Microscale nuclear magnetic resonance gradient chip

Figure 2 illustrates the microfabrication process and the materials involved. The presented cleanroom process employed materials that possessed a magnetic susceptibility close to that of water, in order to minimise the field inhomogeneity [28]. Processing started on bare 4" Borofloat 33 (Schott Glass Malaysia) substrates which were wet-chemically cleaned using piranha etch solution, rinsed in de-ionised water (DIW) and spin dried. The initial plating seed was patterned by a lift-off process (figure 2(a)) by using the ma-N 1440 negative tone resist (Micro Resist Technology GmbH) [29]. In contrast to related coil structuring processes [30], etching of the seed layer was therefore not required, which simplified the processing. The 9 nm thick W/Ti adhesion layer was sputtered at 100 W RF power. Subsequently, an 80 nm thick Pt layer was deposited (300 W). The lift-off was conducted in dimethyl sulfoxide (DMSO), and subsequently three alternating acetone bathes, followed by an isopropyl alcohol (IPA) rinse and quick-dump-rinsing (QDR). Within the application, the applied WTi/Pt metallisation not only served as the seed layer for the electroplating [24], it was also used to implement RTDs to detect thermal heat-up of the gradient coils. In fact, the Pt layer was chemically inert towards our Au, Cu or Cr etching solutions, which were later used for the removal of the second plating seed.

Zoom In Zoom Out Reset image size

In the next step, the SU-8 plating mould was applied in order to achieve isotropic electroplating of the gradient coil and also to route the fluid delivery ports. Before the resist application, an O₂ plasma flash (10 min at 80 W) was performed to clean the glass substrate and to further reduce the surface tension. For an improved resist adhesion, the substrates were pre-baked at 200 ^∘C for around 30 min. After, SU-8 3050 (MicroChem, Westborough, MA, USA) with a targeted thickness of 100 µm was applied by spin coating. In contrast to the SU-8 processing recommendations, the processing parameters were adjusted to our local cleanroom conditions by elevating the rotational speed to 1250 rpm for a spin-coating duration of 35 s. The soft-bake of a SU-8 layer with a desired thickness of 100 µm required a special procedure. The method of Lee et al [31] resulted in a more uniform surface quality of the SU-8 film by performing the soft-bake while keeping an atmosphere of edge bead removal (EBR) fluid above the substrate. The lithography for the first electroplating mould was performed from both sides of the wafer and two shadow masks were employed as illustrated in figure 2. The required doses for the BS and front-side (FS) lithography were determined by an exposure series (550 mJ cm⁻² for the BS exposure in figure 2(b), 740 mJ cm⁻² for the FS exposure in figure 2(c)). An anti-reflection foil (Spectral Black, Acktar Ltd.) was placed between the lithography chuck and the wafer, to suppress back-reflections from the chuck through the transparent glass substrate. The BS lithography rendered excellent sharp edges, defined by the coil structures, as defined into the platinum layer. Excellent resist adhesion was achieved when exposed from the BS, as the incident light got absorbed at the interface between the SU-8 and glass. The mask for the FS lithography was used to define the chip edges and the fluid delivery channels.

After post lithography processing of the SU-8, another O₂ plasma flash was performed to enhance the Cu electrolyte wetting. Without wait time, the substrates were subjected to Cu electroplating (figure 2(d)). Filling up the SU-8 moulds with Cu by means of electroplating required precise control of the electroplating process. We performed DC electroplating based on our self-built setup using a custom designed current source [32]. Figure 3 shows a picture of the first electroplated coil layer. The two RTDs were not electrically connected and were retained on the wafer without any Cu deposition.

Zoom In Zoom Out Reset image size

Before a supplemental coil layer could be added, an electrically insulating, permanent resist layer had to be applied. For encapsulating the first coil layer, hot-roll lamination was performed using a 50 µm thick ADEX^TM (DJ MicroLaminates, USA) DFR. Throughout the entire layer stack, VIAs were required in order to connect the electrically closed coil loops. To interconnect the coil structures to another overlying coil layer, VIA holes of 310 µm were patterned into the ADEX^TM resist by photo lithography (see figure 2(d)). The VIA holes were topped-up by Cu electroplating, in order to compensate for the height difference of the encapsulation resist layer (figure 2(e)).

