Method and device for on-line sensing of laser processing of semiconductors under multi-field coupling of photothermal forces

Abstract

This invention relates to an online sensing method and apparatus for laser processing of semiconductors under photothermal multi-field coupling. The online sensing method includes: acquiring melt depth data and molten pool or processing groove cross-sectional morphology of the processing area via super-resolution OCT; acquiring temperature distribution, temperature fluctuations, and spatter information of the processing area via thermal imaging; and performing time synchronization and spatial registration of OCT data, thermal imaging data, spatter events, and process parameters to form multi-dimensional online observation data of position, time, melt depth, temperature, and spatter. Based on the multi-dimensional online observation data, a photothermal multi-field coupling state characteristic is constructed to identify under-melting, over-melting, penetration risk, spatter anomalies, thermal cracking risk, collapse risk, and processing groove morphology anomalies, and to output defect risk levels and feedback control suggestions. This invention achieves synchronous online sensing of melt depth, temperature field, and spatter behavior in semiconductor laser processing, improving the accuracy of processing state identification and defect early warning.

Description

Technical Field

[0001] This invention belongs to the field of semiconductor laser processing and intelligent manufacturing process sensing technology, and particularly relates to a method and device for realizing online defect identification and closed-loop feedback control through the fusion sensing of super-resolution optical coherence tomography (OCT) and infrared thermal imaging under the coupling effect of multiple physical fields of light, heat and force. Background Technology

[0002] In semiconductor manufacturing processes, precision laser processing applications such as wafer dicing, laser grooving, thin film removal, laser drilling, localized annealing, and packaging material removal are extremely widespread. Laser energy, as a high-density heat source, acts on a localized area of ​​the material for an extremely short time, triggering a series of complex transient processes on the surface and subsurface, including rapid absorption, heating, melting, vaporization, resolidification, and thermal stress release. The combined effect of these processes directly determines the final processing quality, such as grooving depth, melt penetration consistency, heat-affected zone size, and the presence of microcracks and recast layers.

[0003] Currently, the mainstream quality assurance methods in the industry still heavily rely on "post-processing inspection," which involves offline sampling inspection of processed samples using microscopic sections, optical profilometers, or scanning electron microscopes. This approach has significant drawbacks: firstly, information feedback is severely delayed, and once a defect is discovered, a large batch of scrap may have already been generated; secondly, it is impossible to observe the transient evolution during processing, especially the dynamic behavior inside the molten pool (such as changes in melt depth, keyhole morphology, and spatter generation), leading to a reliance on a "trial and error" approach for process optimization, resulting in long cycles and high costs.

[0004] Some advanced solutions attempt to introduce online detection technologies, such as using high-speed cameras to monitor the surface brightness of the molten pool or using infrared thermal imagers to acquire the temperature field of the processing area. However, the former can only reflect two-dimensional surface information and cannot penetrate the molten pool or plasma plume for depth measurement; while the latter can provide thermal field distribution, its measurement results are greatly affected by changes in material emissivity, plasma obstruction, and observation angle, and cannot directly infer the decisive geometric quality indicator—melt depth—from temperature. For example, when thermal imaging shows a low temperature, it may mean insufficient energy input, but it may also mean that focus drift causes energy dissipation. The two have completely different effects on melt depth, and temperature information alone cannot effectively distinguish between them.

[0005] Therefore, existing technologies have long faced a core bottleneck: the lack of a method and device that can directly and online observe the internal geometry of the molten pool (especially the melt depth) and synchronously fuse it with external thermal fields and spatter behavior in a spatiotemporal manner. This results in the inability to establish a multi-field coupled quantitative relationship between "optical field input - thermal field response - molten pool geometry - mechanical disturbance," making it difficult to achieve accurate online prediction and closed-loop control of processing status and defect risks. An online sensing scheme that unifies direct observation of melt depth, online measurement of temperature field, and identification of spatter behavior is needed, transforming the laser processing process from "post-processing inspection" to "direct observation and real-time judgment during processing." Summary of the Invention

[0006] This invention aims to solve the aforementioned technical problems by proposing an online sensing method and device for laser-processed semiconductors under photothermal-mechanical multi-field coupling. The core inventive concept is as follows: using "super-resolution OCT direct observation of melt depth" as an endoscope reflecting the geometric quality of the processing, and "simultaneous observation of thermal imaging temperature field / splash" as a panoramic window reflecting thermal behavior and mechanical disturbances. Through "multi-source spatiotemporal registration," these two methods, along with process parameters, are integrated into a unified multi-dimensional data stream. Based on the "photothermal-mechanical multi-field coupling state recognition" model, transparent perception and intelligent control of the processing process are achieved.

