Physical design support system, physical design support method, and program

Abstract

This invention provides a physical design support system, method, and program that assist in the search for the causes of defects in physical design. [Solution] The system includes: a search state management means that maintains the causal relationships of the target defects as causal hypothesis records and manages the set of said causal hypothesis records as a search state; a search control means that prioritizes the causal hypothesis records based on input information, updates the search state based on comparison with counter-evidence conditions, and determines the next observation target to be verified; an active observation generation means that generates observation parameters for the determined observation target, including at least one of viewpoint change, transparency display, cross-section extraction, layer selection, and magnification change using the internal attributes of the design data, and generates a diagnostic image; a feature extraction means that extracts physical features from the diagnostic image; and an update means that reprojects the physical features in relation to the components of the causal hypothesis records and updates the state of the causal hypothesis records referenced by the search control means.

Description

[Technical Field]

[0001] This invention relates to an information processing technology that supports the investigation of the causes of defects and the evaluation of countermeasures related to the electromagnetic compatibility (EMC) of electronic equipment, and more particularly to a technology that supports the evaluation of design changes through active observation based on design data and search control based on the refutation of causal hypotheses. [Background technology]

[0002] In recent years, attempts have been made to apply AI technology to EDA (Electronic Design Automation) and inspection processes with the aim of improving the efficiency of design and manufacturing. For example, a technology has been proposed that uses AI to detect design clones by imaging design code such as RTL (Register Transfer Level) and reusing design assets and determining similarity in logic design (Patent Document 1). A technology that automatically links design requirements on a circuit diagram to nets and elements and checks the consistency of the design is also known (Patent Document 2). Furthermore, in the manufacturing process, there is a technology that automatically maps wafer inspection images to the coordinates of the design data and identifies the location of physical defects (Patent Document 3). [Prior art documents] [Patent Documents]

[0003] [Patent Document 1] U.S. Patent No. 11429768 [Patent Document 2] Chinese Patent Application Publication No. 121118817 Specification [Patent Document 3] Korean Patent Publication No. 102252326 [Non-patent literature]

[0004] [Non-Patent Document 1] "Practical Aspects of EMC Design - Controlling Radiated Interference Waves," by Michel Maldigian, supervised by Mitsutoshi Hatori, translated by Takehiko Kobayashi, 2nd printing, Maruzen Co., Ltd., published October 20, 2001. [Overview of the Initiative] [Problems that the invention aims to solve]

[0005] However, the prior art described above has the following problems.

[0006] The technology described in Patent Document 1 deals with logic design code and does not explore the causes of electromagnetic malfunctions resulting from physical layout wiring or layered structures. Furthermore, the technology described in Patent Document 2 focuses primarily on attribute matching on circuit diagrams and does not consider the extraction of a "shape context"—how physical shape (geometric shape) relates to electromagnetic risks (antenna properties, coupling properties, etc.)—or the exploration and control of causal relationships based on it. The technology described in Patent Document 3 is a coordinate mapping aimed at identifying manufacturing defects such as scratches and debris, and does not have a framework for elucidating the complex causal relationships resulting from the interaction (SPV) of the source, path, and victim of electromagnetic malfunctions.

[0007] In EMC design aimed at suppressing electromagnetic interference (EMI) and improving immunity (EMS), empirically based diagnostic flows have traditionally been proposed. For example, Non-Patent Literature 1 discloses a detailed flowchart for diagnosing radiated interference and improving prototypes. Non-Patent Literature 1 systematically presents "recipes" and design philosophies for countermeasures based on physical phenomena, such as "if radiation does not stop even after disconnecting the cable, suspect direct radiation from the enclosure" and "insert a ferrite core to block common-mode current."

[0008] However, even with such excellent knowledge (recipes), actual EMC countermeasures are often extremely difficult. Firstly, multiple physical mechanisms (e.g., differential mode conversion, power plane resonance, cable antennaization, etc.) are involved in a single malfunction phenomenon, and it is not easy to identify which mechanism is dominant. Secondly, countermeasures in one area (e.g., strengthening shielding) can easily lead to a "whack-a-mole" situation where countermeasures in one area (e.g., strengthening shielding) change the resonant frequency of another area or cause energy to leak to a different path through reflection, thus creating new malfunctions. In particular, unless the noise energy itself is lost (dissipated as heat) and dissipated, the energy continues to wander within the system in different forms, so countermeasures frequently fail to converge.

[0009] Furthermore, with the increasing density and complexity of electronic devices in recent years, the number of possible countermeasure combinations has become enormous, making it extremely difficult even for experts to select the shortest solution within limited time and cost. Moreover, conventional diagnostic approaches assume the completion of prototypes or design freezes, and end with diagnosing the cause of the failure, lacking a means to pre-evaluate EMC risks at the design stage. For example, methods such as rapidly iterating through a design loop of CAD modification → recalculation → evaluation, quantitatively determining the improvement effects and side effects of design changes without physical prototyping or full-wave simulation of the entire system, and screening at the CAD stage before manufacturing (virtual DRC (Design Rule Check), input is based on virtual sources (full net excitation or critical net excitation) for potential risk exploration) were not provided in conventional technology.

[0010] Thus, conventional AI application technologies lacked a "search control" mechanism that integrates experience, measurement, and analysis into a single logical foundation called "SPV," and actively refutes (prunes) the hypotheses generated by the system based on user input, thereby converging to the true cause via the shortest path.

[0011] This invention was made to solve such problems, and aims to improve efficiency in an auditable and reusable manner while enabling EMC evaluation even during the design process. [Means for solving the problem]

[0012] In one embodiment, the physical design support system is a physical design support system that supports the search for the cause of a defect in physical design, and includes a search state management means that stores the causal relationship of the target defect as a causal hypothesis record including a combination of excitation source, coupling path, and victim, as well as physical state transitions (conversion) in the coupling path, and manages the set of said causal hypothesis records as a search state, and a search control means that prioritizes the causal hypothesis records included in the search state based on input defect information, design data attribute information and / or user input, and updates the search state by rejecting the causal hypothesis records based on observed facts based on comparison with counter-evidence conditions, and determines the next observation target to be verified based on the result of the prioritization or rejection, and the search control means A physical design support system comprising: an active observation generation means that generates observation parameters for an object to be observed, using the internal attributes of design data, including at least one of viewpoint change, transparency display, cross section extraction, layer selection, and magnification change, and generates a diagnostic image according to the observation parameters; a feature extraction means that extracts physical features from the diagnostic image; and an update means that reprojects the physical features in relation to the components of the causal hypothesis record based on geometric coordinates or design attributes, and updates the state of the causal hypothesis record referenced by the search control means, thereby reflecting this in the next prioritization and rejection decision by the search control means, wherein the search control means, the active observation generation means, the feature extraction means, and the update means constitute a closed loop in which hypothesis verification and convergence of the search state are repeated by returning the output of the update means to the next decision of the search control means. In one embodiment, the search state management means is characterized in that, in the causal hypothesis record, it maintains a path sequence as an essential element, which is a sequence of path state transitions in which a plurality of connection paths and state transitions that define physical mode conversion or propagation mode changes between the connection paths are arranged alternately. In one embodiment, the search state management means manages a training data database that stores definition information of the causal hypothesis record and a log database that records the search history and observed evidence as physically or logically independent storage areas, and the causal hypothesis record is characterized by storing reference information to the log database. In one embodiment, the search control means accepts at least one of the facts, constraints, design intent, and verification results input by the user as boundary conditions for updating the search state, and rejects the causal hypothesis record based on a comparison of the boundary conditions with the refutation conditions defined in the causal hypothesis record. In one embodiment, the feature extraction means calculates a physical scale normalization index that indicates the relative scale of the geometric shape on the design data with respect to the wavelength of the physical phenomenon in question when extracting physical features from the diagnostic image, and the update means updates the priority of the causal hypothesis records based on the index. In one embodiment, the search control means virtually modifies the physical parameters corresponding to the connecting paths and / or state transitions included in the causal hypothesis record based on user input without updating the design data, calculates the change in the risk score based on the virtual modification using the physical scale normalization index and / or the physical evaluation formula, constant, or approximation model stored in the causal hypothesis record, and presents the result to the user. In one embodiment, the search control means calculates the amount of risk score reduced by the virtual change as an improvement degree, calculates an increment of the risk score occurring in another causal hypothesis record as a side effect, presents the improvement degree and the side effect in comparison, and updates the design data only when an approval input from a user is received. In one embodiment, the search control means enables selection of the accuracy and calculation load of an evaluation model used in the virtual evaluation in a plurality of stages, and selectively executes at least one of evaluation by a lightweight approximate model and evaluation by detailed analysis. In one embodiment, a physical design support method is a physical design support method for a computer to support searching for a cause of a defect in physical design, the method comprising: (a) a step of holding a causal hypothesis record; (b) a step of performing prioritization and rejection to update a search state; (c) a step of generating a diagnostic image based on an observation target; (d) a step of extracting and re-projecting physical features; and (e) a step of virtually changing physical parameters without updating design data, calculating a change amount of a risk score to present to a user, and updating the design data in response to approval, wherein the search state is converged by repeating the steps (b) to (e). In one embodiment, a program causes a computer to execute the above method. Advantageous Effects of Invention

[0013] According to the present invention, it is possible to provide a physical design support system, a physical design support method, and a program that can perform EMC evaluation even during design and can improve efficiency in an auditable and reusable manner. Brief Description of Drawings

[0014] [Figure 1] It is a diagram showing the overall configuration and data flow of a physical design support system. [Figure 2] It is a flowchart showing a search and evaluation process including a virtual design loop. [Figure 3]This diagram shows the logical structure of causal hypothesis records (SPV hypothesis records) that are managed as exploration states. [Figure 4] This diagram shows a specific application flow using radiated EMI malfunctions as an example. [Figure 5A] This figure shows an example of the configuration of a CAD active observation screen generated by an active observation generation means. [Figure 5B] This figure shows an example of the configuration of a CAD active observation screen generated by an active observation generation means. [Figure 5C] This figure shows an example of the configuration of a CAD active observation screen generated by an active observation generation means. [Figure 6] This figure shows an example of a CAD diagnostic image generated by an active observation and generation means for verifying asymmetry around differential vias in the inner layers of a printed circuit board. [Figure 7] The flowchart shows the details of the priority calculation process. [Figure 8] This is a block diagram showing the information processing process of the state update / research unit 130 (normalization of user input and reconstruction of the search space). [Figure 9] This is a block diagram showing the algorithm for evaluating proposed countermeasures (a comprehensive evaluation flow for improvement and side effects). [Modes for carrying out the invention]

[0015] 0. Definition of Terms The main technical terms used in this specification are defined as follows:

[0016] SPV: A framework that describes the causal relationship of a malfunction using three elements: the excitation source, the connection path, and the victim. It may include physical state transitions (conversion) within the path as needed.

