Study of die break strength and heat-affected zone for laser processing of thin silicon wafers (2024)

https://doi.org/10.2351/1.4916979

There are three factors that dominate the semiconductor technology development process: cost, performance, and form factor. The wafer level performance factor is driven by the front-end semiconductor foundries struggling to keep pace with “Moore's law.” The increased performance also requires thinner and improved “low-k” interlayer dielectric (ILD) materials and 2.5D/3D-IC designs. The often brittle and porous features in the ILD materials reduce the adhesion between stack layers, making the stack inherently weaker1 and susceptible to cracking, chipping, and delaminations.2 Furthermore, with consumer devices targeting thinner, smaller, and more powerful components, technologies such as substrate thinning and through silicon via interconnects are mostly employed to achieve performance with less volume. As an example, Solid state drives (flash-SSD) are contributing to the thinning trend of substrates to ultrathin levels (<50 μm). However, these techniques present challenges for the mechanical integrity of the chips and introduce weakness in dies, which introduce yield loss which greatly impacts the manufacturing costs. Singulation of a substrate by any means, including mechanical saw, laser-mechanical hybrid (laser scribe followed by mechanical saw), or laser full cut, inevitably creates damage to the substrate.3 Understanding and controlling this damage is a key to minimizing yield impacts to the functional device and mechanical defects that reduce the strength of the substrate during the packaging processes.

This paper focuses on identifying the laser characteristics, attributes, and process factors that impact the mechanical flexural strength of ∼50 μm silicon substrates. Heat-affected zone (HAZ) is a term used to encompass the damage caused to a material during and after irradiance by a laser pulse. The impacts to die strength for laser induced HAZ are understood to be related to HAZ size and defect density.4 We use empirical test data from full cut laser dicing technique using nanosecond and ultrafast lasers to quantify the HAZ impact on the mechanical strength of the substrate. We further model a theoretical HAZ for nanosecond laser conditions with a computer simulation and correlate the results with the empirical data. We then extend the laser parameter space to ultrafast lasers in the empirical testing to augment the range of fluences used for full dicing <30 J/cm² due to nanosecond ablation threshold limitations for dicing processes.5 The findings of this paper are applicable to laser processes for both scribing and full dicing of silicon substrates, among other materials.

Die break strength (DBS) is a commonly used and key indicator of a scribing and/or dicing process's capability to produce strong dies with minimum structural impacts. We experimentally determined die strengths with an Admet 5600 universal strength testing machine, which measures the max force (N) prior to catastrophic failure. The method for destructive testing follows the procedure and process outlined in the semi standard G86-0303 for three point bend testing, intended to restrict failures to edge defects of laser cut sidewalls of the material. The G86 standard requires a sample size of at least 25. We operate the Admet tester with a Shimaduz 3pt jig as supporting test fixture. The two test jig support pins and anvil had a radius of 0.3 mm. The test parameters of span and feed speed across the entire data set are 2 mm and 4 mm/min, respectively. We measured the average thickness in microns of each sample set to account for stress differences due to thickness variation. Figure 1 shows the flexural stress calculation equation and parameter definition. The flexural stress can be calculated by the following equation:

where W is the load in kilograms, applied via the anvil in the direction of the orange arrow, perpendicular to the laser exit surface. The span between the two support pins is L in mm. The sample thickness is h in mm. The sample breath is b in mm. The sample was set circuit side or laser incident side face down on the two support points; the background die surface or laser exit side is facing the anvil. The sample size was typically 10 × 10 mm. Sample substrate thickness was 0.050 mm.

We performed the laser processing experiments on a standard laser system architecture:6 the laser beam passes through a combination of mirrors, beam expanders, and an acousto optical modulator for energy control into a set of orthogonal motor driven mirrors (“galvos”) for fast beam steering, through a F-theta tele-centric lens (F/# 6.4) with a 20 mm field size onto the work piece. The work piece was held flat on a glass chuck surface by vacuum; the chuck was mounted onto a linear bearing XY-stage. The substrate processing by laser was executed at atmospheric pressure and room temperature. No gas assists were used for processing in this experiment.

