A double-support selected area epitaxy multi-modal data fusion growth method

Abstract

The application relates to the technical field of epitaxial growth of semiconductor materials, and discloses a double-support selected-area epitaxial multi-modal data fusion growth method, which comprises the following steps: placing a substrate support and a mask support, which are independent and have rotation functions, in an equipment; the mask support is lowered to a near-contact position to form non-contact shielding and is synchronously rotated with the substrate support for selected-area growth; after current area growth, the mask support is raised and independently rotated to switch an opening, and then is lowered for next area growth, thereby forming multiple epitaxial areas with different conditions on the same substrate; in-situ multi-modal data is collected throughout the process, and a performance label is obtained offline. The double-support synchronous rotation ensures the uniformity of growth, the in-situ switching of areas eliminates batch system errors, realizes in-situ monitoring throughout the process without shielding, a mapping model is constructed through high-quality data fusion, and the prediction accuracy of material performance is improved.

Description

Technical Field

[0001] This application relates to the field of semiconductor material epitaxial growth technology, specifically a dual-scaffold selected area epitaxial multimodal data fusion growth method. Background Technology

[0002] Selective epitaxial growth is a core process in semiconductor material preparation. It mainly addresses the need to deposit and form epitaxial layers with specific structures in specific substrate areas. The industry has introduced multimodal data acquisition and monitoring methods during the growth process to control the quality of the epitaxial layer and establish a mapping relationship between process parameters and material properties.

[0003] The core principle of selected area epitaxy is to use a mask to partially block the substrate surface, so that the epitaxial material only nucleates and grows in the exposed opening area. In the specific growth process, with the help of the monitoring device in the cavity, in-situ characteristic data such as changes in crystal structure, thickness evolution and electronic structure of the substrate surface can be acquired in real time. The collected multi-dimensional in-situ monitoring data is combined with the growth adjustment conditions and the final characterization data after the growth is completed to construct a low-level data mapping model, thereby guiding the adjustment and optimization of subsequent process parameters.

[0004] Current selective epitaxy (SEE) techniques employ a structure where a mask is fixed above a substrate. This, coupled with a fixed masking pattern, necessitates stopping substrate rotation, resulting in significant radial gradients and inhomogeneities on the epitaxial growth surface. Furthermore, the fixed mask's distance from the sample surface severely obstructs the optical path and electron beam channel, limiting the operation of in-situ monitoring equipment such as high-energy electron diffractometers and ellipsometrists. Edge effects also occur during growth. The lack of a flexible region-switching mechanism necessitates batch-by-batch experiments with different adjustment conditions, introducing uncontrollable batch-to-batch systematic errors and making it impossible to obtain highly consistent datasets on a single wafer. Even attempts to move the mask are hampered by inherent mechanical design contradictions, failing to simultaneously address collision safety, close-contact masking, and continuous synchronous substrate rotation. This hinders the establishment of an accurate mapping between multimodal data and material properties.

[0005] Therefore, the purpose of this application is to provide a dual-support selected area epitaxial multimodal data fusion growth method to address the shortcomings of the existing technology. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this application provides a dual-support selected epitaxial multimodal data fusion growth method, which solves the problems in existing selected epitaxial technologies, such as uneven growth caused by fixed mask, systematic errors caused by batch experiments on multiple wafers, and the inability to continuously acquire multimodal data and establish accurate mapping relationships due to structural obstruction of the in-situ monitoring path.

[0007] To achieve the above objectives, this application provides the following technical solution:

[0008] A dual-support selected region epitaxial multimodal data fusion and growth method includes: A substrate holder with a fixed substrate and a mask holder supporting a mask are placed in an epitaxial device. The mask holder is placed around the substrate holder. The substrate holder and the mask holder are independent of each other and both have rotation functions. The mask is located on the path from the epitaxial source to the substrate growth surface. Adjust the mask holder to a position close to the working surface so that the mask forms a selected area blockage without contacting the substrate. The mask holder and the substrate holder rotate synchronously at the same angular velocity to perform selected area epitaxial growth. After completing the selected epitaxial growth of the current region, the mask holder is rotated independently relative to the substrate by a preset independent rotation angle to switch the opening region. Then, the selected epitaxial growth of the next region is carried out to form multiple epitaxial regions with different adjustment conditions on the same substrate. During continuous selected area epitaxial growth and region switching, in-situ multimodal data are collected throughout the process by using the space avoidance of the mask holder and the gap between the mask and the substrate. After the growth process is completed, each epitaxial region is characterized offline to obtain the final performance label. The collected in-situ multimodal data is then fused with the corresponding adjustment conditions and the final performance label at the feature level to establish the mapping relationship between the epitaxial process and the material properties.

[0009] This application achieves in-situ close-contact masking and continuous position switching within a closed cavity environment by setting up independent substrate and mask supports, thereby enabling selected area epitaxial growth on the same substrate surface. The independent lifting and rotation mechanism of the dual supports allows the mask to mask at a set working gap and avoids physical collisions during region switching. Combined with the observation channel reserved in the equipment structure, this application extracts surface crystal structure changes, thin film optical evolution, and bandgap information in real time. In terms of data processing, this application maps the collected multimodal data sequences to the same high-dimensional feature space, establishes the correlation features between modes through an attention mechanism, and constructs the calculation logic of the network loss function based on the physical laws of material growth. This ensures that the model conforms to the material evolution and bandgap change laws of crystal growth when updating weights, completing the quantitative mapping from process parameters to material properties.

[0010] Preferably, the substrate support adopts a recessed design so that the upper edge of the substrate support is lower than the substrate growth surface; The acquisition of in-situ multimodal data involves the RHEED electron beam being incident at a low grazing angle, passing through the opening of the mask, the close contact gap between the mask and the substrate, and the reserved electron beam clearance channel to reach the exposed surface of the substrate.

[0011] Preferably, the acquisition of in-situ multimodal data also includes real-time acquisition of in-situ optical characterization signals through a fixed optical window in the cavity; During the selected area epitaxial growth and mask support area switching process, RHEED image sequences and in-situ optical characterization signal sequences were continuously recorded according to a unified timestamp, and time-correlated with the rotation angle and adjustment conditions of the mask.

[0012] Preferably, the acquired in-situ multimodal data is pre-labeled with state information, specifically as follows: Data collected during selective epitaxial growth with the mask support in a close-to-working position are marked as stable growth state data. The data collected during the area switching process of the mask plate support is marked as switching status data; Data collected during the switching process or when the mask support briefly pauses at the observation height to supplement surface state information at the moment of switching is marked as snapshot state data.

[0013] Preferably, when switching regions, the mask holder is adjusted to a safe switching height; The close-contact working position is at a height of 0.2–0.5 mm between the mask and the substrate growth surface; The safe switching height is 5-10 mm above the substrate growth surface to prevent the mask from scratching or colliding with the substrate surface, molybdenum support, or support structure during the rotation switching process.

[0014] Preferably, the mask adopts a 1 / 4 fan-shaped opening pattern or a parallel strip array pattern; when the mask adopts a 1 / 4 fan-shaped opening pattern, the set angle for the mask holder to rotate independently relative to the substrate each time is 90°; when the mask adopts a parallel strip array pattern, the set angle for the mask holder to rotate independently relative to the substrate each time is 60°.

[0015] Preferably, the photomask is a thin plate with a thickness of 200 μm, and the material of the photomask is selected from one of SiN, SiO2, SiC, pyrolytic BN, Mo, Ta or W.

[0016] Preferably, a mask storage chamber is additionally provided in the growth chamber of the epitaxial device. The mask storage chamber contains multiple masks with different circular dot pattern parameters. After the selected area epitaxial growth is completed, the old mask is sent back to the mask storage chamber by a transfer robot inside the chamber, and a new mask is picked up or lifted from the mask storage chamber and transferred to the mask holder, realizing in-situ rapid replacement without opening the chamber.

[0017] Preferably, the specific implementation of feature-level fusion is as follows: extract surface reconstruction feature vectors from the RHEED image sequence to characterize surface reconstruction and nucleation dynamics; extract optical evolution feature vectors to characterize thickness growth and composition changes, and electronic structure-related feature vectors to characterize band gap and reflectivity inflection points from the in-situ optical characterization signal sequence; and use the Cross-Attention mechanism to interactively fuse the surface reconstruction feature vectors, optical evolution feature vectors, and electronic structure-related feature vectors to generate a unified joint feature vector.

[0018] Preferably, establishing the mapping relationship between epitaxial processes and material properties specifically involves: inputting the generated joint feature vector into a multi-task prediction network, outputting predicted material properties including average quantum dot size, quantum dot density, surface roughness, or bandgap-related parameters, and updating the weights of the multi-task prediction network based on a preset physical constraint loss; the specific calculation logic of the physical constraint loss includes obtaining the performance prediction results output by the multi-task prediction network, setting theoretical calculation boundaries based on the surface energy changes of material nucleation and the relationship between thin film growth kinetics, quantifying and analyzing the deviation between the prediction results and the theoretical calculation boundaries, and converting the deviation into a nonlinear penalty term that is superimposed on the total loss function of the multi-task prediction network to guide the updating of network weights.

[0019] This application provides a dual-support selected region epitaxial multimodal data fusion and growth method. It has the following beneficial effects: 1. This application sets up a substrate holder and a mask holder that are independent of each other and both have rotation functions. The mask holder is lowered to a position close to the working area to form a selected area shielding. The mask holder and the substrate holder rotate synchronously at the same angular velocity to perform selected area epitaxial growth. This avoids the problem that the substrate cannot rotate due to the traditional fixed mask. The selected area shielding effect is guaranteed without contacting the substrate, and the uniformity of epitaxial growth in each selected area is improved.

[0020] 2. This application sets up a mask holder that rises to a safe switching height after completing the growth of the current region, and rotates independently relative to the substrate by a preset independent rotation angle each time to switch the opening region. Then it descends again to a close-contact working position to grow the next region. This allows multiple epitaxial regions with different adjustment conditions to be continuously formed on the same substrate without opening the growth chamber. This eliminates the batch-to-batch systematic errors caused by traditional multi-wafer separate experiments and ensures high consistency of data from different epitaxial regions.

[0021] 3. This application, by setting a mask holder clearance space and a gap between the mask and the substrate, collects in-situ multimodal data throughout the continuous selected area epitaxial growth and region switching process. This data is then fused with adjustment conditions and final performance labels at the feature level, enabling continuous monitoring of the growth state without obstructing the monitoring path. Furthermore, the fusion processing establishes a mapping relationship between epitaxial processes and material properties, improving the accuracy of multimodal data in predicting material properties. Attached Figure Description

[0022] Figure 1 This is a comparison diagram of quantum dot densities in Example 1 and Comparative Example 1 of this application under multi-region selected epitaxial growth. Detailed Implementation

[0023] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0024] Examples 1-6: Example 1: This example provides a dual-support selected region epitaxial multimodal data fusion growth method. As a basic implementation method of this application, this paper specifically explains how to achieve uniform selected epitaxy and full-process in-situ multimodal data acquisition of four different growth adjustment regions on the same substrate through the collaboration of mutually independent substrate supports and mask supports.

[0025] The epitaxial equipment used in this embodiment is a commercial MOCVD system, the growth material is InGaAs quantum dots, the substrate is a 2-inch GaAs (001) wafer, and the back of the substrate is processed with an annular anti-rotation structure with a convex bond height of 0.3 mm and a width of 2 mm, which is uniformly distributed along the circumference. The annular anti-rotation structure is fully engaged with the groove on the top of the substrate support to prevent the substrate from being misaligned or drifting at an angle relative to the substrate support during rotation.

[0026] In this embodiment, the substrate is pre-fixed on a substrate support or molybdenum holder. The upper surface of the substrate is the epitaxial growth surface. The epitaxial source beam or reactive gas is incident from above or above the side of the substrate growth surface and epitaxial growth occurs after reaching the exposed surface of the substrate. The working position of the mask is located above the substrate growth surface and in the propagation path of the epitaxial source beam or reactive gas to the substrate surface. The mask allows the epitaxial source beam or reactive gas to reach the exposed area of ​​the substrate through its opening area and blocks the epitaxial source beam or reactive gas from reaching the non-selected area through its blocking area, thereby realizing selective epitaxy.

[0027] The equipment's structural parameters include a recessed substrate support design, with the entire molybdenum substrate holder sunk within the support's groove. The upper edge of the support is lower than the substrate growth surface to reduce obstruction of the RHEED grazing electron beam by the support edge. The substrate support retains its original rotation function and can drive the substrate to rotate stably at a set angular velocity during growth. The mask support is an independent peripheral annular support module, positioned around the substrate support and above the substrate growth surface. The mask support includes an annular support section, an annular support groove, a positioning structure, a Z-axis lifting mechanism, and a rotation drive mechanism. The mask support features independent rotation, synchronous rotation at the same angular velocity as the substrate support, and precise Z-axis height adjustment. The Z-axis travel range is 0–15 mm, with an adjustment accuracy of ±0.01 mm. The annular support section supports the mask from its outer peripheral positioning edge, ensuring that the central area of ​​the mask covers the substrate growth surface. The mask support has clearance spaces in the RHEED incident direction and optical characterization path, allowing the RHEED electron beam, ellipsoidized beam, or edge-characterization beam to reach the substrate exposure area through the corresponding channels. The photomask material is a 200μm thick SiN plate, but SiO2 (silicon dioxide), SiC (silicon carbide), pyrolytic BN (pyrolytic boron nitride), Mo (molybdenum), Ta (tantalum) or other high-temperature resistant, low-outgas, and thermally compatible materials can also be selected according to the epitaxial environment.