[0017] • Active observation: This refers to the process of generating observation parameters for an observation target selected by the search and control means, based on the internal attributes of the design data, including at least one of viewpoint change, transparency display, cross-sectional extraction, layer selection, and magnification change, and then controlling the CAD display with these parameters to obtain diagnostic images used to verify causal hypotheses. It is not limited to a fixed view, and the observation conditions can be updated according to unsatisfied evidence, consistency with counter-evidence conditions, and risk scores.

[0018] • PEL (Physical Evidence Log): A log format that records hypotheses, evidence, and judgment results obtained during the exploration process in chronological order. SPV hypothesis records do not retain PEL entities, but only reference information.

[0019] • Physical characteristics: This refers to a general term for information including geometric features (length, spacing, aperture area, loop area, asymmetry, etc.) extracted from diagnostic images or design data, and their corresponding physical tags (Resonance, Crosstalk, Reflection, etc.).

[0020] • Noise DRC (or EMC Design Rule Check): This is a design rule verification function specifically for electromagnetic compatibility (EMC) performed on the EDA system. Unlike general manufacturing DRC, which determines whether a pattern can be formed, this technology automatically determines compliance with geometric "design guidelines" for noise suppression, such as the distance of bypass capacitors from the vicinity of IC power pins, the presence or absence of wiring crossing plane slits (notches) (continuity of return paths), the equality of differential wiring lengths, or crosstalk between wirings. The output format is generally a binary judgment result of "OK / NG (pass / fail)" for each element, based on a comparison between fixed thresholds set by the user in advance and geometric measurements on the CAD.

[0021] • Reprojection: This refers to the process of recording physical features extracted from a diagnostic image and associating them with one of the Source, PathSequence, or Victim elements in a causal hypothesis record, based on CAD element IDs (net name, part reference, etc.) and coordinate correspondence.

[0022] • Internal attributes of design data: Attribute information held by the design data, including at least one of the following: coordinates, layer, net name, component attributes (rating, part number, etc.), and shape parameters (wiring width, spacing, via diameter, etc.).

[0023] • Exploration State: Defined as a set of causal hypothesis records. Each causal hypothesis record includes at least the following fields: Status, Score, RefutationCondition, and EvidenceRefs.

[0024] • Risk Score: A numerical value indicating the degree of risk in a causal hypothesis record. It is calculated from a physical scale normalization index, an estimate of the joint probability, accuracy coefficients (recipe accuracy and simulation accuracy), and a gain coefficient based on the refutation results. Specific calculation methods are illustrated in the following chapters.

[0025] • Accuracy: A coefficient that represents the reliability of an estimate or evaluation model. A typical example is recipe accuracy C. recipe and simulation accuracy C sim Includes.

[0026] • Physical scale normalization: This refers to the process of evaluating geometric quantities based on the effective wavelength of the target frequency, in order to enable comparison of structures with different dimensional systems using the same standard.

[0027] • Dimensionlessization: This refers to the operation of converting physical quantities such as length into dimensionless indices (e.g., L / (λ / 4)) when performing physical scale normalization.

[0028] • Effective wavelength: The wavelength in the medium at the target frequency f, and the effective relative permittivity ε eff in view of

number

[0029] • Design intent: This refers to design constraints, including the designation of non-modifiable areas, priority of countermeasures, and designation of areas where exploration is prohibited. If the design intent is retained as an attribute in CAD, it is included in the design data (170); if it is given as an additional instruction outside of CAD, it is accepted as user input (180).

[0030] Metamodel evaluation: This refers to an evaluation method that combines recipes (mathematical formulas and constants) with geometric displacements on images to estimate the physical effects of virtual design changes without actually modifying the CAD data. Specifically, it calculates the effect by substituting one of the following into geometric parameters extracted from the recipe and images: a physical scale normalization index (e.g., dimensionless length L / (λ / 4), loop area change, GND impedance change estimation formula, CM-DM conversion coefficient approximation formula, etc.). This allows for rapid evaluation of effects and side effects while avoiding heavy calculations such as full-wave analysis.

[0031] The image feature extraction unit (feature extraction means) in this specification can be implemented using either a pre-trained model (e.g., a visual language model: VLM, etc.) or a rule-based image processing algorithm. Even when using a pre-trained model, the operation of this system does not require updating the weights within the model.

[0032] 1. System Configuration As shown in Figure 1, the physical design support system of the present invention comprises a knowledge structuring unit 110, a priority assignment unit 120, a state update / research unit 130, an active observation instruction unit 140, an image feature extraction unit 160, user design data (170), user input (180), and three physically and logically independent databases (recipe collection DB90, SPV training data DB100, PEL log DB105), and adopts a configuration in which data flows organically connect each module.

[0033] Here, the search state management means in the claim is implemented by the priority assignment unit 120 and the state update / research unit 130 in Figure 1, the search control means and update means mainly correspond to the state update / research unit 130, and the active observation generation means corresponds to the active observation instruction unit 140.

[0034] (1) Definition of the database structure This system has the following three independent databases.

[0035] (a) Recipe collection DB (90): A database that stores expert knowledge in natural language format and is readable by humans.

[0036] (i) SPV training data DB (100): This is a master DB that stores SPV causal records generated by structuring recipes, and is used for priority calculation and search control.

[0037] (c) PEL Log DB (105): This is a log database that stores the search history, evidence, and refuted results for each design project.

[0038] In addition, there is design data (CAD data) for the design unit, and user input data such as measurement data and additional instructions. (e) User's design data (170): This is the CAD data of the design targeted by the user, and is used as input to the priority assignment unit 120 for the selection of risk candidate elements and extraction of physical quantities based on search-specific rules. If the design intent can be retained as a CAD attribute, it is incorporated into this design data (170).

[0039] (O) User input (180): This is additional information other than design data (CAD information), and includes facts about EMC measurement results, placement and wiring constraints not reflected in CAD, designation of non-changeable areas, designation of countermeasure priority, etc. It is used as input to the state update / research unit 130 in the form of normalization / constraint input to limit the search space and synchronize evaluation parameters.

[0040] (2) The nature of this system It should be noted that this invention is not a self-evolving AI. The updating of knowledge in this system is described as updating statistical confidence parameters and updating applicable boundary conditions. Specifically, the following three types of data are updated in this system:

[0041] (a) Status and Score of Causal Hypothesis Records: As prioritization and rejection are performed by the exploration control means, the status value (Active, Refuted, Verified, etc.) and risk score of each record are updated.

[0042] (i) Exclusion conditions (incompatible boundary conditions): Extract the physical conditions of hypotheses that have been repeatedly rejected and add them to the SPV training data DB100 as logical exclusion conditions. This is not a weight update for the learning model.

[0043] (c) Statistical confidence parameter: As verified cases accumulate, the recipe confidence C recipe Statistical parameter values ​​such as these are updated. The update is a recalculation of the number of successful validations and the success rate ratio or frequency count from the total number of validations, and is not an update of the weights of the learning model.

[0044] Furthermore, "generalization" in this specification does not mean complete abstraction, but rather the reuse of causal relationships across different designs based on physical scale similarities such as physical scale normalization (dimensionlessness) and impedance matching.

[0045] Furthermore, by repeatedly performing the aforementioned search control and evaluation, the convergence speed of hypotheses is improved, and the time required for design decisions is shortened.

[0046] 2. Specific Information Processing The information flow in this system consists of four stages: initial setup, exploration execution, log storage, and knowledge recirculation.

[0047] First, in the initial construction phase, the expertise of skilled users accumulated in the recipe database (90) is structured by the knowledge structuring unit (110) and stored in the SPV training data database (100).

[0048] Next, during the search execution phase, the priority assignment unit (120) assigns rough scores to hypotheses using only the user's design data (170) and SPV training data DB (100) as input, and outputs an initial hypothesis list with priorities. Subsequently, the state update / re-search unit (130) receives user input (180) (defect information, EMC measurement results, placement and routing constraints, etc.) and the search results, and by comparing them with the output of 120, it rejects many hypotheses and updates the search status while re-evaluating the remaining hypotheses. Furthermore, during the log saving phase, the search history and verification results by the state update / re-search unit (130) are recorded in the PEL log DB (105).

[0049] Finally, in the knowledge recycling phase, the verified cases (Verified cases) verified by the state update unit (130) are recycled to the knowledge structuring unit (110), the statistical confidence parameters of the SPV training data DB (100) are updated, and new effective boundary conditions are added to the recipe collection DB (90).

[0050] (1) SPV structuring process (Figure 1) The knowledge structuring unit 110 constructs the SPV training data DB 100 (accumulation of knowledge) from external knowledge sources (books, public and private (internal) trouble case collections and EMC countermeasure recipes, etc., hereinafter referred to as recipes) through the following multi-stage process.

[0051] (a) Data collection and preprocessing: Text data is morphologically analyzed and context-dependent technical terms (e.g., "differential pair," "return discontinuity," etc.) are tokenized. The knowledge structuring unit 110 converts the EMC countermeasure "recipes" described in natural language or flowchart format, as exemplified in Non-Patent Document 1, into an SPV structure that includes three elements: excitation source (Source), coupling path (Path), victim, and physical state transitions (Conversion), and stores it in the SPV training data DB 100.