The lasers for the nanosecond processing experiments are commercially available frequency tripled diode pumped solid state Q-switched lasers with pulse widths of 1-154 ns and a wavelength of 355 nm, while the lasers for the ultrafast processing are commercially available mode locked fiber lasers in the 400-800 fs nominal pulse width range with a wavelength of 343 nm. The lasers used for the experiments are detailed in Table I.

In order to expose DBS trends for nanosecond processes and achieve higher DBS, some wafers were processed with XeF₂ etch after dicing. A heated micro-electro-mechanical systems vacuum chamber capable of processing full 300 mm diameter wafers mounted on wafer frames serves as our etch system. The etch process consists of a continuous flow of XeF₂ regulated with a Mass Flow controller to spontaneously etch the sidewalls of the die removing a certain HAZ thickness for a given time, effectively stress relieving the laser-processed area.

Blank silicon substrates with an oxide film (SiO₂) are used for three main purposes in our experiments; (1) to compare the effects of nanosecond and femtosecond laser processes with DBS and HAZ, (2) to contrast the empirical results with the simulated results, and (3) to directly measure the HAZ size as a function of etch time and etch rate. The blank substrates are prime grade 〈100〉 orientation silicon, with a 5 kÅ thermal oxide film. The oxide film acts as a mask, protecting the top surface silicon from the XeF₂ etchant. The SiO₂ substrate thickness is a nominal 50 μm and diameters are 200–300 mm. The dynamic random-access memory (DRAM) device substrates were used in our experiment to compare the nanosecond laser process DBS and HAZ on device wafers. The DRAM device substrates in our experiments are prime grade 〈100〉 orientation silicon. Total DRAM thickness is 60 μm comprised two distinct regions, the silicon substrate is 50 μm, and the device circuitry thickness is 10 μm. DRAM substrate diameters are 300 mm. Both SiO₂ and DRAM substrate samples received a dry polish finish after the grinding process to remove the residual grinding damage on the wafer backside. We processed both SiO₂ and DRAM samples device or film side incident to the laser beam with the samples mounted on polyolefin dicing tape.

We can divide our processes into four categories: laser full dice, laser scribe and full dice, laser full dice with etch, and laser scribe and dice with etch.

Characterization of the effective HAZ and the effect on DBS for our experimental results is an important step in understanding the process impact to the substrate strength.6 The experimental results for this method are displayed and discussed in Sec. III B titled “Effective HAZ for 60 ns multipulse,” Fig. 5. For the purpose of this experiment, we have chosen to use a correlation of material removed by etch in microns and DBS in MPa to characterize the size of the effective HAZ. Figure 2 shows four images, the pre and postetch process on the oxide on silicon substrate. The cartoons 2(a) and 2(b) depict a top down and cross-sectional view of the substrate pre and postprocessing (laser and etch). The microscope images 2(c) and 2(d) are examples of a 60 ns laser processed and etched substrate. With image analysis and Hough transform, we have measured the laser and laser and etch kerf widths across the wafer. By subtracting the respective mean kerf widths across a wafer, we can find the mean material removed by etch from each kerf edge. This is due to the nature of the isotropic etch process, which causes an undercut of HAZ and silicon beneath the SiO₂, the remaining overhanging SiO₂ shelf is then measured. The benefit of the 5 kÅ oxide film is that the HAZ removed by the etch process can be quantified easily by optical microscopy from the laser incident surface. The oxide acts as a mask to the silicon die surface, allowing the XeF₂ etchant to remove only the sidewalls of the silicon due to the selectivity toward Si and not to SiO₂.

