a standard pre-deposition/HF dip/drive-in/deglaze process. I obtained Phosphoric acid in pure form on Amazon and Boric acid from Roach & Ant killer. Since the starting wafer for PMOS here is N-type, I am doing P diffusions of Boron for the source/drain regions and am targeting a sheet resistance in diffused regions of 100 to 250 Ω/sq.
The above mentioned patterning steps are then repeated twice for the gate oxide layer and then the contact layer. The gate oxide must be much thinner (<~750Å) than the field oxide, so the regions between the source/drain are etched away and a thinner oxide is grown there. Then, since the whole wafer has been oxidized during the doping step, contact windows must be etched for the metal layer to make connection with source/drain doped regions.
(click to enlarge)
Now, all the transistors are formed and are ready to be interconnected and broken out to the lead frame. A blanket layer of Aluminum (400-500nm) is sputtered or thermally evaporated onto the wafer. An alternative would be to use the lift-off process in which the photoresist is patterned first and then metal is deposited. To support wire bonding, this metal layer is made thicker (around 2.5µm for Au wire wedge bonding.) These films have a measured bulk resistivity around 5.4e-6 Ω-cm for thermally evaporated films, double the ideal value of 2.7e-6 Ω-cm for Al at 20ºC. The incorporation of Oxygen and other gasses into the Al film during vacuum deposition likely accounts for this difference.
The metal layer is then patterned with photolithography and etched in hot Phosphoric acid (50ºC) to yield the completed IC. The final steps before testing are visual inspection and high temperature annealing of the Aluminum to create ohmic connections.
The finished chip is now ready for packaging and testing.
I don’t have a wire bonder (accepting donations!) so my testing right now is limited to manually probing the wafer with sharp tweezers or using a flip-chip board (difficult to align) to connect it to a curve tracer. The differential amplifier is also tested empirically in-circuit to verify operation.
EDIT: see update at the bottom, I now have a wire bonder!
As you can see above in the PMOS FET Id vs. Vds curves, there is lots of die to die variation and devices made on the same day can have widely different characteristics. Taking 5 traces with -1V Vgs increment requires about a -8V body/substrate bias to overcome fixed charges (positive impurity ions trapped under gate) and lattice defects in the gate region and yield the expected graph.
The chip can also be wired as a 3 stage ring oscillator, the classic test for a new IC fabrication process:
Showing a natural frequency of around 5kHz for 3 stages, limited mainly by excess the gate to source capacitance due to lithography alignment limitations.
Electrical characteristics of Al/Si junctions are characterized as well and show the expected results. We can create three such basic contacts between Aluminum and Silicon. Aluminum is P-type with respect to Silicon so Schottky diodes are formed whenever Aluminum comes into contact with lightly doped N Silicon. Sometimes my devices showed a tunneling characteristic rather than the expected diode, so I theorize that if the same device is processed for a longer time under high temperatures (>1000ºC), increased oxidation at the Si surface causes the Phosphorous at the surface of the wafer to “pile-up” because of the increased solubility of N-type dopant in SiO2. This creates an “N+” region at the surface and the higher dopant concentration creates a diminishing depletion layer which relates to a small potential energy barrier (the electrons can easily tunnel across it), explaining the symmetrical IV curve.
Additionally, the gate oxide dielectric breakdown voltage can be destructively tested. For high quality SiO2, this should be a little over 1V/nm and is easily tested by sweeping Vgs up from 0V and noticing when a large current flows (in normal operation the gate is insulating and no current should be able to flow).
This plot shows gate dielectric breakdown occurring at 21.7V for a 25nm thick gate device, indicating a decent thermally grown oxide quality which could be improved by being grown in an atmosphere with higher Nitrogen content.
The switching and differential amplifier characteristics can also be demonstrated. The trace on the right shows the output of the chip configured as a fully differential amplifier, mixing (adding/subtracting) a 1kHz and 50kHz sine waves together.
