Ductile–Brittle Mode Classification for Micro-End Milling of Nano-FTO Thin Film Using AE Monitoring and CNN

Abstract

This study introduces a real-time acoustic emission (AE) monitoring system for the micro-milling of fluorine-doped tin oxide (FTO) thin films, a critical transparent conductive oxide (TCO) material. The system uses AE sensors to capture high-frequency elastic waves generated during the micro-milling process. We combine experimental and theoretical analyses to investigate how various milling parameters influence the AE signals. To address the crucial challenge of ensuring ductile mode cutting in brittle materials like FTO, we employed a convolutional neural network (CNN) to identify the transition between ductile and brittle machining modes. A CNN was trained on energy-based features extracted from the AE signals, achieving a classification accuracy of 97.37%. This high accuracy demonstrates the effectiveness of integrating AE sensing with deep learning for interpreting complex micro-machining data. The results confirm that this combined approach offers a powerful, non-destructive, and intelligent monitoring solution for improving process control and understanding in the micro-milling of fragile conductive thin films.

1. Introduction

Transparent conductive oxides (TCOs) are a class of materials that are both electrically conductive and optically transparent, making them essential for next-generation transparent electrodes. They are widely used in devices such as touch screens, liquid crystal displays, and solar cells to enhance performance. TCOs can be broadly categorized into metallic thin films and oxide semiconductors. Metallic films have limited use in transparent electronics due to their low optical transparency and poor environmental durability. In contrast, oxide semiconductors, such as indium tin oxide (ITO), offer superior transparency, mechanical robustness, and chemical stability, making them ideal for flexible and transparent electronic applications [1]. While ITO is a popular choice, its performance degrades at temperatures above 500 °C, leading to increased resistance and a reduction in electrical properties, as well as heat and chemical resistance [2]. To overcome these limitations, fluorine-doped tin oxide (FTO) has been developed as a cost-effective alternative. FTO offers exceptional resistance to corrosion, chemicals, and thermal instability, maintaining a constant resistivity at temperatures up to 600 °C [3].

Conventional FTO patterning techniques, such as photolithography, are often complex and expensive. Laser-based methods have been explored as a simpler alternative, but they introduce new challenges, including laser-induced heat, which can compromise machining precision and surface quality, and a lack of control over machining depth [4,5]. Therefore, there is a critical need for developing precise and controllable micro/nanoscale processing techniques for brittle materials like FTO.

Micro-end milling is a high-precision technology used to process thin films like TCO. It can create complex or freeform three-dimensional microstructures, such as micro-grooves, micro-channels, and micro-dimples. By operating within suitable cutting ranges, particularly the ductile mode, micro-end milling can precisely remove material to create desired patterns on a microscopic scale while preserving the integrity of fine and thin materials like TCO [6].

Brittle materials, including oxide films, undergo a transition from brittle fracture to ductile material removal when the uncut chip thickness falls below a critical threshold. This phenomenon, known as the brittle-to-ductile transition (BDT), is a critical aspect of micro-machining. Accurately controlling machining parameters to stay within the ductile regime is essential for achieving high precision without causing cracks or chipping. Recent research has focused on understanding and expanding the BDT window. For instance, Zhang et al. [7] investigated the factors influencing BDT under high-temperature and high-speed conditions, while Lin et al. [8] analyzed the effect of ultrasonic elliptical vibration-assisted cutting on the BDT range of zirconia ceramics. Other studies, such as those by Qiu et al. [9] and Wang et al. [10], have explored critical transition points in SiC ceramic milling and nanomachining of single-crystal silicon, respectively. Yang et al. [11] proposed an ultra-precision diamond cutting process for potassium dihydrogen phosphate to improve BDT depth and machining efficiency.