The simulations in this work are carried out using a multiphysical simulation package developed by Vienna University of Technology (TU-Wien) and Electro Scientific Industries (ESI).7 The simulation package is designed to model nanosecond laser processes involved in many industrial laser cutting, drilling, and scribing applications. Specifically, the model considers the laser beam propagation to the work surface, the interaction of the laser beam with the target, namely, absorption, reflection, and transmission. The coupled heat transfer and Navier–Stokes equations are solved to account for the convective and conductive heat transfer during laser processing as well as the fluid or gas flow when the material is melted or vaporized. The model also takes into account the recoil pressure from the vapor causing deformation of the free surface as it expands into the surrounding medium. Further details of the simulation model can be found in Refs. 8 and 9, and the material properties of silicon used in the simulations are listed in Ref. 10. In this work, we apply the multiphysical simulation model to investigate the HAZ in single and multiple nanosecond laser processing in silicon.

To differentiate between the experimentally determined and computer simulated HAZ in this paper, we refer to “effective HAZ” and “simulated HAZ,” respectively.

In order to understand how the laser pulse width affects the HAZ in silicon, we carried out a series of multiphysical simulations for both single pulse and multiple pulse processes in a 50 μm silicon wafer. Figure 3 shows the results for the width of the simulated HAZ in a range of laser pulse widths from 1 ns to 200 ns at a given pulse energy of 10 μJ and spot size of 8 μm, i.e., an incident fluence of about 20 J/cm². In the simulations, the HAZ in the silicon wafer by the laser processing is identified using the maximum temperature, T_max, in the remaining silicon wafer that rises above the melting temperature of silicon, i.e., T_melt = 1685 K during the simulation. We expect that mechanical properties in the HAZ are different from the single crystalline silicon due to the appearance of crystalline defects and polycrystalline structure during resolidification.11 Figure 3(a) shows that as the pulse width increases from 1 ns to 200 ns, the simulated HAZ increases from 0.3 μm to 1.7 μm and the ablation depth increases from 0.4 μm to 5.8 μm. For pulse widths less than 20 ns, the simulated ablation depth agrees with the theoretical thermal diffusion length $2Dτ,$ where D is the thermal diffusivity of silicon, 0.8 cm²/s and τ is the pulse width of the laser as shown by the plotted black line in Fig. 3(a). This can be understood by the fact that for these pulse widths, most of the absorbed energy within the skin depth is used for ablation. For longer pulse widths, calculation base on thermal diffusion theory overestimates the material removal, as part of the absorbed energy converts into heating the surrounding material, creating a larger HAZ as shown in Fig. 3(a). Figure 3(b) shows the HAZ from a single-pulse 150 ns simulation as defined from T_max > T_melt in 50 μm laser-processed silicon wafer, i.e., only the part of remaining silicon wafer experiencing melting and resolidification is shown. Our multiphysical simulations also show that the largest HAZ in the ablated silicon wafer is found to be near the surface region of the irradiated target [Fig. 3(b)], indicating that the HAZ is likely affected by hydrodynamic effects from the melt flow driven by the vapor pressure during the laser process. The size of the HAZ from the simulation is consistent with previous experiments,12 where a damaged layer on the order of 1 μm containing defects (dislocations) is observed between the ablated boundary and inner single crystalline silicon. Furthermore, flexural bending theory asserts that the tensile and compressive stresses exerted on the substrate during die pick are maximized at the laser incident and laser exit interfaces, hence the maximum stresses are always going to be located in proximity to the weakest edge defects.