[0007] The technical solution of this invention is the online sensing method for laser processing semiconductors under photothermal multi-field coupling, which is characterized by including the following steps: (1) Obtaining the process parameters P(t) during laser processing: Collect process parameters such as laser power, pulse width, repetition frequency, scanning speed, spot diameter, focal position, processing trajectory, material type, processing layer thickness, protective gas flow rate, and motion platform status for the current processing task, and form a process parameter sequence P(t); (2) Using a super-resolution optical coherence tomography (OCT) observation module, tomographic signals of the processing area are acquired in real time, and first data characterizing the internal geometry of the processing area is generated based on a super-resolution reconstruction algorithm. The first data includes at least the real-time melt depth H(t) and the cross-sectional morphology of the processing groove. An OCT observation module is set up near the laser processing head to maintain a stable geometric relationship between the OCT probe light and the laser processing area. The OCT module acquires the tomographic interference signal of the processing area in real time, and obtains the internal cross-sectional information of the processing area through noise reduction, phase compensation, tomographic reconstruction and super-resolution reconstruction. The extractable geometric features include the current melt depth H(t), melt depth change rate dH / dt, melt pool cross-sectional width, bottom morphology of the processing tank, tank wall continuity, boundary roughness, melt depth stability and whether anomalies such as under-melting, over-melting, penetration or local collapse have occurred. (3) Using an infrared thermal imaging observation module, thermal radiation images of the processing area are acquired in real time, and second data characterizing the thermal behavior and mechanical disturbances of the processing area are generated. The second data includes at least the temperature field distribution T(x,y,t), temperature fluctuation characteristics, and splash event information S(t). The system acquires thermal radiation images of the processing area using an off-axis high-speed infrared thermal imaging module, and obtains the surface temperature distribution T(x,y,t) of the molten pool by combining calibration parameters. It processes consecutive frames of thermal images to extract features such as maximum temperature, average temperature, temperature gradient, width of the heat-affected zone, temperature fluctuation amplitude, temperature fluctuation frequency, and cooling rate. For splash events, the system identifies short-lived, high-temperature small targets in the thermal images that are significantly brighter or hotter than the background, located outside the molten pool body, and exhibit high-speed displacement characteristics. It records their number, temperature, velocity, trajectory direction, duration, and frequency of occurrence. (4) The first data, the second data, and the process parameter P(t) are synchronized in time and registered in space to form multi-dimensional online observation data mapped to a unified processing trajectory coordinate system: D(t,x,y)={H(t),T(x,y,t), G_T(t), S(t), P(t)}, In the formula, H(t) represents the melting depth, T(x,y,t) represents the temperature field, G_T(t) represents the temperature gradient or temperature fluctuation characteristics, S(t) represents the splashing state, and P(t) represents the process parameters; Using the target melting depth H0 and the allowable deviation range [Hmin, Hmax] as benchmarks, if H(t) is lower than Hmin for multiple consecutive sampling periods, it is judged as insufficient melting depth; if H(t) is higher than Hmax or dH / dt exceeds the preset upper limit, it is judged as over-melting or penetration risk. Using the highest temperature, average temperature, temperature gradient, heat-affected zone range, and temperature fluctuation amplitude as inputs, the system determines whether the local heat input is stable. If the temperature gradient is too large and the cooling rate is too high, the risk of thermal stress cracking is determined to be increased. If the temperature field continues to expand, the risk of heat accumulation or thermal damage is determined to be increased. The intensity of molten pool disturbance is judged by the number of splashes, splash speed, splash temperature and splash frequency. If the number of splash events increases significantly per unit time and occurs simultaneously with temperature fluctuations and sudden changes in melt depth, the processing is considered unstable. (5) Based on the aforementioned multidimensional online observation data, a multi-field coupled state feature set of photothermal and mechanical fields is constructed. This feature set includes light field input features, thermal field response features, geometric field features, and mechanical disturbance-related features. Based on the registered multi-source data, the coupling state characteristics are constructed as follows: optical field input characteristics include laser power, energy density, spot size, scanning speed, focus offset, and energy input per unit length; thermal field response characteristics include maximum temperature, average temperature, temperature gradient, heat-affected zone range, cooling rate, and heat accumulation degree; geometric field characteristics include melt depth, melt width, melt depth / melt width ratio, bottom flatness, and cross-sectional profile continuity; mechanical disturbance-related characteristics include spatter intensity, temperature oscillation, molten pool surface disturbance, abrupt melt depth changes, and local collapse tendency. (6) Based on the aforementioned photothermal-mechanical multi-field coupling state feature set, the current processing state and defect risk are identified online. The system uses a combination of rule-based thresholds, statistical discrimination models, machine learning models, or physical models and data-driven models for judgment. When the OCT melting depth is consistently below the target range and the thermal imaging temperature is insufficient, it is judged as a risk of under-melting or insufficient melting depth. When the melting depth increases rapidly and the local temperature rises significantly, it is judged as a risk of over-melting, penetration or thermal damage; When the temperature field exhibits high-frequency fluctuations and is accompanied by an increase in splashing events, it is judged as an unstable molten pool or abnormal splashing. When the penetration depth is basically normal but the temperature gradient is too large and the cooling rate is too high, it is judged as a risk of thermal stress cracking. When there is a sudden change in melting depth, discontinuous bottom morphology, and increased spatter intensity, it is judged as an abnormal morphology of the processing tank or a risk of local collapse, and the risk level and control suggestions are output.