Given the importance of BDT in micromachining, there is an increasing demand for monitoring processes that involve extremely small uncut chip thicknesses. Acoustic emission (AE) sensors have emerged as an effective tool for this purpose. AE sensing enables real-time monitoring and classification of machining states, making it a highly suitable technology for delicate micro-scale applications. Previous research has explored the brittle–ductile transition (BDT) in micromachining. For example, Shimada et al. [12] investigated this phenomenon in micro-milling and micro-turning of single-crystal silicon and LiNbO3. Concurrently, Cheng and colleagues [13] developed theoretical models for the micro-milling and micro-turning of brittle materials. Other studies have focused on classifying machining modes. Arif et al. [14] categorized processing modes in glass micro-milling as ductile, brittle, and plowing, while Lee [15] used an AFM tip with an acoustic emission (AE) monitoring setup to analyze machining modes in silicon as a function of scratch depth, confirming that micro-cutting theory is applicable at the nanoscale. Mianet et al. [16] utilized AE signals to determine the minimum chip thickness in micro-milling. These studies highlight that the critical depth of cut, which dictates the machining state of brittle materials, can be effectively estimated using AE signal parameters, demonstrating that AE signal acquisition is a reliable method for analyzing micromachining processes.

Building on these findings, this study employs AE sensor monitoring to process nanoscale thin films of brittle materials, specifically TCO, in ductile mode. We designed an end-milling process for FTO thin films integrated with an AE monitoring system. By examining the AE signals generated in different machining modes using various AE parameters, we aimed to improve process precision. To accurately and efficiently identify signal differences based on the processing mode, we introduced a convolutional neural network (CNN). The CNN was used to distinguish between ductile and brittle modes, leveraging its ability to automatically extract features and achieve high classification performance. This approach enables a more robust and intelligent interpretation of AE data during the micromachining of brittle thin films.

2. Micro-Cutting Theoretical Background

To calculate the critical feed per tooth, which determines the transition between ductile and brittle machining modes, a theoretical model based on the mechanics of chip formation was employed. Since chip formation is only possible when the undeformed chip thickness exceeds a minimum threshold, this threshold must be theoretically estimated as it cannot be directly measured during machining. The critical feed per tooth thus plays a crucial role as a boundary condition that governs whether material removal occurs in a ductile or brittle manner. A detailed mechanical analysis of cutting force generation and undeformed chip thickness is provided in Appendix A, along with Figure A1, which illustrates the chip formation mechanism considering the geometry of the rounded cutting edge and contact forces. These theoretical considerations form the basis for deriving the critical condition, which is calculated as

$$f_c={\lambda }_c{\left({\displaystyle \frac{H}{E}}\right)}^{{\displaystyle \frac{1}{2}}}{\left({\displaystyle \frac{K_{IC}}{H}}\right)}^2$$

(1)

Table 1. Property of film.

Material	$\mathrm{S}\mathrm{n}{\mathrm{O}}_2:\mathrm{F}$
Resistance (ohm/sq)	12~14
Young’s modulus (GPa)	70
Hardness (GPa)	4.61
Fracture toughness (MPa√m)	0.65
Poisson ratio	0.2
Mechanical strength (MPa)	120
Work function (eV)	4.4~4.7
Transmittance (%)	83.5
$\mathrm{Density} (\mathrm{k}\mathrm{g}/{\mathrm{m}}^3$)	2500

presents the physical properties of the FTO thin film, including hardness, elastic modulus, and fracture toughness. Using these parameters, the critical feed per tooth was calculated to be approximately 44.39 nm, as shown in Equation (1).

3. FTO Thin-Film Micro-Milling Process

3.1. Experiment Setup

The experimental sample comprises an approximately 750 nm thick FTO coating on lime glass. Figure 1 depicts the FTO coating layer. Figure 2 presents the experimental configuration of an AE monitoring system for the micro-milling process. A TOOLI 34P model (Boardtech & David, Incheon, Republic of Korea) was used for CNC milling, and its specifications are displayed in

Table 2. Specification for CNC (TOOLI 34P).

Spindle (rpm)	24,000
Maximum feed speed (mm/min)	5000
Repeatability (mm)	±0.01
Position precision (mm)	±0.005

. The runout of the tool is a crucial factor in micromachining and can significantly impact machining. Hence, the model’s runout was rectified before conducting the experiments [17]. A laser displacement sensor was utilized to correct any errors, and it was confirmed that the precision was less than 5 μm at 24,000 rpm, which was the maximum spindle speed used in the experiment. Furthermore, the thickness of the FTO thin film in FTO-coated glass was 750 nm, which required the use of a nanoscale stage in the z-axis direction to supplement the CNC for precise processing. In this study, we utilized the PI P-611.3S Nano positioner, and the detailed specifications for this model are provided in

Table 3. Specification for nanostage.