The heat-affected zone is further analyzed in the case of multiple subsequent pulses where—depending on the scan speed—during each pass several laser pulses can partially overlap in the same location. Here, we quantify the pulse overlapping via a so-called net fluence, defined as an accumulated pulse energy per surface area within a 1/e² spot area and can be calculated by

The summation in Eq. (2) is done over a series of laser pulses, noted by i, that overlap in a given 1/e² spot area. F₀ is the laser fluence defined as $F0=E/πω02,$ where E is the laser pulse energy and ω₀ is the 1/e² spot radius. $N=2×(2ω0/δ)$ is the total number of spots overlapped in the same spot area. The interspot separation, δ = v/PRF, is determined by the laser scan velocity, v, and the repetition rate of the laser, PRF. The geometric factor, g_i, in Eq. (2) accounts for the partial overlapping of a Gaussian pulse within a 1/e² spot area

where for a particular laser pulse i, the integration in Eq. (3) is done over the overlapped area between the Gaussian pulse and the designated 1/e² spot. We believe that the net fluence is a meaningful factor in understanding multipulse processing of substrates for commercial applications utilizing high repetition rate and average power lasers, especially when the heat accumulation effects in thin silicon substrates become important at high laser frequencies on the order of hundreds of kHz. For a given repetition rate and pulse energy, the net fluence generally increases when the scan speed is reduced, because smaller spot-to-spot distance leads to higher accumulated pulse energy within the same spot area.

Figure 4 shows the simulated HAZ as a function of the net fluence obtained in a series of multiphysical simulations carried out for a range of scan velocities from 25 mm/s up to 1500 mm/s. The shared parameters for all the simulations in Fig. 4 are a pulse width of 60 ns, pulse energy of 10 μJ, repetition rate of 200 kHz, and a laser spot size of 8 μm. The simulated HAZ increases in accordance with increasing net fluence, except that at low net fluences of 40 J/cm² or less, which correspond to scan velocities beyond 700 mm/s, the simulated HAZ appears to saturate to 1.1 μm.

Interestingly, as shown in Fig. 4, the HAZ obtained from our multiphysical simulations appears to be about 2–3 times smaller than the actual HAZ determined from the etch test in Figs. 5 and 6. This discrepancy of the widths of HAZ between the simulations and experiments suggest that in addition to the HAZ region identified by our simulations, the DBS is effected by the occurrence of other structure imperfections beyond the simulated HAZ region,13 such as microcracking caused by the thermally induced tensile stress during the cooling process of resolidification in the silicon wafer.14–16 The laser dicing parameters, including repetitions, pulse energy, scan velocity, and spot size, are known to effect the surface morphology and HAZ structure17 which can also impact the DBS.

Figure 5 plots the empirical mean DBS versus effective HAZ removed for SiO₂ substrates processed with 60 ns and a net fluence process of 20 J/cm² process. The process used to characterize the HAZ for Fig. 5 is described in Sec. II B titled “Effective HAZ characterization method,” Fig. 2. From this chart, we see that the DBS begins to saturate at 4–5 μm of HAZ removal. The saturation of the DBS at 4–5 μm implies that laser induced HAZ has been fully removed by the etch process. Therefore, the laser induced HAZ—as it pertains to DBS—is in the range of 4–5 μm for a 60 ns 20 J/cm² process.

Figure 6(a) displays the empirical DBS results for the following substrates; DRAM (“DRAM w/etch”) and SiO₂ (“laser only”). The DRAM w/etch process received a post laser etch to expose the DBS trend. The laser only process serves as a baseline using SiO₂ substrate without an etch. We apply a scribe and dice process to the DRAM substrate for a range of scribe scan velocities from 100 mm/s up to 1500 mm/s, resulting in a net fluence skew. Again, we apply a laser process with a pulse width of 60 ns, scribe pulse energy of 10 μJ, repetition rate of 200 kHz, and a laser spot size of 8 μm. The effective fluence is 20 J/cm². All the DRAM w/etch data points are etched for 80 s in a batch process, i.e., the etch rate and HAZ removed is equivalent between points. The effective etch rate for the 80 s process is 3.4 μm/min, hence we removed an average of 4.5 μm HAZ from all DRAM “w/etch” sidewalls. Due to the combination of measurement error, noise with three-point bending, the defect density, and degree of damage caused by the 60 ns laser process, it is very difficult to resolve the simulated HAZ trends in the effective HAZ by DBS measurements in the range of 200–250 MPa. The green laser only fit line displays no discernable trend in the DBS.

