Methods and devices for identifying faulty electrical connections, and methods for generating a trained computational model

Abstract

A method for identifying defective electrical connections of a substrate (10) having a first surface contact (21) and a first electrical connection (20) extending from the first surface contact through the substrate is provided. The method comprises the steps of: charging the first surface contact (21) by directing an electron beam (111) onto the first surface contact (21); detecting a first secondary electron signal (114) as a function of time during the charging of the first surface contact; inputting input data (510) containing or based thereon to a trained computational model (500), in particular a trained machine learning model; and receiving defect information (511) regarding the first electrical connection (20) as an output from the trained computational model (500). Additionally, an apparatus (100) for identifying defective electrical connections of a substrate and a computer-readable storage medium for storing the trained computational model (500) are described.

Description

Technology Field

[0001] The present disclosure relates to methods and apparatuses for identifying defective electrical connections extending through a substrate, particularly through an advanced packaging (AP) substrate or a panel-level packaging (PLP) substrate. More specifically, the embodiments described herein relate to non-contact testing of electrical connections within a substrate using an electron beam in combination with machine learning techniques and artificial intelligence, particularly to identify and characterize defects such as short circuits, open circuits, and / or leaks. Background Technology

[0002] In many applications, it is necessary to inspect substrates to monitor their quality. Since defects can occur during substrate processing, for example, during the coating of substrates, inspecting the substrates to review defects and monitor the quality of displays can be beneficial.

[0003] Semiconductor substrates and printed circuit boards for the manufacture of complex microelectronic or micromechanical components are typically tested before, during, and / or after manufacturing to determine defects such as "short-circuit defects" or "open-circuit defects" in conductive paths and interconnections extending on or through the substrate layers. For example, substrates for the manufacture of complex microelectronic devices may include multiple interconnection paths intended to connect semiconductor chips to be mounted on the substrate.

[0004] Various methods for testing these components are known. For example, to determine whether a component is defective, the contact pads of the component to be tested can be mechanically brought into contact with a contact probe. However, as components and contact pads are becoming increasingly smaller due to the ongoing miniaturization of components, bringing the contact pads into contact with a contact probe can be difficult, and there may even be a risk of damaging the device under test during the test.

[0005] The complexity of packaging substrates is increasing, while design rules (feature sizes) are substantially decreasing. Within these substrates, surface contacts (for subsequent flip-chip or other chip mounting) are connected to other surface contacts on the packaging substrate to interconnect semiconductor (or other) devices. Standard methods, such as electromechanical probing for electrical testing, cannot meet the requirements of mass production testing because yield decreases (due to a larger number of test points) and contact reliability decreases (due to smaller contact sizes). In addition to the reduction in size and the potential damage to contact pads, the topography of packaging substrates poses difficulties for other test methods, such as those utilizing capacitive or electric field detectors, because such methods advantageously have small mechanical clearances.

[0006] Therefore, it would be beneficial to provide test methods and test devices suitable for reliably and rapidly identifying defects within the electrical connections of complex microelectronic devices.

[0007] In light of the foregoing, methods and apparatus for identifying defective electrical connections of a substrate are provided according to the independent claims. Methods for generating a trained computational model for identifying defective electrical connections are provided, as well as a computer-readable storage medium for storing the trained computational model. Further embodiments, advantages, and beneficial features are apparent from the dependent claims, description, and accompanying drawings.

[0008] According to one embodiment, a method for identifying defective electrical connections of a substrate is provided, wherein the substrate has a first surface contact and a first electrical connection extending from the first surface contact. The method comprises: (a) charging the first surface contact by directing an electron beam onto the first surface contact; (b) detecting a first secondary electron signal as a function of time during the charging of the first surface contact; (c) inputting input data including or based on the first secondary electron signal to a trained computational model, in particular a trained machine learning model; and (d) receiving defect information regarding the first electrical connection as an output from the trained computational model.

[0009] The input data may additionally include location information regarding a first surface contact on a substrate. The trained computational model may be configured to provide defect information as an output based on the input data including a first secondary electronic signal generated during charging of the first surface contact and location information of the first surface contact. The input data may optionally include additional information or additional input parameters, such as information regarding the substrate and / or information regarding the first electrical connection, for example, material information and / or electrical properties of the first electrical connection that may affect the first secondary electronic signal generated during charging.

[0010] In some embodiments, the substrate is an advanced packaging substrate (AP substrate), a panel-level packaging (PLP) substrate, or a wafer-level packaging (WLP) substrate.

[0011] According to another embodiment, a method is provided for generating a trained computational model, in particular a trained machine learning model, for identifying defective electrical connections of a substrate. The trained computational model may then be used to identify defective electrical connections according to any of the methods described herein. The method comprises the steps of: charging a first surface contact by directing an electron beam onto a first surface contact; detecting a secondary electron signal as a function of time during the charging of the first surface contact; determining associated defect information regarding a first electrical connection extending from the first surface contact; providing a first training data set comprising input data and associated defect information including or based on the secondary electron signal; and generating a trained computational model by training the computational model with the first training data set and a plurality of additional training data sets similarly provided for a plurality of additional surface contacts having their respective electrical connections extending from the same.

[0012] The computational model may be a machine learning model trained with multiple training data sets based on machine learning algorithms.

[0013] In some embodiments, the first training data set includes (i) a secondary electronic signal as a function of time generated during the charging of the first surface contact (directly or in a processed form), (ii) positional information regarding the location of the first surface contact, and (iii) associated defect information specifying whether the first electrical connection extending from the first surface contact contains a defect. The associated defect information may be determined through one or more voltage contrast measurements. Multiple additional training data sets generated in a similar manner by testing multiple surface contacts (having defective and / or defect-free electrical connections extending therefrom) may be used to train a computational model, thereby providing a trained computational model capable of identifying defects based on input data including SE signals as a function of time.

[0014] According to another embodiment, an apparatus configured to identify defective electrical connections of a substrate according to any of the methods described herein is provided.

[0015] An apparatus for identifying defective electrical connections of a substrate as described in this specification may include: a vacuum chamber housing a stage for placing the substrate; an electron source configured to generate an electron beam; a scanning deflector for directing the electron beam onto a first surface contact to charge the first surface contact; a secondary electron detector configured to provide a first secondary electron signal as a function of time during the charging of the first surface contact; and a data processing unit having a memory for storing a trained computational model configured to receive input data including or based on the first secondary electron signal and to provide defect information regarding a first electrical connection extending from the first surface contact as an output.

[0016] According to another embodiment, a computer-readable storage medium is provided for storing a trained computational model, in particular a trained machine learning model. The trained computational model is configured to receive input data including or based on a secondary electronic signal detected as a function of time during charging of a surface contact, and in response, to provide fault information regarding an electrical connection extending from the surface contact.