For the galvanic deposition of the second coil, another seed layer was applied on top of the ADEX^TM plateaus. The second seed layer required sufficient adhesion towards the ADEX^TM layer and extensive surface cleaning was necessary to remove any remaining CuSO₄ residues and oxides from the VIA top-up process. The copper oxide formation got reduced while the wafers were dipped for around 2 min–3 min into 15%–20% sulphuric acid solution. Rapid rinsing and spin-drying was performed and the substrates were directly placed into the physical vapour deposition (PVD) vacuum chamber. During the PVD, it is important the ADEX^TM resist layer prevails without weight loss or out-gassing, when penetrated at temperatures up to 180 ^∘C. In previous work, we confirmed the stability of the ADEX^TM and SUEX® resists by performing a combined thermogravimetry (TG) and differential scanning calorimetry (DSC) evaluation. Metal vacuum evaporation of a Cr/Au layer (5 nm/80 nm) was performed on top of the ADEX^TM plateau, as shown in figure 2(e). After the seed layer deposition, an electrical resistance of 10 Ω–30 Ω was measured between the Au covered ADEX^TM plateau and the Cu tracks on the wafer base. The conductivity was adequate to proceed with the lamination of the second plating mould.

Without waiting time, a non-permanent electroplating mould was applied using the ORDYL® FP 450 resist (Elga Europe). To obtain an electroplating mould with a height of at least 95 µm, a dual layer of ORDYL® FP 450 was laminated (figure 2(f)). The DFR was successfully applied at a hot-roll temperature between 80 ^∘C and 90 ^∘C. The required exposure dose for cross-linking the ORDYL® FP 450 was between 900 mJ cm⁻² and 1050 mJ cm⁻², which was determined experimentally. Wet development using a mild alkaline solution (0.8% Na₂CO₃) was performed in a mega-sonic bath. A suitable development time was in the range between 8 min and 9.5 min. Longer development led to resist de-lamination ($\gg$14 min). For the dual-layer ORDYL® FP 450 mould, Cu electroplating of the second coil layers was performed up to a height of 70 µm–75 µm, while rotating the substrates within the middle of the plating-time (see figure 2(f)).

After the electroplating, the ORDYL® FP 450 mould was stripped in a strong alkaline solution, that was heated up to 40 ^∘C (figure 2(g)). The resist did not dissolve in the 4%–5% KOH or NaOH solution, but swelled and detached in the ultrasonic bath. The employed TFAC (Transene Company, Inc.) gold etchant for intermetallic substrates is suitable for the selective removal of an Au seed layer. TFAC showed selectivity against Cu and did not corrode the Cr adhesion layer. During TFAC etching, we observed a porous, zinc-coloured layer formed around the Cu structures. Also the Cr adhesion layer did not corrode in the TFAC etch solution and was removed separately by the Cr-ETCH-200 (NB Technologies GmbH) which offered an etch rate between 12 nm min⁻¹ and 15 nm min⁻¹. Since a compound layer of Au and Cr atoms was formed between the two deposited layers, it was necessary to alternate between the Cr-ETCH-200, adequate DIW rinsing and by repeating the Au etching using the TFAC (see figure 2(g)).

A wafer-level DFR bonding process was developed to encapsulate the second coil layer (see figure 2(h)). Before applying the 100 µm thick SUEX® DFR (DJ MicroLaminates, USA), the Cu structures were cleaned by dipping the wafers for 2 min–3 min in 20% H₂SO₄. Without wait time, QDR, nitrogen drying and subsequent dehydration in a vacuum oven at 55 ^∘C at a pressure below $\lt$0.1 bar were carried out. A substrate bonding machine (SB6 Wafer Bonder 1st gen., SUSS Microtec, Germany) was used for this purpose with the top and bottom chuck heated-up to a maximum temperature of 46 ^∘C–48 ^∘C. The hold time was 5 min and the bonding force was set to the minimum adjustable value of 60 N. Bonding of the SUEX® did not cause much of the DFR to flow inside the fluidic network. Bonding was performed under vacuum and the applied resist encapsulated the coil structures completely.