[0052] (i) Semantic role labeling: This can be achieved using either natural language processing (NLP) techniques or rule-based methods based on keywords and syntactic patterns. Elements corresponding to the excitation source, victim, and propagation medium are identified, and verbal expressions connecting them ("via ~", "converted to ~") are extracted as physical state transitions (Conversion).

[0053] (c) Consistency verification: Check whether the extracted elements form the logical structure of the SPV, and if the Path spans multiple layers, automatically chain them together as a path state transition sequence (PathSequence).

[0054] (e) Data storage format: Each hypothesis is normalized into a schema such as JSON format as shown in Figure 3, and is persisted in the SPV training data DB100 with a search index that enables vector searching.

[0055] As a specific example, if the recipe described in Non-Patent Document 1, "Removing the external cable will allow the system to pass the radiated emission test," is input, the knowledge structuring unit 110 defines the internal signal source as Source, the board wiring and external cable as Path, the coupling at the connector as Conversion, and the radiated noise as Victim, and further generates the following JSON-format causal hypothesis record which includes a causal relationship (counter-evidence condition) that the malfunction is resolved by "removing the cable."

number

[0056] (2) Prioritizer calculation process (Figure 7) The priority assignment unit 120 takes only the design data (170) and the SPV training data DB (100) as inputs and always assigns a rough score (score_pre) to the initial hypothesis candidates (HypothesisList_pre) to be sent for active observation before outputting them. This is an initial screening before user input (180) and functions as a preliminary step before the subsequent state update / re-search unit 130 performs mass rejection and re-evaluation based on user input. The steps shown in Figure 7 will be explained below.

[0057] Step 700 (Input acceptance and preprocessing) The priority assignment unit 120 receives design data (170) (including layout, wiring, layers, enclosure / cable definitions, bill of materials, net attributes, and circuit connections) and SPV training data DB (100). It normalizes the target design to the analysis target graph and generates a design feature reference table that allows referencing of component nodes, nets, physical domains, enclosures / cables, openings, etc.

[0058] Step 710 (Comprehensive enumeration of potential Sources) The system takes CAD bill of materials (BOM) and circuit attributes (clock / interface / power supply / switching information) and component specification links (frequency, rise time, etc.) as input and classifies components and nets into excitation source categories to create a candidate set. If a component specification link does not exist, the representative bandwidth is estimated from the circuit type and net name. Specifically, clock sources, DC / DC converters, motor drivers, high-speed differential interfaces, RF transmitters, switching power supply loops, etc. are output as a Source candidate list (S_list). Each element includes ID, category, representative frequency band including harmonics, intensity index (e.g., default values ​​for switching frequency and circuit type), and associated nets / regions.

[0059] Step 720 (Comprehensive enumeration of victim candidates) The system takes CAD bill of materials and circuit attributes (receiving system / high sensitivity system / threshold system), enclosure / antenna / aperture / external interface information as input and identifies potential victims, including not only components but also phenomena and affected areas. Specifically, it outputs a victim candidate list (V_list) including wireless reception (2.4 / 5GHz band), GNSS, IMU / ADC, reset line, antenna vicinity area, far electromagnetic field (radiation standard), etc. Each element includes ID, category, sensitivity band, tolerance index (e.g., receiving sensitivity threshold, circuit type default value), and location / area.

[0060] Step 730 (SV Pairing: Search Key Generation) This function takes S_list and V_list as input and generates the minimum key required to search the SPV training data DB (100) without creating a Cartesian product. The search key (SV_key) is a combination of S category, V category, frequency band matching label, enclosure / cable type label, proximity label, etc., and is output as a set of search keys (SV_key_list) rather than a Cartesian product of individual (S,V) instances to avoid path combinatorial explosion.

[0061] Step 740 (Extracting and selecting candidates for Path information by referencing DB100) The SV_key_list and SPV training data DB(100) are used as inputs. For each SV_key, the SPV causal records stored in DB(100) are referenced to extract and candidate the top K risky path information (including conversion information as needed). The focus here is not on enumerating specific wiring, but on candidate information including the conditions for fulfillment and observation requirements. The output is a candidate list (H0_list), where each element includes S candidate, V candidate, Path template ID, Conversion template ID, conditions for fulfillment, and observation requirements.

[0062] The template referred to here is a reference format obtained by extracting Path / Conversion information from the SPV causal records stored in the SPV training data DB (100) that is referenced in the initial screening.

[0063] Step 750 (Coarse filtering of feasibility based on design attributes) The system takes the candidate list H0_list and the design feature reference table as input and checks the template's validity conditions against the design attributes to exclude candidates that clearly do not meet the requirements. Examples of exclusions include templates that assume cables but lack cable definitions, templates that are antenna-coupled but lack antennas / apertures, and cases where the bandwidths do not match. The output is the filtered candidate list (H1_list).

[0064] Step 760 (Rough Scoring) The program takes H1_list, CAD attributes, and pre-weights from the SPV training data DB (100) as inputs and always synthesizes a score (score_pre) that can be calculated without user input. Examples of calculation elements include bandwidth matching (degree of agreement between S-bandwidth × V-bandwidth × template bandwidth), geometric scale matching (physical scale normalization index such as cable length, aperture dimensions, and rough loop estimation), proximity (distance between S and V, presence or absence of partitions / enclosures), and DB pre-weights (frequency, risk, generality). The output is a scored candidate list (H2_list), where each element has a score_pre.

[0065] Step 770 (Priority List Output) The H2_list is taken as input, sorted by score_pre, and the top N items are placed in the active observation queue. The output is the initial hypothesis list (HypothesisList_pre) and the observation queue (ObservationQueue). Each element of the initial hypothesis list contains a hypothesis ID, S, V, Path template, Conversion template, score_pre, and evidence summary, while the observation queue contains an observation request and recommended view type (cross-section / transparency / layer / magnification, etc.) as a request to the active observation instruction unit 140.

[0066] Step 780 (Status update / Handover to re-search unit 130) HypothesisList_pre and ObservationQueue are registered with the state update / re-search unit 130 as the initial state of the search state (set of causal hypothesis records). This allows them to be carried over to the next cycle (active observation → feature extraction → re-projection). Most of the passed hypotheses are rejected by the state update / re-search unit 130 (described later) by comparison with user input (180) (defect information, facts about EMC measurement results, etc.), and only the remaining hypotheses are subject to re-evaluation.

[0067] (3) Reconstruction algorithm of the search space in the state update / research unit 130 (Figures 8 and 3) The state update / research unit 130 performs the following information processing procedure (Figure 8) based on the initial hypothesis handed over from the priority assignment unit 120 and the accepted user input (180). It compares the defect information included in the user input (facts of EMC measurement results, placement and wiring constraints, etc.) with the counter-evidence conditions, rejects most of the output of 120, and then re-evaluates the remaining hypotheses and reconstructs the search space.

[0068] (a) User data normalization process (810) First, natural language (e.g., "The connector position cannot be changed," "The heat sink cannot be moved") and raw measurement data received through the user interface are subjected to linguistic analysis or data parsing to generate normalized attribute data 811 (attribute vector format), which are common variables within the system. This 811 is normalized so that its physical dimensions match those of each element (Source, Path, Victim) in record 300.

[0069] (i) Search space filtering process (820): Next, based on the design constraints included in the normalized attribute data 811, the state update / re-search unit 130 updates the search exclusion mask (821) to restrict the search space it manages. Note that "masking" in this step does not refer to image processing, but rather to a calculation that logically sets the risk score of a hypothesis containing design elements that the user has specified as unchangeable to 0.

[0070] • Calculation details: For all candidate hypotheses contained in the search space (a set of 300 causal hypothesis records), identify candidates that conflict with the locations designated as "immutable" in 811, and multiply their risk score by a mask value (e.g., 0).

[0071] • Result: The state update and re-search unit 130 updates the set of causal hypothesis records (300) (search space) it manages. This automatically excludes physically impossible candidates in subsequent search processes, and efficiently extracts only feasible solutions.

[0072] (c) Evaluation parameter synchronization process (830): As an update of calculation variables (831), the physical quantities included in the measurement facts are set as evaluation variables referenced by the state update / research unit 130. The re-execution of the priority assignment unit 120 is initiated only at 850.

[0073] • Calculation details: Measured frequency f extracted from normalized attribute data 811 obs The state update / research unit 130 sets this as the center frequency parameter used for score recalculation.

[0074] • Reflection: This results in the geometric resonant frequency being f obs The risk score of the hypothesis record 300, which is adjacent to the record, is calculated within 130 and automatically increases.

[0075] (e) Record status update process (840): As a state / score update (841) for the causal hypothesis record, the state update / re-search unit 130 compares the normalized attribute data 811 with the causal hypothesis record 300 and persists the results to each field in the record. Specifically, it consists of the following substeps.

[0076] • Logical matching (matching logic): The system determines whether the verification event (e.g., verification action and its result) included in the normalized attribute data 811 matches the trigger and expected value defined in the "Refutation Condition 350" within the record. For example, if record 350 is defined as "Expected Change: MustPass" when the cable is removed, and the event in data 811 is "Result: NoChange" when the cable is removed, then the refutation (rejection) condition is determined to be met.

[0077] • Status update: If the refutation condition is met as a result of the comparison, the status field of record 300 is updated to "Refuted". Conversely, if the expected change is confirmed and the hypothesis is supported, it is updated to "Verified", and if it is unverified, a status value such as "Candidate" is retained.

[0078] • Recalculating the risk score: For affirmed hypotheses, or other hypothesis records that share affirmed elements, the risk score is increased by multiplying them by a gain coefficient G based on the strength of the evidence. On the other hand, the score of rejected hypotheses is set to a minimum value (e.g., 0) and effectively excluded from subsequent investigation.