[0028] The photomask has a positioning edge, a pick-and-place notch, or a fork slot on its outer periphery. The positioning edge mates with the annular support groove of the photomask holder to achieve coaxial and rotational positioning of the photomask. The pick-and-place notch or fork slot is used for lifting, positioning, and retraction by a transfer robot, and provides a margin for the photomask's slight thermal expansion and mechanical assembly errors under thermal conditions. The pick-and-place notch or fork slot is not used as the epitaxial selection pattern, nor is it limited to matching the protruding structure of the substrate holder or heater holder. The photomask pattern adopts a 1 / 4 fan-shaped opening, with the center of the 1 / 4 fan-shaped opening coaxial with the center of the substrate. The opening angle is 90°, and the opening area covers a quarter circle of the substrate surface. There are no other obstructing structures within the opening. The area of ​​the photomask other than the 1 / 4 fan-shaped opening is the shielding area, used to prevent the epitaxial source beam or reactive gas from reaching the non-selected area, including the following steps: During the mask transfer stage, the substrate is pre-loaded and fixed on a substrate holder or molybdenum support, with the upper surface of the substrate serving as the epitaxial growth surface. The transfer robot uses a fork-shaped or tray-type insert structure to enter from the side or lower side of the cavity, delivering the 1 / 4 fan-shaped opening mask to the mask holder receiving position located above the substrate growth surface.

[0029] During the transfer process, the positioning edge of the mask's outer periphery aligns with the annular support groove of the mask holder, and the pick-and-place notch or fork slot provides space for the transfer robot to lift, position, and retract. The mask holder supports the mask at a safe receiving height of 5–10 mm from the substrate growth surface, with a safe receiving height of 8 mm. Subsequently, the transfer robot descends or retracts laterally, transferring the mask from the transfer robot to the support of the mask holder.

[0030] After the handover is completed, the mask holder lowers the mask along the Z-axis to a near-contact working position 0.2–0.5 mm above the substrate growth surface. Preferably, the gap between the mask and the substrate growth surface is 0.3 mm. At this point, the mask does not make physical contact with the substrate, but forms a near-contact shielding structure above the substrate growth surface, allowing the epitaxial source beam or reactive gas to reach the substrate exposure area only through a 1 / 4 fan-shaped opening in the mask.

[0031] First growth cycle, Region 1, First adjustment conditions. The mask holder is precisely lowered so that the mask is positioned 0.3 mm above the substrate growth surface in a near-contact working position, forming a selected area masking without contacting the substrate. The mask holder and substrate holder rotate synchronously at the same angular velocity, at a speed of 10 rpm. During synchronous rotation, the mask opening and the selected area on the substrate remain fixed in relative position, while the mask and substrate as a whole rotate relative to the epitaxial source beam or reactive gas flow field, thereby improving the growth uniformity within the selected area.

[0032] Subsequently, TMI, TMI, and AsH3 source gases were introduced, and the growth temperature was set to 580℃, the V / III ratio to be 80, and the growth time to be 15 min. The epitaxial source beam or reactive gas reached the substrate exposure area through a 1 / 4 fan-shaped opening of the mask, and selected InGaAs quantum dot growth was performed under the first adjustment conditions in the corresponding first quadrant region. On the substrate surface corresponding to the mask-covered area, epitaxial growth did not occur or only negligible weak deposition occurred due to the mask's obstruction.

[0033] Switch to the second growth cycle, region 2, second adjustment conditions. After the first growth cycle is complete, stop or adjust the corresponding growth source so that the mask holder raises the mask to a safe switching height of approximately 8 mm from the substrate growth surface. This safe switching height is used to prevent the mask from scratching or colliding with the substrate surface, molybdenum support, or support structure during rotational switching.

[0034] The mask holder is rotated 90° independently relative to the substrate at a safe switching height, switching the 1 / 4 fan-shaped opening from the first region to the second region. After rotation, the mask holder descends along the Z-axis again, positioning the mask back in a close-contact working position 0.3 mm above the substrate growth surface. The growth parameters are then adjusted to a temperature of 590℃ and a V / III ratio of 100, and the mask holder and substrate holder are brought back into a state of synchronous rotation at the same angular velocity. Selective epitaxial growth under the second adjustment conditions is then performed in the second region for 15 minutes.

[0035] The third and fourth growth cycles are performed by repeating the above steps of "rising to the safe switching height, rotating 90° independently relative to the substrate, descending to the near-contact working position, and rotating synchronously to grow," thus completing the third and fourth selected area epitaxial growth cycles sequentially.

[0036] In the third growth cycle, the mask holder is independently rotated 90° relative to the substrate at a safe switching height, aligning the mask opening with the third region. The mask is then lowered to a near-contact working position 0.3 mm above the substrate growth surface, and selected-area epitaxial growth is performed under the third adjustment conditions while rotating synchronously. The third growth conditions are: temperature 600℃, V / III ratio 120, and growth time 15 min. In the fourth growth cycle, the mask holder is again raised to the safe switching height and independently rotated 90° relative to the substrate, aligning the mask opening with the fourth region. The mask is then lowered to a near-contact working position 0.3 mm above the substrate growth surface, and selected-area epitaxial growth is performed under the fourth adjustment conditions while rotating synchronously. The fourth growth conditions are: temperature 610℃, V / III ratio 140, and growth time 15 min. After four cycles, four physically distinguishable 90° sector regions are formed on the same substrate, each corresponding to different growth adjustment parameters. Since the four regions are located on the same substrate, and the substrate support and mask support can rotate synchronously during the growth process, batch differences between different wafers can be reduced, and the comparability of data between regions can be improved.

[0037] In-situ multimodal characterization: Throughout the growth process, the mask support is maintained with high precision, ensuring a close-contact gap of approximately 0.3 mm between the mask and the substrate growth surface. The RHEED electron beam is incident at a low grazing angle, preferably 2°, and reaches the exposed substrate surface through the mask's fan-shaped opening, the close-contact gap, and the reserved electron beam clearance channel, enabling continuous monitoring of surface reconstruction dynamics. The electron beam clearance channel is defined by the mask opening position, the mask support clearance space, and the recessed structure of the substrate support, reducing the obstruction of the RHEED incident path by the mask support or mask edge.

[0038] Simultaneously, the Ψ / Δ spectrum and band edge reflectance curves of the spectral ellipsometer are acquired in real time through a fixed optical window in the cavity. The spectral ellipsometer is used to obtain information related to thickness, composition, or growth rate during the selected area growth process. The band edge characterization is used to obtain information on band gap changes, reflectance changes, or optical response changes. The RHEED image sequence, ellipsometric spectrum sequence, and band edge curves are all recorded with a unified timestamp and correlated with parameters such as mask angle, region number, growth temperature, V / III ratio, and growth time.

[0039] During each region switch, the mask support rises to a safe switching height and completes independent rotation. If necessary, the mask can be kept at a safe height or intermediate observation height for short-term RHEED snapshots before and after the switch to supplement the surface state information at the moment of switching. After the snapshot is completed, the mask support descends to a position close to the working position to continue the next round of growth. All in-situ data are aligned and stitched according to timestamps to form a continuous record of the multimodal growth process.

[0040] Growth Completion and Region Division: After all four growth cycles are completed, the mask holder raises the mask to a safe removal height of approximately 10 mm from the substrate growth surface. A transfer robot enters from the side or lower side of the cavity and supports the mask, transferring it from the mask holder to the transfer robot for removal. The substrate is then unloaded. After substrate unloading, the sample is region-divided on the wafer coordinate system using ellipsometry, optical microscopy, PL mapping, AFM, or SEM characterization. For the 1 / 4 fan-shaped aperture mask in this embodiment, the same wafer can be automatically divided into four 90° fan-shaped regions. Each fan-shaped region corresponds to one growth cycle and a set of growth adjustment parameters. The corresponding RHEED image sequence, ellipsometry spectrum sequence, band edge reflectance curve, and subsequent offline characterization results are extracted for each region to form a structured multimodal dataset.

[0041] Example 2: This example provides a dual-support selected region epitaxial multimodal data fusion growth method. The epitaxial equipment used in this embodiment is a commercial MOCVD system, the growth material is InAs nanowires, and the substrate is a 2-inch GaAs (001) wafer. A ring-shaped anti-rotation structure is fabricated on the back side of the substrate, with a bond height of 0.3 mm and a width of 2 mm, uniformly distributed circumferentially. The ring-shaped anti-rotation structure fully engages with the groove on the top of the substrate support to prevent angular displacement of the substrate relative to the substrate support during rotation.

[0042] In this embodiment, the substrate is pre-fixed on a substrate support or molybdenum holder. The upper surface of the substrate is the epitaxial growth surface. The epitaxial source beam or reactive gas is incident from above or above the side of the substrate growth surface and epitaxial growth occurs after reaching the exposed surface of the substrate. The working position of the mask is located above the substrate growth surface and in the propagation path of the epitaxial source beam or reactive gas reaching the substrate surface.

[0043] The mask allows the epitaxial source beam or reactive gas to reach the substrate exposed area through its strip-shaped opening area, and blocks the epitaxial source beam or reactive gas from reaching the non-selected area through its shielding area, thereby realizing the selected area epitaxy of the strip array.

[0044] Equipment structural parameters: The substrate support adopts a sunken design, with the entire molybdenum substrate holder sunk within the support groove. The upper edge of the support is lower than the substrate growth surface to reduce the obstruction of the RHEED grazing electron beam by the support edge. The substrate support retains its original rotation function and can drive the substrate to rotate stably at a set angular velocity during growth. The mask support is an independent peripheral annular support module, located on the periphery of the substrate support and above the substrate growth surface. The mask support includes an annular support part, an annular support groove, a positioning structure, a Z-axis lifting mechanism, and a rotation drive mechanism. The mask support has independent rotation function, synchronous rotation function with the substrate support at the same angular velocity, and... The mask features a precision Z-axis height adjustment function with a travel range of 0–15 mm and an adjustment accuracy of ±0.01 mm. An annular support is used to support the mask from its outer periphery, ensuring that the central pattern area of ​​the mask covers the substrate growth surface. The mask holder has clearance spaces in the RHEED incident direction and optical characterization path, allowing the RHEED electron beam, ellipsoidized beam, or edge-characterization beam to reach the substrate exposure area through the corresponding channels. The mask material is a 200 μm thick SiN sheet, but SiO2, SiC, pyrolytic BN, Mo, Ta, or other high-temperature resistant, low-outgas, and thermally compatible materials can also be selected depending on the epitaxial environment.

[0045] The photomask has a positioning edge, a pick-and-place notch, or a fork slot on its outer periphery. The positioning edge mates with the annular support groove of the photomask holder to achieve coaxial and rotational positioning of the photomask. The pick-and-place notch or fork slot is used for lifting, positioning, and retraction by a transfer robot, and allows for tolerances to minor thermal expansion and mechanical assembly errors under thermal conditions. The pick-and-place notch or fork slot is not used as an epitaxial selection pattern, nor is it required to match the protruding structure of the substrate holder or heater holder. The photomask pattern uses a parallel strip array pattern. The strip opening width is 5μm, the strip spacing is 10μm, and the strip direction is distributed radially or near-radially. The strip opening area covers approximately 1 / 6 of the substrate surface area. Based on the symmetrical distribution of 6 groups of strip array openings in the circumferential direction, the preset independent rotation angle is 60° each time, so that after each cycle, the photomask strip opening corresponds to the next 60° sector strip area on the same substrate. The area of ​​the photomask other than the strip array openings is the masking area, used to block the epitaxial source beam or reactive gas from reaching the non-selected area, including the following steps: Mask transfer stage: The substrate is pre-loaded and fixed on the substrate support or molybdenum support. The upper surface of the substrate serves as the epitaxial growth surface. The transfer robot adopts a fork-shaped or tray-type insert structure, enters from the side or lower side of the cavity, and delivers the strip array pattern mask to the mask support receiving position located above the substrate growth surface.

[0046] During the transfer process, the positioning edge of the outer periphery of the mask is aligned with the annular support groove of the mask holder. The pick-and-place notch or fork slot provides space for the transfer robot to lift, position, and withdraw. The mask holder supports the mask at a safe receiving height of 5–10 mm from the substrate growth surface, with a safe receiving height of 8 mm. Subsequently, the transfer robot descends or withdraws laterally, transferring the mask from the transfer robot to the support of the mask holder.

[0047] After the handover is completed, the mask holder drives the mask to descend along the Z-axis to a close-contact working position 0.2–0.5 mm above the substrate growth surface. The gap between the mask and the substrate growth surface is 0.3 mm. At this time, the mask does not make physical contact with the substrate, but forms a close-contact shielding structure above the substrate growth surface, so that the epitaxial source beam or reactive gas can only reach the substrate exposure area through the strip array openings.

[0048] First growth cycle, Region 1, First adjustment conditions. The mask holder is precisely lowered so that the mask is positioned 0.3 mm above the substrate growth surface in a close-contact working position, forming a strip array of selected areas without contacting the substrate. The mask holder and substrate holder rotate synchronously at the same angular velocity of 12 rpm. During synchronous rotation, the strip opening of the mask and the selected area of ​​the substrate remain fixed in relative position. At the same time, the mask and substrate as a whole rotate relative to the epitaxial source beam or reactive gas flow field, thereby improving the growth uniformity within the strip selected area.

[0049] Subsequently, TMI and AsH3 source gases were introduced, and the growth temperature was set to 520℃, the V / III ratio to be 150, and the growth time to be 20 min. The epitaxial source beam or reactive gas reached the substrate exposure area through the opening of the mask strip array. Selective InAs nanowire growth under the first adjustment conditions was carried out in the first 60° sector. Due to the mask blocking area, no epitaxial growth or only negligible weak deposition occurred on the substrate surface.

[0050] Switch to the second growth cycle, region 2, second adjustment conditions. After the first growth cycle is completed, stop or adjust the corresponding growth source so that the mask holder moves the mask to a safe switching height of about 8 mm from the substrate growth surface. The safe switching height is used to prevent the mask from scratching or colliding with the substrate surface, molybdenum support, or support structure during the rotation switching process.