[0017] The embodiments also relate to devices for carrying out the disclosed methods and include parts of the device for carrying out each described method embodiment. These method embodiments may be carried out by hardware components, a computer programmed by suitable software, any combination of the two, or any other method. Furthermore, embodiments according to the present disclosure also relate to methods for operating the described device and methods for manufacturing the devices and devices described herein. Methods for operating the described device include method embodiments for carrying out all functions of the device. Brief explanation of the drawing

[0018] In a manner that allows the features of the above-mentioned disclosure to be understood in detail, a more specific description of the disclosure briefly summarized above can be obtained by referring to the embodiments. The accompanying drawings relate to embodiments of the disclosure and are described below: FIG. 1 shows a schematic diagram of a device for identifying defective electrical connections according to embodiments described in this specification. FIGS. 2a-2c schematically illustrate methods for testing electrical connections according to embodiments described herein. Figure 2d shows some examples of secondary electronic signals detected as a function of time during the charging of each surface contact. FIG. 3 schematically illustrates a method for identifying defective electrical connections according to embodiments described in this specification. FIG. 4 schematically illustrates a method for identifying defective electrical connections according to embodiments described in this specification. FIG. 5 is a block diagram illustrating a method for identifying defective electrical connections using a trained machine learning model according to embodiments described in this specification. FIG. 6 is a flowchart illustrating a method for identifying defective electrical connections according to embodiments described in this specification. FIG. 7 is a flowchart illustrating a method for generating a trained machine learning model to identify defective electrical connections according to embodiments described in this specification. Specific details for implementing the invention

[0019] Various exemplary embodiments will be referenced in detail below, and one or more examples thereof are illustrated in the respective drawings. Each example is provided for illustrative purposes only and is not intended to be limiting. For example, features illustrated or described as part of one embodiment may be used in or together with other embodiments to create other embodiments. The present disclosure is intended to include such modifications and variations. In the following description of the drawings, the same reference numerals refer to the same components. Only differences between individual embodiments are described. The structures illustrated in the drawings are not necessarily depicted in actual proportions but provide a better understanding of the embodiments.

[0020] The complexity of packaging substrates has increased over the years with the aim of reducing the space requirements of semiconductor packages. Conventional semiconductor packages are manufactured from semiconductor wafers before being diced and packaged. Next, the semiconductor package can be mounted on a printed circuit board (PCB) along with other microelectronic components.

[0021] To reduce manufacturing costs and space requirements, packaging technologies have advanced further in recent years, and technologies such as 2.5D ICs, 3D ICs, wafer-level packaging (WLP), e.g., fan-out or fan-in WLP, and panel-level packaging (PLP) have been proposed. In WLP technologies, the integrated circuit is packaged before dicing while remaining part of the wafer. Consequently, the resulting package has nearly the same size as the wafer.

[0022] "2.5D integrated circuits" (2.5D ICs) and "3D integrated circuits" (3D ICs) combine multiple dies into a single integrated package. Here, two or more unpackaged dies are placed on a packaging substrate, for example, a silicon interposer. In 2.5D ICs, the dies are placed side by side on the packaging substrate, whereas in 3D ICs, at least some of the dies are placed on top of each other. The assembly can be packaged as a single component, which reduces cost and size compared to a conventional 2D circuit board assembly.

[0023] A packaging substrate typically includes a plurality of device-to-device electrical interconnection paths intended to provide electrical connections between dies to be disposed on the packaging substrate. The device-to-device electrical interconnection paths may extend vertically (perpendicular to the surface of the packaging substrate) and / or horizontally (parallel to the surface of the packaging substrate) through the body of the packaging substrate in a complex connection network, and endpoints (referred to herein as surface contacts) are exposed on the surface of the substrate.

[0024] To reduce manufacturing costs, panel-level substrates are manufactured configured to integrate multiple devices (e.g., chips / dies that may be heterogeneous, e.g., have different sizes and configurations) into a single integrated package. The panel-level substrate typically provides multiple device-to-device electrical interconnection paths extending through the body of the packaging substrate, as well as chip sites on which multiple chips / dies are disposed on its surface, e.g., on one side or both sides. In particular, the size of the panel-level substrate is not limited to the size of the wafer. For example, the panel-level substrate may be rectangular or have other shapes. Specifically, the panel-level substrate may provide a surface area larger than that of a typical wafer, e.g., 1000 cm² or more. For example, the panel-level substrate may have a size of 30 cm x 30 cm or more, 60 cm x 30 cm or more, 60 cm x 60 cm or more, or larger.

[0025] Advanced packaging (AP) substrates provide device-to-device electrical interconnection paths on or within a wafer, such as a silicon wafer. For example, the AP substrate may include through silicon vias (TSVs) provided, for example, within a silicon interposer, or other conductor lines extending through the AP substrate. Panel-level packaging substrates are typically provided as composite materials, for example, materials for printed circuit boards (PCBs), or other composite materials, for example, ceramic and glass materials.

[0026] Conventional test devices may not be adapted or suitable for testing advanced packaging substrates or panel-level packaging substrates due to the geometry and density of surface contacts and / or the size of the packaging substrate, which may differ from the size of conventional dies or printed circuit boards. The present disclosure relates to methods and apparatuses for identifying defects within a substrate having a plurality of densely arranged surface contacts and a plurality of electrical connections extending between each of two or more surface contacts. In particular, the methods and apparatuses described herein may be suitable for testing packaging substrates configured to incorporate a plurality of devices in a single integrated package and may include at least one device-to-device electrical interconnection path extending between a first surface contact and at least one second surface contact.

[0027] As used herein, "surface contact" may be understood as an endpoint of an electrical interconnection path (hereinafter also referred to as "electrical connection") exposed on the surface of a substrate, thereby allowing an electron beam to be directed toward the surface contact to charge or probe the surface contact non-contactually. The surface contact may be intended to electrically contact a chip / die to be placed on the surface of the substrate, for example, through soldering. For example, the surface contact may be configured as a solder bump.

[0028] FIG. 1 illustrates an apparatus (100) for identifying defects in electronic connections, such as interconnection paths and / or vias, extending through a substrate (10) according to embodiments described herein. The apparatus (100) may include a vacuum chamber (101) which may be a test chamber specifically configured for testing, or a vacuum chamber of a larger vacuum system, for example, a processing chamber of a substrate manufacturing or processing system. For example, the apparatus may be configured as an inline inspection device integrated into a substrate processing system.

[0029] The vacuum chamber (101) may house a stage (105), for example, a movable stage, for placing a substrate (10) thereon. The device (100) further includes an electron source (120) for generating an electron beam, a scanning deflector (130) for deflecting the electron beam to a predetermined position on the substrate, and an electron detector (140) for detecting secondary electrons emitted from the substrate upon collision of the electron beam.

[0030] As schematically shown in FIG. 1, the substrate (10) comprises a first surface contact (21) and a first electrical connection (20) extending from the first surface contact (21) through the substrate to one or more additional surface contacts that may be located, for example, on the same substrate surface as the first surface contact (21) or on an opposite substrate surface (not shown in the drawings). The substrate (10) may comprise a plurality of surface contacts and a plurality of electrical connections extending from the plurality of surface contacts, for example, more than 10,000 electrical connections, particularly more than 100,000 electrical connections, or even more than 1,000,000 electrical connections. According to the methods described herein, a plurality of electrical connections extending from the plurality of surface contacts may be inspected to identify defective connections.