Maximum travel range (μm)	120 × 120 × 120
Resolution (nm)	0.2

. For specimen fixation, a dedicated holder was fabricated to secure the FTO-coated glass and to prevent breakage due to external experimental conditions.

Regarding sensor attachment for monitoring the machining process, the optimal attachment position would be on the specimen itself. However, for future process automation considerations, attaching the sensor directly to the specimen would obstruct the tool path. Therefore, the sensor was affixed to the specimen holder instead of the specimen, allowing unobstructed monitoring of the machining process. The micro-milling process employed a two-flute rib end mill supplied by JJTOOLS. A schematic representation of the end mill is provided in Figure 3, while

Table 4. Specifications for micro-milling tools.

	1	2	3
D	1 mm	0.7 mm	0.5 mm
d	4 mm	4 mm	4 mm
L1	1 mm	1 mm	0.4 mm
L2	2 mm	2 mm	1 mm
L	45 mm	45 mm	40 mm

presents the size details of the tool. The end mill utilized a 0.2 μm ultra-fine-grain cemented carbide with an HRc ranging from 50 to 60. The d tolerance was less than 5 μm for a diameter of 0.1 mm. The nose radius of the cutting edge ($r_e$) measured 1~1.5 μm, and it was coated with TiSiN.

3.2. Micro-Milling Results

The variables with the most influence on the processing modes of micro-milling are the chip load per tooth and the depth of the cut [18,19]. Choices can be made regarding the radial and axial depths of cut. Therefore, in this experiment, tool diameter, spindle speed (rpm), and feed rate were chosen as experimental factors for calculating the chip load per tooth, which is related to the radial depth of cut. The chip load per tooth (feed per tooth, $f$) is defined as the linear distance advanced by each cutting edge per revolution. It is determined by the feed rate ($V_c$), spindle speed ($n$), and the number of cutting edges ($z$), and directly affects cutting forces and chip formation.

$$f={\displaystyle \frac{V_c}{n·z}}$$

(2)

The experiment levels were set based on the mechanical properties of FTO glass and referenced research into the micro-milling of soda-lime glass [14,20,21,22].

Table 5. Condition parameters and level.

Factor	Level
Endmill diameter (mm)	0.5	0.7	1.0
Spindle speed (rpm)	12,000	13,500	15,000	16,500	18,000
Feed rate (μm/s)	5	10	15	25	35
Cutting oil	Dry
Axial depth of cut (nm)	200

illustrates the experimental factors and levels utilized in the experiment. To verify the ductile–brittle machining mode theory based on the radial depth of cut in the case of axial depth of cut, Rodrigues’ chip formation mechanism in micro-milling was calculated to minimize the impact on machining depth [23].

$$h_{min}\approx 0.3·r_e$$

(3)

Here, $h_{min}$ is the minimum axial depth of cut. Accordingly, this representative value was applied to calculate the minimum axial depth of cut using Equation (3), where $r_e$ represents the nose radius of the cutting edge measured in the range of 1~1.5 μm. To ensure conservative estimation of the critical depth, the lower bound of 1 μm was adopted for the calculation. Therefore, in this experiment, the axial depth of cut was set to 200 nm, ensuring it remained less than 300 nm. The adjustment of the milling tip-to-specimen distance was initially approximated via CNC machine operation and further refined precisely using a piezoelectric stage. Upon contact between the micro-milling tip and the specimen, hit signals of AE were recorded, establishing this point as the reference for the z-axis adjustment. In this study, machining conditions were established to analyze the machining mode concerning the feed per tooth. Based on Equation (2), which relates the feed rate, spindle speed, and number of tool edges, the feed per tooth for each condition was calculated and listed in

Table 6. Feed per tooth calculated for different machining conditions.