The initial 100 µm thick SUEX® resist decreased to 90 µm to 95 µm after bonding. To ensure that the fluidic channel height was sufficiently high to allow for adequate flow, an additional 50 µm thick ADEX^TM layer was laminated on top (figure 2(i)). This resulted in a minimum channel height of 140 µm, which was 65 µm above the copper structures. The channel width was 128 µm and self-filling without trapped air bubbles was possible.

Note that in the theoretical design of the gradient coil winding patterns, the added separation by the ADEX^TM layer was not included, and figure 4 illustrates the expected drop in the gradient efficiency according to a Biot–Savart calculation. This calculation also implies an improved gradient homogeneity, however it does not account for near field effects or the modified geometry of the gradient conductor that was necessary for connecting the coil loops in series.

Zoom In Zoom Out Reset image size

To stack the two resist layers, lithography was carried out. The utilised mask was adjusted such that the SUEX® DFR that propagated during bonding into the fluidic channels did not cross-link. The exposure energy was 900 mJ cm⁻². After performing the post exposure bake (PEB) and development, the layer stack rendered a planar surface (see figure 2(i)).

To top-seal the fluidic channels, another 50 µm ADEX^TM was laminated which was structured by lithography (figure 2(j)). The PEB was carried out without development, and another ADEX^TM layer was laminated as a final layer. By another lithography step, alignment structures for assembling the Tx/Rx coil and the glass spacers were added (figure 2(k)). The wafers were subjected to a post exposure bake, development and dicing to retrieve the gradient chips, as illustrated in figure 5.

Zoom In Zoom Out Reset image size

The RF coil considered had an elongated shape to fully enclose the linear region of the gradient coil. The entire Tx/Rx coil was based on two bi-plates with a single electroplated copper layer. First, a Cr/Au seed layer was applied on top of a 200 µm thick borosilicate wafer (MEMpax®, 200 µm, SCHOTT Malaysia) by a lift-off process, using the ma-N 1420 resist. Subsequently, a 30 µm ORDYL® SY 330 (ElgaEurope, Italy) layer was hot-roll laminated, and exposed from the BS to resolve the electroplating mould. Post processing was performed according to a conventional approach [33]. The ORDYL® SY mould was filled-up by Cu electroplating. After wet-chemical cleaning and plasma activation, another layer of ORDYL® SY 330 was applied with the aid of a substrate bonder (SB6, SÜSS Micro Tec, Germany). This layer was used to pattern openings for interconnecting to the Cu structures.

The front side of the wafer was now finished and after cleaning in IPA, rinsing and spin-drying, a seed layer of Cr/Au was deposited on the backside of the wafer. The backside of the Tx/Rx coil glass substrate was stacked-up to a thickness of around 12 µm to add metal shielding to the RF coil in order to attenuate the interaction with the gradient. The thickness of the copper shield was adjusted to a factor of four of the conductive skin depth of Cu, which is around 3 µm at a frequency of 500 MHz. However, this copper shield added additional load capacitance to the Tx/Rx coil, which may also result in additional RF energy deposition in the sample, and therefore heating.

As shown in figure 6, there were two pads for the solder interconnection of the coil bi-plates in the flip-chip configuration. The coil plated on the BottomChip had additional pads for placing the tuning capacitor and to interconnect with the probe holder.

Zoom In Zoom Out Reset image size

To study gradient field homogeneity, grid phantoms with water filled, geometrically predefined sections have been used previously [34, 35]. We considered a micro-fabricated phantom with photo-defined alternating channels. The principle of the design and fabrication of the phantom insert was based on a substrate bonding process that has been presented previously by our group [27], by using two glass substrates bonded together with a photo-defined ORDYL® SY DFR layer in between. Spengler et al [27]. used 200 µm soda-lime glass, whereas here we instead employed 100 µm thick borosilicate glass (D 263® T, SCHOTT, Malaysia). We required thinner glass substrates in order to conserve space. The D 263® T is more rigid than standard soda-lime glassware and thus improves handling, but it shows a larger water contact angle ($\gt$60^∘) and poorer adhesion towards the ORDYL® SY.