[0008] As a preferred option: step (2) of generating the first data based on the super-resolution reconstruction algorithm specifically includes: (2.1) The acquired raw tomographic interferometric signals are preprocessed by denoising and phase compensation; (2.2) Perform tomographic reconstruction on the preprocessed signal to obtain the initial cross-sectional image; (2.3) The initial cross-sectional image is processed by using a super-resolution reconstruction algorithm based on deconvolution or deep learning model to break through the optical diffraction limit of the system and obtain the internal cross-sectional morphology of the processing area with submicron resolution. (2.4) From the reconstructed cross-sectional morphology, the real-time melt depth H(t), the width of the melt pool cross-section, the flatness of the bottom of the pool, and the geometric features of the boundary continuity are accurately extracted; Step (3) involves generating the second data, specifically including: (3.1) Based on the calibration parameters, the thermal radiation image is inverted into a two-dimensional temperature field distribution T(x,y,t) of the molten pool and the heat-affected zone; (3.2) Perform pixel-by-pixel temporal analysis on the temperature field distribution of multiple consecutive frames, and extract the features of the highest temperature, average temperature, temperature gradient, width of the heat-affected zone, temperature fluctuation amplitude and cooling rate to form the temperature fluctuation features. (3.3) Identify small targets that are short-lived, high-temperature and high-speed moving in the thermal image sequence that meet the preset spatiotemporal conditions, mark them as splash events, and record the number, temperature, speed, trajectory direction and frequency of occurrence of the splash events to form the splash event information S(t).

[0009] As a preferred option: step (4) involves time synchronization and spatial registration, specifically including: (4.1) Establish the spatial transformation mapping relationship between the OCT coordinate system, the thermal imaging coordinate system, the laser processing coordinate system, and the motion platform coordinate system; (4.2) By using hardware trigger signals and unified timestamps, the data acquisition times of the OCT module and the thermal imaging module are precisely aligned; (4.3) Using the position information fed back in real time by the motion platform encoder and the predefined processing trajectory information, the first data and the second data at each moment are accurately allocated to the actual processing trajectory coordinate points according to the spatial transformation mapping relationship; (4.4) Form structured multidimensional online observation data with "location-time-melt depth-temperature-splash-process parameters"; Step (5) involves constructing a multi-field coupled state feature set of photothermal and mechanical fields, specifically including: (5.1) The optical field input characteristics include at least the laser power, energy density, scanning speed, and focus offset; (5.2) The thermal field response characteristics include at least the highest temperature, average temperature, temperature gradient, range of the heat-affected zone, and cooling rate; (5.3) The geometric field characteristics include at least the real-time melt depth H(t), melt depth change rate dH / dt, melt depth / melt width ratio and bottom flatness; (5.4) The mechanical disturbance-related characteristics include at least the spatter intensity, temperature oscillation amplitude, frequency of sudden changes in melt depth, and tendency for local collapse; Step 6 involves online identification of the current processing status and defect risks, specifically including: (6.1) Input the photothermal multi-field coupling state feature set into a preset recognition model; (6.2) When the real-time melting depth H(t) is continuously lower than the lower limit of the preset target melting depth range and the average temperature is lower than the preset temperature threshold, it is determined to be a risk of undermelting or insufficient melting depth; (6.3) When the melting depth change rate dH / dt exceeds the preset positive change rate threshold and the highest temperature exceeds the preset high temperature threshold, it is determined to be a risk of over-melting or penetration; (6.4) When the temperature fluctuation amplitude exceeds the preset fluctuation threshold and the frequency of the splashing event exceeds the preset frequency threshold, it is determined to be an unstable molten pool or an abnormal splashing risk. (6.5) When the melting depth is basically normal but the temperature gradient and the cooling rate both exceed the preset range, it is determined to be a risk of thermal stress-induced cracking.

[0010] As a preferred option: the identification model is based on a weighted scoring mechanism, which quantifies defect risk through a comprehensive risk score R, and its calculation formula is as follows: R = w1·RH + w2·RT + w3·RS + w4·RP In the formula, RH represents the melt depth anomaly score, RT represents the temperature anomaly score, RS represents the splashing anomaly score, RP represents the process parameter deviation score, and w1 to w4 are weighting coefficients. Based on the magnitude of R, four risk levels are output: normal, attention, warning, and shutdown. w1, w2, w3, and w4 are weighting coefficients, and they satisfy w1+w2+w3+w4=1. When R exceeds the first threshold, a "warning" signal is output. When R exceeds the second threshold, a "stop and adjust immediately" signal is output.

[0011] As a preferred option: the identification model in step (6.1) is based on a weighted scoring mechanism, which quantifies defect risk through a comprehensive risk score R, and its calculation formula is as follows: R = w1·RH + w2·RT + w3·RS + w4·RP In the formula, RH represents the melt depth anomaly score, RT represents the temperature anomaly score, RS represents the splashing anomaly score, RP represents the process parameter deviation score, and w1 to w4 are weighting coefficients. Based on the magnitude of R, four risk levels are output: normal, attention, warning, and shutdown. w1, w2, w3, and w4 are weighting coefficients, and they satisfy w1+w2+w3+w4=1. When R exceeds the first threshold, a "warning" signal is output. When R exceeds the second threshold, a "stop and adjust immediately" signal is output.

[0012] Preferably, after the step of outputting the risk level and control recommendations, step (7) is also included: (7.1) When the defect risk level exceeds the preset intervention threshold, a feedback control command is generated and sent to the laser processing control unit; (7.2) The feedback control command is used to perform at least one of the following operations: adjust the laser output power, change the scanning speed, correct the focal position, adjust the pulse frequency, adjust the protective gas flow rate, or perform dynamic compensation on the processing path.