No.	Spindle Speed (rpm)	Feed (μm/s)	Feed/Tooth (mm/tooth)
1	12,000	35	0.000088
2	13,500	35	0.000078
3	15,000	35	0.00007
4	16,500	35	0.000064
5	12,000	25	0.000063
6	18,000	35	0.000058
7	13,500	25	0.000056
8	15,000	25	0.00005
9	16,500	25	0.000045
10	12,000	15	0.000042
11	18,000	25	0.000042
12	13,500	15	0.000037
13	15,000	15	0.000033
14	16,500	15	0.00003
15	18,000	15	0.000028
16	12,000	10	0.000025
17	13,500	10	0.000022
18	15,000	10	0.00002
19	16,500	10	0.000018
20	18,000	10	0.000017
21	12,000	5	0.000013
22	13,500	5	0.000011
23	15,000	5	0.00001
24	16,500	5	0.000009
25	18,000	5	0.000008

.

The experiments were conducted five times for each condition. If cracks and fractures were observed in all five trials, the machining mode was classified as brittle fracture. If brittle fracture was observed in some of the five trials (e.g., one or three times), it was categorized as partial brittle. When stable cutting was achieved in all five trials without visible cracks, the mode was defined as ductile fracture. Figure 4, Figure 5 and Figure 6 present 20× magnified optical microscope images of specimens machined at 18,000 rpm. Brittle fracture occurred under the conditions shown in Figure 4e, Figure 5e, and Figure 6d,e, where cracks were frequently observed along the edges of the machined regions.

Figure 7 shows the surface profiles of the machined areas measured using a surface roughness tester, SJ-310 (Mitutoyo, Japan), and the corresponding surface roughness values for Figure 7a–e were 13, 16, 18, 26, and 62 nm, respectively. Notably, under condition Figure 7e, where brittle fracture occurred, a significant increase in surface roughness was observed.

Figure 8 illustrates the machining modes of FTO thin films using the spindle rotational speed and feed rate at a tool diameter of 0.5 mm. It displays the feed per tooth, indicating the radial depth of cut in each machining mode. The experimental radial depth of cut values ranged between 25 and 56 nm, which are comparable to the theoretical critical radial depth of cut value (44.39 nm).

The diverse occurrence of partial brittleness results from differences between the feed per tooth set by experimental conditions and the actual chip load per tooth influenced by CNC errors and tool deflection during practical experiments. The end mill used in the experiment had a diameter of 0.5 mm, the smallest among those employed. The runout error in the CNC used for this experiment was non-zero, suggesting its considerable impact on the results.

Figure 9 depicts the machining modes of FTO thin films using the spindle rotational speed and feed rate with a tool diameter of 0.7 mm, displaying the chip load per tooth for each machining mode. The experimental critical feed per tooth ranged between 25 nm, where initial partial brittle occurrence was observed, and 45 nm, marking complete brittle failure. Hence, the observed experimental critical feed per tooth for a tool diameter of 0.7 mm ranged between 25 and 45 nm. As the tool diameter increased (which enhanced material removal rates), ductile failure occurred at lower chip loads. Increased tool thickness, or diameter, reduced deflection due to tool stiffness, consequently narrowing the range of critical feed per tooth compared to those observed at a tool diameter of 0.5 mm.

Figure 10 illustrates the machining modes of FTO thin films using spindle rotational speed and feed rate with a tool diameter of 1.0 mm, displaying the feed per tooth for each machining mode. The experimental critical chip load per tooth ranged between 22 nm, where initial partial brittle occurrence was noted, and 28 nm, indicating complete brittle failure. Therefore, the observed experimental critical feed per tooth for a tool diameter of 1.0 mm ranged between 22 and 28 nm. Compared to tool diameters of 0.5 and 0.7 mm, an increase in tool diameter resulted in higher material removal rates, leading to ductile failure at a lower feed per tooth. The stiffness of the tool reduced deflection, decreasing the range of critical feed per tooth compared to tool diameters of 0.5 and 0.7 mm. At the same feed rate and rpm, ductile mode machining was observed at a tool diameter of 0.5 mm, whereas partial brittle mode and complete brittle mode were observed at 0.7 mm and 1.0 mm, respectively. This demonstrates that the critical feed per tooth concerning rpm increases as the tool diameter decreases.

4. AE Monitoring Results and CNN Analysis

4.1. AE Monitoring Signal in Machining Mode

An AE sensor was utilized for monitoring to detect the machining mode during the micro-milling of thin films. The WSα wideband sensor from MISTRAS was used, and data were collected using AE signal analysis software called AEwin^TM (USB version E5.33) from PCA.