[0031] The device (100) includes an electron beam column (110) having an electron source (120), such as a thermal field emitter, configured to generate an electron beam (111) that propagates along an electron beam path (115) toward a substrate. The electron beam (111) can be directed, particularly focused, at a predetermined location on the substrate. Specifically, the device (100) may include a scanning deflector (130) configured to direct the electron beam (111) onto a surface contact, for example, onto a first surface contact (21) as depicted in FIG. 1. By directing the electron beam (111) onto the surface contact, the surface contact and the electrical connection extending from the surface contact can be charged, particularly negatively charged.

[0032] In some embodiments, the electron beam (111) may be focused onto the surface of the substrate (10) by, for example, a focusing lens (125), particularly a magnetic and / or electrostatic focusing lens. Specifically, the focusing lens (125) may be configured to focus the electron beam (111) onto the first surface contact (21) in order to charge the first surface contact (21) in a targeted manner.

[0033] Additional beam optical components (171), such as condenser lenses and / or aberration correctors, such as stigmators and / or chromators, may be optionally provided along the electron beam path (115) to affect the electron beam (111).

[0034] In some embodiments, the electron energy of the electron beam (111) (i.e., the landing energy of the electrons of the electron beam (111) on the substrate surface) may be above the neutral charging point. As used herein, “neutral charging point” refers to the electron energy of the electron beam that does not change its charges when the electron beam strikes an uncharged surface contact, because the amount of signal electrons emitted from the substrate upon impact essentially corresponds to the amount of electrons delivered to the surface contact by the electron beam. In some embodiments, the neutral charging point may correspond to an electron energy of the electron beam (111) of 1.5 keV to 3 keV, particularly about 2 keV.

[0035] Electrons colliding with the substrate with landing energy above the neutral charge point may have a reduced probability of generating secondary electrons emitted from the substrate, so when struck by an electron beam having electron energy above the neutral charge point, the substrate becomes negatively charged. Electrons colliding with the substrate with landing energy below the neutral charge point may have an increased probability of generating secondary electrons leaving the substrate, so when struck by an electron beam having electron energy below the neutral charge point, the substrate can be discharged.

[0036] In some embodiments, the electron energy of the electron beam (111) may be greater than 5 keV, particularly about 10 keV, particularly electron energy above the neutral charge point. Therefore, the amount of signal electrons emitted from the substrate upon collision is typically less than the amount of electrons delivered to the substrate by the electron beam (111). Thus, since negative charges can be delivered to the surface contact by the electron beam (111), the surface contact can be negatively charged along with the electrical connection extending therefrom. As used herein, "charging" may specifically relate to applying negative charges, i.e., electrons, to the surface contact to cause a predetermined (negative) electrical potential of the electrical connection extending from the surface contact.

[0037] The device (100) further includes an electron detector (140) configured to detect secondary electrons (113) emitted from the substrate (10), particularly during the collision of the electron beam (111). The electron detector (140) may be configured to detect secondary electrons (SE) emitted while charging the first surface contact (21) with the electron beam (111) to provide a secondary electron signal (114) as a function of time during charging. As used herein, "secondary electron signal" or "SE signal" may refer to the number of secondary electrons emitted by the surface contact as a function of time during the charging process of the surface contact. The number of SE depends on the surface voltage. Specifically, the secondary electron signal (114) contains information regarding the time dependence of the surface voltage during charging. FIG. 1 schematically illustrates the secondary electron signal (114) as a graph showing the SE yield (x-axis) detected as a function of time (y-axis) during charging. The SE yield can be measured in discrete time periods (e.g., every 100 nm until a saturation value is reached) during charging to provide a secondary electronic signal (114) as a (discrete) function of time.

[0038] In some embodiments, the electron detector (140) includes an Everhard-Thornley detector. An energy filter for signal electrons may be arranged in front of the electron detector (140), particularly in front of the Everhard-Thornley detector. The energy filter may include, for example, a grid electrode configured to be set to a predetermined potential. The energy filter may allow for the suppression of low-energy signal electrons. The energy filter may be set for detection relative to an optimized voltage. Thus, the signal intensity detected by the electron detector (140) may depend on the energy of the signal electrons, indicating whether the surface contact point is provided at a predetermined electric potential.

[0039] In particular, when an image of the surface region is captured by a conventional scanning electron microscope, generally only the SE yield per pixel is of interest, and the time dependence of the SE signal during the charging process is not of interest, which distinguishes the embodiments disclosed herein from conventional imaging scanning electron microscopes.

[0040] The secondary electron signal (114) generated during charging with the electron beam (111) is time-dependent because the amount of negative charges on the first surface contact increases during charging, which increases the secondary electron yield due to the increasing negative potential of the first surface contact. Additionally, the temporal progression of the secondary electron signal (114) depends on the electrical characteristics of the first electrical connection (20), for example, the capacitance of the first electrical connection and / or the cross-capacitance associated with neighboring electrical connections. Specifically, electrical connections with small capacitance charge quickly, causing the secondary electron signal (114) to rise rapidly over time, whereas electrical connections with large capacitance charge slowly, causing the secondary electron signal (114) to rise slowly over time. The time-dependent secondary electron signal (114) during the charging of the surface contact can generally follow an S-shaped curve, that is, it rises slowly at the initial collision of the electron beam, rises steeper in the central section, rises slowly toward the saturation value, and the saturation value corresponds to the electrical potential of the first electrical connection which no longer increases due to the collision of the electron beam (since there are almost no secondary electrons left to leave the surface). However, in reality, the time dependence of the secondary electron signal may show deviations from this S-shaped curve due to, for example, cross-capacitance effects, interactions between neighboring surface contacts, charge accumulation effects, defects and / or other effects.

[0041] If an electrical connection is interrupted due to a fault (i.e., there is an "open" fault), the capacitance of each electrical connection is smaller than expected, which will lead to an unexpectedly rapid rise in the secondary electronic signal over time. If an electrical connection is short-circuited to another electrical connection due to a fault (i.e., there is a "short" fault), the capacitance of each electrical connection is larger than expected, which will lead to an unexpectedly slow rise in the secondary electronic signals over time during charging. Therefore, if it is known what the SE signal of each electrical connection looks like in the absence of a fault, the secondary electronic signal generated as a function of time during charging of the surface contact can provide fault information regarding the electrical connection connected to the surface contact.

[0042] It can be difficult to reliably determine defect information regarding electrical connections based on secondary electronic signals generated during charging, because the time dependence of the secondary electronic signals generated during charging may depend on multiple factors, such as the defect class of the defects that may exist and the location of each surface contact on the substrate. For example, surface contacts located near the corners or edges of the substrate may generally generate different secondary electronic signals than surface contacts located in the center region of the substrate during charging, due to density fluctuations of peripheral surface contacts that can cause cross-capacitance effects that alter the temporal behavior of the SE yield. In particular, the electrical characteristics of the electrical connections, and the SE signals generated during charging accordingly, may depend on the location of each surface contact from which the electrical connection extends.