[0051] The mask holder was independently rotated 60° relative to the substrate at a safe switching height, switching the strip array opening from the first 60° sector to the second 60° sector. After rotation, the mask holder descended along the Z-axis again, positioning the mask back in a close-contact working position 0.3 mm above the substrate growth surface. The growth parameters were then adjusted to a temperature of 530℃ and a V / III ratio of 180, and the mask holder and substrate holder were brought back into a state of synchronous rotation at the same angular velocity. Selective InAs nanowire growth under the second adjusted conditions was performed in the second 60° sector for 20 minutes.

[0052] For the third to sixth growth cycles, the above steps of "rising to the safe switching height, rotating independently by 60° relative to the substrate, descending to the near-contact working position, and synchronously rotating for growth" are repeated sequentially to complete the third to sixth selected area epitaxial growth cycles. During the third growth cycle, the mask holder is rotated independently by 60° relative to the substrate again at the safe switching height, aligning the strip array openings with the third 60° sector. The mask is then lowered to a near-contact working position 0.3 mm above the substrate growth surface, and selected area epitaxial growth is performed under the third adjustment conditions while in synchronous rotation. The third growth conditions are: temperature 540℃, V / Ⅲ ratio 210, and growth time 20 min.

[0053] In the fourth growth stage, the mask holder is raised to the safe switching height again and rotated independently by 60° relative to the substrate, aligning the strip array openings with the fourth 60° sector. The mask is then lowered to a near-contact working position 0.3 mm above the substrate growth surface, and selected area epitaxial growth is performed under the fourth adjustment condition while maintaining synchronous rotation. The fourth growth conditions are: temperature 550℃, V / III ratio 240, and growth time 20 min. In the fifth growth stage, the mask holder continues to rotate independently by 60° relative to the substrate in the same manner, aligning the strip array openings with the fifth 60° sector. Selected area epitaxial growth is then performed under the fifth adjustment condition while maintaining a near-contact working position and synchronous rotation. The fifth growth conditions are: temperature 560℃, V / III ratio 270, and growth time 20 min.

[0054] In the sixth growth cycle, the mask holder is rotated independently by 60° relative to the substrate again, aligning the strip array openings with the sixth 60° sector. Selective epitaxial growth is then performed under the sixth adjustment conditions in a near-contact working position and synchronous rotation. The sixth growth conditions are: temperature 570℃, V / III ratio 300, and growth time 20 minutes. After six cycles, six physically distinguishable 60° sector strip regions are formed on the same substrate, each corresponding to different growth adjustment parameters. Because the six regions are located on the same substrate, and the mask holder and substrate holder can rotate synchronously during each growth cycle, batch variations between different wafers can be reduced, and the comparability between different growth adjustment conditions can be improved.

[0055] In-situ multimodal characterization: Throughout the growth process, the mask support is kept under highly precise control, maintaining a close contact gap of approximately 0.3 mm between the mask and the substrate growth surface. The RHEED electron beam is incident at a low grazing angle, preferably 2°, and reaches the substrate exposure surface through the mask strip array opening, the close contact gap, and the reserved electron beam clearance channel. This enables continuous monitoring of the surface reconstruction dynamics. The electron beam clearance channel is defined by the mask opening position, the mask support clearance space, and the recessed structure of the substrate support, in order to reduce the obstruction of the RHEED incident path by the mask support or the mask edge.

[0056] Simultaneously, the Ψ / Δ spectrum and band edge reflectance curves of the spectral ellipsometer are acquired in real time through a fixed optical window in the cavity. The spectral ellipsometer is used to obtain information related to thickness, composition, or growth rate during the selected area growth process. The band edge characterization is used to obtain information on band gap changes, reflectance changes, or optical response changes. The RHEED image sequence, ellipsometric spectrum sequence, and band edge curve are all recorded with a unified timestamp and correlated with parameters such as mask pattern ID, mask rotation angle, region number, growth temperature, V / III ratio, and growth time.

[0057] During each region switch, the mask support rises to a safe switching height and completes independent rotation. If necessary, a short RHEED snapshot can be taken before and after the switch while maintaining the mask at a safe height or intermediate observation height to supplement surface state information at the moment of switch. After the snapshot is completed, the mask support descends to a position close to the working area to continue the next round of growth. All in-situ data are aligned and stitched according to timestamps to form a continuous record of the multimodal growth process.

[0058] Growth Completion and Region Division: After all six growth cycles are completed, the mask holder lifts the mask to a safe removal height of approximately 10 mm from the substrate growth surface. A transfer robot enters from the side or lower side of the cavity and supports the mask, transferring it from the mask holder to the transfer robot for removal. The substrate is then unloaded. After unloading, the sample is region-divided on the wafer coordinate system using ellipsometry, optical microscopy, PL mapping, AFM, or SEM characterization. For the strip array pattern mask in this embodiment, the same wafer can be automatically divided into six 60° sector strip regions. Each region corresponds to one growth cycle and a set of growth adjustment parameters. The corresponding RHEED image sequence, ellipsometry sequence, band edge reflectance curve, and subsequent offline characterization results are extracted to form a structured multimodal dataset.

[0059] Example 3: This example provides a dual-support selected region epitaxial multimodal data fusion growth method. The epitaxial equipment used in this embodiment is a commercial MOCVD system, the growth material is InGaAs quantum dots, the substrate is a 2-inch GaAs (001) wafer, and the back of the substrate is processed with an annular anti-rotation structure with a convex bond height of 0.3 mm and a width of 2 mm, which is uniformly distributed along the circumference. The annular anti-rotation structure is fully engaged with the groove on the top of the substrate support to prevent the substrate from shifting at an angle relative to the substrate support during rotation.

[0060] In this embodiment, the substrate is pre-fixed on a substrate support or molybdenum holder, and the upper surface of the substrate is the epitaxial growth surface. The epitaxial source beam or reactive gas is incident from above or above the side of the substrate growth surface and epitaxial growth occurs after reaching the exposed surface of the substrate. The working position of the mask is located above the substrate growth surface and in the propagation path of the epitaxial source beam or reactive gas to the substrate surface. The mask allows the epitaxial source beam or reactive gas to reach the exposed area of ​​the substrate through its circular dot matrix opening area, and blocks the epitaxial source beam or reactive gas from reaching the non-selected area through its blocking area, thereby realizing the selective epitaxy of the dot matrix pattern.

[0061] Equipment structural parameters: The substrate support adopts a sunken design, with the entire molybdenum substrate holder sunk within the support groove. The upper edge of the support is lower than the substrate growth surface to reduce the obstruction of the RHEED grazing electron beam by the support edge. The substrate support retains its original rotation function and can drive the substrate to rotate stably at a set angular velocity during growth. The mask support is an independent peripheral annular support module, located on the periphery of the substrate support and above the substrate growth surface. The mask support includes an annular support part, an annular support groove, a positioning structure, a Z-axis lifting mechanism, and a rotation drive mechanism. The mask support has independent rotation function, synchronous rotation function with the substrate support at the same angular velocity, and precise Z-axis height adjustment function. The Z-axis travel range is 0–15 mm, and the adjustment accuracy is ±0.01 mm.

[0062] The annular support is used to support the mask from its outer periphery and ensure that the central pattern area of ​​the mask covers the substrate growth surface. The mask holder has clearance spaces in the RHEED incident direction and optical characterization path, allowing the RHEED electron beam, ellipsoidized beam, or edge-characterization beam to reach the substrate exposure area through the corresponding channels. The mask material is a 200μm thick SiN sheet, but SiO2, SiC, pyrolytic BN, Mo, Ta, or other high-temperature resistant, low-outgas, and thermally compatible materials can also be selected depending on the epitaxial environment. The outer periphery of the mask is provided with positioning edges, pick-and-place notches, or fork slots. The positioning edges cooperate with the annular support slots of the mask holder to achieve coaxial and rotational positioning of the mask.

[0063] The pick-and-place notch or fork slot is used for the transfer robot to lift, position, and retract the mask, and to allow for the slight thermal expansion of the mask in the thermal environment and the mechanical assembly error. The pick-and-place notch or fork slot is not used as the epitaxial selection pattern, nor is it limited to matching the protruding structure of the substrate holder or heater holder. An additional mask storage compartment is set in the growth cavity. The mask storage compartment is located in the side or lower side area of ​​the cavity that can be reached by the transfer robot, and forms a preset transfer path between it and the receiving position of the mask holder. The transfer path does not pass through the substrate growth area, and the mask is not required to pass through the substrate holder or heater holder.

[0064] The storage compartment can pre-load multiple photomasks with different circular dot matrix pattern parameters, such as the first photomask, the second photomask, and the third photomask. The storage compartment can adopt a vertical slot type, a horizontal slot type, or a turntable type structure. Each storage position is equipped with a limiting structure that cooperates with the positioning edge of the outer periphery of the photomask to ensure the stability and repeatability of the photomask in the storage state. The transfer robot can transfer the photomask between the storage compartment and the photomask support, realizing in-situ quick replacement without opening the cavity.

[0065] Mask Pattern: This embodiment uses a circular dot matrix pattern mask. The circular dot matrix pattern can be composed of multiple micro-nano-scale circular openings. The diameter of the circular openings and the array spacing are determined according to the target material system, mask processing capability, epitaxial material diffusion length, and target nucleation density. For example, different dot matrix spacings can be set on three masks: the dot matrix spacing of the first mask is 200 nm, the dot matrix spacing of the second mask is 300 nm, and the dot matrix spacing of the third mask is 400 nm. The diameter of the circular openings can be set to nanometer or submicrometer size according to the processing conditions to adjust the local nucleation density, quantum dot distribution, or nanostructure spacing in different selected areas. The area of ​​the mask other than the circular dot matrix openings is the shielding area, used to block the epitaxial source beam or reactive gas from reaching the non-selected area. It should be noted that the circular dot matrix pattern in this embodiment is used to illustrate that the mask pattern parameters are replaceable and adjustable; the specific diameter of the circular holes, array spacing, opening coverage ratio, and pattern distribution range can be adjusted according to the material system and process objectives, and are not limited to covering the entire surface of the substrate, including the following steps: The first mask is transferred during the first growth cycle, in region 1, under the first adjustment conditions. The substrate is pre-loaded and fixed on a substrate support or molybdenum holder, with the upper surface of the substrate serving as the epitaxial growth surface. The transfer robot first moves to the mask storage compartment, grabs or lifts the first circular dot pattern mask from the first storage position, and then moves from the side or lower side of the cavity along a preset transfer path to deliver the first mask to the mask support receiving position located above the substrate growth surface.

[0066] During the transfer process, the positioning edge of the outer periphery of the mask aligns with the annular support groove of the mask holder. The pick-and-place notch or fork slot provides space for the transfer robot to lift, position, and retract. The mask holder supports the first mask at a safe receiving height of 5–10 mm from the substrate growth surface, with a safe receiving height of 8 mm. Subsequently, the transfer robot descends or retracts laterally, transferring the first mask from the robot to the support of the mask holder. After the handover, the mask holder lowers the first mask along the Z-axis to a close-contact working position 0.2–0.5 mm above the substrate growth surface, with a gap of 0.3 mm between the mask and the substrate growth surface.

[0067] At this point, the mask does not physically contact the substrate, but forms a near-contact shielding structure above the substrate growth surface, allowing the epitaxial source beam or reactive gas to reach the substrate exposure area only through the circular lattice opening of the first mask. Subsequently, the mask holder and substrate holder rotate synchronously at the same angular velocity, at a speed of 10 rpm. During synchronous rotation, the circular lattice opening of the mask and the selected area of ​​the substrate remain fixed in relative position, while the mask and substrate as a whole rotate relative to the epitaxial source beam or reactive gas flow field, thereby improving the growth uniformity within the selected lattice area. Selective InGaAs quantum dot growth under the first conditioning conditions is performed by introducing TMI, TMI, and AsH3 source gases, setting the growth temperature to 580℃, the V / III ratio to 80, and the growth time to 12 min.

[0068] The first mask is retrieved and the second mask is switched, in area 2, under the second adjustment condition. After the first growth cycle ends, the corresponding growth source is stopped or adjusted, causing the mask holder to lift the first mask to a safe switching height of approximately 8 mm from the substrate growth surface. This safe switching height is used to prevent the mask from scratching or colliding with the substrate surface, molybdenum support, or support structure during the replacement process. The transfer robot enters the receiving position of the mask holder from the side or below the cavity and supports the first mask. Subsequently, the mask holder releases its support for the first mask or lowers to avoid it, transferring the first mask from the mask holder to the transfer robot. The transfer robot then returns the first mask to the first storage position in the mask storage compartment along a preset transfer path.

[0069] The transfer robot then moves to the second storage position, grasps or lifts the second circular dot pattern mask, and delivers it to the safe receiving height of the mask holder along the same lateral or slightly downward transfer path. After the mask holder supports the second mask, the transfer robot retracts. The mask holder then lowers the second mask to a close-contact working position 0.3 mm above the substrate growth surface, ensuring that the epitaxial source beam or reactive gas reaches the substrate exposure area only through the circular dot pattern openings of the second mask.

[0070] After the mask support and substrate support re-enter a state of synchronous rotation at the same angular velocity, the growth parameters are adjusted to a temperature of 590℃, a V / III ratio of 100, and a growth time of 12 minutes for selected InGaAs quantum dot growth under the second adjustment conditions. Because the second mask has a different circular dot spacing or opening parameters, the nucleation density, quantum dot distribution, or local epitaxial structure in the second adjustment region can differ from those in the first adjustment region.

[0071] The second mask is retrieved, and the third mask is switched in area 3, with the third adjustment condition. After the second growth cycle, the mask holder raises the second mask to a safe switching height of approximately 8 mm from the substrate growth surface. A transfer robot enters from the side or lower side of the cavity and supports the second mask, transferring it from the mask holder to the transfer robot. The transfer robot then places the second mask back into the second storage position of the mask storage compartment along a preset transfer path.