[0079] The risk score in this system is calculated, for example, from the following components. • Physical scale normalization index (e.g., proximity of L / (λ / 4) to 1.0) • Estimated coupling probability (e.g., coupling coefficient between adjacent wirings, mode conversion efficiency) • Accuracy coefficient (recipe accuracy C) recipe and simulation accuracy C sim • Gain coefficient G based on the refutation result (added if affirmative, zeroed if rejected)

[0080] • Accumulation of evidence: Normalized attribute data 811 (measured facts, design constraints, etc.) that serves as the basis for determination is recorded in the PEL log DB 105 in PEL (Physical Evidence Log) format. At the same time, only a reference (pointer) to the relevant log record is retained in record 300, and no actual Evidence data is held, thereby maintaining consistency as the master DB. As a result, the history of the search process is saved not as a mere change in numerical values, but as a chain of physical facts. Furthermore, when Refuted is repeated for the same hypothesis, the system extracts its physical conditions (e.g., a specific frequency band or laminated structure) and registers them in the SPV training data DB 100 as "non-conforming boundary conditions". This avoids the application of the corresponding recipe in future searches. Note that this process is implemented not as weight update of the learning model, but as addition of a logical application exclusion condition.

[0081] (e) Presentation of Regeneration Reason and User Interaction: The state update / research unit 130 presents the physical basis for the regeneration of the search space (rejection of hypotheses or change of priority) performed in (d) above to the user for approval. For example, a description in natural language or illustration such as "Since the observed frequency characteristics of the noise level do not match the cable length resonance, the cable radiation hypothesis is rejected and the priority of the housing resonance hypothesis is raised" is generated and displayed on the user interface.

[0082] (f) Search Space Reconfiguration Step (850): Based on the matching result of (d) above, reconfiguration of the search space is activated when any of the following conditions is satisfied. (a) When the hypothesis set delivered from 120 is widely rejected, (b) Observed facts (measured frequency f obs , removal test results, shield addition results, etc.) are less than or equal to a predetermined threshold, or (c) The user specifies a specific Source candidate or Victim candidate with a request such as "I want you to check this part again". These conditions are determined as an extension of rejection and re-evaluation.

[0083] The input for reconstruction is limited to normalized attribute data 811 and design data (170) (including bill of materials and circuit attributes). Enumeration of SV direct products is not performed. The aim is to shift the focus of the search from observed facts and design attributes when existing templates cannot adequately explain the observed facts, and then return the processing to the priority assignment unit 120, active observation instruction unit 140, CAD subsystem 150, and image feature extraction unit 160.

[0084] The reconstruction process is as follows: Observation center frequency f obtained from 810 and 830 obs Using the test operation results (cable removal, shielding addition, etc.), (i) the categories and bandwidths of the Source candidate set (clock, DC-DC, high-speed interface, etc.) and Victim candidate set (wireless receiver, high-sensitivity line, aperture, etc.) that can be extracted from the design data (170) are updated; (ii) the parameters such as category, bandwidth matching label, enclosure / cable type label, proximity label, etc. used by 120 to generate the DB (100) search key are reset; and (iii) 120 is requested to re-screen (re-execute Figure 7) with the reset key. Instead of keeping SV as a Cartesian product and searching, the assumptions for key generation (category, bandwidth, proximity, etc.) are updated with observed facts, and 120's search and scoring are redone.

[0085] The basis for initiating the reconstruction (reasons for broad rejection, observed facts that resulted in inconsistencies) is recorded in the PEL log DB (105). The hypothesis record side only retains a reference to the log.

[0086] (4) CAD active observation and image feature extraction The active observation instruction unit 140 generates specific observation parameters for controlling the CAD subsystem 150 based on the logical observation instruction received from the state update / re-search unit 130, which is "the next target to be verified (e.g., return path around a via spanning multiple layers)."

[0087] (a) Correspondence between observation trigger conditions and instruction parameters: The state update / research unit 130 checks the risk score, the matching result of the RefutationCondition, and the observation center frequency f. obs Based on the degree of consistency with and the presence or absence of unsatisfied evidence, the target and observation conditions for active observation are determined. The active observation instruction unit 140 instructs the CAD subsystem 150 of the observation parameters according to this determination. Specifically, if the uncertainty of the path is high and unsatisfied evidence remains, the instruction unit will specify conditions to emphasize the net of interest and adjacent GND layers and to display other layers transparently. If resonance or the involvement of a long conductor is suspected, f obs If the degree of consistency is high, observation conditions including multi-magnification zoom and local magnification are specified. If the involvement of mode conversion is suspected and asymmetric elements remain without refutation, visualization conditions using cross-sectional and isometric projections are specified. As a result, the generated diagnostic images satisfy observation conditions suitable for physical feature extraction performed by the subsequent image feature extraction unit 160.

[0088] (i) Generation of observation parameters and API control: The active observation instruction unit 140, in order to optimally visualize the instructed object, refers to the geometric attributes of the object (coordinates, net name, layered structure, etc.) from the CAD data and issues API commands that involve the following information processing.

[0089] • Layer filtering instruction: By displaying only the target net and adjacent GND layers, and setting the transparency of other objects to α ≤ 0.1, a diagnostic image with visual noise eliminated is generated. Note that this threshold is just an example and can be set according to the design target while ensuring screen visibility.

[0090] • Multi-view camera instruction: A set of viewpoints (top view, cross-sectional view, isometric projection) that three-dimensionally encompass the start, end, and via positions of the network are generated simultaneously with an appropriate zoom magnification based on the wavelength λ (e.g., a 2λ square field of view). Furthermore, the viewpoint is dynamically controlled according to the level of detail of the design data. For example, if a 3D model of the connector is available, a cross-sectional viewpoint that allows confirmation of the internal pin structure is automatically instructed. If housing data exists, rendering is automatically instructed at an arbitrary angle (such as an oblique viewpoint) that makes the radiation and intrusion paths from openings and gaps (seams) most visible.

[0091] (c) Image feature extraction and reprojection onto causal records: The image feature extraction unit 160 is a broad functional block that recognizes geometric features from the generated image data and extracts physical risks (evidence). Here, it performs the following processing using a pre-trained model (Visual Language Model: VLM, etc.) and image processing algorithms.

[0092] • Physical scale evaluation: Calculate the physical length L of the elements in the image and determine the dimensionless length L / (λ / 4) using the effective wavelength λ.

[0093] • Reflection in causal hypotheses: Based on the feature extraction results, physical tags such as "Resonance Context" and "Crosstalk Context" are generated and added to the Evidence field of the corresponding 300 causal hypothesis records, and the risk score is updated.

[0094] Here, reprojection is performed as a process that (i) associates feature regions on the diagnostic image with coordinates based on a physical scale bar or CAD coordinate metadata, (ii) identifies CAD element IDs from internal attributes of the design data corresponding to those coordinates (net name, part reference, layer information, etc.), and (iii) associates the obtained physical tag and CAD element ID pair with the corresponding elements of Source, PathSequence, or Victim in the causal hypothesis record 300. As a result, the search control means can consistently refer to the grounds necessary for the next prioritization and rejection decision via EvidenceRefs, which are references to evidence entities recorded in the PEL log DB 105, in addition to updating the Status and Score of the record.

[0095] In this specification, the image feature extraction unit refers to a functional block that recognizes geometric features from input image data, and is a broad concept that includes pre-trained models that have learned the correlation between geometric shapes and physical risks (e.g., visual language models: VLM, etc.) and rule-based image processing algorithms. In other words, the image feature extraction unit can be implemented using either a pre-trained model or rule-based image processing, and even when using a pre-trained model, the operation of this system does not require updating the weights within the model.

[0096] The image feature extraction unit 160 extracts electromagnetic risks from the input image using the following procedure.

[0097] (a) Extraction of geometric elements: A recognition model (e.g., a CNN or Vision Transformer-based detector) is used to segment the image into line segments (wiring), apertures (slots), and closed loops (wiring and return paths).

[0098] (b) Pixel / Length Conversion: Converts pixel measurements to physical length L by referring to either a physical scale bar or CAD coordinate metadata. If a scale bar is available, L is calculated from the number of pixels; if the image is CAD-derived, L is calculated from the coordinate metadata.

[0099] (c) Physical scale evaluation: The dimensionless length L / (λ / 4) is calculated using the effective wavelength λ calculated from the target frequency.

[0100] (d) Reprojection onto causal hypotheses: Based on the results of feature extraction, physical tags are generated, such as "Resonance context" if the dimensionless length is in the range of 1.0 ± 0.2, "Crosstalk context" if the distance between adjacent wirings is below a threshold, or "Reflection context" if impedance mismatch at a specific frequency is confirmed. Here, 1.0 ± 0.2 is just one example for determining whether the target structure falls within the resonance conditions near λ / 4, and can be adjusted according to the target device and frequency band. The image feature extraction unit 160 adds these tags to the attribute fields of the corresponding SPV hypothesis record (300) and updates the risk score of the record according to the strength of the discovered physical evidence (for example, adding +0.3 points if the resonance length matches).

[0101] 3. Design Loop and Quantitative Evaluation (Figures 2 and 9) As shown in Figure 2, this system has a loop structure that virtually verifies the effectiveness of design changes (countermeasures). Specifically, in steps 235 (evaluation of virtual design changes) and 270 (evaluation of iterations), the system executes the evaluation algorithm (900) shown in Figure 9 to quantitatively determine the validity of the proposed countermeasures.

[0102] (1) Steps for calculating the degree of improvement (910): The system calculates the reduction in the risk score of the 300 targeted causal hypothesis records based on the CAD data and attribute information after the design change.

[0103] Input: Modified CAD geometry, modified part attributes (driver capabilities, etc.).

[0104] ·process: • Shape factor evaluation: Re-observation is performed using the image feature extraction unit 160 to measure the decrease in the resonance risk index L / λ due to wiring shortening, etc. • Source Factor Assessment: The reduction in emission intensity due to component replacement (e.g., changing to a driver with a lower slew rate) is calculated based on attribute vectors.

[0105] • Output: Improvement score ΔScore imp .

[0106] (2) Side effect extraction step (920): The system automatically detects any new or increased side effects resulting from the modifications.

[0107] • Processing: Candidate side effects are extracted from the surrounding network of the modified area and re-evaluated within the state update / re-search unit 130. Note that the priority assignment unit 120 is only re-executed when 850 is activated.