[0043] In some embodiments, each surface contact of a plurality of surface contacts of a substrate having a corresponding "position identifier" or of at least a subset of surface contacts may generate their own characteristic secondary electronic signal upon charging—provided there are no defects. The positional dependence of the time dependencies of the secondary electronic signals can make it particularly difficult to reliably determine defect information regarding the electrical connections.

[0044] With the above in mind, the apparatus and methods described herein provide a method for reliably identifying defective electrical connections of a substrate, even when the substrate has a large number of surface contacts capable of generating characteristic secondary electronic signals during charging.

[0045] As schematically depicted in FIG. 1, the device (100) includes a data processing unit (160), such as a computer, which includes, for example, a trained computational model (500) stored in its memory, and the trained computational model is configured to receive input data (510) which includes or is based on a first secondary electronic signal (114) as a function of time, and is configured to provide fault information regarding a first electrical connection (20) as an output.

[0046] In some embodiments, the trained computational model is a machine learning model, in particular a deep learning-based model, previously trained with the respective training data sets described below. In some embodiments, the machine learning model may be based on a neural network, in particular a deep neural network (DNN) having at least one hidden layer, for example, a deep neural network having three or more layers. Additionally, neural networks having fewer layers may be used depending on the complexity of the substrate and the number of parameters affecting SE signals during charging. This description, referring to a deep neural network (DNN), may generally be applied to neural networks (NN) and similarly to other artificial intelligence (AI) computational models, in particular machine learning (ML) computational models. Specifically, the computational model may include an input layer, an output layer, and one or more hidden layers between the input layer and the output layer. However, the computational model used according to the embodiments of this specification is not limited to being based on a DNN having hidden layers, and the relationship between input data (particularly including secondary electronic signals and optionally the locations of their respective surface contacts) and output data (particularly including defect information) may be simpler in some embodiments. For example, the input data and defect information may be linked through an if-else based relationship that provides a mapping between one or more features retrieved from the SE signal, on the one hand, location information of their respective electrical contacts, and on the other hand, their respective defect information. That is, the computational model is not limited to a deep learning-based model having multiple layers.

[0047] The trained computational model (500) is configured to receive input data (510) that includes or is based on a first secondary electronic signal (114). For example, the input data (510) may include the secondary electronic signal (114) in an unprocessed or processed form, for example, one or more portions of the secondary electronic signal (114) (e.g., an initial portion until a predetermined threshold potential of the first surface contact is reached; or a variation or steepness value of the secondary electronic signal (114) at one or more predetermined times during charging). In some embodiments, the input data (510) may include the secondary electronic signal (114) as a function of time from the start of charging to a specific point in time, for example, until a predetermined potential is reached.

[0048] The input data (510) may optionally include additional information, particularly position information regarding the location of the first surface contact (21) on the substrate. For example, the input data (510) may include the (absolute) location of the first surface contact (21) on the substrate (e.g., x and y coordinates on the substrate surface), or the relative location of the first surface contact (21) to one or more additional surface contacts (e.g., "corner location," "edge location," "center location," number of direct neighbor contacts, etc.). In some embodiments, the position information may include a position identifier characterizing the location of the first surface contact among a plurality of surface contacts, and / or characterizing the location of the first electrical connection among a plurality of electrical connections. In the example, the position identifier is "n, m" for the surface contact in the n-th row and m-th column of a 2D array of surface contacts, and the position identifier is "x", where x identifies a subgroup of surface contacts having common electrical characteristics to which the surface contact under test belongs, and n, m, x are integers.

[0049] In some embodiments, both the SE signal as a function of time during charging and the position information of each surface contact can be used as input parameters (i.e., part of the input data) for reliably retrieving defect information as outputs from a computational model. As mentioned above, since each surface contact position can lead to its own characteristic SE signal, or each subgroup of surface contacts can be characterized by its own typical SE signal, if both the "SE signal" and the "position information" are given as input parameters to the computational model, defect information can be reliably retrieved.

[0050] Optionally, input data entered into the computational model may include any one or more of the following information: (1) information regarding the substrate, in particular, the substrate type, the substrate material, the material of one or more substrate layers, the substrate design rules, and / or the substrate identifier. For example, the substrate identifier may define the electrical connections and / or surface contacts present in a specific substrate type, for example, the number and arrangement of electrical connections. If the substrate of the same substrate identifier has already been tested multiple times previously and the computational model is trained with the training data sets retrieved by said tests, the substrate identifier, which is part of the input data, may improve the reliability of the retrieved defect information. (2) information regarding the first electrical connection, in particular, the capacitance value, the cross-capacitance value providing information regarding cross-capacitance in relation to neighboring electrical connections, the number of surface contacts connected to the first electrical connection, the number of neighboring surface contacts, and / or the connection identifier characterizing the type, arrangement, and / or electrical characteristics of the electrical connection. For example, if an electrical connection with the same connection identifier has already been tested multiple times previously, and a computational model is trained on training data sets retrieved by said tests, the connection identifier, which is part of the input data, can improve the reliability of the retrieved fault information.

[0051] The trained computational model (500) may be configured to provide fault information regarding the first electrical connection (20) as an output. In some embodiments that may be combined with other embodiments described herein, the fault information provided by the trained computational model includes information regarding whether the first electrical connection appears to be faulty. In particular, the fault information may include binary fault values ​​(e.g., "0" = "no fault"; "1" = "faulty"). Optionally, if a fault is detected, the fault information may further include a fault class, e.g., a short-circuit fault, an open fault, or a leakage fault. Alternatively or additionally, if a fault is detected, the fault information may include a fault location that locates the location of the fault on or within the substrate. In some embodiments, the fault location may specify the location of the faulty electrical connection on or within the substrate (e.g., a "location identifier" that identifies the faulty electrical connection among a plurality of electrical connections on the substrate). In some embodiments, the fault location may further identify the location of the fault along the faulty electrical connection (e.g., finding the fault location between two specific surface contacts, or, for example, in the case of a short circuit fault, finding the fault location at the intersection between two specific electrical connections). Alternatively or additionally, the fault information may further include reliability information characterizing the reliability of the fault information. For example, the trained computational model may detect faults and / or specific fault classes with a high reliability of 90% or more in some cases. In other cases, for example, when the secondary electronic signal (114) deviates only slightly from the expected secondary electronic signal of a fault-free connection, the trained computational model may not identify the fault with high reliability.For example, a trained machine learning model may provide defect information including a confidence value (e.g., in % units) indicating the confidence that the detected defect and / or the detected defect class or defect location actually exist.