Figure 6(b) displays a scanning electron microscope (SEM) image of the DRAM device diced in the experiment. The substrate is positioned at a die corner, where the laser processed and etched sidewall is facing the viewer, the substrate laser exit surface with die attach film is at the top of the image, and the laser incident side is at the bottom of the image. The caliper measurement is measuring the underside of the device circuitry side, where the etchant has removed HAZ silicon, also known as the “etch-back” or “undercut.” The bulk silicon substrate is the middle gray layer, whose surface texture matches that of a silicon etched to the single crystal boundary.18 By measuring the device undercut width, we approximate the HAZ removed as a result of the etch process, this image shows 4.74 μm removed, the average HAZ removed for all points was found to be 4.5 μm. Thus, we can approximate from the graph that at net fluences of >50 J/cm² the average effective HAZ—as it pertains to DBS—from the 60 ns laser process exceeds 4.5 μm on the DRAM substrate.

Figure 7 plots the empirical mean DBS versus Net Fluence for both ultrafast (red) and nanosecond processes (black) by pulse duration. Due to the ablation thresholds limitations for the nanosecond dicing process, with effective process speeds and energy densities experiencing diminishing returns for net fluences and fluences of <30 J/cm² we could not effectively explore the nanosecond regime under 30 J/cm². Instead we employed ultrafast laser processes to explore the fluences below <30 J/cm². Due to the relative immaturity of commercial high ultraviolet (UV) average power femtosecond lasers in the 300–500 fs pulse width range, the 400 fs data could not be explored above a fluence of >2 J/cm² and net fluence of >22 J/cm². The dashed lines serve as a “guide to the eye” for illustration purposes. The dominant factor of net fluence is evident for the ultrafast pulse widths. However, the large distribution of data points around the 800 fs fit line suggests that there are other important factors playing a role. No statistically significant trend is identified in the 60 ns data set, however, the fit line does inversely correlate somewhat to the simulated HAZ trend from Fig. 4. At 1 ns we begin to see a significant trend as the net fluence decreases. We believe that at effective HAZ > 5 μm the 3 pt measurement technique is insufficient at identifying HAZ versus DBS trends. This is due to a screening effect where the inherent weakness of the substrate due to the nanosecond laser damage causes dies to fail prematurely during die pick and handling, resulting in a bias or skew where only the stronger die are tested for DBS and the weaker die are filtered out of the data. Strangely, we did not see this failure/screening effect for the femtosecond laser processes at high fluences or high net fluence processes. This is most likely due to the impact of the laser process to the die strength for the laser exit surface, which can be different from the laser incident surface die strength, the subject of this study.

An interesting learning from Fig. 7 chart is that under the correct parameters and conditions a nanosecond process can achieve higher or equivalent DBS to the 800 fs processes, implying that the HAZ size or impact of the HAZ on the substrate strength is not solely dependent on the pulse width. If poor process parameters are chosen for either ultrafast or nanosecond lasers, the mechanical damage to the silicon lattice can far exceed the observed HAZ size creating weakness and failures in the substrate.11,19–21

Figure 8 plots the empirical mean DBS versus bite size by fluence for 800 fs (red) and 60 ns (black) processes. The terms bite size and interpulse separation as described in Sec. III A are synonymous. Bite size is the term commonly used in the laser micromachining industry to describe pulse separation. In this chart, we can compare the impact of fluence on DBS for fixed rep rates and equivalent bite sizes. The bite size ranges explored in Fig. 8 were limited by a combination of factors. Factors limiting femtosecond bite size were a combination of >800 kHz rep rate and maximum scan velocity of 4.5 m/s. Factors limiting nanosecond bite size were excessive damage to the silicon at bite sizes of <10 μm resulting in mechanical failures during pick, effectively screening the samples prior to DBS testing. This chart illustrates that reducing the process fluence for both 800 fs and 60 ns will result in a DBS increase. This chart also illustrates the importance of bite size or scan velocity for the ultrafast processes. An example of this effect is a 16 J/cm² femtosecond process at <5 μm bite size, results in lower or equivalent DBS to a 40 J/cm² 60 ns high velocity process (12–20 μm bite size). Figure 8 suggests that fluence is a dependent variable to a higher order parameter. This chart supports the observations from Fig. 7 that net fluence dominates the DBS and HAZ formation over fluence or pulse width.