The final characteristic to test for is a low-leakage, fully insulating gate as one of the main requirements of true MOSFET operation. As you can see, I am able to charge up the gate of the device and turn it on through a high impedance connection through my fingertip, and the 1, 0 states of the FET are “latching” due to charge staying on the gate of the FETs and having no pathway to dissipate.
Update 9/3/18: I got a wire bonder (K&S Al/Au wedge bonder)! It will definitely take some more practice before I can bond to a chip, but results will be posted. This will also allow for more extensive testing. I just moved out to college so progress will hopefully be made on school breaks. A huge thanks to Jeremy Gordon (@JeremySF on Twitter) for the gracious donation.
Update 7/8/19: FET gate length (feature size) reduced to <5µm, brining this project to be state-of-the-art in about 1975 and allows the transistors to operate with much better characteristics. Pictured below is a ~4.5µm Aluminum gate and the corresponding characteristics showing 5 traces, -3V step and a -8V body/substrate bias.
Thanks for following my work and feel free to contact me with your thoughts at email@example.com !
In IC fabrication it is necessary to deposit/grow and etch insulating layers of Silicon Dioxide. This presents a few problems for standard photolithographic patterning because SiO2 is hydrophilic which can cause photoresist adhesion issues and also the HF etchant attacks most photoresists. These issues combine to leave you with poor pattern definition and often complete photoresist lifting during etch.
The steps I have found to mitigate these issues are (in order): dehydration bake, HMDS vapor prime, thick resist coating, hard bake, and buffered oxide etch.
First, SiO2 is thermally grown on a test wafer using a water vapor source on a nearby hotplate to fill the furnace with steam during oxidation. The first step to ensure good resist adhesion is a dehydration bake which creates a hydrophobic wafer surface. This does not need to be done if the wafer recently came out of the furnace but if it has been in storage, then a bake of up to 700C may be necessary to restore the dehydrated surface. The next step is HMDS vapor priming:
Here, the wafer is heated to around 200C in the presence of Hexamethyldisilazane (HMDS) vapor forming a surface monolayer on the wafer that further increases resist adhesion. HMDS can also be spin coated but this often yields a far too thick layer and can lead to incomplete photoresist development.
The final steps before etch are to spin the resist and to hard bake it. Naturally, a thicker resist film allows for a longer etch time. For maximum chemical stability, the hard bake should be conducted for extended periods of time close to the resist softening point which is usually around 145C. This can make the photoresist difficult to remove, so an ultrasonic acetone bath may be necessary unless you have proper stripping chemicals. If difficulty persists, then it is likely that the top layer of resist has cross-linked and you may be unable to remove it. One may try high power Oxygen RIE followed by Piranha solution and N-Methylpyrrolidone (NMP) stripper as is used commercially to remove resists after hard ion implantation.
Instead of a standard HF etch, a buffered oxide etch of NH4F (Ammonium Fluoride) in HF can be used to control the etch rate and photoresist lifting. I use approximately 20-30g of 100% NH4F per 50mL of HF (stock whink rust remover) and etch time for 6000Å SiO2 is 20min at 20C. A couple drops of Triton X-100 nonionic surfactant may be added to the BOE to improve etch uniformity, wetting, and ensure consistency through a thicker resist. A good BOE recipe can be found here but assumes industrial-strength HF.
Details and photos of my newest maskless photolithography stepper have been posted here: Info Page
Automated DLP submicron stepper for 2″ (50mm) wafers with LabView control, computer alignment, and wafer vacuum chuck. Based on an old Nikon microscope with custom optics and in-situ UV-VIS spectroscopy for illumination process control. Diffraction-limited resolution is <250nm with a 365nm light source.
Lift-off is a technique that allows you to patten a metal layer without any etchant chemicals. Photoresist is spun on, exposed, and developed then the metal is sputtered or evaporated on top of the resist. The photoresist is then striped and any metal on top of it is peeled off. This leaves metal in only the areas in which the resist was not present after developing.
