Heterogeneously integrated flexible microwave amplifiers on a cellulose nanofibril substrate

Low-cost flexible microwave circuits with compact size and light weight are highly desirable for flexible wireless communication and other miniaturized microwave systems. However, the prevalent studies on flexible microwave electronics have only focused on individual flexible microwave elements such as transistors, inductors, capacitors, and transmission lines. Thinning down supporting substrate of rigid chip-based monolithic microwave integrated circuits has been the only approach toward flexible microwave integrated circuits. Here, we report a flexible microwave integrated circuit strategy integrating membrane AlGaN/GaN high electron mobility transistor with passive impedance matching networks on cellulose nanofibril paper. The strategy enables a heterogeneously integrated and, to our knowledge, the first flexible microwave amplifier that can output 10 mW power beyond 5 GHz and can also be easily disposed of due to the use of cellulose nanofibril paper as the circuit substrate. The demonstration represents a critical step forward in realizing flexible wireless communication devices.


Supplementary Figures
to expose electrodes of membrane HEMT, the PMMA sacrificial layer was dissolved in hot acetone and the membrane HEMT encapsulated by PI layers was transfer-printed to SU8-coated CNF substrate with spin-cast SU8.   for the HEMT sitting on SU8/CNF with reference to that sitting on Si substrate is resulted from the attacking of the HEMT active region by buffered oxide etchant (BOE, 6:1) when removing SiO2 protection layer in order to expose metal contact pads for characterizations using the etchant 1 . The problem was eliminated when the metal pads of the HEMT was exposed using dry etching for the HEMT on PI/SU8/CNF. In this case, the threshold voltage only positively shifted by 0.09 V with reference to that of the HEMT sitting on Si substrate. The 0.09 V shift may be due to the annealing effects of the Ni/AlGaN Schottky gate of the HEMT when curing the PI 2 .      The fabrication process began with transfer-printing the membrane GaN HEMT on a PDMS stamp (from the very last step of Supplementary Figure 1) to PI/PMMA/Si substrate with partially cured PI as the adhesive layer. The yellow dashed line in magnified view depicts the electrodes of GaN HEMT. A PI encapsulation layer was spin-cast on the transfer-printed HEMT and via holes were opened above metal pads of the HEMT by dry etching. The 1 st metal layer, which serves as the inductor's spiral metal lines and the coplanar ground plane, was formed using contact photolithography, electron beam evaporation, and liftoff procedures. Then the capacitor's dielectric layer and the top electrode of MIM capacitor were subsequently deposited. A PI layer was spin-cast on the sample as the insulating layer. After opening via holes on the specific locations, the 2 nd metal layer was deposited to form interconnects. A final encapsulating layer was spin-cast on the sample and via holes on GSG pads were opened using dry etching to expose RF pads for characterizations. The fMIC amplifier was picked up using a PDMS stamp after dissolving sacrificial PMMA layer in hot acetone. The amplifier, encapsulated in two PI layers, was attached to flexible CNF substrate with spin-cast SU8 as the adhesive layer and exposed to UV light for curing of the SU8 layer. After detaching the PDMS stamp from the CNF substrate, the amplifier was fixed to the flexible CNF substrate.

Supplementary Note 1: Rational of choosing of GaN HEMT as the active device
Thermal issue is one of the biggest issues faced by microwave flexible electronics due to highfrequency operation of the active devices (transistors), from which a certain output power is needed and static DC power dissipation can cause transistor self-heating and thus substrate heating when organic substrates with very low thermal conductivity are used. Therefore, transistors that are resistant to selfheating are preferred for these applications. Special attention also needs to be paid on the influence of the heating of transistor on the substrate. Evaluation of substrate heat dissipation to avoid possible damage to the substrate is also necessary.
Based on the above considerations, GaN-based HEMT was chosen in this work for its good thermal stability as a result of its wide bandgap and superior microwave properties 15

Cost-effectiveness:
In general, the cost of fabricating a microwave circuit consists of material cost and processing cost 16 , the latter of which includes the processing cost of active transistors and passive elements. Furthermore, the processing cost for passive elements is in general lower than that of processing active transistors because passive elements have a much simpler structure and larger feature size.