We patterned multiple, parallel channels aligned across the imaging region into a single layer of ORDYL® SY 390. The width of the channels on the lithographic mask was 80 µm which was near the resist's resolution limit. By the exposure through the glass substrate from the BS, a triangular channel cross section was achieved. Improved adhesion of the glass-ORDYL® SY interface was achieved by the aforementioned backside lithography, since the incident light got absorbed first at the respective interface. After lithography and wet chemical development, another glass wafer was bonded onto the ORDYL® SY 390 structures [27]. The thickness of the insert ($t_\mathrm{insert}$), as sketched in figure 1 (see also figure 10(b)), measured 288 µm for the entire stack.

To mount and electrically connect to the glass chips, a custom-designed PCB with a milled insert was purchased. A low temperature solder alloy Bi₅₇Sn₄₂Ag₁ (Stannol GmbH, Velbert, Germany) was used to interconnect the gradient flip-chip, the Tx/Rx micro coil and also to establish the solder connection to the PCB, as shown in figure 6. A semiautomatic Al wirebonder was used to interconnect the RTDs to the PCB. A tuning capacitor was soldered to the Cu pads on the Tx/Rx coil glass chip (case code 0805, 2.7 pF, SRT-MICROCERAMIQUE, Vendome, France). The PCB was then attached to a Micro5 probe base (Bruker BioSpin, Germany).

The equivalent circuit of a gradient coil can be considered as a resistance R and in series an inductance L. A gradient coil driven by an ideal current source leads to an exponential rise of the magnetic flux of the coil. The time constant for such an RL circuit is then given by $\tau = \frac{L_G}{R_G}$ to reach 63.2% of the final current. In the NMR experiment, capacitive effects can be neglected and rise time of the gradient current amplifiers is generally in the range of 4 µs–6 µs [2].

The impedance of the TopChip, BottomChip and for the flip-chip assembly was measured independently using a digital LCR-meter, as listed in table 1.

Table 1. Electrical parameters of the assembled gradient coil, measured by four-terminal sensing (LCR-821, Good Will Instrument Co., Ltd).

TopChip		BottomChip		Flip-chip assembly
$R_{z,\mathrm{TopChip}}$	$L_{z,\mathrm{TopChip}}$	$R_{z,\mathrm{BottomChip}}$	$L_{z,\mathrm{BottomChip}}$	$R_{z,\mathrm{grad}}$	$L_{z,\mathrm{grad}}$
134.5 mΩ	0.14 µH	133.9 mΩ	0.13 µH	262.8 mΩ	0.34 µH

The thermal dissipation was investigated for various current loads. The analysis was performed for the BottomChip, without the Tx/Rx coil attached, since a half-chip allows for visual access to the gradient coil. The on-chip RTDs had a DC resistance of 92.15 Ω and 93.20 Ω and were driven by an LT1880 OP-AMP (Linear Technology) which delivered an exact current of 1.00 mA. First, the temperature coefficient α₀ was derived by measuring the RTD when heated in an oven at temperatures between 23 ^∘C and 50 ^∘C. A value of $(2.60\pm0.04)\times 10^{-3}\,\mathrm{K}^{-1}$ for α₀ was obtained. In comparison to a conventional Pt100 sensor ($\alpha_{\mathrm{Pt}} = 3.92\times 10^{-3}\,\mathrm{K}^{-1}$) the obtained temperature coefficient of the RTDs was smaller. Generally, the compound layer between the WTi and Pt due to re-sputtering and contamination within the vacuum chamber from the other targets affected α₀. Furthermore, variations from the lithography that rendered the RTDs pattern and plasma cleaning affected the α₀.

In addition, the surface temperature of the chip was captured by an infrared (IR) camera (PI-160, Optris GmbH, Berlin, Germany). The raw data of the IR images was post processed by a two point correction procedure to adjust gain and offset [36]. The RTDs also served for calibrating the thermographic imaging.