[0013] Another technical solution of the present invention is the online sensing device for laser processing semiconductors under photothermal multi-field coupling, used to implement any of the aforementioned online sensing methods for laser processing semiconductors under photothermal multi-field coupling, characterized in that the device comprises: A laser processing module that generates processing lasers to process semiconductor materials; The laser processing module includes a laser, a beam shaping unit, a scanning galvanometer or motion platform, a focusing lens, and a protective gas unit, and is used to process semiconductor materials; The super-resolution OCT observation module has its detection optical path arranged coaxially with the processing laser to acquire first data inside the processing area. The first data includes at least the real-time melt depth and the cross-sectional morphology of the processing groove. The super-resolution OCT observation module includes an OCT light source, a reference arm, a sample arm, an interference signal acquisition unit, a scanning unit, and a super-resolution reconstruction unit, used to acquire melt depth, melt pool cross-sectional profile, and machining groove geometry; An infrared thermal imaging observation module is tilted off-axis, with its field of view covering the processing area, and is used to acquire second data of the processing area. The second data includes at least temperature field distribution, temperature fluctuation and splash event information. The infrared thermal imaging observation module includes a high-speed infrared camera, an infrared lens, a filter, a blackbody calibration unit, a temperature inversion unit, and a splash identification unit, used to acquire temperature field, temperature fluctuation, and splash information. The spatiotemporal synchronization module is connected to the controllers of the super-resolution OCT observation module, the infrared thermal imaging observation module, and the laser processing module, respectively, and is used to precisely align the first data and the second data with the processing position and processing timestamp. The spatiotemporal synchronization module includes a hardware triggering unit, a timestamp synchronization unit, an encoder acquisition unit, and a coordinate transformation unit, which are used to realize the fusion of multi-source data under the same time axis and coordinate system; The multi-source fusion analysis module receives multi-dimensional online observation data from the spatiotemporal synchronization module and is configured to construct a set of optical-thermal-mechanical multi-field coupled state features. The multi-source fusion analysis module is used to fuse OCT melt depth, thermal imaging temperature, spatter events and process parameters to form photothermal-mechanical multi-field coupling state characteristics; The online defect identification module is configured to identify the processing status and calculate the defect risk level based on the photothermal multi-field coupling state feature set. The feedback control module is configured to output control commands to the laser processing module based on the defect risk level.

[0014] The feedback control module is used to output adjustment suggestions or control commands for laser power, scanning speed, focal position, pulse parameters and protective gas parameters based on the recognition results.

[0015] Preferably, the super-resolution OCT observation module is a spectral domain OCT or a swept frequency OCT, which integrates a super-resolution reconstruction unit. The super-resolution reconstruction unit breaks through the axial and lateral resolution limits of the OCT system by executing a deconvolution algorithm or calling a pre-trained deep learning model.

[0016] Preferably, the infrared thermal imaging observation module integrates a splash recognition unit. This unit identifies short-lived clusters of high-temperature points in the sequence of images that have a temperature more than three times the average temperature of the outer periphery of the molten pool and have a directional motion trajectory, using frame difference or optical flow methods, and marks them as splash events.

[0017] Preferably, the multi-source fusion analysis module and the online defect identification module are integrated in the same industrial computer or edge computing node, and can process data in real time at a speed faster than the processing cycle, ensuring that the identification results are output within a preset time window after the first and second data are generated.

[0018] Preferably, the spatiotemporal synchronization module receives encoder pulse signals from the motion platform and uses these signals as a reference to synchronously trigger the acquisition actions of the super-resolution OCT observation module and the infrared thermal imaging observation module, thereby achieving registration based on equal spatial spacing sampling.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) To address the problem that "the geometry of the melt depth cannot be directly observed online," this invention employs a super-resolution OCT module coaxially integrated with the processing laser as a technical means. Its technical advantage is that it directly penetrates the plasma plume and the molten pool to acquire sub-micron resolution cross-sectional tomographic images of the processing area, thereby accurately and in real-time measuring the melt depth, melt width, and bottom morphology, transforming the invisible internal processing quality into directly quantifiable process data.

[0020] (2) To address the problem that "a single thermal imaging sensor cannot comprehensively characterize the processing state," this invention employs a high-speed infrared thermal imaging module and uses a proprietary algorithm to simultaneously extract dual information from the thermal image: first, the two-dimensional temperature field and temperature fluctuations of the molten pool surface and heat-affected zone; and second, the quantitative characteristics (quantity, velocity, temperature, etc.) of spatter behavior, which is usually considered interference. The technical effect is that it elevates the temperature field from a simple surface measurement to a thermal response index that can characterize heat accumulation and cooling rate, while simultaneously elevating spatter behavior from a phenomenon observation to a quantitative index that can characterize the intensity of mechanical disturbances within the molten pool, greatly enriching the dimensions of perception.

[0021] (3) To address the problem of inaccurate defect identification due to the lack of effective fusion of multi-source heterogeneous sensor information, this invention employs a spatiotemporal synchronization module combining hardware triggering and coordinate mapping, as well as an identification method based on a multi-field coupling model. Its technical effect is that it is the first to construct a unified state feature set including optical field input, thermal field response, geometric field (molten pool), and force field (splash, oscillation) disturbances. This enables the system to, like an experienced expert, comprehensively "see" (melt depth) and "feel" (temperature, splash) information to accurately classify and predict the causes of various complex defects such as under-melting, over-melting, penetration, and hot cracking, significantly improving the accuracy and robustness of defect identification and early warning.