The demonstrated flexible microwave integrated circuit (fMIC) is cost-effective if one compares
fMIC fabrication with that of monolithic microwave integrated circuits (MMICs) made with GaN or other III-V materials. To make MMICs from these materials, entire substrates were used to carry a MMIC circuit.
These substrates are much more expensive than Si substrates. Furthermore, the majority of these expensive substrates were occupied by large-area passive components, such as inductors, capacitors, and transmission lines and only a very small fraction of the substrate real-estate was used/occupied by active transistors.
When converting these III-V based MMICs into fMICs using substrate thin-down processes such as polishing or etching, the expensive substrates were wasted 17 . It is noted that such a conversion process from MMIC into flexible MMIC made by thinning the rigid MMIC substrates adds extra cost on top of the cost of making the rigid-chip based MMICs. Moreover, because processing large-area chips leads to low manufacturing yield 18 , the cost of the thin-down approach further increases.
For the fMIC approach described, the wafer/epi area of GaN and III-V substrates is much more effectively used than MMIC (13 times more efficient according to Supplementary Equation 1) as we made a dense array of active transistors on these substrates. Our large-area passive components were not fabricated on native III-V substrates, but eventually reside on a CNF substrate, of which the cost is almost negligible in comparison to III-V epi wafers. On the other hand, the processing cost of fabricating passive components on the temporary Si substrate is the same as that of fabricating on the III-V rigid substrates.
Finally, the processing cost of removing substrate for our dense array of active devices is essentially the same as that for thinning down the substrate of a rigid MMIC. Based on the above comparisons, the cost effectiveness of our fMIC approach mainly comes from much more efficiently use of III-V substrates/epi materials.
The only additional cost incurred in our approach is the cost associated with the transfer printing step used to transfer our circuits from the temporary handling substrate to the CNF. It is noted that transfer printing has already been commercialized for making large-area micro LEDs displays. It is shown 19 that the transfer printing step accounts for only a very small fraction of the total cost of fabricated micro LEDs due to the high throughput and yield 20 of the automated transfer printing equipment, even though the LED fabrication process is much easier than that of HEMTs.
It is noted that for Si-based MMIC 14 , although the substrate thin-down process added extra cost to that of the rigid MMIC, the inexpensive substrate of Si in the MMIC made it a cost-effective approach in addition to the advantage of maintaining the high-level metal interconnects associated with Si-based MMIC.
For fabricating Si-based MMICs, our fMIC approach may not be a better approach than the one demonstrated in Ref. 14 .

Repeatability:
With regard to the fabrication process of fMIC, with the exception of the transfer-printing steps using polydimethylsiloxane (PDMS) stamps, all other processing steps were performed using conventional semiconductor processing tools. It is noted that PDMS stamp based transfer-printing process has been adopted by some semiconductor industry proving its repeatability and high fidelity (see Supplementary Note 4). The overall process for the fMIC fabrication is very repeatable.

Timeline:
The fabrication methods of rigid Si-based MMIC and rigid III-V based MMIC are very different.
The Si-based MMICs are featured with high integration levels (e.g., CMOS, HBT, passives, etc.) along with many levels of metal interconnects, of which the fabrication may take many weeks, depending on the complexity of the chips and the integration levels. In contrast, the turn-around time of III-V based MMICs is much shorter than Si-based MMICs. Converting Si-based rigid chip MMICs into flexible ones via chip thin-down simply added a few more steps to the turn-around time of the rigid ones. For the fMIC approach described herein, the short turn-around time of the rigid III-V MMICs is largely inherited. The processing time for the active transistors and that for the passives are essentially the same as that for their rigid counterparts. In this fMIC approach, substrate removal also added extra processing time in comparison to the III-V rigid ones. Moreover, our approach added additional processing time for transfer-printing.
Because of the high speed of transfer-printing, the time shared by each chip is rather short. Overall, the fMIC processing still maintains a relatively short timeline. Since CNF substrate has a much lower thermal conductivity than Si, we had to reduce the bias voltage down to 10 V (see below). To compensate the reduced device power due to the reduced bias, we increased the gate width to 90 µm.

Supplementary Note 3: Considerations of GaN HEMT device size and bias voltages
GaN HEMT can sustain a very high operation voltage due to the wide bandgap of GaN. The choice of drain bias is typically determined by applications. AlGaN/GaN HEMTs with similar structure to our devices are typically biased at drain voltage beyond 10 V 27,28 . However, the CNF substrate has limited thermal conductivity. To accommodate the substrate's thermal limit, we chose to lower the bias and thus to lower the operation power in order to reduce the generation of heat from the devices. The drain bias voltage could be increased to higher values if a better heat dissipation scheme was employed.

