Lock-In-Based TMR Magnetic Sensing for Feasibility Evaluation of ACF-Bonded FFC Interconnections

Anisotropic conductive film (ACF) bonding is widely used for fine-pitch electronic interconnections because conductive particles trapped between opposing electrodes provide through-thickness electrical paths while the polymer matrix maintains lateral insulation. Joint resistance depends on particle capture, bonding pressure, particle deformation, and local contact resistance [1]–[3]. Nonuniform bonding can therefore increase resistance, create partial opens, and reduce reliability. Conventional resistance measurements are effective for global continuity but provide little spatial information about current transfer across a bonded region. Optical methods have been used to assess particle deformation, conductive area, and resistance-related morphology in ACF assemblies [4]–[6], but morphology remains an indirect proxy for the actual current path. Magnetic sensing provides a complementary observable because conduction current produces a surrounding magnetic field. TMR devices are attractive for weak-field current sensing because of their sensitivity and compact size [7], [8], while magnetic current imaging has also been used to localize current paths in electronic structures [9], [10]. This study asks whether a phase-synchronous TMR readout can recover a repeatable current-synchronous magnetic response from an ACF-bonded FFC specimen. It is intentionally framed as a feasibility evaluation rather than particle-state reconstruction or a validated pass/fail classifier. The main contributions are a differential TMR/shunt acquisition architecture referenced to a common digital excitation phase, an explicit separation of the shunt’s global current role from the TMR probe’s spatial role, and a conservative repeatability analysis with the limits of the retained dataset stated explicitly.

The shunt and TMR channels provide complementary observables, summarized in Table I. The shunt measures total series current and is therefore a global excitation reference; it cannot localize where the current distribution changes. A scanned TMR probe instead provides a position-dependent magnetic response. For a straight-conductor approximation, the field at probe distance r is given by Eq. (1). Here μ₀ is the permeability of free space. Even after current normalization, B/I retains an approximately 1/r geometric dependence, so normalization cannot remove standoff, conductor-width, return-path, or probe-orientation sensitivity. The TMR approach is therefore meaningful when probe geometry is controlled and spatial information is required; for a single lumped resistance value, a conventional four-wire electrical measurement remains simpler and more directly calibrated.

The specimen is driven from a current-limited 5 V source through the implemented 22 Ω/2 W series resistor and a low-side IRLZ44N MOSFET; the return path includes the 1 Ω shunt. The ideal circuit-level current ceiling is therefore about 0.217 A before specimen and MOSFET voltage drops, so the excitation is reported conservatively as approximately 0.1–0.2 A. A Portenta H7 provides a 1 kHz PWM-based excitation. Because no reconstruction filter produces a pure analog sinusoid at the specimen, the signal contains a component at f₀ together with PWM-related harmonics and is modeled by Eq. (2). In Eq. (2), A₁ and φ₁ denote the component at f₀, Aₖ and φₖ denote harmonics at kf₀, and η[n] denotes residual noncoherent noise and interference. Coherent integration at f₀ attenuates components outside the detection band, although switching-edge coupling and coherent harmonic leakage remain experimental limitations.

The CT220 TMR sensor [11] is positioned approximately 0.3 mm above the bonded region. With the circuit-consistent current range, Eq. (1) gives a first-order straight-conductor field estimate of roughly 67–145 µT at this standoff; the estimate is only a sanity check because the real FFC has finite conductor width, a nearby return path, and orientation-dependent coupling. The ADS1256 [13] measures the TMR signal differentially as AIN0–AIN1 and the shunt as AIN2–AIN3. Because one delta-sigma conversion path is multiplexed between the pairs, they are selected sequentially rather than sampled simultaneously. The shunt is consequently used as a global current magnitude/calibration reference, not as a strictly sample-coincident cross-channel phase reference. The shunt path includes the implemented 1 kΩ/0.1 µF input network described by Eq. (3). For R_f = 1 kΩ and C_f = 0.1 µF, f_c ≈ 1.59 kHz. At 1 kHz, |H_sh| ≈ 0.847 (−1.45 dB) with a phase shift of approximately −32.1°, so inverse correction of the raw complex shunt phasor requires about 1.18× magnitude and +32.1° phase. This correction applies only to the shunt path. Because the archived spatial profiles do not retain the complete raw time-stamped TMR/shunt calibration chain, Fig. 3 is not retroactively assigned an absolute V/A scale

For each differential sequence q ∈ {T, sh}, coherent in-phase and quadrature components at f₀ = 1 kHz are computed using the same digital excitation phase reference [12]. Eq. (4) uses the acquisition time t_q,n and sample count N_q belonging to each multiplexed sequence, so it does not assume simultaneous TMR and shunt samples. The retained TMR processing record gives an effective 15 kS/s sequence rate, corresponding to 15 TMR samples per excitation cycle and 750 samples over 50 cycles, i.e., a 50 ms coherent-integration interval. Both channels use the same digital excitation reference, but no zero inter-channel sampling delay is assumed. Equation (5) defines the complex differential-voltage phasor, the excitation-current magnitude obtained from the RC-corrected shunt amplitude, and the signed TMR projection S_ref referenced to the digital excitation phase. Unlike the nonnegative phasor magnitude, S_ref can have either polarity. A calibrated current-normalized signed response can be formed when complete complex data and the voltage calibration chain are retained; the archived Fig. 3 profiles are therefore reported only in arbitrary units.