[0052] In some embodiments that may be combined with other embodiments described herein, the substrate has a plurality of surface contacts having their respective electrical connections extending therefrom, particularly more than 10,000, more than 100,000, or even more than 1,000,000 surface contacts, which are tested sequentially according to the test methods described herein. In particular, each of the plurality of surface contacts may be tested as follows: the surface contacts are charged by directing an electron beam onto them by deflection using, for example, a scanning deflector (130) (stage (a)). The SE signal generated during the charging of each surface contact is detected as a function of time by an electron detector (140) (stage (b)). Input data based on the SE signal and optionally the position of each surface contact is generated, and the input data is input into a trained computational model (500) (stage (c)). Fault information regarding electrical connections extending from each surface contact is received as the output of a trained computational model (500) (stage (d)). Thus, multiple electrical connections can be tested in succession by deflecting an electron beam onto the surface contact connected to each electrical connection, detecting the SE signal as a function of time during charging, and analyzing the SE signal using the trained computational model (500), specifically a trained machine learning model, more specifically a trained deep learning-based model.

[0053] In some embodiments, the substrate (10) is an advanced packaging substrate or a panel-level packaging substrate configured to provide multi-device in-package-interconnection, and the first electrical connection is a device-to-device electrical interconnection path. For example, the substrate may be a panel-level packaging (PLP) substrate, a wafer-level packaging (WLP) substrate, or a micro LED substrate. The packaging substrate may include a plurality of 1,000,000 or more surface contacts having their respective electrical connections extending therefrom, all of which can be tested in succession.

[0054] In some embodiments, multiple surface contacts are distributed over a surface area of ​​a substrate, or a sub-surface area of ​​a substrate of 16 cm² or more, specifically 25 cm² or more, more specifically 100 cm² or more, or even 225 cm² or more. The method may include continuously deflecting an electron beam (111) to multiple surface contacts with a scanning deflector (130) to continuously charge and test electrical connections extending therefrom using a trained computational model.

[0055] In some embodiments, the scanning deflector (130) may provide a deflection area of ​​16 cm² or more, specifically 100 cm² or more, more specifically 225 cm² or more (i.e., a surface area of ​​the substrate that can be reached by deflecting an electron beam with the scanning deflector (130) without moving the stage (105). A large deflection area allows for rapid and accurate testing of a large number of electrical connections because there is no need to move the substrate to test multiple electrical connections. In particular, at least one complete chip site on the surface of the substrate can be tested by deflecting an electron beam onto the surface contacts of the chip site without moving the stage.

[0056] A trained computational model (500) may be generated in a preceding stage of training as follows: the computational model, in particular a machine learning model, is trained by a machine learning algorithm with a plurality of training data sets. Each training data set includes input data (containing or based on SE signals detected as a function of time during charging of the surface contacts), and fault information regarding electrical connections extending from said surface contacts associated with said input data. The input data may optionally include any of the additional input parameters mentioned above, in particular position information regarding the location of each surface contact.

[0057] After the creation of the trained computational model (500), the trained computational model (500) can be used to reliably and quickly determine whether there are faults in specific electrical connections based on SE signals generated during charging.

[0058] In some embodiments that may be combined with other embodiments described herein, the device further comprises a discharge device for discharging at least a portion of the substrate, particularly for discharging a first surface contact before and / or after charging. For example, to ensure that charging starts from a predetermined electrical potential of the first surface contact, the first surface contact may be discharged before charging. For example, if multiple charges already exist on the first surface contact before charging, the secondary electronic signal (114) will appear differently than expected due to the existing charges affecting the SE yield. Therefore, it may be beneficial to discharge the first surface contact before charging and testing. In particular, each of the multiple surface contacts may be discharged before charging and testing.

[0059] Alternatively or additionally, the first surface contact may be discharged after charging and testing the first surface contact. Discharging after testing reduces or prevents the accumulation of charges on the substrate surface that could distort the results of subsequent testing measurements. In particular, charges on the substrate surface can deflect the electron beam and / or affect the SE signal detected by the electron detector, for example, if the charged surface contact is positioned near the first surface contact currently being tested. Therefore, it may be beneficial to discharge the first surface contact after charging and testing. In particular, each of the plurality of surface contacts may be discharged after charging and testing. In particular, each of the plurality of surface contacts may be discharged both before and after testing.

[0060] In some embodiments, the discharge device may include any of a second electron source configured to generate a second electron beam for discharge, an electron flood gun configured to discharge a large surface area of ​​the substrate, and / or a UV discharge lamp. The second electron source may be configured to generate a second electron beam having a second electron energy different from the electron energy of the electron beam (111). In particular, since the second electron energy of the second electron beam may be at a neutral charge point, for example, 2 keV or lower, for example, about 1.5 keV or lower, the second electron beam may be used to remove charges from the substrate.

[0061] FIGS. 2a–2c schematically illustrate methods for identifying defective electrical connections. FIG. 2a illustrates a test of a first electrical connection (20) extending from a first surface contact (21), wherein the first electrical connection (20) is free of defects. An electron beam (111) is directed onto the first surface contact (21) to charge the first surface contact (21), and secondary electrons (113) emitted from the first surface contact are detected by an electron detector (140) during charging. A first secondary electron signal (114) corresponding to an SE yield as a function of time during charging is detected and input (directly or in a processed form) as part of the input data to a trained computational model. The input data may additionally include a location identifier or other location information regarding the first surface contact (21) or the first electrical connection (20) under test. In the example of FIG. 2a, the first electrical connection (20) is fault-free, which results in a characteristic time dependency of the first secondary electronic signal (114) (Note: The time dependency of SE yield typically depends on the location of the electrical connection tested), and thereby, the trained computational model can determine that the first electrical connection is fault-free ("fault-free") based on the input data.

[0062] In particular, the trained computational model may have been previously trained with training data sets retrieved by testing multiple fault-free (and faulty) electrical connections having the same location identifier as the first electrical connection, thereby enabling the trained computational model to recognize whether the first electrical connection (20) is faulty based on the input data. In some embodiments, one or more absolute SE yields during charging, one or more slopes of the SE curve during charging, final potential, and / or characteristic shapes of the SE signal over time may provide the trained computational model with indications of whether the first electrical connection is faulty and what fault class may exist.

[0063] In some embodiments, if the secondary electronic signal (114) is partially and / or completely out of a predetermined range (224) which is characteristic of SE signals of defect-free electrical connections extending from surface contacts having the same location identifier and / or the same electrical characteristics, the trained computational model may recognize that the first electrical connection (20) is defective. The predetermined range (224) associated with a specific location identifier may be defined by the computational model by training with a plurality of training data sets having the same location identifier. For example, the predetermined range (224) may be defined by the computational model as a range around the typical (e.g., average) secondary electronic signal over time of defect-free electrical connections having the same location identifier.

[0064] FIG. 2b shows a test of the first electrical connection (20) extending from the first surface contact (21) in the case where the first electrical connection (20) is defective. Specifically, the first electrical connection (20) depicted in FIG. 2b is interrupted, i.e., contains an open fault (31). Consequently, the capacitance of the first electrical connection is reduced, which leads to a faster rise in the detected secondary electronic signal (114') because the first electrical connection charges faster. From input data containing the secondary electronic signal (114'), the trained computational model can recognize that the first electrical connection (20) has a defect and contains an open fault (31). Additionally, defect information and reliability of the defect location may be optionally included in the output data provided by the trained computational model.