Typically, negative photoresist is preferred for lift-off for a number of reasons but I did not want to change my existing process so I attempted it with positive AZ4210 resist yielding decent results.
Photoresist can be removed in acetone or developer solution (assuming it is exposed by ambient light during processing) but a lower vapor-pressure solvent is preferred. An ultrasonic bath can also improve lift-off.
If lift-off is difficult, a thicker resist film or thinner metal coating can help. Also, substrate heating during deposition leads to resist softening and side wall coating, making lift-off impossible. I used a Peltier cooler in my chamber to prevent this.
Using a 50W Ytterbium fiber laser to scribe and cut Silicon wafers. Laser wavelength is 1062nm and since Silicon has poor transmission at this low IR wavelength, enough energy is absorbed to make laser marks.
Scribe marks are made at higher speed and must follow perpendicular or parallel to a flat of the <100> wafer so that the scribe lies along a crystal lattice line. When cutting all the way through the wafer, higher power is used and arbitrary shapes that do not follow the crystal lines are possible.
There are a number of open source layout and design tools for ICs however here I will just focus on Magic VLSI. Below are the steps to take a design from Verilog through synthesis and layout to the physical mask that is used for fabrication, via the Qflow digital synthesis flow.
The Verilog digital design for this example is a UART interface from www.asic-world.com and is synthesized/routed with the standard SCMOS rules, however a custom .tech file can be specified containing process details and design rules for DRC.
Once Qflow executes successfully, the design can be viewed in Magic after loading the appropriate cells and a GDS file is generated. This is opened in OwlVision GDSII Viewer or similar which can be used to generate the individual mask image files for each layer (active, poly, metal, etc.)
Using the same tools mentioned above, I also designed a simple PMOS chip to test my process. It is scalable to any reasonable size and contains 2 differential amplifier circuits (seen on right and middle) and a number of diodes/resistors and other test features on the left.
The design requires 4 masks for fabrication: active/diffusion, gate oxide, contact, and metal.
Fiducials should be added for subsequent layer alignment. Note: Grateful Dead bears are necessary for the circuit to function correctly.
Thermal evaporation deposition of Al and Ge from Tungsten and Tantalum boats, respectively. A blind hole is drilled and tapped in the chamber bottom plate for a center tap feedthrough (common ground) for the boats. Deposition starts at around 7e-7Torr and ends around 5e-6Torr due to outgassing. Aluminum alloys with Tungsten at the high temperature and causes boat failure, a thicker gauge boat will be used in the future or one made of TiB2-BN or BN. Update: W 0.015″ boat thickness seems OK.
Approx. deposition rate throughout the run was 2.2A/s, with total accumulation of 500A. Much faster than my sputtering setup but yields a worse film.
In situ plasma cleaning is via the red ICP coil seen in the 8th picture. A Quartz Crystal Microbalance (QCM) is used to measure the thickness of the deposited films and current is supplied by a rewound microwave oven transformer. The UV-VIS spectrometer is used to monitor the emission spectra of O2 plasma. O2 is flowed into the chamber via a Mass Flow Controller (MFC) until the pressure is 75-100mTorr and the substrate is plasma cleaned for 5 minutes with 100W RF prior to depositions.
As current pass through the boats, they heat up to 1000 – 1800C and subsequently heat up much of the surrounding chamber and mounting parts. This starts serious outgassing in the chamber and without prior cleaning and bake out quickly raises the pressure to non-workable pressures and the deposition rate slows.
I added a second turbo pump to raise the pumping speed/gas throughput (previously 110L/s and now an additional 50L/s) and to tolerate higher outgassing.
It was also noted that the evaporation of Al with lots of H2O vapor in the chamber (no baking) leads to a reduction of chamber pressure (presumably the formation of Al2O3 with H2O) and the production of H2 as seen on an RGA.