Figure 7 illustrates the heat-up without liquid cooling. The maximum temperature from the thermal camera was extracted for the ROI at the sample insert. For the experiment in figure 8, a holder made of PMMA was placed above the BottomChip to deliver DIW for the cooling. The thermal image was captured through the porthole in the centre of the PMMA holder. The peristaltic pump was adjusted to its maximum flow rate of 4.3 ml min⁻¹. The RTDs at the inflow and outflow position of the coolant are highlighted in figure 3. For an applied constant current of 4 A, the gradient coil can generate a gradient strength of 12.8 T m⁻¹ for the assembled chip, causing Joule heating of the assembly of up to 18 K above ambient within the sample ROI.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

When performing an NMR imaging experiment, the gradient coil is not constantly switched on and is generally operated in pulsed mode. The resulting duty cycle is typically much less than 2%. The duty cycle D is defined as the squared ratio of the root-mean-square current I_RMS by the peak current I_p, given by $D = (\frac{I_{\mathrm{RMS}}}{I_{\mathrm{p}}})^2$. To rapidly pulse an electric load, an adjustable current source was constructed as shown in figure 9(a). The circuit was based on a single OPA549 operational amplifier (Texas Instruments) and consequently required a bipolar power supply of ±15 V to deliver an output current of up to 10 A for a maximum resistive load $R_{\mathrm{Grad}}$ of 0.8 Ω. An adjustable RC-snubber network permitted trimming the overshoot to an acceptable value without continuously creating an infinite oscillation, in order to keep the rise time short. The measured overshoot was 24%, the rise time to 90% was 1.72 µs and the settling time to 5% was 9.54 µs as plotted in figure 9(b).

Zoom In Zoom Out Reset image size

Figure 9(c) shows the heat up from a pulse current experiment with a peak current of 10 A for the z-gradient without liquid cooling. A period time T of 400 ms was selected to closely match with the repetition time TR of previously performed MRM experiments [26, 27]. Figure 9(d) shows the thermal characteristic with active liquid cooling. At the subjected current of 10 A and a duty-cycle of 2%, a temperature rise of approximately 3 ^∘C can be expected in the sample region.

We performed the NMR experiment within a superconducting magnet that possessed a main magnetic field strength $\left|\boldsymbol{B}_0\right|$ of 11.7 T (Avance III, wide-bore, Bruker BioSpin, Germany), which corresponds to a $^{1}\mathrm{H}$ centre frequency of 500.13 MHz. An NMR imaging experiment requires a functioning RF pickup coil. Our approach was firstly to investigate the RF coil performance in a separate assembly without the gradient coils attached. The return loss (S11) was measured using an network analyser (NWA), which resulted in a dip of −37.8 dB at the respective resonance frequency. The coil was placed into the magnet and the coil's B ₁ field uniformity was examined by a nutation experiment. A rectangular glass capillary (Product 5012, sample space depth 100 µm, width 2 mm, glass thickness 100 µm, VitroCom, USA) was used as the sample container, and was filled with DIW. The maximum spectral amplitude was found at a pulse length of 27.75 µs with a pulse power of 0.23 W.

In a second experiment, a fresh capillary tube was filled with 0.5 M nicotinamide (C₆H₆N−2O) in D₂O (deuterium oxide) and we recorded a spectrum with the singlet-peaks detected at $\delta^{1}\mathrm{H} = 8.82$ ppm, 8.64 ppm, 8.14 ppm, 7.52 ppm and 4.87 ppm. The dominant peak at 4.87 ppm had a full width at half maximum (FWHM) of 9.84 Hz after moderate shimming, which was more than adequate to proceed with an imaging experiment.

In a third experiment, the gradient field homogeneity was evaluated using the micro-fabricated grid phantom. First, a reference magnetic resonance (MR) image of our phantom was recorded using our Tx/Rx coil and the commercial tri-axial gradient system Micro5 (Bruker BioSpin, Germany) that offered a gradient strength per axis of 0.05 $\mathrm{T}\,\mathrm{m}^{-1}\,\mathrm{A}^{-1}$. The MRI image of the channel's cross-section rendered sharp triangular shaped peaks as shown in figure 10(a). We filled blue ink into a fresh micro phantom from the same batch, to illustrate the NMR active region in figure 10(b). However, the microscope image was recorded several months after recording the MRI image and we experienced adhesion loss of the inner polymer towards the glass substrate. The channel shape resulted from the fabrication specific BS exposure. The matrix size of the MR image was 128 × 128, the anisotropic in-plane resolution (x–z) was 27.3 × 3.9 µm, the slice thickness was 0.25 mm, the echo time $T_{\mathrm{E}}$ was 4.98 ms, the number of scans was 500 and the scan time was 1 h 46 min. A signal-to-noise ratio (SNR) of 19.3 was calculated from the average signal intensities as extracted from the highlighted areas in figure 10(a). The 2D imaging experiment gave sufficient insight and allowed us to evaluate the micro gradient by recording a 1D MR image.