[0072] Next, the robotic arm picks up or lifts the third circular dot pattern mask from the third storage location and delivers it to the safe receiving height of the mask holder. After the mask holder supports the third mask, the robotic arm retracts, and the mask holder then lowers the third mask to a near-contact working position 0.3 mm above the substrate growth surface, forming a third type of dot pattern selective masking without contacting the substrate. Subsequently, the mask holder and substrate holder rotate synchronously at the same angular velocity, adjusting the growth parameters to a temperature of 600℃, a V / Ⅲ ratio of 120, and a growth time of 12 minutes, to perform selective InGaAs quantum dot growth under the third adjusted conditions. The entire mask switching process is completed inside the cavity, without opening the growth cavity or disrupting the growth environment. The time for a single mask change can be determined based on the robotic arm's movement speed, the storage compartment position, and the holder's lifting stroke, and can be controlled within approximately 3 minutes.

[0073] In-situ multimodal characterization: Throughout the growth and mask switching process, the mask support is maintained with high precision. During the growth stage, a close-contact gap of approximately 0.3 mm is maintained between the mask and the substrate growth surface. The RHEED electron beam is incident at a low grazing angle, preferably 2°, and reaches the substrate exposure surface through the circular lattice opening, the close-contact gap, and the reserved electron beam clearance channel, enabling continuous monitoring of surface reconstruction dynamics. The electron beam clearance channel is defined by the mask opening position, the mask support clearance space, and the recessed structure of the substrate support, in order to reduce the obstruction of the RHEED incident path by the mask support or the mask edge.

[0074] Simultaneously, the spectral ellipsometer Ψ / Δ spectrum and band edge reflectance curve are acquired in real time through a fixed optical window in the cavity. The spectral ellipsometer is used to obtain information related to thickness, composition, or growth rate during the selected area growth process. The band edge characterization is used to obtain information on band gap changes, reflectance changes, or optical response changes. The RHEED image sequence, ellipsometric spectrum sequence, and band edge curve are all recorded with a unified timestamp and correlated with parameters such as mask pattern ID, mask number, circular lattice parameters, region number, growth temperature, V / III ratio, and growth time.

[0075] Each time the mask is changed, the mask holder rises to a safe switching height, and a transfer robot retrieves the old mask and transfers the new one. If necessary, a short RHEED scan can be performed before and after the change, maintaining the mask at a safe height or intermediate observation height to supplement surface state information before and after the mask change. After the scan, the mask holder descends to a near-contact working position to continue the next round of growth. All in-situ data are aligned and stitched together according to timestamps to form a continuous record of the multimodal growth process.

[0076] Growth Completion and Region Division: After all three mask switching and growth cycles are completed, the mask holder lifts the last mask to a safe removal height of approximately 10 mm from the substrate growth surface. A transfer robot enters from the side or lower side of the cavity, supports the last mask, transfers it from the mask holder to the transfer robot, and then places it back into the mask storage compartment. Subsequently, the substrate is unloaded.

[0077] After unloading the substrate, the sample is region-divided on the wafer coordinate system using ellipsometry, optical microscopy, PL mapping, AFM, or SEM characterization. For the multiple circular dot-matrix pattern masks in this embodiment, the same wafer can be divided into three regions corresponding to different mask pattern parameters and growth adjustment conditions based on the mask number, dot-matrix parameters, timestamp, and growth parameters. Each region corresponds to one growth cycle and a set of growth adjustment parameters. The RHEED image sequence, ellipsometry spectrum sequence, band edge reflectance curve, and subsequent offline characterization results corresponding to each region are extracted to form a structured multimodal dataset. It should be noted that when the pattern coverage positions of different circular dot-matrix masks are independent, multiple spatially distinguishable selected regions can be formed; when the pattern positions of different circular dot-matrix masks partially overlap, pattern superposition regions can also be formed. The specific spatial partitioning mode or pattern superposition mode can be determined according to the material growth target, mask pattern design, and subsequent dataset construction requirements.

[0078] Example 4: This example provides a dual-support selected region epitaxial multimodal data fusion growth method. The epitaxial equipment used in this embodiment is a commercial MOCVD system, the growth material is InGaAs quantum dots, and the substrate is a 2-inch GaAs (001) wafer. A ring-shaped anti-rotation structure is fabricated on the back side of the substrate, with a bond height of 0.3 mm and a width of 2 mm, evenly distributed circumferentially. The ring-shaped anti-rotation structure fully engages with the groove on the top of the substrate support to prevent angular displacement of the substrate relative to the substrate support during rotation. In this embodiment, the substrate is pre-fixed to the substrate support or molybdenum holder, and the upper surface of the substrate is the epitaxial growth surface. The epitaxial source beam or reactive gas is incident from above or slightly above the substrate growth surface and epitaxial growth occurs after reaching the exposed surface of the substrate. The mask is positioned above the substrate growth surface and in the propagation path of the epitaxial source beam or reactive gas reaching the substrate surface. The mask allows the epitaxial source beam or reactive gas to reach the exposed area of ​​the substrate through its opening area and blocks the epitaxial source beam or reactive gas from reaching the non-selected area through its blocking area, thereby achieving selected epitaxy.

[0079] Equipment structural parameters: The substrate support adopts a sunken design, with the entire molybdenum substrate holder sunk within the support groove. The upper edge of the support is lower than the substrate growth surface to reduce the obstruction of the RHEED grazing electron beam by the support edge. The substrate support retains its original rotation function and can drive the substrate to rotate stably at a set angular velocity during growth. The mask support is an independent peripheral annular support module, located on the periphery of the substrate support and above the substrate growth surface. The mask support includes an annular support section, an annular support groove, a positioning structure, a Z-axis lifting mechanism, and a rotation drive mechanism.

[0080] The mask holder features independent rotation, synchronous rotation at the same angular velocity as the substrate holder, and precise Z-axis height adjustment with a travel range of 0–15 mm and an adjustment accuracy of ±0.01 mm. An annular support supports the mask from its outer periphery, ensuring that the central opening of the mask covers the substrate growth surface. The mask holder incorporates an electron beam clearance channel along the RHEED electron beam incident direction. This clearance channel is defined by the mask opening position, the clearance space within the mask holder, the close-contact gap between the mask and the substrate, and the recessed structure of the substrate holder, enabling the RHEED electron beam to reach the substrate exposure surface through the mask opening at a low grazing angle.

[0081] The height, width, and circumferential position of the electron beam clearance channel are determined based on the RHEED incident azimuth, grazing angle, mask opening boundary, and effective characterization area of ​​the substrate. The RHEED grazing angle is 2°, and the mask material is a 200μm thick SiN sheet. Alternatively, SiO2, SiC, pyrolytic BN, Mo, Ta, or other high-temperature resistant, low-outgas, and thermally compatible materials can be selected depending on the epitaxial environment. Positioning edges, pick-and-place notches, or fork slots are provided on the outer periphery of the mask. The positioning edges cooperate with the annular support groove of the mask holder to achieve coaxial and rotational positioning of the mask. The pick-and-place notches or fork slots are used for lifting, positioning, and retraction by a transfer robot, and allowance is provided for minor thermal expansion and mechanical assembly errors of the mask under thermal conditions. The pick-and-place notches or fork slots are not used as epitaxial selection patterns, nor are they limited to matching the protruding structures of the substrate holder or heater holder.

[0082] The mask pattern uses a 1 / 4 fan-shaped opening, with the center of the opening coaxial with the center of the substrate. The opening angle is 90°, and the opening area covers a quarter-circle area of ​​the substrate surface. There are no other obstructing structures within the opening. The area of ​​the mask other than the 1 / 4 fan-shaped opening is the blocking area, used to prevent the epitaxial source beam or reactive gas from reaching the non-selected area. The process includes the following steps: Mask transfer and initial near-contact positioning stage: The substrate is pre-loaded and fixed on the substrate holder or molybdenum support, with the upper surface of the substrate serving as the epitaxial growth surface. The transfer robot uses a fork-shaped or tray-type insert structure to enter from the side or lower side of the cavity, delivering the 1 / 4 fan-shaped opening mask to the mask holder receiving position located above the substrate growth surface.

[0083] During the transfer process, the positioning edge of the mask's outer periphery aligns with the annular support groove of the mask holder. The pick-and-place notch or fork slot provides space for the transfer robot to lift, position, and retract. The mask holder supports the mask at a safe receiving height of 5–10 mm from the substrate growth surface, with a safe receiving height of 8 mm. Subsequently, the transfer robot descends or retracts laterally, transferring the mask from the robot to the support of the mask holder.

[0084] After the handover is completed, the mask holder drives the mask to descend along the Z-axis to a close-contact working position 0.2–0.5 mm above the substrate growth surface. The gap between the mask and the substrate growth surface is 0.3 mm. At this time, the mask does not make physical contact with the substrate, but forms a close-contact shielding structure above the substrate growth surface, so that the epitaxial source beam or reactive gas can only reach the substrate exposure area through the 1 / 4 fan-shaped opening of the mask.

[0085] After the mask is lowered to the near-contact working position, the system verifies the correspondence between the mask opening, the mask holder clearance channel, and the substrate exposure area according to the RHEED incident direction, so that the RHEED electron beam can reach the substrate exposure surface through the opening area at the set grazing angle. If necessary, edge occlusion can be eliminated by small angle compensation or height fine adjustment of the mask holder.

[0086] In the first growth cycle, continuous in-situ characterization was performed. The mask holder and the substrate holder rotated synchronously at the same angular velocity of 10 rpm. During the synchronous rotation, the mask opening and the selected area of ​​the substrate remained fixed in relative position. At the same time, the mask and the substrate as a whole rotated relative to the epitaxial source beam or the reactive gas flow field, thereby improving the growth uniformity in the selected area.

[0087] Subsequently, TMI, TMI, and AsH3 source gases were introduced, with the growth temperature set at 580℃, V / III ratio at 80, and growth time at 15 min. The epitaxial source beam or reactive gas reached the substrate exposure area through a 1 / 4 fan-shaped opening in the mask, and selected InGaAs quantum dot growth was performed in the first quadrant under the first adjustment conditions. In the first growth cycle, the RHEED electron beam was incident at a low grazing angle, preferably 2°, and reached the substrate exposure surface through the fan-shaped opening in the mask, the close contact gap between the mask and the substrate, and the reserved electron beam clearance channel. Surface reconstruction, surface flatness, and changes in growth mode were continuously monitored during the epitaxial process. RHEED images were continuously acquired at a set frame rate and recorded synchronously with the growth temperature, source flux, V / III ratio, mask angle, and region number.

[0088] Simultaneously, the cavity's fixed optical window acquires the Ψ / Δ spectrum and band edge reflectance curves of the spectral ellipsometer in real time. The spectral ellipsometer is used to obtain information related to thickness, composition, or growth rate during the selected area growth process. The band edge characterization is used to obtain information on band gap changes, reflectance changes, or optical response changes. The RHEED image sequence, ellipsometric spectrum sequence, and band edge curve are all recorded according to a unified timestamp, thus forming in-situ multimodal time series data of the first growth region.

[0089] Ensuring continuity of characterization during mask switching: After the first growth is completed, the corresponding growth source is stopped or adjusted so that the mask holder drives the mask to a safe switching height of about 8 mm from the substrate growth surface. The safe switching height is used to avoid the mask from scratching or colliding with the substrate surface, molybdenum support or holder structure during the rotation switching process.

[0090] The mask holder is rotated 90° independently relative to the substrate at a safe switching height, switching the 1 / 4 fan-shaped opening from the first region to the second region. After rotation, the mask holder descends along the Z-axis again, positioning the mask back in a close-contact working position 0.3 mm above the substrate growth surface. The growth parameters are then adjusted to a temperature of 590℃ and a V / III ratio of 100, and the mask holder and substrate holder are brought back into a state of synchronous rotation at the same angular velocity. Selective epitaxial growth under the second adjustment conditions is performed in the second region for 15 minutes.

[0091] During the mask's ascent, independent rotation, and descent, RHEED, ellipticity, and band edge characterization data are continuously recorded according to timestamps. For RHEED, since the mask height and opening position change instantaneously during switching, this time period can be marked as a "switching state" or "transition state," distinguishing it from stable growth state data. If necessary, the mask support can briefly pause at a safe switching height or intermediate observation height to perform RHEED quick snapshots or surface state confirmation. After completing the quick snapshot, it descends to a 0.3mm close-contact working position to continue the next round of growth. This ensures that surface state information before and after switching is continuously recorded on the timeline and avoids mistaking transitional data at the moment of switching for stable growth data.

[0092] Continuous characterization in the second to fourth growth cycles: The second, third, and fourth growth cycles were completed sequentially according to the switching method described above. The second growth conditions were a temperature of 590℃ and a V / III ratio of 100; the third growth conditions were a temperature of 600℃ and a V / III ratio of 120; and the fourth growth conditions were a temperature of 610℃ and a V / III ratio of 140. The growth time for each cycle was 15 minutes.

[0093] Before each growth cycle, the mask support is lowered to a near-contact working position approximately 0.3 mm above the substrate growth surface and rotates synchronously with the substrate support at the same angular velocity. During each growth cycle, the mask opening remains relatively fixed to the corresponding substrate region. The RHEED electron beam reaches the exposed substrate surface through the corresponding opening, the near-contact gap, and the electron beam clearance channel, continuously acquiring a sequence of surface reconstruction images of the corresponding region. The spectroscopic ellipsometer and band edge characterization system synchronously acquire the Ψ / Δ spectrum and band edge reflectance curves through a fixed optical window.

[0094] After four cycles, four physically distinguishable 90° sector regions are formed on the same substrate. Each region corresponds to different growth adjustment parameters and has independent RHEED image sequences, ellipsometry sequences, and band edge curve data. Since all regions are located on the same substrate and each growth process is completed under the same equipment conditions, batch differences between different wafers can be reduced, and the comparability between multimodal data can be improved.