One ACF-bonded FFC specimen was fixed on a mechanical fixture and scanned four times (t1–t4) under nominally identical conditions. The probe was translated in 0.1 mm increments while the nominal standoff was maintained at approximately 0.3 mm. A star-grounding scheme kept the high-current return out of the sensitive analog reference path. Because absolute stage origin and bond-zone endpoints were not retained, the stored profiles are reported on a normalized 0–1 coordinate. The principal hardware and processing settings are summarized in Table II.

Direct DC readings near the bonded region were strongly influenced by CT220 offset, geomagnetic background, slow drift, and environmental interference. The phase-synchronous detector instead isolates the component coherent with the 1 kHz excitation. A quantitative matched comparison would normally use the amplitude-domain SNR in Eq. (6). In Eq. (6), A_sig is the recovered coherent signal amplitude and σ_noise is the RMS noise in the same units. A valid numerical DC-to-lock-in comparison requires matched analog gain, bandwidth, observation interval, differential input configuration, and probe position. The retained four-scan dataset does not contain a time-matched raw DC trace, so no numerical SNR-improvement claim is made; the evidence supports qualitative baseline separability only.

Fig. 3 shows the signed, phase-referenced response from four repeated scans. The ordinate is not the nonnegative lock-in magnitude; it is a scaled signed projection referenced to the digital excitation phase, so both positive and negative values are physically allowed. The archived traces were lightly smoothed for visualization, but the exact smoothing-window metadata were not retained; consequently, no quantitative claim is based on an individual local peak. The result supports recoverable current-synchronous spatial structure, not particle-level current mapping or a unique defect-localization peak.

For each scan, three scalar descriptors are extracted: the mean-absolute response, the RMS response, and the peak-to-peak response. The mean-absolute response is used instead of the signed spatial mean because the signed response fluctuates about zero. Repeatability across the \(M\) repeated scans is summarized using the sample mean, sample standard deviation, and coefficient of variation (CV), as defined in Eq. (7). Here, M = 4. Table III summarizes the resulting descriptive CV estimates. Because only four repeated scans are available, these CVs should be interpreted as preliminary feasibility-level estimates rather than statistically validated repeatability metrics. No inferential claim is made that the differences among the descriptor CVs are statistically significant.

The shunt is not used as the primary inspection metric because it measures only the total series current, whereas the scanned TMR probe provides position-dependent spatial information. Current normalization also cannot eliminate the approximately 1/r dependence on probe standoff described in Eq. (1). The present evidence is therefore limited by the use of a single specimen with four repeated scans, normalized position metadata, sensitivity to probe alignment and standoff, sequentially multiplexed rather than simultaneous ADC acquisition, incomplete retention of the raw complex TMR/shunt data, the absence of matched DC and explicit current-ON/current-OFF controls, and the lack of retained smoothing-window metadata. Future validation should retain raw time-stamped differential waveforms and preprocessing parameters, record absolute spatial position, employ simultaneous acquisition when cross-channel phase comparison is required, quantify geometric uncertainty and measurement repeatability, and evaluate multiple known-good and intentionally defective specimens using confidence intervals, effect sizes, and classification statistics.

A lock-in-based TMR magnetic sensing architecture was evaluated as a feasibility approach for ACF-bonded FFC interconnections. The system combines a 1 kHz PWM excitation, a CT220 probe, a 1 Ω shunt, and two multiplexed ADS1256 differential input pairs to recover a current-synchronous spatial response while explicitly separating the shunt’s global calibration role from the TMR probe’s spatial sensing role. The formulation distinguishes the signed phase-referenced response from the nonnegative I/Q magnitude and accounts for the non-negligible response of the 1 kΩ/0.1 µF shunt input network at 1 kHz. Four repeated scans yielded descriptive CVs of 2.6%, 4.8%, and 9.3% for the mean-absolute, RMS, and peak-to-peak response descriptors, respectively (Table III). Because the present dataset contains only one specimen and does not retain the complete raw calibration chain or matched control data, the results should be interpreted as feasibility-level evidence rather than as validated inspection performance. Further validation using multiple known-good and intentionally defective specimens, rigid probe positioning, raw time-stamped data retention, and calibrated processing is required before statistically grounded defect classification or production-level inspection capability can be established.

This work is financially supported by the Ministry of Trade, Industry & ENERGY(MOTIE) through the fostering project of project for industry-university convergence district build-up.