[0022] (4) The three technical means mentioned in (1) to (3) above are not simply superimposed, but have a profound synergistic effect, together constituting a systematic technical solution. If only OCT is used without thermal imaging, the risk of cracks caused by thermal stress cannot be assessed; if only thermal imaging is used without OCT, it is impossible to confirm whether the melt depth meets the standard, and the root cause of abnormal spatter (whether it is excessive energy or focus shift) cannot be determined. When the two are precisely registered in time and space, a systemic effect of "one plus one is greater than two" is generated: for example, if OCT detects a sudden decrease in melt depth, and at the same time thermal imaging detects a sudden increase in the surface temperature of the molten pool and a sudden increase in spatter, the system can immediately determine that it is a "keyhole collapse type" defect, rather than a normal energy decrease, thereby triggering the most appropriate control strategy (such as adjusting the focus position rather than simply increasing the power). It is by constructing and analyzing the dynamic correlation of different sensor characteristics in time and space that this invention has achieved a leap from single-dimensional "monitoring" to multi-dimensional "understanding".

[0023] (5) For the first time, online, direct, and high-precision observation of melt depth in semiconductor laser processing was achieved through coaxial super-resolution OCT, changing the status quo of relying on indirect inference from temperature or post-processing detection.

[0024] ⑹ By performing rigorous spatiotemporal registration and multi-field coupling analysis on the melt depth and cross-sectional geometry information obtained by OCT and the temperature field and splash dynamics information obtained by thermal imaging, a complete "process fingerprint" from energy input, thermal response, geometric shaping to mechanical disturbance is formed, which greatly improves the accuracy and completeness of processing status identification.

[0025] (7) It can identify and distinguish various defects such as under-melting, over-melting, hot cracking, and spattering abnormalities online, and differentiate the risk levels. This provides a direct and quantitative basis for the adaptive closed-loop control of the process, significantly reducing the reliance on manual sampling and improving the yield and efficiency of semiconductor precision processing. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the overall system structure of the present invention; Figure 2This is a flowchart of the online sensing method of the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and two specific embodiments.

[0028] Example 1: Laser Grooving of Semiconductor Wafers This embodiment describes the application of the present invention to the laser grooving scenario of low-k dielectric wafers to control the uniformity of groove depth and prevent damage to the underlying structure.

[0029] Reference Figure 1 and Figure 2 The device described in this invention is constructed as follows: The laser is a 355nm ultraviolet nanosecond laser, and the super-resolution OCT module uses a spectral domain OCT with a center wavelength of 850nm. Its probe light is coupled to the processing laser through a dichroic mirror to form a coaxial optical path, and the focused spot diameter is about 10μm, slightly smaller than the processing spot (about 15μm). The high-speed infrared thermal imaging camera (InSb detector, response band 3-5μm) is installed off-axis, and the acquisition frame rate is 1kHz.

[0030] During processing, the motion platform moves the wafer along an "S" shaped trajectory, and the laser operates in pulse mode, with a target groove depth of H0 = 23.0 μm ± 1.5 μm. The spatiotemporal synchronization module receives the encoder signal from the platform and uses it as a reference to send trigger pulses to the OCT and thermal imaging cameras at equal spatial intervals, ensuring that each acquisition is accurately located on the actual processing trajectory point, forming spatial registration; at the same time, it records the absolute timestamp of each pulse to achieve time synchronization.

[0031] After processing begins, the OCT module acquires the interferometric signal from the bottom of the tank in real time at a line scan rate of 100 kHz. This raw data is first subjected to spectral shaping and numerical dispersion compensation for phase correction, and then tomographic reconstruction is performed using Fourier transform. Next, the super-resolution reconstruction unit employs the Lucy-Richardson deconvolution algorithm, iteratively deconvolving the initial tomographic image with a pre-measured system point spread function (PSF), effectively sharpening image edges. From the reconstructed clear image, the air-material interface and the reflection peak at the bottom of the tank are automatically identified; the distance between these two points is the real-time melt depth H(t).

[0032] Meanwhile, the thermal imaging camera, based on a blackbody radiation calibration model, inverts the grayscale values ​​of each frame into a two-dimensional temperature field T(x,y,t). The system calculates the average temperature T_avg within a 1μm wide annular region at the edge of the slot in real time as the temperature of the heat-affected zone. The splash recognition unit extracts fast-moving bright spot targets using the frame difference method, filters out point noise with an area smaller than 3 pixels, and identifies targets with a duration of less than 5 frames and a temperature higher than T_avg+500℃ as splash events, recording their splash velocity vector.

[0033] When processing a corner region of the wafer, the acceleration and deceleration of the motion platform caused a localized increase in energy density. The system detected the following synchronization anomalies in multi-dimensional data: First data (OCT): The real-time melt depth H(t) jumped rapidly from 23.1 μm to 26.8 μm, exceeding the preset upper limit H_max=24.5 μm, and the melt depth change rate dH / dt increased sharply.

[0034] Second data (thermal imaging): The highest temperature T_max jumped from 1420℃ to 1680℃, and the surface of the molten pool showed violent high-frequency temperature fluctuations, with the fluctuation amplitude exceeding three times the normal value. At the same time, the spatter recognition unit detected that the frequency of spatter events increased dramatically from an average of 10 times / second to 55 times / second, and the average velocity of the spatter particles increased by 40%.