Table 7. AE signal data acquisition condition.

Parameter	Value
Threshold	44 dB
Gain	20 dB
Bandpass filter	20 kHz–1 MHz
Sample rate	5 MSPS
Signal length	2048
PDT	200
HDT	800
HLT	1000

presents the data acquisition condition of the AE sensor, and Figure 11 shows the raw AE signals collected during the ductile and brittle machining modes of the FTO micro-milling process under these conditions.

Appendix B compares the RMS and FFT results of AE signals to analyze the characteristics of different machining modes. Figure A2 presents the RMS and FFT analyses of AE signals acquired at feed rates of 5 μm/s and 35 μm/s for ductile and brittle modes, respectively. In brittle machining, the RMS values exhibited irregular distribution due to fracture-related signals, whereas in ductile machining, the RMS distribution remained relatively stable. However, it was challenging to clearly distinguish between the two modes based solely on these signal differences. Similarly, although slight amplitude differences were observed in the high-frequency range of the FFT results, the frequency patterns and magnitudes were not distinct enough to reliably differentiate the machining modes.

Micro-milling continuously removes material, and the above experimental results show that more material is removed in brittle mode than in ductile mode. Consequently, the energy increase rate will be higher for brittle processing than for ductile processing as time increases. The AE power is related to the material removal rate ($MRR$) during machining [24]:

$$A{E}_{power} = a\times MRR = a\times nfZ{a}_p{a}_e$$

(4)

In Equation (4), a is a constant affected by parameters such as tool conditions, geometric shape, and material properties; $n$ is the spindle speed; $f$ is the feed per tooth; $Z$ is the number of teeth; $a_p$ is the axial depth of cut; and $a_e$ is the radial depth of cut. The feed rate of the tool can be calculated as $nfZ$. AE power is known to be equivalent to the square of the RMS value of the AE signal [25]:

$$A{E}_{power} = {\left(A{E}_{RMS}\right)}^2 = {\displaystyle \frac{1}{∆T}}{\int }_0^{∆T}V{\left(t\right)}^{2}dt$$

(5)

In Equation (5), $V\left(t\right)$ represents the voltage of the AE signal as a function of time, and $∆T$ denotes the duration of the signal segment used for the calculation. As the feed per tooth of the tool and the feed rate increase, the material removal rate increases. This shows that the amplitude of AE energy value increases as the material removal rate increases. AE energy is expressed as in Equation (6) [26].

$$A{E}_{Energy}={\int }_0^{∆T}V{\left(t\right)}^{2}dt$$

(6)

AE energy is mathematically expressed as in Equation (7), which indicates that it is directly proportional to the material removal rate ($MRR$). This relationship suggests that AE energy can serve as an effective indicator of the amount of material removed during micro-milling. Figure 12 displays the AE energy of the signal acquired through the AE sensor over time.

$$A{E}_{Energy}=∆T\times a\times MRR$$

(7)

The AE signals obtained by the AE sensor demonstrate an energy difference when machining in the ductile regime versus machining in the brittle region. The ductile region possesses significantly less energy than the brittle region. AE energy images can be used to distinguish processing modes of FTO thin films based on the energy contrast of ductile and brittle regions. The signal selected for comparison is located at the extremes of the ductile and brittle regions. If the difference is small, it should be checked whether it can be determined by AE energy. Therefore, to classify the processing mode, a CNN was employed using the AE energy graph as input data.

