Magnetic Random Access Memory (MRAM) is a promising non-volatile memory technology poised to replace existing solutions like eFlash. However, achieving high performance and yield in MRAM necessitates precise control over device properties, particularly given the shrinking size of magnetic bits (often <60 nm). Traditional metrology techniques, such as magneto-optical Kerr effect (MOKE) and current in-plane tunneling (CIPT), often suffer from limitations in spatial resolution and the ability to characterize individual bits. This work explores the use of scanning nitrogen-vacancy magnetometry (SNVM), a quantum sensing technique, to overcome these limitations. SNVM offers high spatial resolution and sensitivity, enabling non-contact characterization of individual bits within MRAM arrays. The authors argue that SNVM's ability to probe individual bits, providing detailed information on their magnetic properties, switching behavior and thermal stability, makes it uniquely suitable for early-stage failure analysis and process optimization in MRAM manufacturing. STT-MRAM, a specific type of MRAM, is used as an ideal candidate to demonstrate the potential of SNVM due to its industrial relevance and current production status for embedded flash replacement. The interfacial nature of the tunneling magnetoresistance (TMR) and spin-transfer torque (STT) effects in STT-MRAM makes uniformity a critical challenge, especially considering the nanoscale dimensions of the magnetic layers. Variations in the etch process significantly impact the energy barrier for magnetic switching, affecting data retention and bit-to-bit uniformity. Existing metrology techniques lack the single-bit resolution needed for comprehensive failure analysis, making SNVM's capabilities particularly valuable.

The paper reviews existing MRAM characterization methods, highlighting the limitations of conventional techniques. Magneto-optical Kerr effect (MOKE) magnetometry, while offering information on magnetic properties, suffers from insufficient resolution to probe individual bits, resulting in ensemble averaging. Current in-plane tunneling (CIPT) provides information on film properties before patterning, but individual bit characterization is only possible after the complete integration process—a long feedback loop. Other methods like conductive atomic force microscopy (cAFM) and magnetic force microscopy (MFM) have limitations in terms of resolution, invasiveness, and reproducibility, making them less suitable for in-line metrology. The authors position SNVM as a superior alternative, capable of addressing the shortcomings of existing methods by providing single-bit resolution, non-contact measurement, and high magnetic field sensitivity. The review emphasizes the need for a metrology tool capable of characterizing individual bits immediately after the patterning stage to significantly reduce development time and costs, contrasting the weeks-long thermal bake experiments currently used for array-level retention assessment with the approximately 1.5-hour SNVM scan of a 2 kB array.

The study employs SNVM using a Qnami ProteusQ microscope. A nitrogen-vacancy (NV) center in a diamond tip is used as a highly sensitive magnetic field sensor. By scanning the NV center over an MRAM array, the local magnetic stray field of each bit is measured and mapped. The signal corresponds to the projection of the stray field vector onto the NV quantization axis, allowing distinction between parallel (P) and antiparallel (AP) bit states, which appear bright and dark, respectively, in the SNVM images. Two wafers were fabricated using different etch processes (Process 1 and Process 2), the latter including an additional etchback and gentle oxidation step. The MRAM pillars had a nominal 60 nm diameter (reduced to ~45 nm after patterning) and 200 nm pitch, consistent with current and near-future applications. The MTJ tri-layer consisted of CoFeB free and reference layers and an MgO tunnel barrier. The samples were encapsulated with SiN to prevent oxygen exposure. Bit states were manipulated by applying out-of-plane magnetic fields. The SNVM measurements were performed after encapsulation. Data analysis involved image processing to identify individual bits and classify them based on their magnetic state. A statistical method was used to analyze switching uniformity, and the domain wall mediated reversal (DWMR) model was employed to fit the SNVM data and extract parameters like the energy barrier (Δ) and anisotropy fields (Hk). Simulations using Magpylib were also performed to model the magnetic stray fields generated by the pillars, considering variations in magnetization, pillar diameter, and magnetization direction. These simulations helped interpret the experimental data and understand the origin of bit-to-bit variations.

SNVM successfully imaged and distinguished individual bits in an MRAM array with high resolution (even with encapsulation). Analysis of two different etch processes showed that Process 2 (with the additional etchback and oxidation) resulted in improved bit-to-bit uniformity compared to Process 1. This was evident in the narrower distribution of the maximum stray field measured in the P state (σ = 136 µT for Process 2 vs. σ = 250 µT for Process 1). The SNVM data also revealed a preferred orientation (P state) and atypical behavior at the edges of the array. Analysis of the switching behavior revealed that Process 2 showed more uniform switching characteristics than Process 1, exhibiting a narrower distribution of switching events over 10 repetitive measurements. By fitting the SNVM data using the DWMR model, the researchers extracted thermal stability values (Δ = 49 for Process 1, Δ = 54 for Process 2), indicating higher thermal stability for the optimized process. Simulations validated that the non-binomial switching behavior was mainly due to variations in the magnetization, magnetization direction, and critical dimensions of the MRAM pillars. There was a strong correlation observed between the average stray field and the number of successful switching events for each bit, illustrating the link between stray field distributions and switching uniformity. The simulations accurately reproduced the experimental stray field histograms and switching probability, corroborating that bit-to-bit variations were a primary source of non-uniform behavior.

The findings demonstrate that SNVM provides a powerful, high-resolution tool for characterizing individual MRAM bits early in the manufacturing process, directly after the etching step. This is a significant advantage over existing techniques that lack single-bit resolution or require electrical connection and complete integration. The ability to identify out-of-distribution bits and analyze their magnetic properties allows for early detection of process failures and optimization of fabrication steps. The consistency between the experimental results and simulations underscores the understanding of the factors contributing to bit-to-bit variability in MRAM. The correlation between stray field distribution, switching uniformity, and device parameters like thermal stability makes SNVM a valuable tool for process monitoring and improvement. The results suggest that SNVM has the potential to replace current slower and less-sensitive methods and become integrated into industrial MRAM manufacturing lines.

This study successfully demonstrated the use of SNVM as a highly effective metrology technique for characterizing individual MRAM bits. SNVM's high spatial resolution, non-contact nature, and high sensitivity enables early-stage failure analysis and process optimization. The observed improved uniformity in Process 2, confirmed by both experimental SNVM data and simulations, highlights the technique's potential for real-world application in MRAM manufacturing. Future work could focus on increasing the measurement speed of SNVM for in-line process control, exploring the technique's applicability to other types of magnetic memory architectures and investigating the effects of localized heating or perturbation fields on data retention.

While the study demonstrates the efficacy of SNVM, limitations include the current scan speed of approximately 1.5 hours for a 2 kB array. While future improvements are expected, this speed might limit the technique's immediate application for high-throughput in-line process control. Another limitation is the assumption of certain parameters in the DWMR model fitting and simulations. The accuracy of the extracted parameters (like thermal stability) depends on the accuracy of these assumptions. However, the study demonstrates strong correlations between the observed behaviour and the simulations, increasing the confidence in the conclusions reached.