Supplementary Note 4: PDMS-based transfer-printing techniques used in this work
Following the same principle of automated transfer-printing machines already commercially available 20 and adopted for micro LED fabrication, a contact-mode mask aligner (MJB-3, Karl Suss) was employed to carry out the transfer-printing process. As shown in Supplementary Figure 5, in our setup a PDMS stamp was placed on a glass slide and the glass slide was mounted on a mask holder via vacuum.
The destination substrate (e.g., PI/PMMA/Si in Supplementary Fig. 5) was mounted on the sample holder on the stage of the mask aligner. By moving up the sample holder, which is similar to contacting a photoresist-coated wafer with a photo mask in contact mode photolithography, the membrane HEMT on the PDMS stamp will be attached/transferred to the destination substrate. Due to a stronger bonding force between the membrane HEMT and the destination substrate by an adhesive layer 29,30 , the HEMT is transferprinted on the destination substrate after moving down the sample stage to detach the destination substrate from the PDMS stamp. The force applied during the transfer process can be precisely adjusted by controlling the distance and speed that the sample stage was moved up by applying the needed number of turns on the adjusting knob. A transfer yield of 90%-100% can be achieved using the contact aligner-based transfer-printing method. The occasional failure of transfer-printing is typically a result of overuse of PDMS stamps, which becomes less sticky and can be "re-conditioned" with oxygen plasma treatment or simply replaced with a new one. Overall, the demonstrated fMIC approach in this work can be adopted for mass production using the commercially available equipment 12 in a straightforward way.

Supplementary Note 5: Comparison between flexible GaN HEMTs
The GaN-based HEMTs fabricated and used in the amplifier circuit in this work employed a similar method to what we showed before (Ref. 10 , T-gate), except that a planar gate and a much smaller area of the GaN membrane were used in this work. A T-gate can achieve low gate resistance. A larger area of intrinsic GaN membrane helped to dissipate heat in the HEMT 10 . As a result, the prior HEMT 10 showed higher RF performance than that reported in this paper. However, T-gate is more fragile than a planar gate under mechanical bending. A much larger area of GaN membrane led to reduced use efficiency of AlGaN/GaN epi area and could also limit the radius of mechanical bending. A comparison of RF performance among published work and our flexible GaN HEMT is shown in Supplementary Table 3. As can be seen, with a comparable gate length range of 120 nm-300 nm, our HEMT showed the highest RF performances. Despite the slightly higher RF performance, the uniqueness of this demonstration is that we used the deterministic transfer-printing method to spread a dense array of HEMTs fabricated from their original expensive host epi substrates into a very sparse array on a CNF substrate, which has substantially improved the usage of epitaxial GaN materials and thus reduced the fabrication cost of microwave amplifiers based on GaN HEMTs. Of more importance, the deterministic transfer-printing method is scalable to volume production as already demonstrated in thin-film LEDs in mini/micro-LED industry 20,31 .

Supplementary Note 6: Simulations of flexible amplifier under bending
The RF performance of the flexible amplifier under bending was studied through circuit simulations using ADS. The circuit diagram shown in Supplementary Figure 12  conditions as this study were used. The empirical fitting curves were obtained in a previous study ( Figure   5 in Ref. 32 .). Based on the calculations, the inductance values decrease from the flat conditions by 2.5% and 4.0% when the bending radii are 38.5 mm and 28.5 mm, respectively. Similarly, the capacitance value was calculated to increase by 4.1% and 9.0% when the bending radii are 38.5 mm and 28.5 mm, respectively.
Supplementary Table 6 summarizes the inductance and capacitance values used in the ADS simulations.
Supplementary Figure 13 shows the simulation results of the circuit under flat and bending conditions. The decreasing trend of the small-signal gain of the amplifier under increased mechanical bending aligns with the measurement results shown in Figure 5j, which qualitatively indicates that the slight degradation of small-signal gain of the flexible amplifier under bending was a result of degraded HEMT's RF performance and the change in values of passive components due to bending.

Supplementary Note 7: Equations
The ratio between the number of transistors made on one wafer and the number of circuits made on sanme size wafer can be calculated using equation as shown below: (1) The large-signal power gain can be calculated using equation as shown below: Large-signal power gain = POUT/PIN The power-added efficiency (PAE) can be calculated using equation as shown below: Where, POUT is output microwave power, PIN is input microwave power, PDC is DC consumption power.