[0035] The multi-source fusion analysis module takes the first and second data mentioned above, as well as the current process parameters P(t) (laser power, scanning speed, etc.) as input, and calculates them according to the formula. R=w1·RH+w2·RT+w3·RS+w4·RP, In the formula, RH represents the melt depth anomaly score, RT represents the temperature anomaly score, RS represents the splashing anomaly score, RP represents the process parameter deviation score, and w1 to w4 are weighting coefficients; based on the magnitude of R, four risk levels are output: normal, watch out, warning, and shutdown. Based on this feature set, the online defect identification module determined a risk of "over-melting / near penetration," accompanied by a secondary risk of "violent oscillation of the molten pool." This conclusion and risk level were immediately transmitted to the feedback control module. The feedback control module then automatically generated and executed two levels of control commands: the first level immediately reduced the laser pulse energy by 15%; the second level commanded the motion platform to increase the scanning speed by 10% at the corner. Within 0.2 seconds after the control commands were executed, the system monitored that the melt depth H(t) steadily dropped to 23.3 μm, T_max returned to the normal range, the spatter frequency dropped sharply, and the comprehensive risk score R dropped to the "normal" range of 0.25, successfully avoiding a potential wafer scrap accident.

[0036] Example 2: Localized Removal of Semiconductor Packaging Material This embodiment describes the laser ablation process for removing excess molding compound on a chip package, where the challenge lies in the material inhomogeneity.

[0037] The device configuration is similar to that of Example 1, but the OCT module plays a more crucial role in material interface identification in this scenario. The target removal thickness is 50 μm, but the molding compound may exhibit localized differences in density and absorptivity due to multiple injection molding processes. During processing, the system operates stably. At a certain moment, the thermal imaging module detected a slight overall decrease in the temperature field T(x,y,t), with T_max decreasing by approximately 3%. Following conventional single-sensor logic, this might be interpreted as insufficient energy.

[0038] However, the method of this invention reveals a completely different truth through multidimensional data synchronized in time and space. At the same moment, OCT monitored the crucial first data: the melt depth H(t) not only did not decrease, but increased slightly; more importantly, in the depth measurement curve, it was clearly identified that the reflection signal at the leading edge of the removal interface was disordered and split, no longer a clear single peak, indicating that the processed surface had become extremely rough.

[0039] The multi-source fusion analysis module captured these seemingly contradictory characteristics: a slight increase in penetration depth but aggravated surface roughness, while the temperature decreased slightly. The system called upon its internal physical model for analysis and concluded that the temperature drop was not due to a decrease in input energy, but rather an abnormal change in the local material absorptivity. This caused the laser energy to no longer be effectively absorbed for layer-by-layer vaporization, but instead to be converted into irregular thermomechanical stress fragmentation, resulting in blocky material peeling rather than vaporization. This leads to an extremely rough processed surface and the potential for microcracks.

[0040] Based on this, the online defect identification module determines that "abnormal material absorption rate / surface contamination leads to deterioration of processing quality" and outputs a "warning" signal, indicating the risk of "uneven thermal stress causing cracking." The feedback control module then issues instructions, not simply increasing power (which would exacerbate the problem), but adjusting the laser pulse frequency and scanning strategy, and suggesting checking the incoming material or adding a pre-cleaning process. This precise diagnostic and control capability is entirely due to the profound "understanding" of the processing process obtained through multi-field coupled sensing in this invention.

[0041] Creative discourse: Compared with the prior art, the technical solution of this invention has outstanding substantive features and significant progress, which are discussed in detail below: Prominent substantive features (non-obviousness) The non-obviousness of the technical concept: In existing technologies, melt depth detection and temperature field monitoring are considered two independent or loosely coupled technical problems. This invention innovatively proposes a collaborative sensing architecture with "coaxial super-resolution OCT direct observation of melt depth" as the core physical quantity and "thermal imaging temperature field / splash observation" as a multi-field coupling extension. Its core concept is not a simple superposition of "sensor A + sensor B," but rather a sensing design based on a causal chain of the complete physical process of "optical field input - thermal field response - geometric field shaping - mechanical disturbance." OCT provides the processing and shaping results (geometry), while thermal imaging provides the thermodynamic state and dynamic disturbances of the process. The two complement and corroborate each other in terms of physical mechanism. This design based on a multi-field coupling physical mechanism is unprecedented in existing technologies.

[0042] The non-obviousness of key technologies: Achieving micron-level precision coaxial super-resolution OCT online measurement and performing millisecond- and micron-level spatiotemporal registration with high-speed thermal imaging under the strong interference environment of laser processing presents significant technical challenges. This invention solves this problem through specific optical path design (coaxial coupling), super-resolution algorithms (breaking the diffraction limit to distinguish minute molten pool features), and a unified spatiotemporal reference mapping. In particular, the deep information fusion method—correlation analysis of rapid spatter identification's dynamic characteristics with geometric features such as abrupt changes in melt depth to diagnose instability mechanisms—is not a conventional approach in this field.

[0043] Significant progress (beneficial technological effects) A qualitative leap in perception capabilities: This invention is the first to transform the melting depth of semiconductor laser processing from an "invisible" inference state to an "online direct visual measurement," completely changing the traditional mode of relying on indirect temperature estimation or destructive post-processing detection, and realizing the "transparency" of core quality indicators in the processing process.

[0044] A significant leap in defect identification accuracy: By employing multi-field coupling criteria, such as "high spatter frequency + abrupt changes in melt depth + temperature spikes" to jointly determine over-melting, and "large temperature gradient + high cooling rate" to determine the risk of hot cracking, its accuracy and robustness far surpass schemes relying on single temperature or image features. This physical mechanism-based cross-validation effectively reduces false alarm and false negative rates, an effect that cannot be achieved with single sensing or simple multi-sensor fusion.