[0108] • Detection examples: Waveform blunting (signal quality deterioration) due to the addition of damping resistors, and the occurrence of self-resonance at a different frequency due to the addition of capacitors are generated as new causal hypothesis records 300.

[0109] Output: Side effect score ΔScore side .

[0110] (3) Overall evaluation score calculation steps (930) and judgment The system integrates the degree of improvement described in (1) above and the side effects described in (2), and calculates an overall score (931) that takes into account the accuracy of empirical rules (recipes) and physical calculations (simulations).

number

[0111] Here, each variable has the following ranges and units as an example. C recipe (Recipe accuracy): Past success rate (0.0~1.0). S recipe (Recipe effect): General effect value (e.g., dB). C sim(Simulation accuracy): A variable parameter (0.0 to 1.0) depending on the accuracy of the calculation method. For example, it is set low (e.g., 0.6) when only image feature inference is performed, and high (e.g., 0.95) when high-precision validation such as partial full-wave analysis is performed. S sim (Simulation effect): The calculated change in physical quantities (e.g., dB).

[0112] The execution of partial full-wave analysis (925) is a process in which full-wave analysis (partial analysis) is applied to a limited area that has been bounded by the image feature extraction unit as a "dangerous region". This avoids the computational cost of full-system analysis (several hours to several days) and enables high-precision electromagnetic verification (calculation of S-parameters and physical quantities of the near field) in a computation time of several minutes to several hours. The obtained analysis results show the simulation effect S sim It is substituted as, and at the same time, the simulation accuracy C sim This triggers the update to the maximum value.

[0113] (4) Judgment and data output: Calculated ΔScore imp and ΔScore side The system compares the effects and determines that an action is "effective" if the improvement outweighs the side effects. This determination result and the overall score (931) are stored in temporary memory or a database and output as indicators for the next step, "presentation and prioritization of proposed countermeasures (250)".

[0114] 4. Detailed definition of causal hypothesis record (SPV hypothesis record) (Figure 3) In this system, the causal relationships of malfunctions are structured as "SPV hypothesis records" and managed using data structures such as JSON format.

[0115] This chapter explains how the data entities passed between the functional blocks shown in Figure 1, primarily the knowledge structuring unit 110, the priority assignment unit 120, the state update / re-search unit 130, and the image feature extraction unit 160, in steps 200-250 of Figure 2, correspond to the record items in Figure 3.

[0116] (Overview of the item-by-item data collection flow) To clarify the feasibility of implementing SPV hypothesis records and PEL logs, the "collection source," "preprocessing," and "storage location" for key items are summarized below.

[0117] (1) Structure and logical definition of the SPV hypothesis record (Figure 3) Figure 3 shows the logical structure of a causal hypothesis record (SPV hypothesis record) 300 managed as a search state in the present invention. This record consists of multiple fields for strictly defining physical causal relationships.

[0118] The core component of this record is the PathSequence (path state transition sequence) 320. This does not treat the path from the excitation source to the damaged side as a single element, but rather as an alternating sequence of multiple Path elements and Conversions (physical transition types) 321 that physically transition them between paths. For example, the process by which a differential mode signal transitions to a common mode signal via an unbalanced path, or the process by which conducted energy is radiated into space, are explicitly defined in this Conversion 321. This structuring allows the system to logically track the causal relationship of "which physical phenomenon is causing the malfunction."

[0119] Furthermore, this record 300 holds the refutation criteria 350, which serve as the basis for rejecting each hypothesis. The refutation criteria 350 are compared with the contents of the Evidence field obtained through active observation or with verification results entered by the user. If the observed facts satisfy the refutation criteria 350, the system updates the Status field of the record to "Refuted" and dynamically converges the search space.

[0120] (a) Collection source, preprocessing, and storage location for each SPV item (a) The HypothesisID is assigned when the initial hypothesis list is generated in step 770. As a preprocessing step, uniqueness verification and duplicate removal are performed, and it is stored as the primary key of the record in the SPV training data DB100.

[0121] (b) Source / Victim data is collected from design data 170, bill of materials, net attributes, and user input 180. Preprocessing includes attribute normalization, categorization, and bandwidth estimation, and the data is stored in the Source / Victim field of record 300.

[0122] (c) PathSequence / Conversion is collected from the extraction and candidate selection of Path information by referencing DB100 (740) and the design attribute matching results (750). As a preprocessing step, the conditions for fulfillment and path consistency are checked and saved to the PathSequence array of record 300.

[0123] (d) The score is generated from rough scoring (760) and recalculation (840). Preprocessing includes weight merging, gain coefficient application, and zeroing, and the score is stored in record 300.

[0124] (e) The Status is determined from the refutation condition matching (840) and the user approval result (240). As a preprocessing step, a refutation logic evaluation (Refuted / Verified / Candidate) is performed and saved in the Status of record 300.

[0125] (f)RefutationCondition is collected from the structured recipe (knowledge structuring unit 110) and the example definition as conditions indicating "under what observational conditions the hypothesis should be rejected". As a preprocessing step, the conditional statements are normalized into pairs of triggers (observation operations) and expected values ​​(expected changes), and stored in the RefutationCondition of record 300.

[0126] (g) EvidenceRefs are collected from the record IDs in the PEL log DB105. As a preprocessing step, referential integrity checks (missing and duplicates) are performed and saved to EvidenceRefs for record 300.

[0127] Below is a specific example of a JSON schema for an SPV hypothesis record.

number

[0128] (2) PEL log (Physical Evidence Log) and PEL log DB105 The exploration history is recorded in PEL format in the PEL log DB105. This format adds "evidence" and "outcome" to a "hypothese." The SPV hypothesis record (DB100) does not directly store these logs, but manages only the reference ID to maintain the independence of the master data.

[0129] Furthermore, to compensate for the lack of recipes in the initial stages of operation, recipe candidates and SPV causal record candidates may be generated from past logs and design documents using a natural language processing system or a large-scale language model, and registered in the database after approval by the designer or administrator.

[0130] (a) Generation triggers, recorded content, and SPV reflection for each PEL item (a) Hypothesis is generated when 840 matching starts or when 850 reconstruction is initiated. The recorded content is the target hypothesis ID and the target Path / Conversion element, and the SPV side references the log ID as EvidenceRefs.

[0131] (b) Evidence is generated when active observation results are input and when the normalization of user input 180 is completed. The recorded contents are measured values, constraints, image extraction features, and confidence levels, and the SPV does not retain the Evidence entity but only retains a reference.

[0132] (c) The Outcome is generated when the refutation or re-evaluation is completed. The recorded contents are the status transition, the amount of change in score, and the reason for rejection, and this is reflected on the SPV side as the status / score update result.

[0133] (d) The Timestamp is added when the PEL record is generated. The recorded content is the recording time and process number (e.g., 840 / 850), and it is not reflected on the SPV side but is kept on the PEL side for audit tracking purposes.

[0134] (e) The equivalent of an Actor is generated when the processing entity is determined. The recorded content is either from automated processing (sections 120 / 130 / 160) or user input, and is not stored on the SPV side but is audited and managed on the PEL side.

[0135] ·Hypothesis:“Connector unbalance causes CM current.” ·Evidence:“Image Feature Extractor detected standard USB connector without shield handling.”(Confidence:0.9) ·Outcome:“Status updated to Active. Priority increased.”

[0136] 5. Example 1: Multimodal causal exploration in BCI testing (Figures 4 and 6) This embodiment deals with a case in which a malfunction occurred during BCI (Bulk Current Injection) testing of automotive electronic equipment, where the communication error rate increased sharply at a specific frequency, leading to communication failure. The target device is housed in a metal enclosure, and an STP (Shielded Twisted Pair) cable is connected via a board-mounted connector, with a gap between the connector opening and the enclosure. On the internal circuit board, the connector GND (STP shield potential) and the internal logic GND are separated, and a common mode choke coil (CMC) is mounted across this separation gap. These are configured via CAD data import 200 or user input 220.

[0137] (Example 1 Step Summary: Input, Processing, Output, DB Change) In step 410, CAD import information and defect information are taken as input, the defect phenomenon and design outline are normalized, and the initial conditions for the analysis target are output. At this time, the SPV side prepares a candidate set for starting the search, and no changes occur on the PEL side.

[0138] In step 420, design data 170 and DB100 are taken as input, and hypothesis enumeration and initial weighting are performed to output the initial set of hypotheses. On the SPV side, the initial value of Status and score_pre are assigned, and on the PEL side, the candidate generation history is recorded as needed.

[0139] In step 430, ObservationQueue and CAD attributes are used as input, and active observation and feature extraction are performed to output diagnostic images, features, and confidence scores. On the SPV side, evidence candidates are added and the information used for score re-evaluation is updated, while on the PEL side, observational evidence is recorded.

[0140] Step 440 takes the normalized attribute data 811 and the refutation condition 350 as input, performs a refutation condition check, and outputs a rejection / support decision. The SPV side updates the Status (Refuted / Verified, etc.), and the PEL side records the basis for the decision.

[0141] In step 450, the judgment result from step 440 and the user approval input are used as input to update the search state and redistribute priorities, and the updated search state is output. On the SPV side, the score is updated and the reconstruction conditions are determined as needed, and on the PEL side, the update history is appended.

[0142] Step 460 takes proposed countermeasures and evaluation results as input, performs countermeasure proposal, approval, and branching to the next process, and outputs an approved policy or a request for re-evaluation. The SPV side retains the approval results as boundary conditions, and the PEL side records the approval / rejection log.

[0143] (Correspondence with Figures 1 and 2) Steps 410 to 460 of this embodiment shown in Figure 4 correspond to one search cycle of the search and evaluation process (200 to 270) shown in Figure 2, and the functional blocks operating in this cycle are the priority assignment unit 120, state update and re-search unit 130, active observation instruction unit 140, CAD subsystem 150, and image feature extraction unit 160 shown in Figure 1.