Zoom In Zoom Out Reset image size

To avoid additional modifications to the probe head and to simplify the installation, active liquid cooling of the here developed gradient system was not performed during the following NMR experiment. Nevertheless, the gradient linearity can also be tested at a low gradient strength and, based on our thermal examination, the gradient could be operated safely with currents below 100 mA. For driving the micro gradient, the single-channel GAP amplifier (Bruker BioSpin, Germany) was used, as it was directly available without requiring fundamental hardware modifications of the spectrometer. The amplifier allowed for a maximum current of 10 A at rise- and fall-time smaller than 10 µs (i.e. for 10%–90% amplitude transition). An additional resistance of 1 ohm was added to the current path, in series to the gradient, to inhibit damage to the amplifier through overheating while pulsing the low-ohmic gradient coil.

The pulse sequence for recording the 1D profiles was implemented using the spectrometer's control software TOPSPIN™ (Bruker BioSpin, Germany). The pulse sequence was inspired by Mansfield [37] by applying the z-gradient during signal readout. We obtained the signal directly from the FID without generating an echo. A pre-phasing gradient pulse as performed by Topgaard et al [38] was omitted, and rather we applied the phase correction as part of the signal post-processing.

To test the pulse sequence and post-processing, a reference profile using the Micro5 gradient coil was captured as shown in figure 11(a). After computing the fast Fourier transform (FFT) of the time-domain FID signal, the spin isochromats as a function of frequencies map to the position of the water filled channels, and revealed the projected cross-section of the phantom. The mapping of the Micro5 gradient confirmed the advertised gradient strength of 0.05 $\mathrm{T}\,\mathrm{m}^{-1}\,\mathrm{A}^{-1}$.

Zoom In Zoom Out Reset image size

The FID raw data was post processed using the nmrglue python module [39]. Within the relevant frequency range, peaks in the NMR spectrum with a selected height threshold were extracted using the module nmrglue.analysis. peakpick. The local gradient strength G between the neighbouring peaks was calculated by,

$$\begin{equation} {G}_{P_{n+1,n}} = \dfrac{\omega_{P_{n+1}}^{^{\prime}} - \omega_{P_{n}}^{^{\prime}}}{\gamma_{^{1}\mathrm{H}} \cdot (P_{n+1} - P_{n})}, \end{equation} \tag{ 1 }$$

where ω is the angular frequency of precession, $\gamma_{^{1}\mathrm{H}}$ is the gyromagnetic ratio, and P is the z position of the detected channel peak. For plotting the gradient linearity in figures 11(a) and (b), the lithographically defined channels of the phantom were considered to have an equidistant spacing to each other. From the layout data, a fixed value of 200 µm was considered as the channel-to-channel spacing $(P_{n+1}-P_{n})$, which was used to determine the gradient linearity of each segment in the plots.

For mapping the micro-gradient, several profiles with different gradient currents were recorded to detect irregularities in the gradient field. Two selected profiles are illustrated in figure 11(b). The acquisition time per scan for each profile was 0.17 s and 0.25 s, the number of scans was 100, and the maximum gradient current i_Gz was 12 mA and 23 mA, respectively. The FID resolution was 6 Hz and 4 Hz and the maximum FID resolution error was ±1.9% and ±0.8%. A gradient efficiency of 3.15 $\mathrm{T}\,\mathrm{m}^{-1}\,\mathrm{A}^{-1}$ was achieved for the micro gradient which was extracted from the peaks $P_{0}-P_{4}$, for a profiled length of 1 mm. The measurement at 23 mA shows insufficient signal from peak 5 onwards (see P₅ in figure 11(b)), which is attributable to inadequate sample holder filling, drying out of the liquid or structuring defects. The gradient strength deviated somewhat from the initial theoretical calculations of 4.0 $\mathrm{T}\,\mathrm{m}^{-1}\,\mathrm{A}^{-1}$ because of the aforementioned modifications to the layer stack, as estimated in figure 4, along with other deviations from the original optimised winding pattern that were necessary during fabrication.