[0065] The secondary electronic signal (114') has an unexpectedly high slope or gradient, which may be an indication to the trained computational model that the first electrical connection is defective, particularly interrupted. Alternatively or additionally, the trained computational model may recognize that the secondary electronic signal (114') extends beyond a predetermined range (224) defined by the computational model for defect-free electrical connections associated with the same location identifier, based on training with multiple training data sets.

[0066] FIG. 2c shows a test of the first electrical connection (20) extending from the first surface contact (21) in a case where the first electrical connection has other defects. Specifically, the first electrical connection (20) depicted in FIG. 2c is short-circuited to the second electrical connection (24), i.e., contains a short-circuit defect (32). Consequently, the capacitance of the first electrical connection increases, which leads to a slower rise in the detected secondary electronic signal (114") because the first electrical connection charges more slowly. A trained computational model can recognize from input data containing the secondary electronic signal (114") that the first electrical connection (20) has defects and contains a "short-circuit" defect (31).

[0067] The secondary electronic signal (114) has an unexpectedly slow slope or gradient in the initial section, which may be an indication to the trained computational model that the first electrical connection is defective, particularly short-circuited. Alternatively or additionally, the trained computational model may recognize that the secondary electronic signal (114) extends beyond a predetermined range (224) defined by the computational model for defect-free electrical connections associated with the same location identifier, based on training with multiple training data sets.

[0068] If a short-circuit fault is detected, for example, to further characterize the fault to determine the fault location, one or more surface contacts of adjacent electrical connections may subsequently be charged with an electron beam (111). For example, if a secondary electron signal subsequently generated upon charging of the second surface contact (22) starts with an unexpectedly high SE yield, a short-circuit fault (31) may be identified as being located between the first electrical connection (20) and the second electrical connection (24) extending from the second surface contact (22). Alternatively or additionally, a confidence value may be included in the output data provided by the trained computational model. For example, if the SE signal expands significantly away from a predetermined range (224) or has a slope that is significantly different from the expected slope of the SE signals of the fault-free electrical connections, the confidence value of the detected fault may be high (e.g., close to 100%).

[0069] FIG. 2d shows some additional examples of secondary electronic signals (214) detected as functions of time t during the charging of surface contacts, each connected to a defective electrical connection. Different defects each lead to different SE signals during charging. In some embodiments, if the secondary electronic signal as a function of time extends beyond a predetermined range (224) (indicated by the gray area in FIG. 2d), the trained computational model can identify that there is a defect in the respective electrical connection. The predetermined range (224) may be defined by the computational model based on training the computational model with a plurality of training data sets.

[0070] In some embodiments, each predetermined range may be defined by a computational model trained for each of a plurality of position identifiers characterizing the location of a surface contact. The predetermined ranges may be defined by a computational model as a result of training using a plurality of training data sets, wherein each training data set includes an SE signal, position information (e.g., a position identifier) ​​identifying the location of each surface contact on a substrate, and defect information specifying whether each electrical connection has a defect.

[0071] FIG. 3 schematically illustrates a method for testing a second electrical connection (24) extending from a second surface contact (22). The second surface contact (22) has a different position identifier than the first surface contact (21) tested in FIG. 2a. Specifically, the second surface contact (22) is located between two adjacent surface contacts—in the cross-sectional plane of FIG. 3—which leads to an increase in cross-capacitance experienced by the second electrical connection (24) compared to the first electrical connection (20). As a result, assuming that neither the first nor the second electrical connection is defective, the secondary electronic signal (314) generated as a function of time during charging of the second surface contact (22) is different from the secondary electronic signal (114) generated as a function of time during charging of the first surface contact (21) (see FIG. 2a).

[0072] A secondary electronic signal (314) is provided as part of input data to a trained computational model, along with a position identifier characterizing the location of the second surface contact (22), and the trained computational model provides fault information (here, “no fault”) regarding the second electrical connection (24) as output. In particular, if the secondary electronic signal (314) is within a predetermined range defined for the surface contacts having their respective position identifiers, the trained computational model can identify that no fault exists.

[0073] In some embodiments that may be combined with other embodiments described herein, the computational model is configured to define each predetermined range that can be compared with a secondary electronic signal measured during charging of a surface contact having each position identifier, based on training data sets for each of a plurality of position identifiers each characterizing the location of one or more surface contacts, in order to determine whether the surface contact is connected to a defective electrical connection.

[0074] FIG. 4 schematically illustrates a method for identifying defective electrical connections of a substrate (10) according to embodiments described herein. The substrate (10) may be a PLP substrate comprising a plurality of chip sites (401, 402, 403) each configured for the placement of their respective chips, and more than 1,000 device-to-device electrical interconnection paths may extend from each chip site through the PLP substrate to one or more other chip sites. For example, the substrate may include two pairs, three pairs, or more associated chip sites, and more than 1,000 device-to-device electrical interconnection paths (optional, exactly two surface contact points, i.e., one surface contact point at each chip site of a pair) may extend through the packaging substrate between each pair of chip sites, as schematically depicted for the upper chip site pair in FIG. 4.

[0075] Each of the plurality of surface contacts located at the first chip site (401) may have a corresponding position identifier. Surface contacts arranged at corresponding positions in different chip sites may have corresponding position identifiers. For example, three surface contacts illustrated as black circles in FIG. 4, each arranged at the upper left corners of their respective chip sites, may have corresponding position identifiers because the electrical characteristics of the electrical connections extending from them may be similar or essentially identical. Alternatively, each surface contact of a specific substrate type may have its own position identifier. The secondary electronic signals generated during charging of surface contacts having the same position identifier may have generally similar or identical behavior as a function of time during charging.

[0076] If surface contacts and their respective electrical connections are grouped according to their respective electrical properties that produce similar or identical SE signals during charging, and each group has its own location identifier, training of the computational model is facilitated, and the reliability of the fault information provided as output by the computational model is improved. Specifically, the trained computational model can define the characteristic temporal behavior of the SE signal associated with its respective location identifier representing a specific fault or fault-free electrical connection (specifically, a "predetermined range" as described above).

[0077] FIG. 5 is a block diagram illustrating a method for identifying defective electrical connections using a trained computational model (500) according to embodiments described herein. The trained computational model (500) may be stored in a computer-readable storage medium, for example, computer memory.

[0078] According to the embodiments described in this specification, the trained computational model (500) is configured to receive input data (510) including or based on a secondary electronic signal (114) detected as a function of time during charging of the surface contact, and in response, to provide fault information (511) regarding an electrical connection extending from the surface contact as an output.

[0079] As schematically depicted in FIG. 5, secondary electrons (113) are detected by an electron detector (140) while charging a surface contact with an electron beam, and the detected secondary electron signal (114) as a function of time during charging is transmitted to an input data generator (515) that provides input data (510) based on the secondary electron signal (114). The input data generator (515) can generate input data (510), which is based on the secondary electron signal (114) and optionally additional input parameters, in particular location information regarding the location of the surface contact under test, in particular location identifiers. The input data (510) may optionally include additional information, for example, information regarding the substrate and / or information regarding electrical connections that can facilitate the identification of defects by a trained computational model (500).