[0095] Growth Completion and Multimodal Data Extraction: After all growth is complete, the mask holder raises the mask to a safe removal height of approximately 10 mm from the substrate growth surface. A transfer robot enters from the side or below the cavity, supporting the mask and transferring it from the mask holder to the transfer robot for removal. The substrate is then unloaded. After substrate unloading, the sample is characterized on a wafer coordinate system using ellipsometry, optical microscopy, PL mapping, AFM, or SEM for region division.

[0096] For the 1 / 4 fan-shaped aperture mask in this embodiment, the same wafer can be automatically divided into 4 90° fan-shaped regions. Each fan-shaped region corresponds to one growth cycle and a set of growth adjustment parameters. The RHEED image sequence, elliptic Ψ / Δ spectrum sequence, band edge reflectance curve and subsequent offline characterization results corresponding to the region are extracted to form a structured multimodal dataset.

[0097] During data processing, data from the stable growth stage, mask switching stage, and RHEED snapshots can be labeled separately. The stable growth stage data is used to characterize the continuous growth process of the material; the switching stage data is used to record the transition state caused by the region switching; and the RHEED snapshots are used to supplement the instantaneous reconstruction state of the substrate surface before and after the switching. All of the above data are associated with a unified timestamp, mask angle, region number, and growth parameters.

[0098] Example 5: This example provides a dual-support selected region epitaxial multimodal data fusion growth method. The epitaxial equipment used in this embodiment is a commercial MOCVD system, the growth material is InGaAs quantum dots, and the substrate is a 2-inch GaAs (001) wafer. A ring-shaped anti-rotation structure is fabricated on the back side of the substrate, with a bond height of 0.3 mm and a width of 2 mm, evenly distributed circumferentially. The ring-shaped anti-rotation structure fully engages with the groove on the top of the substrate support to prevent angular displacement of the substrate relative to the substrate support during rotation. In this embodiment, the substrate is pre-fixed to the substrate support or molybdenum holder, and the upper surface of the substrate is the epitaxial growth surface. The epitaxial source beam or reactive gas is incident from above or slightly above the substrate growth surface and epitaxial growth occurs after reaching the exposed surface of the substrate. The mask is positioned above the substrate growth surface and along the propagation path of the epitaxial source beam or reactive gas to the substrate surface. The mask allows the epitaxial source beam or reactive gas to reach the exposed area of ​​the substrate through its opening area and blocks the epitaxial source beam or reactive gas from reaching the non-selected area through its blocking area, thereby achieving selected epitaxy.

[0099] Equipment structural parameters: The substrate support adopts a sunken design, with the entire molybdenum substrate holder sunk within the support groove. The upper edge of the support is lower than the substrate growth surface to reduce the obstruction of the RHEED grazing electron beam by the support edge. The substrate support retains its original rotation function and can drive the substrate to rotate stably at a set angular velocity during growth. The mask support is an independent peripheral annular support module, located on the periphery of the substrate support and above the substrate growth surface. The mask support includes an annular support part, an annular support groove, a positioning structure, a Z-axis lifting mechanism, and a rotation drive mechanism. The mask support has independent rotation function, synchronous rotation function with the substrate support at the same angular velocity, and precise Z-axis height adjustment function. The Z-axis travel range is 0–15 mm, and the adjustment accuracy is ±0.01 mm. The annular support part is used to support the mask from the outer peripheral positioning edge and ensure that the central pattern area of ​​the mask covers the substrate growth surface. The mask support has clearance spaces in the RHEED incident direction and optical characterization path, allowing the RHEED electron beam, ellipsoidal beam, or edge-characterization beam to reach the substrate exposure area through the corresponding channel. The RHEED electron beam is incident at a low grazing angle, preferably 2°, and reaches the exposed surface of the substrate through the mask opening, the close contact gap between the mask and the substrate, and the electron beam clearance channel.

[0100] The photomask material is a 200μm thick SiN sheet, but SiO2, SiC, pyrolytic BN, Mo, Ta, or other high-temperature resistant, low-outgas, and thermally compatible materials can also be selected depending on the epitaxial environment. The photomask has positioning edges, pick-and-place notches, or fork slots on its outer periphery. The positioning edges mate with the annular support groove of the photomask holder to achieve coaxial and rotational positioning of the photomask. The pick-and-place notches or fork slots are used for lifting, positioning, and retraction by a transfer robot, and also allow for tolerances to minor thermal expansion and mechanical assembly errors under thermal conditions. The pick-and-place notches or fork slots are not used as epitaxial selection patterns, nor are they limited to matching the protruding structures of the substrate holder or heater holder.

[0101] This embodiment uses a 1 / 4-fan-shaped aperture mask. The center of the 1 / 4-fan-shaped aperture is coaxial with the center of the substrate, the aperture angle is 90°, and the aperture area covers a quarter-circle area of ​​the substrate surface. By independently rotating the mask holder 90° relative to the substrate each time, four physically distinguishable 90° fan-shaped regions can be sequentially formed on the same substrate, including the following steps: Multimodal data real-time acquisition stage: During the four growth cycles in Example 1, the mask holder keeps the mask at a close-contact working position 0.3 mm above the substrate growth surface in each stable growth stage. The close-contact working position is within the range of 0.2–0.5 mm. The mask does not make physical contact with the substrate, but forms a selected area masking above the substrate growth surface.

[0102] During the growth process, the mask holder and the substrate holder rotate synchronously at the same angular velocity, with a rotation speed of 10 rpm. This synchronous rotation keeps the mask opening and the selected area of ​​the substrate in a fixed relative position, while also causing the mask and substrate as a whole to rotate relative to the epitaxial source beam or reactive gas flow field, thereby improving the growth uniformity within the same selected area.

[0103] The four regions correspond to four growth cycles. Region 1 corresponds to the first growth cycle, with a growth temperature of 580℃ and a V / III ratio of 80; Region 2 corresponds to the second growth cycle, with a growth temperature of 590℃ and a V / III ratio of 100; Region 3 corresponds to the third growth cycle, with a growth temperature of 600℃ and a V / III ratio of 120; and Region 4 corresponds to the fourth growth cycle, with a growth temperature of 610℃ and a V / III ratio of 140. Each growth cycle lasted 15 minutes. Throughout the growth process, surface reconstruction image sequences were continuously acquired using a RHEED system at a low grazing angle of 2°, a sampling frame rate of 30 frames per second, and an image resolution of 1024×1024. The RHEED electron beam reached the exposed substrate surface through a mask opening, a near-contact gap, and an electron beam clearance channel, used to characterize the surface reconstruction, surface smoothness, nucleation process, and growth mode evolution in different regions.

[0104] Simultaneously, spectral ellipsometry Ψ / Δ spectral sequences are acquired in real time through a fixed optical window. Preferably, the ellipsometry acquisition wavelength range is 200–1700 nm, and the sampling frequency is 1 Hz. The ellipsometry data is used to extract information related to thickness, composition, changes in optical constants, or growth rate during the growth process of each region. Furthermore, a band edge reflectance curve or other band edge response curves are acquired in real time through a band edge characterization system. The band edge reflectance curve acquisition wavelength range is 800–1200 nm, and the sampling frequency is 1 Hz. The band edge data is used to extract band gap changes, absorption edge shifts, reflectance changes, or electronic structure-related features.

[0105] All three types of in-situ data are recorded using a unified clock and automatically associated with the mask pattern ID, mask angle, region number, growth temperature, V / III ratio, source flow rate, growth time, synchronous rotation speed, and mask height status. Transitional data during mask rising, independent rotation, and falling are separately labeled "Switching Status Data." Data acquired when the mask is in a 0.3mm close-contact working position and undergoing stable epitaxial growth is labeled "Stable Growth Status Data." Data acquired via short-term RHEED snapshots before and after switching is labeled "Snapshot Status Data." These status labels prevent the misuse of transitional signals during mask switching as stable growth data.

[0106] Growth Completion and Region Division: After all four cycles are completed, the mask holder raises the mask to a safe removal height of approximately 10 mm from the substrate growth surface. A transfer robot enters from the side or lower side of the cavity, supports the mask, and transfers it from the mask holder to the transfer robot for removal. The substrate is then unloaded.

[0107] After unloading the substrate, an ellipsomer is used to scan the entire 2-inch wafer in 1mm increments on the wafer coordinate system to obtain information on wafer surface thickness, optical constants, or compositional distribution. Combining mask rotation angle records and growth timestamps, the same wafer is automatically divided into four 90° sector regions. Each region corresponds to one growth cycle and a specific set of growth conditioning conditions. Region 1: 580℃, V / III=80; Region 2: 590℃, V / III=100; Region 3: 600℃, V / III=120; Region 4: 610℃, V / III=140.

[0108] In addition to elliptic mapping, the region boundary can be corrected by optical microscopy, PL mapping, AFM or SEM characterization results. For the transition region near the edge of the mask, a boundary buffer zone can be set, which is not included in the model training samples, or it can be marked as a boundary sample separately to reduce the impact of mask edge diffusion, occlusion transition and local deposition unevenness on model training.

[0109] Multimodal dataset extraction for each region: For each segmented sector, a complete subset of the corresponding growth cycle is extracted based on a unified timestamp and region number. Each region subset includes in-situ process data, growth control parameters, mask state parameters, and offline performance labels. The first category is RHEED image sequence data. Each region corresponds to approximately 15 minutes of stable growth process. Acquiring images at 30 frames per second, approximately 27,000 RHEED images are obtained per region. If only valid frames from the stable growth phase are selected or time downsampling is performed, approximately 13,500 valid image sequences can be formed. RHEED data is used to extract features such as surface reconstruction stripe intensity, stripe spacing, stripe clarity, spot morphology, surface roughening trend, and nucleation dynamics.

[0110] The second category is ellipsometry Ψ / Δ spectral sequence data. Each region corresponds to a stable growth process of approximately 15 minutes. Sampling at 1 Hz, approximately 900 Ψ / Δ spectral sequences are obtained per region. Ellipsometry data is used to extract thickness growth curves, changes in optical constants, compositional trends, growth rate changes, and equivalent parameters of the surface roughness layer. The third category is band-edge reflectance or band-edge response curve data. Each region corresponds to a stable growth process of approximately 15 minutes. Sampling at 1 Hz, approximately 900 band-edge curves are obtained per region. Band-edge data is used to extract absorption edge positions, reflectance inflection points, bandgap trends, changes in optical response intensity, and electronic structure-related features. The fourth category is growth control parameters and mask status parameters. Growth control parameters include growth temperature, V / III ratio, source flow rate, growth time, rotation speed, and cavity pressure. Mask status parameters include mask pattern ID, mask angle, region number, mask height, whether it is in synchronous rotation state, and whether it is in switching or fast-shot state.

[0111] The fifth category is offline characterization labels. After growth, each region is characterized using SEM, AFM, PL, or other offline methods to obtain quantum dot density, average diameter, height distribution, size uniformity, surface roughness, PL peak position, PL full width at half maximum (FWHM), and bandgap-related parameters. These offline characterization results serve as the final performance labels for model training. This forms a structured sample set. Each region sample includes: a RHEED image sequence, an ellipsometry Ψ / Δ spectrum sequence, a band edge curve, growth parameters, mask state parameters, and the final performance label. In this embodiment, at least four region-level independent samples are formed; simultaneously, the stable growth stage of each region can be further sliced ​​according to a time window, thereby expanding into multiple time-series sub-samples for training time-series prediction models or process state recognition models.

[0112] Data preprocessing and feature extraction: Before model training, time alignment, spatial alignment, and quality screening are performed on data from different modalities. First, the RHEED image sequence, ellipsometry Ψ / Δ spectrum sequence, and band edge curve are time-synchronized based on a unified timestamp. For modalities with different sampling frequencies, resampling, interpolation, or time window aggregation can be used to align them to a unified time axis. For example, using 1 second as the basic time step, 30 frames of RHEED images are aggregated into a single RHEED temporal feature, and the ellipsometry spectrum and band edge curve at the same time point are used as the corresponding optical features.

[0113] Secondly, spatial mapping of region locations is performed based on the wafer coordinate system. The mask rotation angle, the projection area of ​​the opening pattern, and the elliptic mapping coordinates are unified into the same wafer coordinate system, establishing a correspondence of "time period - mask angle - region number - spatial location". For region edge locations, mask occlusion transition zones, or suspected contamination points, rejection rules or individual labels can be set to improve the quality of training data. Thirdly, quality screening is performed on each modality of data. In RHEED images, excessively dark stripes, severe occlusion, unstable electron beams, or frames with switching states can be individually marked or rejected; in elliptic data, signal anomalies and spectral segments with excessively large fitting residuals can be quality labeled; and time periods with excessive noise or significant signal drift in the banded curves can be smoothed, normalized, or anomaly detected.

[0114] After quality screening, features of each modality are extracted. For the RHEED modality, surface reconstruction image features can be extracted using CNN, visual Transformer, or temporal convolutional networks. For the elliptic modality, thickness, composition, growth rate, and optical constant evolution features can be extracted using LSTM, Transformer, or one-dimensional convolutional networks. For the bandgap modality, bandgap, absorption edge displacement, and electronic structure-related features can be extracted using MLP, one-dimensional convolutional networks, or sequence models.

[0115] Feature-level fusion model construction and training: This embodiment uses a feature-level multimodal fusion model for data fusion. First, surface reconstruction feature vectors are extracted from the RHEED image sequence. Then, a CNN or visual Transformer is used to extract a 512-dimensional image representation vector from the RHEED image sequence. This vector represents surface reconstruction, surface roughening, nucleation dynamics, and changes in growth patterns.

[0116] Secondly, optical evolution feature vectors are extracted from the ellipsoidal Ψ / Δ spectral sequence. For example, LSTM, Transformer, or a one-dimensional convolutional network can be used to extract ellipsoidal feature vectors with a dimension of 256, which characterize thickness growth, compositional changes, optical constant changes, and growth rate evolution. Thirdly, electronic structure-related feature vectors are extracted from the bandgap reflectance curves. For example, MLP or a one-dimensional convolutional network can be used to extract bandgap feature vectors with a dimension of 128, which characterize bandgap changes, absorption edge shifts, reflectance inflection points, and electronic structure changes.