[0144] Specifically, step 410 corresponds to CAD data loading and defect definition 200 in Figure 2, and step 420 corresponds to SPV search calculation 210 in Figure 2 (the process in which the priority assignment unit 120 refers to the SPV training data DB 100 to enumerate hypotheses and assign weights). Furthermore, step 430 corresponds to active observation and feature extraction in the SPV search calculation 210 in Figure 2, and the active observation instruction unit 140, CAD subsystem 150, and image feature extraction unit 160 in Figure 1 are activated. Steps 440-450 correspond to the determination of completion of matching with counter-evidence conditions 230 and user approval / hypothesis update 240 in Figure 2, and the state update / re-search unit 130 in Figure 1 updates the search status while recording the search history in the PEL log DB 105. Step 460 corresponds to the presentation and prioritization of countermeasures 250 in Figure 2 (transitioning to CAD data update 260 and improvement / side effect determination 270 as needed).

[0145] (1) Initial Candidate Listing (Steps 410-420 = Figure 2: 200 → 210, Figure 1: Prioritization Unit 120) Step 410 corresponds to 200 in Figure 2, where the symptoms of the BCI test and the design outline are normalized as input for the search. In step 410, failure information (communication failure in the BCI test) and the design outline (STP, metal housing, differential transmission) are input. Step 420 corresponds to 210 in Figure 2, where the prioritization unit 120 extracts and candidates Path / Conversion candidates by referring to the SPV training data DB 100 and assigns an initial score. In step 420, the prioritization unit 120 selects physically feasible Path / Conversion candidates as an initial screening based on the design data and DB 100 reference. In the initial stage, the following path candidates are generated by referring to known Path / Conversion patterns stored in the SPV training data DB 100.

[0146] (a) Aperture space coupling: Electromagnetic field intrusion through the opening gap between the connector and the housing. (i) Internal enclosure resonance: Concentration of electromagnetic fields at specific points due to cavity resonance inside the metal enclosure. (c) PCB plane resonance: Resonance of the ground plane / power plane of a PCB at a specific frequency. (e) Shield leakage: Common-mode current in the STP cable's outer sheath (Shield) couples to the inner sheath (Inner) at the end. (e) Shield termination unbalance: Conversion from common mode to differential mode (CM-DM conversion) at connector joints (pigtails, etc.). (c) Connector pin imbalance: CM-DM conversion due to geometric asymmetry of the internal pins of the connector. (Ki) Via GND asymmetry: Mode conversion due to asymmetry in the GND via arrangement around differential vias. (c) Asymmetry of parallel wiring: Asymmetry in the length, width, or distance from surrounding conductors of adjacent differential wiring on a PCB. (Ke) CMC pattern asymmetry: Asymmetry in the land pattern or wiring outlet of a common mode choke coil (CMC). (c) Incomplete GND isolation: Coupling between the connector GND and the internal circuit GND due to wiring that crosses the isolation slit or due to stray capacitance. (S) Slit-crossing return disconnection: Mode conversion caused by the signal line crossing the GND slit. (c) GND plane impedance: Fluctuations (bounces) in the GND potential due to the finite impedance of the plane. (S) Power plane bounce: The effect of noise superimposed on the power plane on IC operation. (C) IC package bonding: Nearby bonding between lead frames or wire bonds. (So) Heatsink resonance: The heatsink on the IC acts as an antenna, radiating and coupling to the surroundings. (T) Internal cable coupling: Spatial proximity coupling (Crosstalk) with other internal wiring (flat cables, etc.). (C) Slot antenna radiation: The seams of the housing function as slot antennas at specific frequencies. (T) Insufficient decoupling: Noise between power supply and ground due to malfunction of bypass capacitors at high frequencies. (Te) Improper wiring termination: Generation of standing waves due to superposition of reflected waves, and the resulting amplification of mode conversion. (T) Housing-circuit board joint connection: Current concentration and connection around the screw holes (GND boss) that secure the housing and the circuit board.

[0147] (2) Initial Culling based on physical constraints (Continued from Step 420 = Figure 2:210, Figure 1: Prioritization Unit 120) The priority assignment unit 120 performs initial weighting (pruning) on ​​the above list based on the physical attribute that the victim is a "differential signal input". Since differential receivers have a high common-mode rejection ratio (CMRR), pure "GND / power bounce" type paths that do not have a conversion (mode conversion) element (such as 12, 13, and 18 in the above list) are given a low initial risk score and their Status is set to "Pending". On the other hand, paths that include CM-DM conversion due to "asymmetry" (such as 5 to 11 above) are given a high priority and their Status is set to "Active" because they are highly likely to convert common-mode noise into differential-mode noise and directly cause communication errors.

[0148] (3) Specific example of risk score calculation by the image feature extraction unit (asymmetry around vias) (Step 430 = Figure 2: 210 (active observation and feature extraction), Figure 1: 140 → 150 → 160) Step 430 corresponds to the active observation process included in 210 in Figure 2, where the active observation instruction unit 140, CAD subsystem 150, and image feature extraction unit 160 in Figure 1 work together to generate diagnostic images and extract physical features. In step 430, the active observation instruction unit 140 instructs the CAD subsystem 150 to perform active observation in order to verify "Hypothesis 7: Mode conversion due to via-GND asymmetry," which is one of the high-priority hypotheses.

[0149] (a) Execution of CAD active observation The active observation instruction unit 140 focuses on the via portion where interlayer movement occurs in the differential signal path from the connector to the CMC, and issues the following command.

number

[0150] Based on the instructions, the CAD subsystem generates an inner layer diagnostic image (Figure 6) that highlights the condition of the differential signal vias and the surrounding GND plane.

[0151] (i) Feature extraction and risk score calculation by the image feature extraction unit (see Figure 6) Figure 6 shows an example of the generated CAD active observation screen (inner layer view) 600. The image feature extraction unit 160 extracts the following visual features from this image. Identifying differential vias: Identify the differential pair signal vias (S1, S2) 610 of interest. These are vias on the board that correspond to the board mounting pins G1, S1, S2, and G2 of the STP connector 601.

[0152] • Detecting asymmetry in the return path: Differential traces Sig1 and Sig2 are drawn from signal vias S1 and S2. Near S1, multiple GND1 vias 602 are located, which are part of the return current path and are firmly connected to the inner layer GND plane 630. On the other hand, near the signal via S2 in the GND layer 630, the GND plane's splitting clearance 603 is nearby, and another signal trace 640 is located in the signal layer 620, so it is detected that the arrangement of the GND via is sparser (asymmetrical) compared to the S1 side.

[0153] • Physical interpretation and risk score calculation: The image feature extraction unit interprets this as an "asymmetric structure in which the return path impedance of one side (S2) of the differential pair is higher than that of the other side (S1)." This provides strong evidence of the physical phenomenon in which the common-mode current injected by the BCI causes an unbalanced voltage drop and is converted into differential-mode voltage (noise) (CM-DM conversion). Based on the degree of this asymmetry, the image feature extraction unit calculates a high risk score (Score) for this hypothesis (e.g., 0.85) and visualizes the reasoning on the image as annotation 650 and score display 651.

[0154] (4) Dynamic updating of the search space (Steps 440-460 = Figure 2: 230 → 235 → 240 → 250, Figure 1: State update / research unit 130 + PEL log DB 105) Step 440 corresponds to 230 in Figure 2, where the state update / research unit 130 compares the refutation conditions 350 with the observed facts to determine whether to reject or support the hypothesis. Between steps 440 and 450, the process may also pass through 235 in Figure 2 (evaluation of the effects and side effects of the virtual design change). Step 450 corresponds to 240 in Figure 2, where the state update / research unit 130 updates the search state and determines the reconstruction conditions of 850 as necessary. Only if 850 is activated, the priority assignment unit 120 re-screens (re-execution in Figure 7). Step 460 corresponds to 250 in Figure 2, and the process transitions to 260 and 270 as necessary to evaluate the effects and side effects of the design change.

[0155] In step 450, based on the high risk score calculated by the image feature extraction unit, the state update / re-search unit 130 updates the PEL record for hypothesis 7 (via GND asymmetry), significantly increasing the score. At the same time, it lowers the relative priority of other hypotheses that physically compete with this hypothesis (e.g., simple power bounce) and controls the search to concentrate the next verification resources (detailed simulation, etc.) on the analysis around the via.

[0156] If the CAD data lacks information about vias or inner layer GND, the system will notify the user of the incompleteness and, at the same time, continue the search by using alternative means such as referencing estimated parasitic parameters for typical via structures from the SPV causal records (including template reference information) in the SPV training data DB100, or, if a photograph of the substrate cross-section is provided, having the image feature extraction unit estimate the structure from the image.

[0157] 6. Example 2: Radiated noise search in a drone video transmission system (Figures 4 and 5) This embodiment illustrates a process that supports the decision-making process for design changes through dialogue, focusing on the processing of the state update / re-search unit 130. Here, "dialogue" refers not to the UI itself, but to the input means for the search space control. That is, user input is treated as boundary conditions for the search control, which virtually manipulates Path elements based on the state update of the hypothesis record and performs a physical evaluation. This embodiment details a process in which, before costly actual CAD changes (260), the user performs low-computational-cost virtual evaluations and iteratively verifies the design policy, thereby rapidly rejecting hypotheses and converging the design policy to one with high accuracy.

[0158] This embodiment deals with a case where radiated noise in the 2.4GHz band degrades the wireless communication rate in a configuration where a camera sensor board PCB1 and a drone control / signal processing circuit board PCB2 are connected by a USB cable. The process by which the root cause is identified and the design policy is determined through the user-presented screen 500 of the CAD active observation results shown in Figures 5A to 5C is detailed. The basic search flow follows the same steps (410-460) as in Figure 4.

[0159] (Correspondence with Figures 1 and 2) Steps 410 to 460 of this embodiment shown in Figure 4 correspond to one search cycle of the search and evaluation process (200 to 270) shown in Figure 2. The functional blocks operating in this cycle are the priority assignment unit 120, the state update and re-search unit 130, the active observation instruction unit 140, the CAD subsystem 150, and the image feature extraction unit 160, as shown in Figure 1.