[0080] The trained computational model (500) provides fault information (511) as output, and the fault information (511) indicates whether the electrical connection has a fault. The fault information may also include a fault class, a fault location, and / or a reliability value.

[0081] A trained computational model (500) may be generated by a preceding stage of training. Training may include generating a plurality of training data sets (551) and training the computational model (501) with the plurality of training data sets (551) in a training process (550) that utilizes a machine learning algorithm that, for example, may depend on the computational model (501) to be trained.

[0082] Each of the multiple training data sets (551) may include input data and differently determined associated fault information. The input data may include or be based on secondary electronic signals detected during the charging of each surface contact and may additionally include location information. The associated fault information may be determined, for example, by performing generally known voltage-to-voltage measurements.

[0083] For example, associated fault information regarding a first electrical connection extending from a first surface contact may be determined by probing with an electron beam or a second electron beam any of (i) the first surface contact, (ii) one or more second surface contacts to be electrically connected to the first surface contact, and (iii) one or more third surface contacts to be electrically disconnected from the first surface contact, through one or more voltage-comparison measurements. Alternatively or additionally, associated fault information may be determined "manually" by operator classification.

[0084] The input data of each training data set may optionally additionally include positional information regarding the location of each surface contact on the substrate, in particular, a position identifier. The input data of each training data set may optionally additionally include additional information regarding the substrate and / or the respective electrical connection under test, for example, a substrate identifier and / or a connection identifier.

[0085] In some embodiments that may be combined with other embodiments described herein, a plurality of training data sets are provided for surface contacts having corresponding location information or corresponding location identifiers. In particular, multiple surface contacts arranged at corresponding locations of different subregions of a substrate (e.g., different chip sites) or at corresponding locations of different substrates of the same type are tested to provide a plurality of training data sets for positionally related surface contacts.

[0086] Grouping surface contacts according to their respective location identifiers can be beneficial because surface contacts with corresponding location identifiers generally possess similar electrical properties that can lead to similar or identical SE signals upon charging. If training data sets include not only their respective SE signals but also corresponding location information—and if, when used for testing, location information regarding the electrical connection under test can be used as input parameters for a trained computational model, training can be facilitated, and subsequently, more reliable fault information can be provided by the trained computational model.

[0087] The computational model (501) may be a machine learning model, particularly a deep learning-based model. In some embodiments, the machine learning model may be based on a neural network, particularly a deep neural network (DNN) having at least one hidden layer, for example, a deep neural network having three or more layers. Additionally, neural networks having a smaller number of layers may be used.

[0088] Each layer of a DNN may include multiple basic computational elements (CEs), typically referred to in the art as dimensions, neurons, or nodes. The computational elements of a given layer are connected to the CEs of subsequent layers by connections. Each connection between the CEs of a previous layer and the CEs of a subsequent layer may be associated with a weight value.

[0089] A given hidden CE can receive inputs from CEs of the preceding layer through their respective connections, and each given connection is associated with a weight value that can be applied to the input of the given connection. The weight values ​​can determine the relative strength of the connections and, thus, the relative influence of each input on the output of the given CE. A given hidden CE can be configured to calculate an activation value (e.g., a weighted sum of inputs) and further derive an output by applying an activation function to the calculated activation value. The activation function can be, for example, an identity function, a deterministic function (e.g., linear, sigmoid, threshold, or similar), a stochastic function, or another suitable function. The output from a given hidden CE can be transmitted to CEs of the subsequent layer through their respective connections. Likewise, each connection at the output of the CE can be associated with a weight value that can be applied to the output of the CE before being received as an input to the CE of the subsequent layer. In addition to the weight values, there may be threshold values ​​(including limiting functions) associated with the connections and CEs.

[0090] The weights and / or thresholds of a neural network may be initially selected prior to training and may be further iteratively adjusted or modified during training to achieve an optimal set of weights and / or thresholds in the trained computational model. The set of DNN input data used to adjust the weights / thresholds of the computational model is referred to herein as the training data set.

[0091] It should be noted that the teachings of the test methods disclosed herein are not constrained by the number of hidden layers and / or the DNN architecture. By example, the layers within the DNN may be convolutional, fully connected, locally connected, pooling / subsampling, recurrent, etc. In some examples, the computational model is an if-else based model. For example, for each location identifier, the computational model may define a predetermined range in which each SE signal is expected to lie if each electrical connection is fault-free. If the measured SE signal falls within the predetermined range—during the test—the trained computational model recognizes "fault-free," and if not, the trained computational model recognizes "fault."

[0092] FIG. 6 is a flowchart illustrating a method for identifying defective electrical connections according to embodiments described in this specification.

[0093] In the box (601), the first surface contact under test is charged by directing an electron beam onto the first surface contact.

[0094] In the box (602), the first and second electronic signals are detected as a function of time during charging.

[0095] In box (603), input data is generated based on the detected first electronic signal. Specifically, the input data may include not only the first electronic signal (or parts or characteristics thereof) but also additional information, such as location information regarding the first surface contact, for example, a location identifier characterizing the location of the first surface contact on the substrate surface. The input data is input into a trained computational model, in particular a trained machine learning model.

[0096] In box (604), fault information regarding the first electrical connection is received as output from a trained computational model. The fault information may indicate whether the first electrical connection is faulty, and—if it is faulty—may additionally include any of a fault class, a fault location, and / or a reliability value.

[0097] The method may proceed by deflecting an electron beam and striking the second surface contact to charge and test the second electrical connection extending from the second surface contact in the box (605). The second electrical connection may be tested similarly to the first electrical connection.

[0098] Optionally, any of the surface contacts may be discharged before and / or after charging to avoid the accumulation of charges on the substrate that could have a negative effect on defect detection.

[0099] FIG. 7 is a flowchart illustrating a method for generating a trained computational model to identify defective electrical connections according to embodiments described in this specification.

[0100] In the box (701), the first surface contact is charged by directing an electron beam onto the first surface contact.

[0101] In the box (702), the first and second electronic signals are detected as a function of time during charging.

[0102] In box (703), associated fault information regarding a first electrical connection extending from a first surface contact is determined. The associated fault information can be retrieved through one or more voltage contrast measurements. For example, the first surface contact, one or more second surface contacts that must be electrically connected to the first surface contact, and / or one or more third surface contacts that must be electrically disconnected from the first surface contact can be probed, for example, with an electron beam or a second electron beam. If the first surface contact and one or more second surface contacts are not provided with the same electrical potential, an "open" fault can be determined. If the first surface contact and one or more third surface contacts are provided with the same electrical potential, a "short" fault can be determined.

[0103] In box (704), a first training data set is provided, comprising input data and associated defect information determined in box (703). The input data includes or is based on a secondary electronic signal and optionally additional information, such as location information regarding a first surface contact.

[0104] In box (705), the trained computational model is generated by training the computational model with the first training data set provided in box (704), specifically using a machine learning algorithm.