[0117] The three types of feature vectors are interactively fused through a Cross-Attention mechanism, enabling key features in one modality to modulate the feature selection of another modality. For example, changes in RHEED surface reconstruction can serve as a query vector, guiding the model to focus on the corresponding thickness growth changes in the elliptic spectrum; the absorbing edge displacement in the banded curve can inversely modulate the model's weight allocation for RHEED nucleation states. Through this cross-modal attention mechanism, the model generates a unified joint feature vector, the dimension of which can be 896 or further compressed depending on the specific network structure.

[0118] The fused joint feature vector is input into a multi-task prediction network or a multi-task Transformer decoder, simultaneously outputting multiple predicted material properties. Prediction targets include, but are not limited to, average quantum dot size, quantum dot density, size uniformity, surface roughness, PL peak position, PL full width at half maximum (FWHM), and bandgap-related parameters. The model loss function may include mean squared error loss, contrast loss, region consistency loss, and physical constraint loss. The mean squared error loss is used to constrain predicted values ​​to closely approximate offline characterization labels; the contrast loss is used to enhance the separability between regions under different growth conditions; the region consistency loss is used to constrain the continuity of features in adjacent time windows within the same region; and the physical constraint loss is used to limit the prediction results from contradicting known growth patterns, such as the reasonable relationship between temperature, V / III ratio, and growth rate or bandgap changes.

[0119] During model training, a combination of region-level samples and time-window-level samples can be used. Region-level samples are used to learn the correspondence between the final material properties and the complete growth process; time-window-level samples are used to learn the state evolution law during the growth process. Since the four regions on the same wafer in this embodiment share the same substrate, the same equipment state, and similar environmental conditions, the differences between different regions mainly come from the artificially set growth adjustment parameters, which helps to reduce the interference of batch fluctuations on model training.

[0120] Fusion Result Validation: After model training, model performance is evaluated using leave-out region, cross-validation, or time window validation. For the average diameter of quantum dots, the model prediction error can be less than 4 nm, and the relative error less than 5%; for quantum dot density, the model prediction error can be less than 0.2 × 10⁻⁶. 10 cm -2 For parameters related to band gap or PL peak position, the prediction error can be less than 8 meV. Overall prediction R 2 The value can reach 0.97.

[0121] Further comparison of the contributions of different modes to the prediction results reveals that when using the RHEED mode alone, the model is sensitive to surface reconstruction and nucleation states, but has limited ability to predict the final optical properties. When using the ellipticization mode alone, the model can well characterize thickness and composition variations, but is insufficient in recognizing local nucleation morphology. When using the band-edge mode alone, the model can reflect changes in optical response, but lacks direct characterization of the surface growth process. After feature-level fusion, the model can simultaneously utilize the surface structure information from RHEED, the thickness and composition information from ellipticization, and the electronic structure information characterized by the band edge, thus significantly improving the prediction accuracy compared to the single-mode model.

[0122] Meanwhile, by analyzing the consistency of feature distribution in different regions within the same wafer, it can be verified that the dataset in this embodiment has lower batch fluctuations compared to traditional multi-wafer experiments. Since each region shares the same substrate, the same equipment environment, and the continuous growth process, the model is more likely to learn the true mapping relationship between "growth adjustment parameters - process state - final performance" rather than learning the systematic errors between different wafer batches.

[0123] Example 6: This example provides a dual-support selected region epitaxial multimodal data fusion growth method. The epitaxial device used in this embodiment is a commercial solid-source MBE system, and the vacuum level of the growth chamber is maintained at 5 × 10⁻⁶. -10 The growth material is InGaAs quantum dots, and the substrate is a 2-inch GaAs (001) wafer. The back of the substrate is processed with an annular anti-rotation structure with a convex bond height of 0.3 mm and a width of 2 mm, which is uniformly distributed along the circumference. The annular anti-rotation structure is fully engaged with the groove on the top of the substrate support to prevent the substrate from shifting at an angle relative to the substrate support during rotation.

[0124] In this embodiment, the substrate is pre-fixed on a substrate support or molybdenum holder. The upper surface of the substrate is the epitaxial growth surface. In, Ga, and As molecular beams generated by the MBE source furnace arrive at the substrate surface from above and to the side of the substrate growth surface at an oblique incidence, and epitaxial growth occurs in the exposed area of ​​the substrate. The working position of the mask is located above the substrate growth surface and in the propagation path of the molecular beams to the substrate growth surface. The mask allows the molecular beams to reach the exposed area of ​​the substrate through its opening area, and blocks or weakens the molecular beams from reaching the non-selected area through its blocking area, thereby realizing selective epitaxy under MBE conditions.

[0125] Equipment structural parameters: The substrate support adopts a sunken design, with the entire molybdenum substrate holder sunk within the support groove. The upper edge of the support is lower than the substrate growth surface to reduce the obstruction of the RHEED grazing electron beam by the support edge. The substrate support retains its original rotation function and can drive the substrate to rotate stably at a set angular velocity during growth. The mask support is an independent peripheral annular support module, located on the periphery of the substrate support and above the substrate growth surface. The mask support includes an annular support part, an annular support groove, a positioning structure, a Z-axis lifting mechanism, and a rotation drive mechanism. The mask support has independent rotation function, synchronous rotation function with the substrate support at the same angular velocity, and precise Z-axis height adjustment function. The Z-axis travel range is 0–15 mm, and the adjustment accuracy is ±0.01 mm.

[0126] The annular support section is used to support the mask from its outer periphery and ensure that the central opening area of ​​the mask covers the substrate growth surface. The mask holder, Z-axis lifting mechanism, rotary drive mechanism, positioning structure, fasteners, and all exposed components are made of ultra-high vacuum compatible materials and structural designs. Moving parts preferably use low outgassing, high temperature resistance, oil-free lubrication or solid lubrication compatible solutions to avoid introducing organic contamination, particulate contamination or outgassing contamination in the MBE ultra-high vacuum environment. The mask holder and the mask itself undergo high-vacuum baking treatment before use to meet the cleanliness and vacuum requirements of the MBE growth chamber. The mask holder is provided with an electron beam clearance channel in the RHEED incident direction. The electron beam clearance channel is defined by the mask opening position, the clearance space of the mask holder, the close contact gap between the mask and the substrate, and the recessed structure of the substrate holder, so that the RHEED electron beam can reach the substrate exposure surface through the mask opening at a low grazing angle of 2°.

[0127] The height, width, and circumferential orientation of the electron beam clearance channel are determined based on the RHEED gun, substrate center, phosphor screen position, and mask opening boundary. The mask material is a 200μm thick SiN sheet, but SiO2, SiC, pyrolytic BN, Mo, Ta, W (tungsten), or other ultra-high vacuum compatible, high temperature resistant, low outgassing, and low-pollution materials can also be selected depending on the MBE environment. Positioning edges, pick-and-place notches, or fork slots are provided on the outer periphery of the mask. The positioning edges cooperate with the annular support groove of the mask holder to achieve coaxial and rotational positioning of the mask.

[0128] The pick-and-place notch or fork slot is used for the transfer robot to lift, position, and retract the mask, and to allow for minor thermal expansion and mechanical assembly errors under thermal conditions. The pick-and-place notch or fork slot is not used as the epitaxial selection pattern, nor is it required to match the protruding structure of the substrate support or heater support. The mask pattern uses a 1 / 4 fan-shaped opening, with the center of the 1 / 4 fan-shaped opening coaxial with the center of the substrate. The opening angle is 90°, and the opening area covers a quarter-circle of the substrate surface. There are no other obstructing structures within the opening. The area of ​​the mask other than the 1 / 4 fan-shaped opening is the blocking area, used to prevent In, Ga, and As molecular beams from reaching the non-selected area, including the following steps: Mask transfer stage: The substrate is pre-loaded and fixed on the substrate support or molybdenum support. The upper surface of the substrate serves as the epitaxial growth surface. The transfer robot adopts an ultra-high vacuum compatible fork-shaped or tray-type insert structure, enters from the side or lower side of the cavity, and delivers the 1 / 4 fan-shaped opening mask to the mask support receiving position located above the substrate growth surface.

[0129] During the transfer process, the positioning edge of the outer periphery of the mask is aligned with the annular support groove of the mask holder. The pick-and-place notch or fork slot provides space for the transfer robot to lift, position, and withdraw. The mask holder supports the mask at a safe receiving height of 5–10 mm from the substrate growth surface, with a safe receiving height of 8 mm. Subsequently, the transfer robot descends or withdraws laterally, transferring the mask from the transfer robot to the support of the mask holder.

[0130] The entire mask transfer process is completed in the MBE ultra-high vacuum environment. It does not require opening the growth chamber or disrupting the vacuum state of the growth chamber. During the transfer process, a safe gap is always maintained between the mask and the substrate growth surface to prevent the mask from colliding with the substrate, molybdenum support, heater, or support structure.

[0131] After the handover is completed, the mask holder drives the mask to descend along the Z-axis to a close-contact working position 0.2–0.5 mm above the substrate growth surface. The gap between the mask and the substrate growth surface is 0.3 mm. At this time, the mask does not make physical contact with the substrate, but forms a close-contact shielding structure above the substrate growth surface, so that the In, Ga and As molecular beams reach the substrate exposure area only through a 1 / 4 fan-shaped opening.

[0132] First growth cycle, Region 1, First adjustment conditions. The mask holder is precisely lowered so that the mask is positioned 0.3 mm above the substrate growth surface in a near-contact working position, forming a selected area masking without contacting the substrate. The mask holder and substrate holder rotate synchronously at the same angular velocity, at a speed of 8 rpm. During synchronous rotation, the mask opening and the selected area on the substrate remain fixed in relative position, while the mask and substrate as a whole rotate relative to the MBE source furnace beam distribution, thereby improving the growth uniformity within the selected area.

[0133] Subsequently, the In, Ga, and As pyrolysis source furnace or solid-state beam source is turned on, allowing the In, Ga, and As molecular beams to enter obliquely from the upper side and pass through a 1 / 4 fan-shaped opening in the mask to reach the exposed substrate surface. Under the first adjustment conditions, the beam equivalent pressure ratio is set to In / Ga = 0.8 and As BEP = 1.2 × 10⁻⁶. -6 Torr, with a growth temperature of 500℃ and a growth time of 10 min, under these conditions, In, Ga and As molecular beams form selected InGaAs quantum dot growth in the first quadrant exposed region. The substrate surface corresponding to the mask-covered region does not undergo epitaxial growth or only undergoes negligible weak deposition due to the mask's obstruction.

[0134] Switch to the second growth cycle, region 2, second adjustment conditions. After the first growth cycle, close or adjust the corresponding source furnace shutter to raise the mask holder to a safe switching height of approximately 8 mm from the substrate growth surface. This safe switching height prevents the mask from scratching or colliding with the substrate surface, molybdenum support, or holder structure during rotation and reduces the risk of thermomechanical interference between the mask and substrate at high temperatures. The mask holder rotates independently 90° relative to the substrate from this safe switching height, switching the 1 / 4 fan-shaped opening from the first region to the second region. After rotation, the mask holder descends again along the Z-axis, positioning the mask back in a close-contact working position 0.3 mm above the substrate growth surface. The growth conditions are then adjusted as follows: growth temperature 510℃, In / Ga BEP = 1.0, As BEP = 1.5 × 10⁻⁶. -6 Torr, growth time 10 min. After the mask scaffold and substrate scaffold re-enter the synchronous rotation state at the same angular velocity, selected InGaAs quantum dot growth under the second adjustment condition is carried out in the second region.

[0135] The third and fourth growth cycles repeat the above steps of "rising to the safe switching height, rotating independently by 90° relative to the substrate, descending to the near-contact working position, and synchronously rotating for growth," sequentially completing the third and fourth selected-area epitaxial growth. During the third growth cycle, the mask holder is rotated independently by 90° relative to the substrate again at the safe switching height, aligning the mask opening with the third region. The mask is then lowered to a near-contact working position 0.3 mm above the substrate growth surface, and selected-area epitaxial growth is performed under the third adjustment conditions while in synchronous rotation. The third growth conditions are: growth temperature 520℃, In / Ga BEP = 1.2, As BEP = 1.8 × 10⁻⁶. -6 Torr, growth time 10 min.

[0136] During the fourth growth stage, the mask holder is raised to the safe switching height again and rotated 90° independently relative to the substrate to align the mask opening with the fourth region. The mask is then lowered to a near-contact working position 0.3 mm above the substrate growth surface, and selected area epitaxial growth is performed under the fourth adjustment conditions while rotating synchronously. The fourth growth conditions are: growth temperature 530℃, In / Ga BEP = 1.4, As BEP = 2.0 × 10⁻⁶. -6 Torr, growth time 10 min.

[0137] After four cycles, four physically distinguishable 90° sector regions are formed on the same substrate. Each region corresponds to different MBE growth adjustment parameters. Since the four regions are located on the same substrate and the mask holder and substrate holder can rotate synchronously during each growth process, batch differences between different wafers can be reduced, and the comparability between different beam parameters, growth temperatures and regional performance can be improved.

[0138] In-situ multimodal characterization: Throughout the growth process, the mask support is maintained with high precision, ensuring a close-contact gap of approximately 0.3 mm between the mask and the substrate growth surface. The RHEED electron beam is incident at a low grazing angle, preferably 2°, and reaches the exposed substrate surface through the mask's fan-shaped opening, the close-contact gap, and the reserved electron beam clearance channel. This enables continuous monitoring of surface reconstruction dynamics, surface smoothness, nucleation process, and changes in growth mode. The electron beam clearance channel is defined by the mask opening position, the mask support clearance space, and the recessed structure of the substrate support, reducing the obstruction of the RHEED incident path by the mask support or mask edge.