[0160] Separation of roles: The active observation instruction unit 140, the CAD subsystem 150, and the image feature extraction unit 160 operate in phase 210 (active observation / feature extraction) in Figure 2, and are responsible for generating diagnostic images and extracting physical features necessary for hypothesis verification. On the other hand, the state update / re-search unit 130 is responsible for presenting hypotheses, receiving and analyzing user input, evaluating the effects and side effects of virtual design changes (235), and presenting and approving results to the user (240). 235 is a lightweight virtual evaluation loop executed within the search control cycle, which involves updating the state of hypothesis records but not updating CAD data. That is, after the refutation and support of multiple hypotheses (H-001~) are completed by 230 and the candidates are narrowed down, the SPV path from V (wireless antenna) to S (excitation source) is specifically determined for the remaining hypotheses. Starting from V has the effect of suppressing combinatorial explosion. Subsequently, in step 235, led by step 130, the narrowed-down candidate paths are virtually modified (by changing constants, bypassing routes, etc.), and the effects and side effects are calculated through metamodel evaluation without updating the CAD data. Through repeated dialogue with the user, the design policy converges. In step 240, step 130 presents the results to the user and obtains their approval, after which step 260 (CAD data update) is executed.

[0161] The following (1) to (4) illustrate how these functional blocks work together through each panel (Figure 5) of the CAD active observation screen 500. The CAD active observation screen 500 consists of the "SPV hypothesis list 510" on the left panel, the "multi-view area 520" in the center, and the "image feature extraction unit diagnosis, evaluation, and countermeasures 530" on the right panel.

[0162] (1) SPV Hypothesis Estimation and Prioritization (510) - Hypothesis Presentation by State Update / Research Unit 130 The state update / re-search unit 130, after completing the refutation and support of multiple hypotheses (H-001~) by 230, specifically determines the SPV path from V (wireless antenna) to S (excitation source) for the remaining hypotheses and reconstructs the search space. The "SPV Hypothesis List 510" on the left panel presents the most promising hypotheses, calculated through the priority calculation process (Figure 7), to the user by 130.

[0163] • Phenomenon (item): Based on the user's observations, interference from video transmission noise in the 2.4GHz wireless band (the wireless communication error rate worsened after video transmission began) is input, and the target is displayed at the top of the hypothetical list 510 as some kind of item. If there are multiple problems, the item is assumed to be selected on a different screen.

[0164] • Target Hypothesis: The most likely hypothesis, "H-001: USB cable radiation caused by video output signal," is displayed. Other hypotheses can be viewed by scrolling.

[0165] • Identification of Source(S): The state update / research unit 130 re-evaluates the source candidates based on frequency compatibility with the victim (2.4GHz radio interference), and selects the video signal driver IC (5210) as the source based on the fact that the harmonic components included in the clock frequency of the video signal driver IC (5210) are close to the 2.4GHz band.

[0166] • Status and score: During the evaluation parameter synchronization process (830), the cable length is close to the 2.4GHz λ / 4 resonance condition (L / (λ / 4)≈1.0). Based on the image analysis of the board layout, the risk score is displayed as "Example: 0.98" and the status as "Active".

[0167] • SPV Display: The noise source (S), path (P), conversion (C), and victim (V) estimated from the phenomenon and image analysis are displayed as a sequence, with a brief display of the actual part name, net name, location, and mechanism of each part. This can be collapsed and expanded to show the SPV for the next hypothesis H-002.

[0168] • SPV Selection: When you select a hypothesis corresponding to an item, the SPV of the analysis result will be displayed in the multi-view area 520, and the diagnosis result (estimated cause) and proposed countermeasures for that hypothesis will be displayed in 530. If you select a different hypothesis, the displays in 520 and 530 will also change accordingly.

[0169] (2) Multi-view area (520) - Diagnostic view generation by active observation instruction unit 140 and CAD subsystem 150 The active observation instruction unit 140 instructs the CAD subsystem 150 to visualize the area from the video vibration output IC (driver IC) 5210 to the USB connector (5220) on PCB1 in the multi-view area 520 in order to verify the coupling path of hypothesis H-001. This operates in phase 210 (active observation) in Figure 2. The results may be displayed in real time via the CAD API, or only the analysis results acquired in advance may be shown.

[0170] • Route View (525): Displays a route view including the entire path of the selected estimated H-001 SPV. The USB connector 5220 and connected USB cable 5230, which connect to the driver IC 5210 mounted on PCB1 inside the camera, are connected to the receiver board PCB2 via a gimbal. This entire path and the physical placement of the Victim's wireless antenna and communication PCB are displayed. The viewpoint can be changed in 2D or 3D, and rotation and zooming are also possible.

[0171] • Area of ​​focus (local magnified view) (522) and cross-sectional view (523): The layout of the signal layer and GND layer, including the discontinuity in the return path from driver IC 5210 to connector 5220, is shown overlaid with highlights 550 and 551.

[0172] • Semantic emphasis (550): Within the area of ​​interest 522, the indicator 140 displays the signal network from the output of driver IC 5210 to 5220 as "net of interest 541," and also displays the layout of the originally intended return path, i.e., the GND layer. Irrelevant layers are hidden by background transparency processing.

[0173] (3) Verification of hypotheses by active image analysis - Physical feature extraction by image feature extraction unit 160 In phase 210, the image feature extraction unit 160 performs geometric element extraction (segmentation) on the generated multi-view images and calculates the following physical risks during scoring.

[0174] • Physical scale evaluation: The length of the USB cable 5230 is calculated from the component information, and based on the dimensionless physical scale normalization index (dimensionless length), it is calculated that the resonant length is approximately close to the 2.4GHz wireless band. It is then calculated that if noise is superimposed on the GND of the connector 5220 to which the cable 5230 is connected, there is a high probability of noise interference to the Victim's wireless antenna (Victim-Path coupling probability).

[0175] • Source frequency matching evaluation: Furthermore, it was confirmed that the harmonic components included in the clock frequency of the video signal driver IC (5210) are close to the 2.4GHz band, and this was reflected in the calculation of the risk score for the hypothesis as the basis for frequency matching on the source side.

[0176] • Hypothesis formulation based on image feature extraction: On the other hand, analysis of the local magnified view 522 and the cross-sectional view 523 revealed that only the differential signal D+ is bypassed by a via hole, and that the GND of the driver IC 5210 and the connector 5220 are not connected by pattern, but are connected with high impedance by the flight bead (FB) 5240 from the circuit information. Using the driver IC 5210 as the source, a hypothesis (H-001) was generated that the differential signal is Path1, the return path of the differential signal is Path2, the common-mode conversion from the wiring length difference of the differential signal is Conversion1, the common-mode is excited by the FB5240 to the GND of the connector 5220 in Path3, and the outer sheath GND of the USB cable 5230 is excited from the GND of the connector 5220 in Path4, and that this noise causes interference (spatial coupling) to the victim's wireless antenna.

[0177] • Scoring: For this reason, it is hypothesized that if the common mode of differential signal 5410 is present, the GND of connector 5220 will generate a pigtail effect.

[0178] (4) Probable cause presentation (560, 561) • Within the user display screen 500, area 530 is the area that displays the diagnostic results and proposed countermeasures from the image feature extraction unit. First, in the estimated cause view 531, for the item (wireless interference), hypothesis H-001 summarizes the estimated cause in a human-readable format, taking into account the paths and conversions of V (antenna) and S (driver IC for video transmission). For example, "1. Mode conversion occurs because the differential transmission signal wiring length from the driver IC to the connector is D+ / D-", "2. A high-impedance connection is made between GND and the slit in the ground layer of board 1 and the flight bead", "3. GND noise from the board excites the outer sheath of the USB cable", "4. This enters the wireless antenna", etc.

[0179] In the diagnostic results and countermeasures 530, estimated causes 560 and 561 regarding "mode conversion risk due to mismatch in signal wiring length," "common mode conversion risk due to return discontinuity between driver circuit and connector," and "pigtail effect (cable radiation of GND noise) risk" are presented to the designer as "physical critical points" requiring countermeasures. It is also possible to have the designer suggest countermeasures from general recipes at each stage of the analysis SPV. For example, in addition to resolving the wiring investigation in Path1, it is suggested that common mode noise itself be reduced by adding a CMC (common mode choke) to the driver side of the differential signal wiring in Conversion1, and that GND stabilization be improved by eliminating the GND gap in Path3.

[0180] In response to this, the countermeasure display view 532 presents countermeasures, their effects, and risks for each estimated cause. For example, for a wiring length difference of 1 (D+ / D-), countermeasure 1 is to eliminate the wiring length difference, the effect is that the mode conversion coefficient will decrease, and the risks include securing wiring area and the fact that it is not effective in reducing common mode noise caused by other factors.

[0181] Based on this organized information, information for user decision-making can be displayed, and the user input unit 533 can then accept the next instruction.

[0182] (5) User input area (533) and internal calculations • User input is provided through repeated, low-computational-cost dialogues and evaluations of the effects and side effects of hypothetical design changes, before making costly design change decisions. For example, users can input things like: "A: I want to eliminate the GND gap on the camera transmission board," "B: Due to lightning induction countermeasures, the GND gap and FB cannot be eliminated. Instead, please evaluate the case where a common mode choke or isolation transformer is inserted into the signal path of the differential wiring D+ / D-," "C: I want to place a shielding plate between the antenna and the USB cable," and "D: Compare the effects, side effects, difficulty of design changes, and cost increase of the countermeasures A-C."

[0183] • Deployment to the design loop (460a (virtual evaluation), 460b (final approval), 235, 240, 260): Step 460 in Figure 4, in correspondence with Figure 2, is internally divided into 460a (virtual evaluation) and 460b (final approval), although not shown in the figure. Definition of virtual design change in this embodiment: In this embodiment, a virtual design change refers to a process that temporarily changes the physical parameters of the Path element or Conversion element in the SPV causal hypothesis record without rewriting the CAD data itself, and calculates the effect by substituting that change into the physical scale normalization index and the coupling coefficient approximation formula. This is not merely a UI operation, but a search control based on updating the internal model (causal hypothesis record).