4.2. Machining Mode Classification Using an AE Signal-Based CNN

The use of artificial intelligence in automating the distinction of ductile and brittle machining modes is important. Therefore, a CNN algorithm was developed to classify the machining modes. The CNN was selected for its ability to classify images, especially those with intricate patterns and features. Although AE signals are not traditional images, they can be processed as image data due to their complex and rich patterns. These signals, when transformed into spectrograms or other visual representations, exhibit patterns and nuances similar to those of images. Also, CNNs are utilized to automatically learn and distinguish complex features and patterns within acoustic signals. Furthermore, this model offers computational advantages when processing image-like data. Convolution and pooling operations extract crucial information while simultaneously minimizing overfitting, reducing computational load. These properties are particularly useful when processing large amounts of data generated by acoustic emission sensors during machining. Therefore, the use of a CNN, a deep learning model utilized for image classification, involves constructing a neural network by extracting features from images through convolutional and pooling layers and adds non-linearity through activation functions [27,28,29,30]. The convolutional layers extract image features, while the pooling layers decrease resolution to minimize computations and prevent overfitting. This structure aims to minimize the loss function by setting optimal parameters, processing images as 3D tensors, and reducing them to vectors for the final output. CNN demonstrates exceptional performance in image processing. Converting these signals into images for classification is more efficient for rapid processing than classifying raw data, especially with AE signals, which often contain a sizable quantity of data. Figure 13 depicts the training procedure of signal images using a CNN algorithm that is composed of three convolutional layers.

To prevent potential confusion caused by graph axes, the AE energy data images were preprocessed by removing all axis labels. In addition, to improve training efficiency, all images were converted to grayscale and resized to 360 × 360 pixels before being labeled as either ductile or brittle for CNN training. As shown in Figure 14, the CNN architecture consists of two convolution–pooling blocks. Each block includes batch normalization and a ReLU activation function, which work together to extract key features while effectively reducing the spatial dimensions.

The extracted features are then passed through a fully connected layer, a softmax layer, and an output layer to classify the machining mode as either ductile or brittle.

Table 8. Hyperparameters of CNN.

Solver	ADAM Optimizer
Learning rate	0.001
Mini-batch size	128
Epoch	30

displays the hyperparameters utilized in the CNN algorithm. The ADAM optimizer was chosen to prevent overfitting, as opposed to gradient descent, which tends to cause overfitting. Figure 15 presents the validation accuracy of the CNN algorithm designed in this study. For analysis, 70% of the total data were used for training, while the remaining 30% were used for validation. The total training time took 37 min 48 s, achieving a validation accuracy of 97.37%. This demonstrates the ability of the CNN algorithm to differentiate between ductile and brittle regions effectively.

5. Conclusions

This study proposes a new method that combines AE signals and CNNs to accurately classify ductile and brittle machining modes in the micro-milling of FTO thin films. The purpose of this study is to investigate the influence of tool diameter, feed rate, and spindle speed during the micro-milling of FTO thin films and identify ductile and brittle regions by analyzing the AE signals collected during the process. The main conclusions are the following:

• The experimental result that most closely matched the theoretical critical feed per tooth value of 44.39 nm was observed with the 0.7 mm diameter tool, where a clear transition to brittle fracture occurred at a feed per tooth of 45 nm.
• Determining the boundaries where partial brittle modes appear depends on tool diameter, feed, and spindle speed. However, even if the diameter is increased, ductile processing is possible if the feed rate is as low as 5 μm/s.
• The parameter that stands out the most in distinguishing between modes is AE energy. There are also some differences in AE RMS and frequency. However, it is difficult to distinguish between brittle and ductile modes at the boundary of partial brittleness.
• A high classification accuracy of 97.37% was achieved using the CNN algorithm, confirming its effectiveness in analyzing AE signals for distinguishing between ductile and brittle machining modes.

The effectiveness of AE monitoring for accurately classifying brittle and ductile modes in micro-milling was demonstrated with an accuracy of 97.37%. The proposed machining-mode classification model, which achieves a high accuracy in distinguishing between ductile and brittle regimes, holds strong potential for practical application in the selective patterning of FTO thin films. This includes manufacturing processes such as electrode fabrication in solar cells, microchannel formation in microfluidic chips, and optical component production in micro-displays. By enabling the real-time monitoring of AE signals during micro-milling, the system can minimize damage to the FTO layer while maintaining the desired product quality. Furthermore, this real-time process monitoring approach can be integrated with a closed-loop feedback control system, allowing dynamic adjustment of machining parameters to optimize the surface roughness and dimensional accuracy of the thin films. Although micro-milling currently lags behind conventional photolithography and etching techniques in terms of absolute precision, it is expected to become increasingly viable as a controllable and flexible process. With continued advancements in tool accuracy and motion control systems, micro-milling is anticipated to offer a robust alternative for high-precision fabrication in next-generation microdevice manufacturing.