Figure 9 shows the highest mean DBS across the given pulse widths for the optimal samples set. This DBS represents the true process potential in terms of mean DBS, and the guide to the eye illustrates a clear pulse width dependency across the lasers tested, but only for the correct combination of laser conditions (pulse energy, scan velocity, and spot size).13 Under optimal conditions, the low thermal damage and minimal HAZ benefits of ultrafast lasers are readily apparent in this chart. The optimal conditions for Fig. 9 are detailed in Table II.

Using multiphysical simulations for nanosecond laser processing, we demonstrate that the HAZ in silicon depends on both the pulse width and net fluence. The simulated ablation depth is in a good agreement with the theoretical thermal diffusion length for pulse widths in the range of 1–50 ns, indicating that for these pulse widths, most of the absorbed laser energy is used for ablating silicon. For longer pulse widths, part of the absorbed energy converts into heating the surrounding material, resulting in a large HAZ in the remaining silicon. The simulation results also illustrate that the thickest HAZ section during laser scribing is typically located at the boundary of the laser incident surface and could further influence the die pick process in which the maximal stresses is realized near the wafer edge. The multipulse simulations for 60 ns scribe process show that the HAZ becomes larger when the net fluence increases as a consequence of an increasing accumulated energy deposition at a smaller interpulse separation.

Our multipulse experimental tests show similar HAZ trends for net fluence for 1 ns, 60 ns with etch and both 400 fs and 800 fs femtosecond lasers. Net fluence is found to have the largest impact on DBS and therefore is an important factor in minimizing the effective HAZ and balancing the residual tensile and compressive stresses in substrates. We have also shown that net fluence is a meaningful process metric for damage induced by HAZ to compare lasers processes with a large variety of pulse widths, laser scan speed, average powers, and repetition rates. Fluence and scan velocity or bite size are the key drivers of net fluences, however, due to observed diminishing returns effect for low fluences, a medium fluence with high scan velocity process was found to produce a reasonable balance of DBS and machining speeds. Pulse width also showed a significant DBS increase when comparing means DBS across each pulse width and for the optimal conditions, as expected ultrafast processes had higher DBS than nanosecond processes, resulting in less HAZ. Fluence was found to have little impact on the overall DBS trend as an independent variable, but when the velocity and pulse width were fixed and fluences were reduced, the anticipated DBS improvement was observed, but at the expense of machining speeds. On the other hand, we also found two femtosecond processes that resulted in lower DBS than the mean DBS for nanosecond processes, and many femtosecond conditions resulting in equivalent DBS to mean DBS for nanosecond processes. This observation suggests that in addition to HAZ, mechanical damage to the silicon wafer during femtosecond processing could significantly influence the DBS and result in weakness and failures in the substrate.

Generally, nanosecond processes gave higher machining speeds with low DBS, femtosecond processes gave higher DBS with low machining speeds. This quality, throughput tradeoff is an important consideration for commercialization of ultrafast scribing and dicing techniques. That said, our data show that a solid understanding of factors beyond just the pulse width is necessary to identify an optimal laser solution and process condition.

The authors thank Andreas Otto, Rodrigo Gómez Vázquez, and Robert Bielak at TU-Wien for helpful discussion of analyzing the multiphysical simulation results.

© 2015 Laser Institute of America All article content, except where otherwise noted, is licensed under a Creative Commons Attribution 3.0 Unported License.

2015

Laser Institute of America