[0184] ·460a (=235 in Figure 2): The state update / research unit 130 executes a loop for evaluating the effects and side effects of virtual design changes based on user input (220). In this loop, the system virtually modifies the candidate paths narrowed down in 230 (changing constants, bypassing paths, etc.) without updating the actual CAD data, and calculates the effects and side effects through metamodel evaluation. The calculation results are repeatedly presented to the user (re-evaluation corresponding to instruction AC, comparative display corresponding to instruction D). In addition, if the user indicates that partial electromagnetic field analysis is acceptable, quantitative evaluation can also be added. By presenting the user with this trade-off between accuracy and computational cost, the policy can be decided with greater certainty before actual design changes are made.

[0185] • Search axis transition algorithm: If the user rejects a specific countermeasure as a boundary condition, the state update / research unit 130 transitions the causal hypothesis record containing the Path element to the rejected state and redistributes priority to the hypothesis group containing the alternative Path. This allows "dialogue" to function as an input means for controlling the search space, clarifying the technical concept (search control based on hypothesis records).

[0186] (6) Identification of potential causes and user approval 460b (Figure 2, 240) • In the final approval, the state update / re-search unit 130 presents the evaluation results to the user and obtains approval for the design policy. Based on the design policy approved by the user, 260 (CAD data update) is executed. For example, a CAD modification is performed to connect the connector GND at the shortest distance. The system then performs an iteration evaluation (270) on this modification and confirms that the improvement (910) outweighs the side effects (920).

[0187] Furthermore, even if the system suggests "changing the layout of the IC circuit" as a countermeasure based on active observation, if the designer inputs the constraint "the layout of the IC cannot be changed due to the board outline constraint" as a boundary condition, the search control means (state update / research unit 130) accepts this input as a formal boundary condition, transitions the causal hypothesis record containing the Path element to a rejected state, and redistributes priority to the group of hypotheses containing alternative paths.

[0188] (7) Extended perspectives on multiphysics and lifecycle This system, by using the same SPV / PEL structure, can be applied to physical design in general, not just EMC. For example, it can perform integrated evaluation through hypothesis structuring, exploration control, and virtual path manipulation for thermal design (e.g., cooling path obstruction by shielding plates), over-design correction (VE), and modeling of manufacturing variations and operating environments. [Explanation of Symbols]

[0189] 90 Recipe Collection DB 100 SPV training data database 105 PEL Log DB 110 Knowledge Structuring Department 120 Priority Assignment Section 130 State Update / Research Unit 140 Active Observation Instruction Unit 150 CAD subsystems 160 Image Feature Extraction Unit 170 user design data 180 User Inputs (Facts, constraints, EMC measurement results, placement and routing constraints, etc., excluding CAD information) 200 CAD data import and defect definition 210 SPV Search Calculation (Active Observation and Feature Extraction) 220 User Input (Facts / Constraints) 230 Determination of Completion of Verification Conditions 235 Virtual Design Change Effect and Side Effect Evaluation Loop (Dialogue-based policy decision-making, preliminary stage to CAD update 260) 240 User Approval / Result Presentation 250 proposed solutions and their prioritization. 260 CAD data update (design change) 270 Improvement / Side effect assessment (Iteration evaluation) 700 Input acceptance and preprocessing (normalization of design data and DB100) Comprehensive enumeration of 710 potential sources Comprehensive enumeration of 720 victim candidates 730 SV pairing (search key generation, Cartesian product avoidance) 740 Extraction and candidate generation of Path information by referencing DB100 Coarse filter of feasibility based on 750 CAD attributes 760 Rough scoring (score_pre) 770 Priority List Output (Active Observation Queue) 780 Handover to the state update / research unit 300 Causal Hypothesis Records 320 PathSequence 321 Physical State Transition (Conversion) 350 Refutation Condition 410 Input of defect / risk definition 420 Generation of the initial SPV hypothesis 430 SPV search calculation (active observation) and image feature extraction 440 Physical Scale Evaluation 450 Verification results and counter-evidence conditions 460 Identification of potential causes (internally divided into 460a virtual evaluation and 460b final approval) 500 CAD Active Observation Screen List of 510 SPV hypotheses 520 Multi-view Area 522 Area of ​​Interest (Local Enlarged View) 523 layer structure (cross-sectional view) 525 Route View 530 Image Feature Extraction Unit Diagnostic Results and Countermeasures 531 Estimated Cause View 532 Areas for Presenting Countermeasures 533 User input area 5210 Video signal driver IC 5220 USB connector 5230 USB cable 5410 Attention Net 550 Semantic highlighting 551 Semantic highlighting 560 Probable cause 561 Probable cause 600 CAD Active Observation Screen (Inner Layer View) 601 STP connector 602 GND via 603 GND plane split clearance 610 Differential pair signal vias (S1, S2) 620 signal layers 630 Inner layer GND plane (solid) 640 Adjacent signal wiring 650 Extraction Annotation (Asymmetrical Highlighting) 651 Risk Score Display 800 State update / research process 810 User Data Normalization Process 811 Normalized attribute data 820 Search Space Filtering Process 821 Update of Exclusion Mask for Search 830 Evaluation parameter synchronization process 831 Updating arithmetic variables

Claims

1. A physical design support system that assists in the search for the cause of defects in physical design, A search state management means that maintains the causal relationship of the target defect as a causal hypothesis record including the combination of excitation source (Source), coupling path (Path), and victim, as well as the physical state transition (Conversion) in the coupling path, and manages the set of said causal hypothesis records as a search state, A search control means that, based on the input defect information, design data attribute information and / or user input, prioritizes the causal hypothesis records included in the search state, updates the search state by rejecting the causal hypothesis records based on observed facts after matching with counter-evidence conditions, and determines the next object to be verified based on the results of the prioritization or rejection. An active observation generation means generates observation parameters for the observation target determined by the search control means, using the internal attributes of the design data, including at least one of viewpoint change, transparency display, cross section extraction, layer selection, and magnification change, and generates a diagnostic image according to the observation parameters. A feature extraction means for extracting physical features from the aforementioned diagnostic image, An update means that reprojects the aforementioned physical characteristics in relation to the components of the causal hypothesis record based on geometric coordinates or design attributes, updates the state of the causal hypothesis record referenced by the search control means, and reflects this in the search control means's next decision on prioritization and rejection; The search control means, the active observation generation means, the feature extraction means, and the update means constitute a closed loop that repeatedly verifies hypotheses and converges the search state by returning the output of the update means to the next decision of the search control means. Physical design support system.

2. The search state management means is characterized in that, in the causal hypothesis record, it maintains a path state transition sequence (PathSequence) as an essential element, which is an alternate arrangement of a plurality of connection paths and state transitions that define physical mode conversion or propagation mode changes between the connection paths. The physical design support system according to claim 1.

3. The search state management means manages a training data database that stores definition information of the causal hypothesis records and a log database that records the history of the search and observed evidence as physically or logically independent storage areas, and the causal hypothesis records are characterized by storing reference information to the log database. A physical design support system according to claim 1 or 2.

4. The search control means is characterized by receiving at least one of the facts, constraints, design intent, and verification results input by the user as boundary conditions for updating the search state, and rejecting the causal hypothesis record based on a comparison of the boundary conditions with the Refutation Condition defined in the causal hypothesis record. The physical design support system according to claim 1.

5. When the feature extraction means extracts physical features from the diagnostic image, it calculates a physical scale normalization index that shows the relative scale of the geometric shape on the design data with respect to the wavelength of the physical phenomenon in question. The update means is characterized by updating the priority of the causal hypothesis records based on the indicator. The physical design support system according to claim 1.

6. The search control means is characterized by virtually changing the physical parameters corresponding to the connecting paths and / or state transitions included in the causal hypothesis record based on user input, without updating the design data, calculating the change in the risk score based on the virtual change using the physical scale normalization index and / or the physical evaluation formula, constant, or approximation model stored in the causal hypothesis record, and presenting the result to the user. The physical design support system according to claim 1.

7. The exploration control means calculates the amount of risk score reduction due to the virtual change as the degree of improvement, calculates the increase in risk score occurring in other causal hypothesis records as a side effect, presents the degree of improvement and the side effect side effect side by side, and updates the design data only when approval input is received from the user. The physical design support system according to claim 6.

8. The search control means is characterized by being able to select the accuracy and computational load of the evaluation model used in the virtual evaluation in multiple stages, and selectively performing at least one of evaluation by a lightweight approximation model and evaluation by detailed analysis. The physical design support system according to claim 6.

9. A physical design support method in which a computer assists in searching for the cause of defects in physical design, (a) A step of storing the causal relationship of the target defect as a causal hypothesis record that includes the combination of excitation source (Source), coupling path (Path), and victim, as well as the physical state transition (Conversion) in the coupling path, and managing the set of said causal hypothesis records as a search state, (b) A step of prioritizing causal hypothesis records included in the search state based on the input defect information, design data attribute information and / or user input, updating the search state by rejecting the causal hypothesis records based on observed facts based on comparison with counter-evidence conditions, and determining the next object to be verified based on the results of the prioritization or rejection, (c) For the determined observation target, generate observation parameters using the internal attributes of the design data, including at least one of viewpoint change, transparency display, cross section extraction, layer selection, and magnification change, and generate a diagnostic image according to said observation parameters, (d) A step of extracting physical features from the diagnostic image and reprojecting the physical features in relation to the components of the causal hypothesis record based on geometric coordinates or design attributes, (e) The process includes: (e) Based on user input, virtually changing the physical parameters corresponding to the connecting paths and / or state transitions included in the causal hypothesis record without updating the design data; calculating the change in the risk score based on the virtual change using the physical evaluation formula, constants, or approximation model stored in the causal hypothesis record; presenting the results to the user; and updating the design data upon approval. The search state is characterized by repeating steps (b) through (e) above. Physical design support method.

10. A program for causing a computer to perform the method described in claim 9.

Images