[0105] Multiple additional training data sets may be provided similarly to the first training data set by testing multiple additional surface contacts having their respective electrical connections extending therefrom, and a machine learning model may be trained with the multiple additional training data sets. In particular, multiple surface contacts having the same location identifier may be tested to provide their respective training data sets that include said location identifier as a parameter. Using location identifiers as input parameters for the computational model trained during testing improves the reliability of defect information because the electrical characteristics of the electrical connections (and thus the time dependence of SE yield) can depend significantly on the location of the tested surface contact on the substrate.

[0106] A trained machine learning model can be further improved over time, for example by using a self-learning algorithm, by training with additional training data sets that may be generated while using the trained computational model for defect inspection and classification.

[0107] According to the embodiments described herein, methods for rapidly and reliably identifying defective electrical connections on a substrate are provided, using trained computational models, in particular trained machine learning models and deep learning-based algorithms. Secondary electron signals, measured as a function of time during charging of each surface contact, are used as input parameters for the trained computational model. Additional input parameters, such as location information regarding the surface contact under test, may be provided to the trained machine learning model. The use of artificial intelligence and machine learning techniques for detecting defective electrical connections improves defect detection capabilities and provides higher defect classification accuracy and purity compared to conventional electron beam inspection techniques based on voltage-to-voltage measurements. Since the computational model can be improved over time by training with additional training data sets, self-learning and self-improving algorithms for defect detection may be provided. The methods described herein may be applied to the testing of complex substrates having a vast number of electrical connections, in particular AP substrates and / or PLP substrates, wherein SE signals generated during charging may depend on a plurality of parameters, including the location of the surface contact under test on the substrate.

[0108] Although the foregoing description relates to some embodiments, other additional embodiments may be made without departing from the basic scope, and their scope is determined by the following claims.

Claims

Claim 1 A method for identifying defective electrical connections of a substrate (10) having a first surface contact (21) and a first electrical connection (20) extending from the first surface contact, comprising: (a) charging the first surface contact by directing an electron beam (111) onto the first surface contact; (b) detecting a first secondary electron signal (114) as a function of time during the charging of the first surface contact; (c) inputting input data (510) containing or based thereon the first secondary electron signal (114) into a trained computational model (500); and (d) receiving defect information (511) regarding the first electrical connection (20) as an output from the trained computational model (500). Claim 2 A method according to claim 1, wherein the input data (510) input to the trained computational model (500) further includes position information regarding the first surface contact (21). Claim 3 A method according to paragraph 2, wherein the position information comprises at least one of the absolute position of the first surface contact on the substrate, the relative position of the first surface contact with respect to one or more additional surface contacts, and a position identifier. Claim 4 A method according to any one of claims 1 to 3, wherein the input data input to the trained computational model comprises: information relating to the substrate (10), comprising at least one of the following: a substrate type, a substrate material, one or more substrate layer materials, a substrate design rule, and a substrate identifier; and additionally comprising any one or more of the information relating to the first electrical connection (20), comprising at least one of the following: a capacitance value, a cross-capacitance value, the number of surface contacts connected to the first electrical connection, the number of neighboring surface contacts, and a connection identifier. Claim 5 A method according to any one of claims 1 to 3, wherein the defect information regarding the first electrical connection comprises: information regarding whether the first electrical connection (20) includes a defect, a defect class, a defect location, and a reliability value, wherein the defect information includes any one or more of the above. Claim 6 A method according to any one of claims 1 to 3, wherein the substrate has a plurality of surface contacts having respective electrical connections extending therefrom, and (a), (b), (c), and (d) are performed on each of the plurality of surface contacts to obtain defect information regarding each of the respective electrical connections. Claim 7 In claim 6, the substrate is a panel-level packaging substrate or an advanced packaging substrate, and the plurality of surface contacts comprises 1,000,000 or more surface contacts. Claim 8 In claim 6, the plurality of surface contacts are distributed over a surface area of ​​the substrate of 16 cm² or more, and the method further comprises the step of deflecting the electron beam to the plurality of surface contacts with a scan deflector (130) to continuously charge and test the plurality of surface contacts. Claim 9 A method according to any one of claims 1 to 3, further comprising the step of generating the trained computational model (500) by training the computational model (501) with a plurality of training data sets (551) each comprising input data and associated fault information regarding their respective electrical connections, wherein the input data each comprises or is based on a secondary electronic signal detected as a function of time during the charging of a surface contact from which each electrical connection extends. Claim 10 A method according to claim 9, wherein the input data each additionally includes position information regarding the surface contact. Claim 11 A method for generating a trained computational model (500) for identifying defective electrical connections of a substrate, comprising: charging the first surface contact (21) by directing an electron beam (111) onto the first surface contact; detecting a secondary electron signal (114) as a function of time during the charging of the first surface contact; determining associated defect information regarding a first electrical connection (20) extending from the first surface contact; providing a first training data set including input data containing or based on the secondary electron signal and the associated defect information; and generating the trained computational model by training the computational model with the first training data set and a plurality of additional training data sets similarly provided for a plurality of additional surface contacts having their respective electrical connections extending from the same. Claim 12 A method according to claim 11, wherein the associated fault information regarding the first electrical connection is determined through one or more voltage contrast measurements. Claim 13 A method according to claim 11 or 12, wherein the input data of the first training data set further includes position information regarding the position of the first surface contact on the substrate. Claim 14 A method according to claim 11 or 12, wherein a plurality of training data sets are provided for surface contacts having corresponding location information or corresponding location identifiers. Claim 15 The method of claim 1 or 11, wherein the trained computational model is a trained machine learning model, or a trained deep learning-based model based on a deep neural network having at least one hidden layer. Claim 16 A device (100) for identifying defective electrical connections of a substrate (10), comprising: a vacuum chamber (101) housing a stage (105) for placing the substrate; an electron source (120) configured to generate an electron beam (111); a scanning deflector (130) for directing the electron beam onto a first surface contact (21) to charge a first surface contact (21); an electron detector (140) configured to detect a first secondary electron signal (114) as a function of time during the charging of the first surface contact; and a data processing unit (160) having a memory for storing a trained computational model (500) configured to receive input data (510) including or based on the first secondary electron signal (114) and provide as an output defect information regarding a first electrical connection (20) extending from the first surface contact (21). Claim 17 In claim 16, the above-mentioned scanning deflector (130) is configured to provide a deflection area of ​​16 cm² or more. Claim 18 An apparatus according to claim 16 or 17, further comprising a discharge device for discharging at least a portion of the substrate. Claim 19 In paragraph 16, the device is configured to perform the method of paragraph 1 or 11. Claim 20 A computer-readable storage medium storing a trained computational model (500), wherein the trained computational model is configured to receive input data (510) comprising or based on a secondary electronic signal (114) detected as a function of time during charging of a surface contact, and in response thereto provide fault information regarding an electrical connection extending from the surface contact as an output. Claim 21 In paragraph 20, the above-mentioned trained computational model (500) is a computer-readable storage medium generated according to paragraph 11.

Images