[0045] The above description is only a preferred embodiment of the present invention. All equivalent changes and modifications made within the scope of the claims of the present invention should be covered by the claims of the present invention.

Claims

1. A method for online sensing of semiconductors in laser processing under photothermal-mechanical multi-field coupling, characterized in that, Includes the following steps: (1) Obtain the process parameters P(t) during laser processing; (2) The super-resolution optical coherence tomography (OCT) observation module is used to acquire the tomographic signal of the processing area in real time, and the first data characterizing the internal geometry of the processing area is generated based on the super-resolution reconstruction algorithm. The first data includes at least the real-time melt depth H(t) and the cross-sectional morphology of the processing groove. (3) The thermal radiation image of the processing area is acquired in real time through the infrared thermal imaging observation module, and second data characterizing the thermal behavior and mechanical disturbance of the processing area is generated. The second data includes at least the temperature field distribution T(x,y,t), temperature fluctuation characteristics and splash event information S(t). (4) The first data, the second data, and the process parameter P(t) are synchronized in time and registered in space to form multi-dimensional online observation data mapped to a unified processing trajectory coordinate system: D(t,x,y)={H(t),T(x,y,t), G_T(t), S(t), P(t)} In the formula, H(t) represents the melting depth, T(x,y,t) represents the temperature field, G_T(t) represents the temperature gradient or temperature fluctuation characteristics, S(t) represents the splashing state, and P(t) represents the process parameters; Using the target melting depth H0 and the allowable deviation range [Hmin, Hmax] as benchmarks, if H(t) is lower than Hmin for multiple consecutive sampling periods, it is judged as insufficient melting depth; if H(t) is higher than Hmax or dH / dt exceeds the preset upper limit, it is judged as over-melting or penetration risk. Using the highest temperature, average temperature, temperature gradient, heat-affected zone range, and temperature fluctuation amplitude as inputs, the system determines whether the local heat input is stable. If the temperature gradient is too large and the cooling rate is too high, the risk of thermal stress cracking is determined to be increased. If the temperature field continues to expand, the risk of heat accumulation or thermal damage is determined to be increased. The intensity of molten pool disturbance is judged by the number of splashes, splash speed, splash temperature and splash frequency. If the number of splash events increases significantly per unit time and occurs simultaneously with temperature fluctuations and sudden changes in melt depth, the processing is considered unstable. (5) Based on the multidimensional online observation data, a set of optical-thermal-mechanical multi-field coupled state features is constructed. The feature set includes optical field input features, thermal field response features, geometric field features, and mechanical disturbance-related features. (6) Based on the aforementioned photothermal-mechanical multi-field coupling state feature set, the current processing status and defect risks are identified online, and the risk level and control suggestions are output.

2. The online sensing method for laser processing semiconductors under photothermal-mechanical multi-field coupling according to claim 1, characterized in that, Step (2) involves generating the first data based on the super-resolution reconstruction algorithm, specifically including: (2.1) The acquired raw tomographic interferometric signals are preprocessed by denoising and phase compensation; (2.2) Perform tomographic reconstruction on the preprocessed signal to obtain the initial cross-sectional image; (2.3) The initial cross-sectional image is processed by using a super-resolution reconstruction algorithm based on deconvolution or deep learning model to break through the optical diffraction limit of the system and obtain the internal cross-sectional morphology of the processing area with submicron resolution. (2.4) From the reconstructed cross-sectional morphology, the real-time melt depth H(t), the width of the melt pool cross-section, the flatness of the bottom of the pool, and the geometric features of the boundary continuity are accurately extracted; Step (3) involves generating the second data, specifically including: (3.1) Based on the calibration parameters, the thermal radiation image is inverted into a two-dimensional temperature field distribution T(x,y,t) of the molten pool and the heat-affected zone; (3.2) Perform pixel-by-pixel temporal analysis on the temperature field distribution of multiple consecutive frames, and extract the features of the highest temperature, average temperature, temperature gradient, width of the heat-affected zone, temperature fluctuation amplitude and cooling rate to form temperature fluctuation features. (3.3) Identify small targets that are short-lived, high-temperature and high-speed moving in the thermal image sequence that meet the preset spatiotemporal conditions, mark them as splash events, and record the number, temperature, speed, trajectory direction and frequency of occurrence of the splash events to form the splash event information S(t).