[0139] Simultaneously, the Ψ / Δ spectrum and band edge reflectance curves of the spectral ellipsometer are acquired in real time through a fixed optical window in the cavity. The spectral ellipsometer is used to obtain information related to thickness, composition, or growth rate during the selected area growth process. The band edge characterization is used to obtain information on band gap changes, reflectance changes, or optical response changes. The RHEED image sequence, ellipsometric spectrum sequence, and band edge curve are all recorded with a unified timestamp and correlated with parameters such as mask pattern ID, mask angle, region number, growth temperature, In / Ga BEP, AsBEP, growth time, source furnace shutter status, and synchronous rotation speed.

[0140] During each region switch, the mask holder rises to a safe switching height and completes independent rotation. For RHEED data, the data during the mask's rise, rotation, and descent can be labeled as "switching state data" and distinguished from the data in the stable growth phase. If necessary, short RHEED snapshots can be taken before and after the switch while maintaining the mask at a safe height or intermediate observation height to supplement the instantaneous reconstruction state of the substrate surface before and after the switch. After the snapshots are completed, the mask holder descends to a near-contact working position to continue the next round of growth. All in-situ data are aligned and stitched according to timestamps to form a continuous record of the multimodal growth process.

[0141] Growth Completion and Region Division: After all four growth cycles are completed, the mask holder lifts the mask to a safe removal height of approximately 10 mm from the substrate growth surface. A transfer robot enters from the side or lower side of the cavity and supports the mask, transferring it from the mask holder to the transfer robot for removal. The substrate is then unloaded. After unloading, the sample is region-divided on the wafer coordinate system using ellipsometry, optical microscopy, PL mapping, AFM, or SEM characterization. For the 1 / 4 fan-shaped opening mask in this embodiment, the same wafer can be automatically divided into four 90° fan-shaped regions. Each region corresponds to one growth cycle and a set of MBE growth adjustment parameters. The corresponding RHEED image sequence, ellipsometry Ψ / Δ spectrum sequence, band edge reflectance curve, and subsequent offline characterization results are extracted to form a structured multimodal dataset.

[0142] During data processing, data from the stable growth stage, mask switching stage, and RHEED snapshots are labeled separately. The stable growth stage data is used to characterize the continuous growth process of MBE quantum dots; the switching stage data is used to record the transition state caused by mask angle switching; and the RHEED snapshots are used to supplement the instantaneous reconstruction state of the substrate surface before and after switching. All of the above data are correlated through a unified timestamp, mask angle, region number, and growth parameters.

[0143] Comparative Example 1: Compared with Example 1, the differences are as follows: This comparative example uses a traditional fixed mask holder for selected area epitaxial growth, and the back side of the substrate does not have an anti-rotation positioning structure that fully engages with the substrate holder; the mask holder is a fixed or limited-stroke holder, which does not have the function of synchronous rotation with the substrate holder at the same angular velocity, does not have the function of independent angle switching within the cavity, and does not have the ability to precisely switch between transfer height, switching height, and near-contact working height; the distance between the mask and the substrate growth surface is fixed at 2.5 mm; the traditional holder does not have an electron beam clearance channel for RHEED low grazing angle electron beams, and all other aspects are the same.

[0144] Test Examples 1-5: Test Example 1: Preparation of experimental objects and loading: A 2-inch GaAs (001) wafer with a ring-shaped anti-rotation structure on the back is fixed on a recessed substrate support as an experimental object. The mask is configured according to the specific verification object. In Examples 1 and 6, a 1 / 4 fan-shaped opening mask is used to prepare InGaAs quantum dots. In Example 2, a parallel strip array pattern mask is used to prepare InAs nanowires. In Example 3, a circular dot pattern mask with different dot spacing is used to prepare InGaAs quantum dots.

[0145] Mask positioning and height adjustment: The corresponding mask is delivered to the mask holder above the substrate using a transfer robot. After receiving the mask, the mask holder drives the mask to descend vertically to a close-contact working position 0.3mm above the substrate surface, forming a selection area masking without contacting the substrate.

[0146] Synchronous rotational growth cycle: The mask holder and the substrate holder are set to enter a synchronous rotation state with the same angular velocity. The synchronous rotation speed is set to 10 rpm in Example 1 and Example 3, 12 rpm in Example 2, and 8 rpm in Example 6. The epitaxial source is introduced synchronously, and epitaxial growth is performed under the first adjustment condition in the corresponding exposed area.

[0147] In-situ switching and multi-region growth: After the initial growth is completed, the mask holder rises to a safe switching height. In Examples 1, 2, and 6, the mask holder is independently rotated by a preset angle to switch to the next selected area. In Example 3, a robotic arm returns the mask to the cavity storage compartment and replaces it with a mask of a different pattern. After the switch is completed, the mask descends back to a near-contact working height of 0.3 mm, and the growth parameters are adjusted to continue synchronous rotational growth. This process is repeated to complete the selective growth of 4 to 6 physically distinguishable regions on the same wafer.

[0148] Offline characterization and uniformity data extraction: After the growth process is completed, the substrate is removed. Using offline methods such as scanning electron microscopy and atomic force microscopy, the quantum dot density or nanowire length is measured inside each sector region on the wafer, and the radial uniformity deviation of size and density in each region is calculated.

[0149] Table 1. Test results of radial uniformity deviation of epitaxial growth in each embodiment

[0150] Conclusions and Analysis: According to the data in Table 1, the radial uniformity deviation of the epitaxial growth regions prepared in Examples 1, 2, 3 and 6 is at a low level, controlled within the ranges of less than 3.0%, less than 4.0%, less than 3.5% and less than 3.2%, respectively. The test data confirms the effectiveness of the dual-support synchronous rotation mechanism adopted in this application in improving the uniformity of selected epitaxial materials.

[0151] During material growth, the mask support is adjusted downwards to a near-contact height of 0.3 mm to form a localized selective masking area. The mask and the wafer at the bottom rotate synchronously at the same angular velocity through a dual-support system. By maintaining synchronized angular velocities, the relative spatial geometry of the mask opening pattern and the corresponding area on the wafer surface remains static, defining a clear boundary for the deposition region. Simultaneously, the mask and wafer, as a whole, continuously rotate physically relative to the fixed epitaxial source gas flow field (such as the gas flow in MOCVD) and molecular beam distribution (such as the beam in MBE) inside the growth chamber. This process induces a rotational averaging effect on the reaction gas flow and temperature field distribution, balancing the precursor concentration and heat distribution received at different orientations on the wafer surface.

[0152] As shown in Table 1, the dual-support structure design overcomes the technical limitation of traditional single-support fixed mask methods, which require the substrate to stop rotating during selected-area growth, thus resolving the significant radial growth gradient problem caused by a static state. Furthermore, the micron-level near-contact gap reduces edge diffusion and shadowing probability of the epitaxial source beam at the mask opening boundary. The data demonstrate that, regardless of whether it's in metal-organic chemical vapor deposition or molecular beam epitaxy equipment, or with mask patterns of different geometries, this scheme can output highly consistent locally grown materials, meeting the physical requirements for high-quality growth when constructing multi-region consistent datasets on the same substrate.

[0153] Test Example 2: The experimental subjects were the InGaAs quantum dot multi-region selected epitaxial sample prepared in Example 1 and the InGaAs quantum dot multi-region selected epitaxial sample prepared in the comparative example. Both were grown using a commercial MOCVD system on a 2-inch GaAs (001) wafer for four selected epitaxial growth cycles. The corresponding growth temperatures and V / III ratios were 580℃ (V / III=80), 590℃ (V / III=100), 600℃ (V / III=120), and 610℃ (V / III=140), respectively.

[0154] In Example 1, independent substrate and mask supports were used during the growth process. The mask support was lowered to a near-contact position 0.3 mm from the substrate surface and rotated synchronously with the substrate support at the same angular velocity of 10 rpm for selective epitaxy of each region. When switching regions, the mask support was raised to a safe height of 8 mm, rotated independently by 90°, and then lowered back to 0.3 mm. The entire switching process was performed in situ without disrupting the cavity vacuum.

[0155] In the comparative example, a traditional fixed mask support was used during the growth process. The distance between the mask and the substrate surface was fixed at 2.5 mm. Since the mask could not rotate synchronously, the substrate support stopped rotating (0 rpm) during the growth process. When switching regions, the source gas input was stopped and the temperature was lowered. The growth chamber was opened to break the vacuum. The mask was manually rotated 90°. Then, the vacuum was re-evacuated and the temperature was raised to stabilize the chamber before the growth of the next region.

[0156] After the growth process was completed, the substrates of both experimental groups were unloaded, and the quantum dot density of the two samples in four different regions was measured using offline characterization methods. The radial uniformity deviation in each region was calculated, and the time required for switching between each region and the data consistency R calculated by the system were recorded. 2 value.

[0157] Table 2. Comparison of multi-region selected epitaxial growth performance between Example 1 and comparative examples

[0158] Conclusions and Analysis: Based on Table 2 and Figure 1Data shows that Example 1 exhibits good dimensional and density uniformity in all regions, with a radial uniformity deviation of less than 3%. In contrast, the comparative example shows a radial uniformity deviation of 16%-22%, and the systematic density fluctuations between different regions exceed the expected range of normal process adjustments. The difference between the two sets of data stems from the structural and operational logic design of this scheme. The comparative example, limited by a fixed mask support, had to stop substrate rotation to prevent selective sweeping, causing the reactive gas flow field and temperature field to lose their rotational averaging effect. Simultaneously, the fixed 2.5mm shielding distance expanded the lateral diffusion space of the source gas, resulting in significant edge effects and radial gradients. Example 1, by decoupling the mask support and substrate support, allows them to rotate synchronously with the same angular velocity relative to the flow field while maintaining their relative positions. Coupled with a precisely controlled near-contact gap of 0.2-0.5mm, lateral gas diffusion is limited, thereby reducing local gradient differences.

[0159] In terms of experimental efficiency and data quality, the comparative example requires approximately 90 minutes for each region switch, involving cooling, vacuum breaking, and manual rotation. This not only prolongs the experimental cycle but also exposes the substrates in different regions to inconsistent surface oxidation and contamination environments, resulting in uncontrollable systematic errors. Example 1 utilizes a mask holder with independent rotation and precise lifting functions. Within the cavity, the switching is completed smoothly through a series of actions: lifting to a safe height, rotating, and lowering. The entire process is vacuum-free and requires no human intervention. This fully in-situ automatic switching control eliminates historical differences in cavity background, temperature drift, and atmospheric fluctuations between different experimental regions, ensuring consistent R-values ​​across the multi-region, multi-modal datasets on the same wafer in Example 1. 2 The value reached 0.98, indicating that the dual-scaffold collaborative working mechanism reduced the data variability between batches and can provide a reliable data foundation for data fusion and growth result prediction.

[0160] Test Example 3: A molybdenum holder containing a wafer substrate is fixed inside a recessed substrate support. A mask with a 1 / 4 fan-shaped opening pattern is transported to a mask support via an in-cavity transfer robot. The mask support is then adjusted to move downward along the Z-axis so that the mask stops at a distance of 0.3 mm from the wafer surface, forming a non-contact selection masking surface.

[0161] The mask holder and substrate holder are controlled to enter a synchronous rotation state at an angular velocity of 10 rpm. Then, the corresponding growth source is introduced and the epitaxial growth of the first region is completed for 15 minutes at 580℃ and V / III ratio of 80. After the first growth is completed, the mask is raised, rotated independently by 90 degrees and lowered back to 0.3 mm. The selected area position is switched sequentially and synchronous rotation is restored. The remaining three growths are completed in sequence, with growth conditions of 590℃ (V / III ratio of 100), 600℃ (V / III ratio of 120) and 610℃ (V / III ratio of 140), respectively.

[0162] During the growth process, the electron beam channel reserved in the mask opening and the gap between the mask and the wafer surface is used to control the RHEED system to continuously acquire the reconstructed dynamic image sequence of the surface at a low grazing angle of 2 degrees. The optical spectrum and band edge reflectivity curve of the spectroradiometer are continuously recorded using a fixed optical window. The system binds various data with specific growth control parameters and mask status according to a unified timestamp.

[0163] After growth, the same wafer was divided into four sector-shaped data acquisition areas. Stable subsets of data were extracted from each area, and surface reconstruction image features, optical evolution features, and bandgap electronic structure features were obtained using an image extraction network and a sequence model, respectively. Subsequently, the cross-attention mechanism in the multimodal fusion model was used to merge the three types of feature vectors, and finally, they were input into a multi-task prediction network to perform mapping calculations on the quantum dot size, density, and bandgap physical quantities of each area. The results were then evaluated and compared with the actual offline characterization results.

[0164] Table 3. Test data on the quality of the multimodal dataset and the prediction performance of feature fusion in Test Example 3.

[0165] Conclusions and Analysis: According to the data in Table 3, under the design architecture of synchronous rotation of the two supports at the same angular velocity, combined with the feature-level fusion network, it is possible to realize the reverse index calculation from the in-situ monitoring signal to the growth physical performance, and the error between the derived value and the actual growth state is maintained within a controllable range.

[0166] The data results rely on the mutual adaptation of physical and mechanical design and algorithms. Employing an independent dual-support system with a 0.3mm recess space, the mask not only achieves contactless region division but also moves within the same rotational field as the substrate. During the transition between growth steps, alignment of new regions is achieved solely through the lifting and independent rotation of the support itself. The entire experiment is conducted under constant vacuum and in-situ conditions. This operational mode avoids the boundary thermal field disturbances and contamination interference caused by traditional methods of opening cavities to change masks or stopping substrate rotation. Consequently, the constructed dataset exhibits a high consistency R-value of 0.98. 2value.

[0167] After acquiring the underlying dataset, the grazing angle channel retained by the device allows for continuous intervention from in-situ probes such as RHEED, ensuring that the timestamped roughness evolution and thickness evolution data are fully preserved. When using the cross-attention mechanism to coordinate the reconstruction of image features, optical evolution features, and electronic structure parameters, the algorithm not only captures individual surface states but also correlates them with the dynamic thermodynamic environment. Therefore, the relative error of the output average size prediction is limited to within 5%, and the estimation error of the bandgap value is within 8 meV. The comprehensive indicators demonstrate that the above-mentioned in-situ data acquisition mode can construct low-noise basic data samples and support data-driven algorithms to uncover the true correlation between epitaxial growth processes and final material properties.

[0168] Test Example 4: Experimental preparation stage: The experimental subjects used 2-inch GaAs (001) wafers as substrates and selected area epitaxial growth of InGaAs quantum dots was performed in a commercial MOCVD epitaxial equipment. Example 1 was configured with an independent sinking substrate support and a mask support with Z-axis height precision adjustment and independent rotation function; Comparative Example 1 was configured with a traditional fixed mask support, which did not have the function of synchronous rotation with the substrate support and independent angle switching within the cavity.

[0169] Initial setup for selected area epitaxial growth: In the first selected area epitaxial growth, the mask holder in Example 1 lowers the mask to a near-contact working position 0.3 mm above the substrate growth surface, and rotates synchronously with the substrate holder at the same angular velocity for epitaxial growth. In Comparative Example 1, the mask is held 2.5 mm above the substrate surface by a fixed holder. To avoid misalignment of the mask opening relative to the substrate area, substrate rotation is stopped during the growth process.

[0170] The process of switching the selected growth area: After the first growth cycle, for area switching, Example 1 controls the mask holder to raise the mask to a safe switching height of approximately 8 mm from the substrate growth surface. Without disrupting the vacuum and temperature of the growth chamber, it automatically rotates 90° independently relative to the substrate in situ, and then lowers back to the 0.3 mm position to continue the next growth cycle. Comparative Example 1 requires stopping the source gas and lowering the temperature to a suitable opening state. The growth chamber is opened to disrupt the vacuum environment. The fixed holder is manually loosened or the mask assembly is removed, rotated 90°, and then re-fixed. Subsequently, the vacuum is evacuated again, the temperature is raised, and the chamber state is stabilized for the next growth cycle.

[0171] In-situ characterization recording and multi-region dataset construction: In Example 1, during the entire growth and in-situ switching process, the sunken design of the mask support and the reserved electron beam clearance channel were utilized to obtain continuous surface reconstruction, thickness, and composition-related multimodal data through RHEED, spectral ellipsometer, and band edge characterization techniques, and the data was recorded aligned with timestamps. In Comparative Example 1, due to the disruption of process continuity caused by multiple cavity openings and the tendency of the fixed support to obstruct the low grazing angle optical path, stable and complete in-situ multimodal data could not be obtained.

[0172] End of growth and region analysis: After multiple cycles of operation, the substrate is unloaded, and data are extracted from multiple physically distinguishable regions on the wafer coordinate system using ellipsometry, optical microscopy and other methods. The radial uniformity deviation, single switching time, batch variability and consistency of multimodal datasets of the two methods are compared and analyzed.

[0173] Table 4. Comprehensive Comparison of Experimental Period, Pollution Control, and Data Consistency

[0174] Conclusions and Analysis: According to the data in Table 4, Example 1 significantly outperforms Comparative Example 1 in terms of single-switch time, radial uniformity deviation, and data consistency. This verification result relies on the dual-support decoupling and sunken structure mechanism of this scheme. In this scheme, the substrate support and mask support are independently mechanically decoupled, and the mask support is given precise Z-axis height adjustment and independent rotation capabilities. Since the mask can be raised to a safe height and rotated within the cavity to a near-contact working position, the manual replacement operation requiring cooling and opening the cavity to disrupt the high vacuum state, as required in Comparative Example 1, is avoided. This not only eliminates the approximately 90-minute switching time per cycle but also avoids thermal stress fluctuations, external contamination, and uncontrollable oxidation phenomena caused by cavity opening. The ability for stable continuous in-situ growth reduces the systematic errors caused by different growth backgrounds experienced by various physical regions, ensuring consistent R-values ​​of the multimodal datasets acquired on the same wafer. 2 The value reached 0.98, and batch variability was reduced by more than 65%.

[0175] Meanwhile, the mask support in this application has the function of rotating synchronously with the substrate support at the same angular velocity, so that the selected area boundary remains relatively fixed in the synchronous state. This ensures that the substrate as a whole receives a continuous and dynamically uniform source in the reaction flow field and temperature field, overcoming the severe edge effect caused by the stop rotation in Comparative Example 1, and controlling the radial uniformity deviation within the region from 16%-22% to within 3%. Data shows that the physical structure and control process of this application can effectively eliminate systematic interference caused by operational intervention, providing a reliable data acquisition method for constructing multi-region datasets with high consistency and low noise characteristics.

[0176] Test Example 5: Experimental subjects: multimodal time series datasets grown in multiple regions of the same wafer and feature-level multimodal fusion models.

[0177] During four growth cycles on the same substrate, the mask holder is controlled to descend to a close-contact working position 0.2-0.5 mm above the substrate growth surface, and the mask holder and the substrate holder are rotated synchronously at the same angular velocity. The growth temperature and V / III ratio are adjusted in different cycle periods to form four physically distinguishable growth adjustment regions.

[0178] Throughout the epitaxial growth process, a high-energy electron diffraction system was used to continuously acquire surface reconstruction image sequences of each region at a low grazing angle. A fixed optical window was used to acquire spectral sequences from a spectrometer in real time. A band edge characterization system was used to acquire band edge reflectivity curves. All data were recorded at a unified clock and automatically associated with mask status, region number, and growth parameter markings.

[0179] After all growth cycles are completed, the mask is raised to a safe height and removed. Then the substrate is unloaded, and the wafer coordinate system is scanned using an ellipsometry and other equipment. Combining the mask rotation angle record and growth timestamp, the same wafer is divided into multiple sector regions corresponding to different growth adjustment conditions. The corresponding offline characterization results are extracted as the final performance label for model training.

[0180] Complete multimodal subsets of data corresponding to the growth cycle are extracted based on unified timestamps and region IDs. Time alignment and spatial mapping are performed on different modalities, and anomalous or transitional state data are labeled and filtered. Neural networks are used to extract image representation vectors, optical evolution feature vectors, and electronic structure feature vectors, respectively.

[0181] A feature-level multimodal fusion model is used for data fusion. The three types of feature vectors are fused through a cross-modal interaction mechanism and then input into a multi-task prediction network. During training, loss functions such as mean square error, regional consistency, and physical constraints are applied to enable the model to learn the correlation mapping relationship between the final material properties and the complete growth process.

[0182] After the model is trained, cross-validation or leave-out method is used to verify and evaluate the network performance. The prediction error data and related determination coefficients of the micromorphological features and optical properties of each region are calculated and recorded by comparing with the actual offline characterization labels.

[0183] Table 5. Test data on multimodal dataset quality and feature fusion prediction performance.

[0184] Conclusions and Analysis: Based on the data in Table 1, the consistency R of the same wafer dataset is... 2The value reached 0.98, and the overall prediction R-value of the model feature fusion was [value missing]. 2 The value reached 0.97, while the relative error of the average diameter of the quantum dots was less than 5%, and the density error was less than 0.2 × 10⁻⁶. 10 cm -2 The bandgap correlation parameter prediction error is less than 8 meV. The test results objectively verify the decisive role of the dual-support collaborative system in the construction of high-precision datasets.

[0185] As can be seen from the innovative mechanism of epitaxial growth, the decoupling and sinking design of the substrate support and mask support eliminates the physical obstruction to the reflection of high-energy electron diffraction rays in the conventional fixed support mode. This ensures that multimodal characterization can achieve unobstructed acquisition throughout the process while maintaining close contact height control and continuous rotation. Full in-situ monitoring eliminates the temperature field inhomogeneity and batch system fluctuations introduced by stopping rotation and opening cavity switching, thus providing a data consistency of up to 0.98. Based on the acquisition of high-quality continuous data, the model integrates surface reconstruction features, optical thickness features and electronic structure features through cross-modal interaction mechanism, effectively filling the information gaps of single monitoring methods.

[0186] The intermodulation of multidimensional data enables the model to accurately reconstruct the mapping network between growth regulation parameters and final structural performance, keeping all prediction errors within extremely low levels. Test results confirm that this scheme not only has superior physical mechanisms in solving regional uniformity and in-situ characterization defects, but also has extremely high data processing accuracy and experimental reliability in materials genome data-driven process optimization.

Claims

1. A dual-scaffold selected area epitaxial multimodal data fusion and growth method, characterized in that, include: A substrate holder with a fixed substrate and a mask holder supporting a mask are disposed in an epitaxial device. The mask holder is disposed around the substrate holder. The substrate holder and the mask holder are independent of each other and both have rotation functions. The mask is located on the path from which the epitaxial source is incident on the growth surface of the substrate. Adjust the mask holder to a near-contact working position so that the mask forms a selected area blockage without contacting the substrate. The mask holder and the substrate holder rotate synchronously at the same angular velocity to perform selected area epitaxial growth. After completing the selected area epitaxial growth in the current region, the mask holder is rotated independently relative to the substrate by a preset independent rotation angle to switch the opening region. Then, the selected area epitaxial growth in the next region is performed to form multiple epitaxial regions with different adjustment conditions on the same substrate. During the continuous selected area epitaxial growth and region switching process, in-situ multimodal data is collected throughout the process by using the mask holder to avoid space and the gap between the mask and the substrate. After the growth is completed, each epitaxial region is characterized offline to obtain the final performance label. The collected in-situ multimodal data is then fused with the corresponding adjustment conditions and the final performance label at the feature level to establish a mapping relationship between the epitaxial process and the material properties.

2. The dual-support selected area epitaxial multimodal data fusion and growth method according to claim 1, characterized in that, The substrate support adopts a recessed design so that the upper edge of the substrate support is lower than the substrate growth surface; The acquisition of in-situ multimodal data includes the RHEED electron beam incident at a low grazing angle, passing through the opening of the mask, the close contact gap between the mask and the substrate, and the reserved electron beam clearance channel to reach the exposed surface of the substrate.

3. The dual-support selected area epitaxial multimodal data fusion and growth method according to claim 2, characterized in that, The acquisition of in-situ multimodal data also includes the real-time acquisition of in-situ optical characterization signals through a fixed optical window in the cavity; During the selected area epitaxial growth and the switching of the mask support area, the RHEED image sequence and the in-situ optical characterization signal sequence are continuously recorded according to a unified timestamp, and are time-correlated with the rotation angle of the mask and the adjustment conditions.

4. The dual-support selected area epitaxial multimodal data fusion and growth method according to claim 3, characterized in that, The acquired in-situ multimodal data is pre-labeled with state information, specifically as follows: The data collected when the mask support is in the near-contact working position for the selected area epitaxial growth is marked as stable growth state data; The data collected by the mask plate support during the area switching process is marked as switching status data; The data collected when the mask support pauses briefly at the switching or intermediate observation height to supplement the surface state information at the moment of switching is marked as quick-capture state data.

5. The dual-support selected area epitaxial multimodal data fusion growth method according to claim 4, characterized in that, When switching regions, adjust the mask plate support to a safe switching height; The close-contact working position is the distance between the mask and the substrate growth surface, which is 0.2 to 0.5 mm. The safe switching height is 5-10 mm above the substrate growth surface, which is used to prevent the mask from scratching or colliding with the substrate surface, molybdenum support or support structure during the rotation switching process.

6. The dual-support selected area epitaxial multimodal data fusion and growth method according to claim 1, characterized in that, The mask adopts a 1 / 4 fan-shaped opening pattern or a parallel strip array pattern; When the mask adopts a 1 / 4 fan-shaped opening pattern, the set angle for the mask holder to rotate independently relative to the substrate each time is 90°; When the mask adopts a parallel strip array pattern, the mask holder rotates independently relative to the substrate by a set angle of 60° each time.

7. The dual-support selected area epitaxial multimodal data fusion and growth method according to claim 1, characterized in that, The photomask is a thin plate with a thickness of 200 μm, and the material of the photomask is selected from one of SiN, SiO2, SiC, pyrolytic BN, Mo, Ta or W.

8. The dual-support selected area epitaxial multimodal data fusion and growth method according to claim 1, characterized in that, An additional mask storage chamber is provided inside the growth cavity of the epitaxial device, and multiple masks with different circular dot matrix pattern parameters are pre-placed in the mask storage chamber. After the selected area epitaxial growth in the current region is completed, the old mask is sent back to the mask storage bin by the transfer robot inside the cavity, and a new mask is picked up or lifted from the mask storage bin and transferred to the mask holder, realizing in-situ rapid replacement without opening the cavity.

9. The dual-support selected area epitaxial multimodal data fusion and growth method according to claim 1, characterized in that, The specific implementation method of the feature-level fusion is as follows: Surface reconstruction feature vectors are extracted from RHEED image sequences to characterize surface reconstruction and nucleation dynamics. From the in-situ optical characterization signal sequence, extract optical evolution feature vectors for characterizing thickness growth and composition changes, and electronic structure-related feature vectors for characterizing band gap and reflectivity inflection points; The surface reconstruction feature vector, the optical evolution feature vector, and the electronic structure-related feature vector are interactively fused using the Cross-Attention mechanism to generate a unified joint feature vector.

10. The dual-support selected area epitaxial multimodal data fusion growth method according to claim 9, characterized in that, The specific steps for establishing the mapping relationship between epitaxial processes and material properties are as follows: The generated joint feature vector is input into the multi-task prediction network, which outputs predicted material properties including average quantum dot size, quantum dot density, surface roughness, or bandgap-related parameters, and updates the weights of the multi-task prediction network based on a preset physical constraint loss.

Images