3. The online sensing method for laser processing semiconductors under photothermal-mechanical multi-field coupling according to claim 1, characterized in that, Step (4) involves time synchronization and spatial registration, specifically including: (4.1) Establish the spatial transformation mapping relationship between the OCT coordinate system, the thermal imaging coordinate system, the laser processing coordinate system, and the motion platform coordinate system; (4.2) By using hardware trigger signals and unified timestamps, the data acquisition times of the OCT module and the thermal imaging module are precisely aligned; (4.3) Using the position information fed back in real time by the motion platform encoder and the predefined processing trajectory information, the first data and the second data at each moment are accurately allocated to the actual processing trajectory coordinate points according to the spatial transformation mapping relationship; (4.4) Form structured multidimensional online observation data with "location-time-melt depth-temperature-splash-process parameters"; Step (5) involves constructing a multi-field coupled state feature set of photothermal and mechanical fields, specifically including: (5.1) The optical field input characteristics include at least the laser power, energy density, scanning speed, and focus offset; (5.2) The thermal field response characteristics include at least the highest temperature, average temperature, temperature gradient, range of the heat-affected zone, and cooling rate; (5.3) The geometric field characteristics include at least the real-time melt depth H(t), melt depth change rate dH / dt, melt depth / melt width ratio and bottom flatness; (5.4) The mechanical disturbance-related characteristics include at least the spatter intensity, temperature oscillation amplitude, frequency of sudden changes in melt depth, and tendency for local collapse; Step 6 involves online identification of the current processing status and defect risks, specifically including: (6.1) Input the photothermal multi-field coupling state feature set into a preset recognition model; (6.2) When the real-time melting depth H(t) is continuously lower than the lower limit of the preset target melting depth range and the average temperature is lower than the preset temperature threshold, it is determined to be a risk of undermelting or insufficient melting depth; (6.3) When the melting depth change rate dH / dt exceeds the preset positive change rate threshold and the highest temperature exceeds the preset high temperature threshold, it is determined to be a risk of over-melting or penetration; (6.4) When the temperature fluctuation amplitude exceeds the preset fluctuation threshold and the frequency of the splashing event exceeds the preset frequency threshold, it is determined to be an unstable molten pool or an abnormal splashing risk. (6.5) When the melting depth is basically normal but the temperature gradient and the cooling rate both exceed the preset range, it is determined to be a risk of thermal stress-induced cracking.

4. The online sensing method for laser-processed semiconductors under photothermal multi-field coupling according to claim 3, Its features are, The identification model in step (6.1) is based on a weighted scoring mechanism, which quantifies defect risk through a comprehensive risk score R. The calculation formula is as follows: R = w1·RH + w2·RT + w3·RS + w4·RP In the formula, RH represents the melt depth anomaly score, RT represents the temperature anomaly score, RS represents the splashing anomaly score, RP represents the process parameter deviation score, and w1 to w4 are weighting coefficients. Based on the magnitude of R, four risk levels are output: normal, attention, warning, and shutdown. w1, w2, w3, and w4 are weighting coefficients, and they satisfy w1+w2+w3+w4=1. When R exceeds the first threshold, a "warning" signal is output. When R exceeds the second threshold, a "stop and adjust immediately" signal is output.

5. The online sensing method for laser processing semiconductors under photothermal-mechanical multi-field coupling according to any one of claims 1 to 4, characterized in that, Following the steps of outputting risk levels and control recommendations, step (7) is also included: (7.1) When the defect risk level exceeds the preset intervention threshold, a feedback control command is generated and sent to the laser processing control unit; (7.2) The feedback control command is used to perform at least one of the following operations: adjust the laser output power, change the scanning speed, correct the focal position, adjust the pulse frequency, adjust the protective gas flow rate, or perform dynamic compensation on the processing path.

6. An online sensing device for laser processing semiconductors under photothermal multi-field coupling, used to implement the online sensing method for laser processing semiconductors under photothermal multi-field coupling as described in any one of claims 1 to 4, characterized in that, The device includes: A laser processing module that generates processing lasers to process semiconductor materials; The super-resolution OCT observation module has its detection optical path arranged coaxially with the processing laser to acquire first data inside the processing area. The first data includes at least the real-time melt depth and the cross-sectional morphology of the processing groove. An infrared thermal imaging observation module is tilted off-axis, with its field of view covering the processing area, and is used to acquire second data of the processing area. The second data includes at least temperature field distribution, temperature fluctuation and splash event information. The spatiotemporal synchronization module is connected to the controllers of the super-resolution OCT observation module, the infrared thermal imaging observation module, and the laser processing module, respectively, and is used to precisely align the first data and the second data with the processing position and processing timestamp. The multi-source fusion analysis module receives multi-dimensional online observation data from the spatiotemporal synchronization module and is configured to construct a set of optical-thermal-mechanical multi-field coupled state features. The online defect identification module is configured to identify the processing status and calculate the defect risk level based on the photothermal multi-field coupling state feature set. The feedback control module is configured to output control commands to the laser processing module based on the defect risk level.

7. The online sensing device for laser processing semiconductors under photothermal multi-field coupling according to claim 6, characterized in that, The super-resolution OCT observation module is a spectral domain OCT or a swept frequency OCT, which integrates a super-resolution reconstruction unit. The super-resolution reconstruction unit breaks through the axial and lateral resolution limits of the OCT system by executing a deconvolution algorithm or calling a pre-trained deep learning model.

8. The online sensing device for laser processing semiconductors under photothermal multi-field coupling according to claim 6, characterized in that, The infrared thermal imaging observation module integrates a splash identification unit. The splash identification unit identifies short-term high-temperature point clusters in the sequence images that have a temperature more than three times the average temperature of the outer periphery of the molten pool and have a directional motion trajectory by using the frame difference method or optical flow method, and marks them as splash events.

9. The online sensing device for laser processing semiconductors under photothermal multi-field coupling according to claim 6, characterized in that, The multi-source fusion analysis module and the online defect identification module are integrated in the same industrial computer or edge computing node, and can process data in real time at a speed faster than the processing cycle, ensuring that the identification results are output within a preset time window after the first and second data are generated.

10. The online sensing device for laser processing semiconductors under photothermal multi-field coupling according to claim 6, characterized in that, The spatiotemporal synchronization module receives encoder pulse signals from the motion platform and uses these signals as a reference to synchronously trigger the acquisition actions of the super-resolution OCT observation module and the infrared thermal imaging observation module, thereby achieving registration based on sampling at equal spatial intervals.

Images